Transcription
Okay, let's begin. We are on the fifteenth episode of the GoGet podcast. We are talking about AI again. The topics are roughly the same as we discussed last time in the previous episode about agents. The first half will be more practical. And in the second half, there will be more philosophical topics. Like, have you noticed AI, where are we going, how is everything developing, some news. Well, it will go as it goes. So. Let's introduce ourselves first. Well, everyone on the channel probably knows me. Vlad, shall we start with you, right? Hello everyone. Can you tell us a little about yourself? Of course. I am Vladten from the Vladten channel. So, I sometimes film something on YouTube, sometimes I delete something. So, usually, you've probably seen me, I talked about algorithms on my usual channel. I conducted interviews. So. What else can I say? Ah, I worked at Microsoft. Once, they invited me to Open AI. I thought it was a company that dealt with Dota, and I said, "No, thank you," and since then, it's been a "telto sir." That's how it is. Did you confuse it with DeepMind or something that deals with it? No, no, no, no, no, no. Long story short, there was Open eCodeCleengge, and it was around Christmas before the launch of GPT, back in some 2021 year. So. And I passed it, like, I got on the leaderboard there. And then they sent me an email, like, here's a t-shirt for you. And like, yo-yo-yo, let's get to work. But I had already received an offer from Microsoft. I thought, "What, I'll go work, I'll make Dota, I'll assemble some Rubik's cubes, like reinforcement learning." That's what I would be doing there. So. But that was an epic fail, the biggest epic fail of my life. Yes, it's a shame. Well, it turns out we have an expert who almost got into Open AI. That's already much closer than us. No, don't say that, please. Okay. So, yes, I won't delete it, don't worry. I'll make a bunch of backups just in case Vlad gets to it. Ildar, will you tell us a little about yourself too? Yes, I'm Ildar. I have a YouTube channel and a Telegram channel too. So, I work as a Go developer at Platon company. So, with Kolya, we are in neighboring teams. So, no one has ever invited me to Open AI. That's about it. Excellent. Yes, unfortunately, no one has invited me to Open AI either. So, let's briefly go over who uses what in terms of AI agents. Ildar, I guess we'll start with you. You use them more, as far as I understand. Vlad said he rarely writes code through AI agents. Tell us what you use, how you use it, where it's well applicable for you, and where it's not. Mhm. Well, even before Copilot appeared, I tried to use things like Tab and Kite, which do roughly the same thing, only worse. So, I played around with it then, and, well, okay, I abandoned it. Copilot appeared. I was skeptical about it, installed it, and was pleasantly surprised. So. And for a long time, that was like my main tool. Well, naturally, I also went to ChatGPT with questions and so on. And when there was still version three or 3.5, the quality was severely lacking in terms of results. So. And basically, I didn't actively use anything other than Copilot back then. So. But at some point, I saw Cursor, tried to use it, wrote a couple of pet projects, what I would have done in a week, it assembled it for me in an evening. That pleasantly surprised me. Well, you just shouldn't have looked at the code, because later, when I needed to add a new feature, I initially didn't understand why it was so difficult for it. It's easier than the rest of the project that it wrote before, but then I found out why. So. Well, we'll probably talk about that later, about the code that agents write and how to work with it. So. But then Anthropic also launched its own agent, Code. And in the end, that's probably my main one now. Periodically, out of interest, I switch to Juni, which is in JetBrains, and to the agent mode in the Z editor. It's also quite good there. I can tell you about it later too. So. Well, yes, by the way, that's also interesting. I glanced at it, but I wasted my trial on nothing, I just didn't have time. Well, so. Well, yes, well, we'll talk about its features. So. Uh, and if we're talking not about agent programming, then, well, code, as, well, as a chat, as a desktop, uh, mainly for technical and semi-programming questions, and for everything else, ChatGPT, sometimes Perplexity. Perplexity is more or less the same for me. Everything is like that. I'm also a search engine. So. Yes. I have a similar story, more or less, as you said. Only. Did you use Goland before Cursor, or VS Code? On Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient. But for comparison, Claude Code, especially with, oh, Claude, just Claude with the recent update, when the fourth version models appeared, like OPUS, it's good at analysis, and although it does it slightly worse than ChatGPT, and at the same time, it's also good at the human readability of this text. Although it's also done slightly worse than ChatGPT, you don't need to switch between models. I really liked working with it in this regard. So, regarding agents, is there anything else to add? Well, I'll just add that a debugger has appeared in Zed, by the way. Perhaps not everyone gives it, yes? But I've already tinkered with it. Cool. And let's briefly explain what Zed is. It's a newfangled thing. Uh, I hear more and more enthusiastic reviews about it. Well, I understood the main thing, I just haven't gotten around to fully testing it myself yet. I installed it, launched it a couple of times, but still not there yet. So. I understood the main excitement is due to the fact that it's lightweight, written in Rust, and doesn't have any complex legacy code like what even JetBrains itself has accumulated, and so on. And they recently attached something to it. It had something like Copilot in some form, I think, attached, and not very long ago, but now they've released a full agent mode. Well, roughly like Cursor, I understand. So. Uh, did I say everything correctly about Zed, or does it have any other features besides AI, and why is everyone switching to it? Initially, their focus was on collaborative editing, like, you know, like in Google Docs, only they also have separate rooms, where you can even have calls inside. That's their idea, because they, I think, launched their first public beta during COVID. So. Uh, but that doesn't seem to be catching on much now. Well, and now it is catching on, and I. And that's why they made a pivot in that direction. It seemed to me that this calling feature wasn't that important to people. They just liked the lightness of Zed. Well, it seemed important. I, I was hooked, actually, by the fact that it's super smooth and everything. Well, you know, when you type text, when you type text like this, when you scroll, it's like it flows like this, and there are no delays, like everything is just so smooth, smooth, smooth, back and forth. I, I was hooked by that, I just liked that. Well, in the end, well, yes, I know that I know that many people find it very important. I also remember a long time ago I had a colleague who wrote code in Sublime, when everyone already had all these powerful IDEs. And I asked why, and he said that it irritates him when he types text and it doesn't appear instantly, that there's some delay. Well, for me, that's cool, of course. It's pleasant when you type text, and your eye really notices that it appears instantly on the screen, and somewhere with a slight delay. But it's not like it's critical. Especially since not all of us have monitors with 120 Hz or more. To notice something like that, you really have to try, I think. But, uh, no one sponsors JetBrains, right? Well, like, we are not from JetBrains and nothing, no one is sponsored. Okay. Okay. Well, for example, as much as I've tried, when I used to write in Go, Goland, it really didn't suit me at all. Precisely for that reason, it seems to attack me, it's too overloaded, and I don't like it. I'm like, "I want some lightness." And that's why I usually, I don't know, keep three tabs in the terminal, somewhere, and use IDE or something lighter, like Zed. I like that more too. Without IDE, you used VS Code and terminals, right? Yes, on Goland. Well, and now on Goland, so I practically don't use Cursor anymore. Well, yes, that's exactly why I didn't like it, simply because, well, it's very painful to switch from Goland to something VS Code-like. But when Code-Code appeared, then yes, then finally it was possible to start using it fully. And you also said that you use both ChatGPT and Claude in terms of chat. What do you use each for? Why? Why do you need two? Well, I have two as well, but why do you need them? Well, it feels like Claude handles technical questions better. So. And at the same time, it doesn't insert a bunch of emojis into its answers when it gives you another listing. So. And ChatGPT, especially if you use GPT-4, it does micro-research quite well. That is, you ask it a question, and it goes to a bunch of sources, collects information, and answers based on them, rather than just from memory. So. And that's very convenient. I try to figure out all sorts of tax-related, especially complex legal questions with it. So. Naturally, I then go and check, but it still helps a lot in such things. And I also use ChatGPT for image generation and have even experimented with audio analysis. That is, you can suddenly upload an audio file to ChatGPT, ask it to analyze it, it will write code within itself, which it will then run itself, and extract some facts. So. So it's cool for general use, ChatGPT, for technical questions, I have Claude, probably. So, well, it's more or less the same for me, yes, it's just that ChatGPT shows some kind of wavy quality for me. In some things, it's very bad. In some things, it's very good. But Claude gives consistently good results. It's worse in some places than ChatGPT, but it's at least stable. So, in terms of technical stuff, it's awesome. I only came to this thought now when I was thinking about why I use them this way. Well, it should also be noted that Claude seems to have worse multimodality. That is, it cannot handle a large number of different input types at once. You also cannot upload an audio file for it to do something with it. You can upload a PDF, but it won't do anything with an audio file. So, it's a more focused model, I would say. Mhm. So, well, Vlad, tell us about your experience. Yes, well, I remember when you were using it, when Copilot just came out in beta, I tried it, I liked it, it was cool, but then I started noticing more and more that, well, I spent time editing what it generated, and that started to trigger me badly. So. And after that, I turned it off. Then ChatGPT was released. And in general, if we take it on a daily basis, well, I probably use ChatGPT most often as Grammarly. I ask it to rewrite, correct punctuation, do something else, proofread something. So. But otherwise, I just encounter hallucinations or similar things too often, and it really annoys me, and that's why I can't use it normally. So. But otherwise, otherwise, I try everything. That is, when it came out, or when Claude Code came out, or the same Z. I really like Z, by the way. The only pity is that there's no debugger yet. So. Uh, yes. Well, well, in general, that's the picture. That is, I try to use them, but I tried Cursor-Cursor directly, to fully embrace this agent flow. Uh, well, it's really not for me. That is, I really, I really didn't like it. I don't know, maybe this is the point where I've gotten old, like, it used to be better and all that. But, uh, I find it very difficult to use. I, I, it often irritates me. Did you try Claude Code? Ah, Claude Code, is that the CLI one? Yes, yes, yes, yes, I tried it. Claude Code didn't work for me either. Ah, but it works for me, but maybe the problem is that I write very little business code now, or practically none at all, and what I do write, it's like I don't need it there. Probably, probably that's the main problem. Well, I understand, yes. Well, I also had such thoughts about Cursor at first, that it's all somewhat intimidating. It asks you to redo something, it gives you a bunch of results, you need to apply here, apply there, and somehow, well, it's hard to describe, but it put me off a bit, and I realized that it's still more convenient for me to use the old-fashioned way through chat. That is, generate code for me, here's a piece of code, generate another piece of code for me, I just copy and paste. But when I tried Code-Code, I quickly moved away from that concept. Everything was organized simply and conveniently for me. Well, that's even besides the fact that the idea itself is unfamiliar to me. They're writing here that they're asking Vlad to turn up the volume, but I can only turn everyone down, not up. Do you have anything to adjust there? Well, I do, yes, I don't know if it will get better. I adjusted it. Well, okay. So, regarding ChatGPT and Claude, what Ildar was saying, and about how it's sometimes good and sometimes bad. For example, when you use GPT-4, it does a very good analysis, collects a lot of information, gives a lot of details, but the phone interface is not very convenient for displaying the result, for example, there are some long tables that, you know, you have to scroll, it's all not very convenient to read. You also need to analyze this analysis. And then I just switch to some model 4.1 4.5 and say, "And now explain it in simple, understandable language, so that it's readable like this, without tables." And then it does a summary of that analysis for you, and then you read it. In this regard, it's not very convenient.
It is much simpler to make them than graphical ones, even as a plugin for editors. There. And there, well, that is, in the agent written in the article, a lot is missing. There are no checkpoints, there is no smart work with context, so as not to send everything at once to the model and not exhaust the entire context. Uh-huh. So, a lot, in fact, also depends on the prompt. This, well, I don't know, probably, how much the cursor costs, 9 billion. Of these, 4 12, I think, have a system prompt. In fact, I think that the system prompt is some kind of secret thing that they invest such a cost in. I'm just not particularly into it, well, in the sense, I understand what a system prompt is, but you said it's expensive. Can you elaborate on this thought? Well, yes. So they, of course, leak periodically, but the point is that how to properly hack this model, how to properly instruct it, and not just the system prompt, but also, if necessary, for each tool. A tool, if anything, is reading a file, searching files, that is, everything that, well, makes an agent an agent, that is, all its capabilities to navigate file systems or interact with the external world. Each tool, in essence, also has its own small prompt, which describes when, in what cases to call it, how to interpret its results, well, calls to this tool, and so on. So I think that all of this together, in fact, well, significantly affects the quality of the agent's work. Yes, I can confirm that, because I, for example, have subscriptions, yes, to ChatGPT, to Claude, well, chat versions. And at the same time, I have a subscription to Re:Caz, that is, all these advanced models I have, well, Re:Caz is a utility for Macs, which, among other things, also has a chat interface with all models. So, it would seem, why should I pay separately for ChatGPT and Claude, if I have Re:Caz, which has both for the same price. But the thing is that communicating with it, the model seems to be the same, but it seems much dumber. That is, you need to explain more to it, give it some context. Well, the trick, as I initially assumed, is precisely in the system prompt. Well, besides that, there are other moments, of course, that their interface is not so convenient, but that's all. Well, the most important thing is the system prompt, apparently. Well, one of the most important, yes? So it's also important how you, in fact, give access to the code. That is, you can, you can send the entire file, you can, as Code-Code does, say that it should grep what it needs. Only that part of the code, well, which is needed, was provided. There are many different techniques, yes, and it is precisely all agents that are now competing in who works better with context. Yes. Because the models are beyond their control, that's all that's left for them for the most part. Yes. Well, regarding what he has simplified, that's a good way to illustrate what these agents are actually cool for. By the way, speaking of what you said, is it difficult to write? They say Anthropic says that 90% of Claude-Code is written by Claude itself. They wrote it in JavaScript. I don't know how much, well, that is, 10% the person did some refinement, tuning, and all that. I don't know how true that is, but in principle, it might be true. So, it would seem that even a cool developer is not needed. But how expandable will it be later? That's a good question, because, yes, it turns into a mess, especially if you trust so much. Well, it's not even about that, that is, I mean that technically, the code itself that executes all this, yes, it's quite simple. The idea is that you need to figure out how to work with it. That is, the work is already at the level of the architecture of this agent, in principle, its approaches, its concepts. That is, it's really very simple to get into. And the article shows this well, yes, you read the article, and it's limited to just a couple of commands: get the entire file content, get the list of files, and edit something in the file. But what simplifications are there? In the sense that, well, firstly, context is not some bottomless thing, and it happens that a couple of your files can fill the entire context. Well, depending on the files. Claude has about 200,000 tokens. Maybe only Gemini can do that, but even then it will eat a lot from you because it has a million tokens, but it will simply be expensive in the end. Besides, you need to send it back and forth. Ah, well, again, for context, for context. Maybe someone doesn't understand what context is. When you communicate with an LLM, it has no state. That is, it doesn't remember what you just talked about. Every new message is like a novelty for it. And for there to be context, we send it the entire conversation context. Everything we wrote, everything the agent or it wrote itself. And, in particular, if we also attach some files, system prompts, all of this is also sent back and forth. And, accordingly, you talked to it, it researched a dozen files, everything went back and forth, and now you're sending, I don't know, 200,000 tokens, and when there are even more, then it will hit its limits for the context window, and plus it will be expensive. Therefore, advanced agents know how to compactify, first of all, that's how it works in Claude-Code. The context honestly grows, they don't do anything with it, they don't delete it. But when it becomes too large, they launch auto-compact. That is, they throw out everything unnecessary from your conversation, leaving only the important. Well, it works, by the way, not badly, because it doesn't interfere much with work. Sometimes it does, it's dumber, it forgets something important, but in principle, you can live with it. This is probably better than what can be invented now at all. Especially if it's done well. Cursor, as far as I know, has a different approach. And by the way, ChatGPT also has the same nonsense in its chat interface, that they simply forget the first messages. That is, they remember how many of the last ones. And this is much worse, because you might have given important instructions at the very beginning, and closer to the end of the conversation, they will simply forget about it. You will have to remind them every time. But there is no compact that makes them sing this. It's not quite like that. In Cursor, it's still closer to the end of this context window. It offers you to create a new chat, but summarize the old one. That is, in essence, it's practically the same compact. And if you refuse, will it forget? Well, yes, yes. That is, well, in fact, in Code-Code it's the same. That is, there is simply auto-compact now, but in the first versions, well, you just say that it offers, as if there is a choice. Well, yes, yes, you can refuse, in principle. That's true. Well, firstly, this, well, that is, you need to teach your agent to do at least auto-compact. Well, probably, this is also not a super difficult feature, but, but there are already not 300-400 lines of code. Plus, direct file editing. Based on this example, it essentially gets access to your entire system and can not only edit something in the project but also outside of it. If, well, an LLM agent is a very random thing, and it depends on luck. Accordingly, this also needs to be foreseen, so that it asks, so that it is clearly limited, where it can, where it cannot. Yes. So, what do I have? I'll look at my notes, what else I wrote down. Ah, yes, in reason models, and in principle, I understand, in these utilities there are built-in tools. Let's say, self-reflection, what they think, whether what they wrote is good or bad, not just, they replaced a line and okay, this might be enough for some simple tasks, but in general, to see how well the task was performed, whether there are any errors, to run tests, this also needs to be taught to your tool, so that it's useful. Well, and again, regarding security, that not only access to files, but also actions in general need to be strictly limited, because if it leaks some of your keys somewhere or they leak somewhere, well, even an LLM, for example, Google's Gemini, it clearly states in their license that they can and will read your chats, not just use them for model training, but a live operator will look and use your chat for something. Accordingly, they write, if you don't want your data to leak, just don't write it in the chat. Did anyone try, sorry, sorry, no, I'm almost finished. Yes, sorry, sorry, finished. Sorry. In corporate licenses, there is no such thing. They promise that they don't train their model, and a person won't read this data. They are lying. Maybe, I have a question: has anyone tried to set up, like, a locally tuned construct and connect to it and set up their own setup like that? How much worse does it get? I tried to run Deepseek when it was hyped, specifically the distilled model. Well, it seemed to me that it thought for a very long time, because, obviously, I don't have the same power as in the cloud. But on what did you run it? On a Mac. Yes. But the quality, well, it's also far from it, but again, that's generally normal. In the end, it's like it's not worth it. So I would consider this option as, well, in case you need to fly on a plane for a long time or be somewhere without internet, but as a main tool, as a main model, probably not. Uh-huh. Well, I've tried it too. I have the same impressions. I've just tried a lot of different, well, not a lot, but different ones. And those that are heavy, like smart ones, that think for a long time, and those that are faster, they respond quite briskly, but they respond poorly. It's like a joke here that I type very fast, but it turns out to be nonsense, pretty much the same. Well, it's understandable that Macs are not something for running models, but I recently built a gaming computer with an RTX 5090, and there you can already run something more powerful, yes. And so far, honestly, I've been too lazy to bother much. I ran a couple of models. It briskly runs those that worked slowly on the Mac. But I also tried to run some Llama, one of the most advanced ones, but it was very slow and for some reason it didn't load my video card. That is, even medium models loaded the video card to 100%, but this thing loaded it somewhere, I don't know, by 20%. And on the processor, I think the load was higher. I started to figure out why. And in the end, I became lazy, and I just keep postponing it. There are some nuances. Yes. But someday I'll get to it. Using it as a replacement, of course, won't work, because, well, no matter what RTX, not RTX, even that won't be enough. In the sense that 5090 will not be enough to run something serious on your own. But here, perhaps, there is room for application if you have some things that, well, some very simple requests, but very many. I tried to think, to come up with such a case for myself. And it's just very inspiring. You think: "Oh, I have a thing that's free, I can run it as much as I want, any number of times, it won't cost anything." Well, yes, it's dumb, but, for example, you can attach a utility for some smart, super-smart autocompletes in your terminal or somewhere else. That is, so that it analyzes every symbol, and you type quickly, and it sends a request to the local model for every change, and then comes up with autocompletes. Something in that direction. But that's the only thing I came up with. I even tried to get ideas from others by researching the topic, but they come up with complete nonsense. That's why, by the way, I haven't figured it out yet. In the era before GitHub Copilot, I was talking at the beginning about Tab9 and Kite. I don't remember which one, it seems Tab9. They ran a local model on your end, for autocompletion. And I had a weak laptop then, with 8 GB of RAM, something like an i5 processor, and it was generally enough, well, no, it was 16 GB of RAM, and it needed about 4 GB of RAM and some very simple autocompletion, this model already did. So, in principle, for such tasks, local models are really good. And here they are writing about RAG. Do you know what RAG is? I'm looking at you, Vlad, right away. I think you'll explain better what it is. It's like you said, you delve deeper into all this. Well, I understand it intuitively. And speaking of this article by Orstan Bola, he specifically lacks RAG. And here Alex Belushkin writes that he wants to set up local RAG for Claude. Maybe it will be easier to understand the project. As far as I understand, RAG is built-in there. Well, naturally, otherwise how would it all work so well? But someone can deploy and explain what this thing is. It's fine. Imagine I'm a RAG expert. I'm not even close. Well, here, well, let's. I wanted to joke, but as far as I understood, RAG is when you have some source of truth, and when the model starts to output some data, it can go and retrieve your specific source of truth. But that's the most basic case that is often mentioned, it's customer support, that when you need to answer a specific case, you have a specific document, and it takes a piece from it, not just makes it up out of nowhere. In simple terms. And how do you decipher it? Does anyone remember? Retrieval Augmented Generation, I think. Yes, something like that, I think. Well, it's, I understand, it's just an architectural approach. In the sense, the architecture is called that, that instead of making up some information or just sending a whole file, it first retrieves what it needs, then incorporates it into the prompt itself, and sends it as a single request. As far as I understood the concept. And this is precisely what is missing. That is, I said that compactors are missing, well, his utility, but RAG is also missing, because you don't necessarily need to send files whole, it's enough to send some embeddings, that is, to embed the necessary pieces into your prompt. That's why I didn't quite understand what Alex Belushkin means when he wants to set up RAG for Claude, because it's already in Claude-Code. So, and did any of you try to set up local models yourself, Vlad? Ah, no, I tried, but I can't say that I saw anything exciting, but worse than worse than what OpenAI sells. Who, who among you has written their own MCP server? Oh, I want to, I'm planning and postponing it. I want to make a video, including on writing an MCP server. And now I also want to write my own agent after this article, it really inspires me to do something similar, but in my own way, but I haven't written it yet. Cool. Well, I understand that there's nothing complicated there. It's a simple protocol. You just need to find the time and figure out why you need it at all, because a lot has already been written. You haven't written. No, well, my hands are itching too, yes. They said they built MCP into Windows, did you hear about that in some versions, yes, on Kinote. And in what sense built-in? Does anyone know the details? That is, how it works and why? Well, yes, MCP, it provides you. Well, that is, MCP is to standardize, roughly speaking, to standardize the use of tools by agents. And the Windows MCP server means that they will expose some Windows functionality as an MCP server, and the models will know that they can call such functionality. Some models are already on board in Windows, like. Well, I know they have some Copilot. On Windows, everything is already on board. Yes, they already, I'm sure, have some advertising model running there, and more, and more, there, there, anything can run there, and KMS, which activates you, but there, everything runs there. Yes. So, well, advertising, you mean what they show, like a list of news or something like that? Yes, yes, yes, I'm joking, but remember they had a scandal when they were like: "We will screenshot your screen and send it to you like this is better." I haven't heard of that, but regarding advertising, yes, well, they do it somewhat tolerably. That is, I recently, well, with the purchase of a new computer, I went to Windows 11, it told me on my old one: "You will have to update." And I was so lazy to do it. And, well, yes, there is this window on the left, you click, a huge pop-up appears, and there are a lot of news, some promotions, stocks, and news. Well, here, of course, there is a lot of room for advertising, but it's cool that they at least limited themselves to this window, more or less. And, by the way, when you search for your applications, they might suggest, like, maybe you'll like this. Uh-huh. Uh-huh. It's done worse, for example, I don't know how it is now, I haven't used it for a long time, phones, for example, Xiaomi. You buy them, and in every folder, when you create folders for applications, in each folder, a pop-up appears like, "And also this." This application will fit here, and some advertisement. And you think, damn, I paid for the phone, but I can't even disable ads in it. It's a trade-off. If you want a cheaper smartphone, you pay at least for the ads. But that's okay, at least it's clear why the price is low. Maybe it's already factored into the price. But I have a TV, it's like, how much, 70 or something, something around that inches. Well, it's not cheap, to put it mildly, a little less than this computer I bought. And you turn it on. The first thing you see is a huge Ozon banner. And okay, if they at least let you enjoy the first impressions, like, well, first use it, and then we'll show you ads unobtrusively, but there's an Ozon banner on the top half of the screen, then all your applications. I looked and realized that I didn't buy an Apple TV box for nothing. I use this TV simply as a monitor for Apple TV. All its smart features, apparently, come down to this. Well, roughly speaking, they, in general, want to expose Windows functionality as tools for agents, so that it's good and convenient for everyone. So, cases, that's already something to think about for applications. Yes, yes, I can't think of anything offhand, I don't know. I, well, I would click "deny," honestly. I, well, if you asked me, like, do you want it, I'd say: "No, no, no." Well, yes, I agree. I usually do that too. Apple also seems to be promising something with its Apple Intelligence, which also seems to run locally. And it's interesting, well, that is, the concept itself is interesting, that you will have some built-in model in your iPhone and Mac. And they also give developers access to the API of this thing, respectively. Of course, their hands are untied, because if they go to some cloud model, then you need to come up with monetization. You can't do that for free. But when you have such a local model, essentially unlimited, then a lot opens up. It's clear that it will be dumb, but let's see what it leads to. True, I don't know, they've been promising their own for how many years, well, they're fiddling with something, yes, they have some private cloud, they're making something for each user, something else. That is, they have a lot of movements. By the way, have you seen that Apple has a lot of Go developers? I recently found out, I was surprised. No, I haven't heard of that. Was there any news about it? No, there's like a guy, Sean, Sean Di, that's his name, he's the one who makes the VbertC thing. Have you seen him? Haven't you seen such a guy? Ah, I only heard about it because you often talk about it. Ah, yes, yes, yes. Well, he, in short, went to Twitch, did something on his Pion, then went to Apple, did something on his Pion for FaceTime, it seems, and now he's at OpenAI. Everything we talk about with ChatGPT, these new things, when real-time dialogues happen, all this is most likely all in Go, in Pion, because he's the main one working on it now. I was like: "Whoa, well, cool, yes." Well, Go is often used to write things that are written by people. By the way, I wanted to touch upon this, they reminded me in the chat about Gemini Claude. Has anyone among you tried it? You, what about it, haven't gotten around to it yet? Well, yes, me too. Well, in short, Google released a thing very similar to Claude-Code. There's also a similar thing, Codex CLI from OpenAI. But Gemini is closer to Claude-Code. Well, all the early reviews I've heard say that it's very raw, it works very poorly. Its only advantage is that they give some huge limits for free and that there are a million tokens in context. The only application I found for it, I haven't used it yet, but I also want to try it. I just read the idea that you can make your own Claude-Code use Gemini to save context. Like, when you need to analyze a large project, it has a million tokens, and plus it's free to talk to Claude, like, you can just run Gemini with parameters, not like a full-fledged chat. And summarize something like that. Plus, Gemini itself is a very smart thing. Gemini 2 with. Maybe it's even smarter than Claude somewhere, so you can use them. Yes. And why did I remember this, that all these things, all three that I mentioned, are written in JavaScript. And, apparently, because they are mostly written by the AI themselves. And in JavaScript, there is apparently a larger base. And, well, I've heard more than once that LLM agents work better. Well, in general, any LLM works better with JavaScript than with Go. But I haven't noticed that myself yet. It works perfectly fast for me. Well, I also write very little in JavaScript. Here, I don't know, do you have any confirmation of this, or is it just rumors? I have more experience not with JavaScript, but with Python. That is, I tried to generate roughly the same code with agents on Python and on Go. And it seems like the models feel a bit more confident in Python, and the code is more, well, more adequate, I would say. Yes. Well, there is such a feeling, yes, with Go, it sometimes starts to get confused, the nesting levels of functions start to grow significantly, and all that. Yes. In Python, as I initially thought, Go is almost made for AI. I had such an initial view. Well, because it's the simplest. There are few ways to implement something. In the sense, when you want to do something, you can do it either in one way, or not at all, like in JavaScript, in a million different ways, and each one does it differently. In this regard, it's better, but no, apparently, because the base is smaller. Well, and when people write these things, I forgot what it's called, I'll say it exactly now. Open CLI, I'll send a link now. In general, there is another CLI, but it's written by the community. By the way, Gemini is also open source. Written in Go and also using this bubble tea. Does anyone know these things? With a lot of various bells and whistles. And you can switch models. Well, this, for example, is written in Go, not some kind of, you choose, and not, God forbid, JavaScript. Well, it looks cool. I also want to try it. Well, the only thing is that here you will have to use tokens paid, and this turns out to be expensive with some models. So, well, we also wanted to discuss. What amounts are we talking about? Expensive. I just don't use it as actively, like. Ah, expensive. Well, look, I haven't used it. It's just that Claude-Code had a mode where before they included it in the subscription, you could only pay for tokens. And I listened to a podcast, people said that they could spend up to $1,000 a month with active use. When you use it fully for work. For me, that's expensive. It can probably be even more, but you have to try. For example, Codex was much cheaper, according to stories. Well, I'm also a bit skeptical about that, that it's much less, because Code-Codex only works by tokens. Yes, although I heard that a subscription model has already appeared, but I haven't checked yet. And it seems, it seems not. But I've tried it a bit, and it has already cost me about $5. Although I just tested it. I didn't even, essentially, I wouldn't have completed the task with that amount of tokens. So, and if you do it yourself and use Claude models, and they are much more expensive, then it will be expensive. It might even exceed $1,000 a month. Uh-huh. Regarding Gemini, by the way, I don't know, maybe it's somehow more profitable. Gemini is very cheap, actually. That is, if not the one that is maximum, I don't remember what it's called, but their basic model is quite cheap. And again, the thing I mentioned is called Open Code. You can either Google it, or I'll write a post in my channel after we finish the podcast with a list of all these links we mentioned. I'm marking files here. Well, I'm sending them to the chat now. Not everyone is listening to us live. And some will listen on audio platforms without a screen. So, you can go to my channel later and see all the links there. So, well, shall we move on? We can go through some interesting news. Well, while we were talking about Claude-Code, there was recently a loud news that some guy, I think, one of the, damn, either CEO or whatever, who was in charge of Claude-Code, was lured away to Cursor. Haven't you heard about it? Ah, I hoped you would now, like football transfers. Ah, well, yes. So Cursor lured away two key executives from Claude-Code. Ah, yes, how could I forget him, the guy? His name is Boris. Boris Cherny. Well, judging by his last name, probably Polish. I don't know who he is, but I remember it well. And Katwo, I think, is a product manager or something from Claude-Code, as far as I remember. Well, in general, so, I don't know, possibly after this, Claude-Code will significantly degrade, because two key guys who actually built what we all like so much have now gone to Cursor. Well, we'll see, we'll see what it leads to. I'm a little sad about such news. I think there will be more and more such news. That they are luring them away. Well, yes, Vlad also mentioned Meta. Can you tell us the news about luring away guys? Yes, by the way, if I were Cursor, I might buy Zed, because now Code is trying to tighten the screws a bit, yes, on the fork and something like that. I think it would be cool. They would buy Zed. Maybe I would use it. No, really, really, why not? I think they have enough money to buy 10 Zeds. Well, they can fork it, by the way, because it's open source. Either way, they can stand by it. Or they can fork it and lure away key guys. It will be cheaper. Because I don't like that Cursor, I don't know about you, but when I opened Cursor, maybe I was in some beta version again, it's really heavy. It irritates me. I click, and I see these strange delays, windows, and I'm like: "Come on, hurry up!" Text, it's like Johnatan Blow likes to rage, that like. And it works worse than VS Code? Yes, as if worse. Honestly, it seemed so to me. It's so slow, so disgusting. And like Johnatan Blow likes to rage, that like in the fifth Battlefield, a house is being destroyed in parallel, and you're being insulted in chat, and they're shouting in voice. And like, the text editor is hanging. Ah, yes, gossip, in short. Sam Altman went on a podcast with his brother and says that Zuckerberg is trying to lure away our key employees, and no one will go because we are all for the mission. And then news comes out that he has poached 20 people already, and these are key people. You earned on Do3, Do1, Do4, image generation. In general, he's poaching hard, and he's poaching many directly. That is, Zuckerberg calls you on your phone, Zuckerberg sends you an email, and one of the offers, like the loudest one, that they are talking about now, is like total compensation $100 million a year. That is, it's more than Cristiano Ronaldo. That is, $100 million a year, in short, that's what the guy paid. That is, $50, as far as I remember, it's not compensation, it's a one-time payment, like for switching. Yes, like a footballer, a transfer. Well, well, yes. So, Zuckerberg calls you, says: "You have 6 hours, I'll give you $100 million, come, and that's it." Well, just like, he, and who, some of the management or Yes, from key research from OpenAI, and he also lured away. How much do they earn themselves, I wonder? I think a lot too, but probably not $100 million, otherwise they wouldn't agree. This is, like, an important point, that if you're offered such an offer, maybe it's not much more. Then he also took a guy from SSI, did you see Sutskeever left to start his own company with one guy? So he lured away his co-founder too. Then he bought Alexander Wang. He made him the director of Meta Research. Alexander Wang is from Scale AI, which does all the dirty work for all the models, roughly speaking, cleaning, data labeling. So, like, Scale AI prepares all the datasets for everyone. So he bought him. And there's another interesting one. Something I'm talking so much now that I've forgotten myself. Drowned in gossip. Yes. Well, well, in general, he's making such moves. He's just showering everyone with money and wants to take everyone for himself. Well, you can think about it if you want. I'll put on some nice music, people will wait. Yes, no, no, well, well, well, in general, that's the gist of it, that Sam was saying that no, no, we are not for sale. Ah, I remembered, I remembered. Also, because of this, an interesting moment arises. Well, as you know, our favorite Sam Altman, he's a bit of a braggart, like he likes to, well, like to scare everyone. Well, and he's not starting to remind Elon Musk of something in some resources. I don't know, I don't know. But he really likes to scare people, like, like, the apocalypse is coming. And many people have asked themselves, like, let's imagine, you are a key, you are a key researcher at OpenAI, according to Sam Altman, like, we'll fine-tune Stargate resources now, and then we'll have a god from the machine. And like, you are a key employee of OpenAI, you know that the Stargate god from the machine is being built there, Zuckerberg calls you, gives you $100 million, you're like: "Well, why not?" Well, well, well, it clearly doesn't add up, agree, such $100 million. No, well, well, like, yes, but if you have a god from the machine under you, well, probably, like, 100 is more than $100 million is worth. I don't know, I don't know. It all looks very strange. And it increasingly resembles a bubble. That is, some already. That is, if you open, I just follow AI, and if you open AI, this batch, well, it's just comical, honestly, it's just comical. Have you seen this cheating tool Kluly, which was initially supposed to hack LeetCode, and now hacking any LeetCode is not tracked. There's such a young eccentric founder. How much did they raise? $16 million from A16. And their software is a wrapper around models and some fork of something that is not tracked in Zoom. That is, I don't know, it's a circus, in my opinion. Well, I think this often happens when something suddenly shoots up, and then doesn't sink. That is, cryptocurrencies also shot up, and around them, there was a lot of expensive and loud stuff, but they've kind of deflated a bit. That is, they are not talked about as much now, and not as many people are interested in them. Maybe they were pushed out in this regard. I love Ed Zitron, you know, Ed Zitron, there's such a guy. It's Zitron. He has, in short, a blog called "Where is your at?" And he constantly grills where the money went, like who gave how much money to whom, who promised what and what didn't work out. And there, well, almost nothing ever works out in terms of numbers, in terms of timings, and in everything. Well, that's also Sam Altman's favorite trick is to talk in very abstract timeframes. That is, he says 1000 days. That's how his timeframes look. He says, "A year," he says, "400 days." Something like that, he likes to spread it out like that. And well, Zitron constantly grills these things. I like that. Yes. And now, let's. There was also some loud news. Maybe you read some guy on Reddit, Nikita, wrote some blockbuster about what's happening at OpenAI. Uh-huh. Uh-huh. Did you read it, did you get into the essence of this thing? Well, yes, there's even a website that opened, called OpenAI cases or OpenAI files, and there's a harsh grilling of where things went wrong. Can you briefly explain what the essence is, what he wrote, and why it's so important to everyone? I just haven't gotten around to it yet. Maybe you, I'm not sure if we're talking about the same thing. I'm talking about OpenAI Files. Ah, ah, are you talking about that too? I'm not entirely sure. I can't right now. Ah, are you talking about this? OpenAI Files. No, I don't think so. It was about some internal squabbles. Yes, that's about internal squabbles. I'll send a link to the chat now. Do you have the chat open on YouTube? Yes. You can see it there. You are also a fan of Alltrade. Yes, yes, that's about the same thing. Well, the joke is that, like, they had these jokes with Microsoft, that Microsoft always believed in them, yes, gave them Azure credits, and as a result, they are now a bit tied to Microsoft, that they have exclusivity on their product, and they want to get out of this exclusivity, and they are getting out of it with a very clever scheme, by saying that Altman hangs out with Trump, with the government, and like AI becomes a key national security resource in the fight against China. And because of this, he wants to manipulate it so that Microsoft releases its initial conditions on exclusivity and on limiting the use of technology. Well, so, in fact, he took money from him, launched, everything is d. And now, like, he realized that he is more than he could ever be. And why, why does Microsoft need it? Well, in the sense that the significance of these companies, I think, is somewhat inflated. That is, when we, as ordinary observers, think that something terrible is happening, they also like to scare us sometimes with such thrillers, that something is happening there. Someone deceived someone. Remember, there was also a loud news before this, already a bit old, that someone, I think Sam Altman himself, they tried to kick him out of there, but something there.
It happened, in the end, they didn't throw it out, and on the contrary, those who were there, yes, he threw everyone out. Although, yes, yes, yes. And and what there is, he constantly tells something terrible, that we are developing something, that they wanted to throw it out almost because he is inventing something dangerous for humanity, and he has spoken little on this topic. And you think at first, damn, some kind of terrible corporations are connected with AI, something terrible. They are all cooking, cooking, yes. Okay, well, it's scarier here. They are inventing something like in Hyperion, there was some kind of internet there too, Warhammer starts, and then some kind of psychic appears or these things are similar, and it's like, well, it's clear that this is an incomplete replacement, it's still a question who learned what, but there seems to be a small corporation and for relatively little money, they made something that is at least comparable to GPT in some use cases. There are people, I often meet them, who use deepseek as their main, main LLM. What do you think about this? And nothing that I'm babbling so much. I just love this gossip and speculation. No, no, it's fine, it's fine. You were mostly silent in the first half while we were discussing tools. You're compensating for your reality show. Yes, yes. What I have, I have this opinion on this matter, that it seems to me this bubble will continue to inflate. And moreover, it is inflating at some absolutely disgusting rates. Here, I will now, but for those who are interested, open it. Summer BCH YC and AI for Language. AI for trees, AI for CPA, AI for this, and I for this, and I for this. I am sure, I don't know, half of these startups are definitely a wrapper around ChatGPT. That is, and you open all the batches, I don't know, of this and last year, and there, each one is a wrapper around GPT 100% or around code. Okay. It seems to me that this hysteria will continue until the first business or serious customer loses money. Like they say Fortune 500. Well, until it gets to the point where some tool is deployed, say, I don't know, WindсёрF deploys something into production and money just evaporates into a vacuum. Well, that's it, and then the market will change a little, I think. Or someone will book a ticket somewhere else. Well, as soon as these hallucinations seep into business, they are not, as it were, on us, right, that we write code and it annoys us that the wrong function is being called or not an existing one. As soon as this seeps out with the help of agents into the outside world, when you order ravioli, and they bring you, I don't know, borscht, then that's it. Well, then, I think, it will start to come down to earth a little. But I don't see it, because in reality, as cool as it is, if you go off to the side a little, hallucinations don't go away at all. For me, they have remained at the same level. From the last thing I did, where it really annoyed me, there is such a thing as hyperlog, you know? No, I don't think so. Well, in short, it doesn't matter. What kind of data structure? There is such a thing as hyperlog. Okay. And it's something like a bloom filter, if you want to put it that way. Okay. And it's quite a well-known thing. A lot has been written about it, a lot has been said about it, there is a lot of information about it. There is a whisper where it is almost step-by-step how to cook it, yes? You write, you write, you write, you write, you come to a conditional, I don't know, copilot, the most, the most [ __ ] there from 3 Pro, or something else. Well, it hallucinates constantly. Well, that is, although this is not sacred knowledge. Well, that is, this is accessible knowledge. It's just that there might not be that much of it on GitHub. And the more you go off to the side, the more these things start to happen. And it drives me crazy. In general, it's like, you need to prepare them correctly. And these LLMs, well, I just often, when they talk about hallucinations, sometimes the first thought is like, well, it seems like they are not there anymore. Well, I understand that this is a deceptive thought, and most likely it arises simply because I constantly use it at work and am very used to it. That is, where it will hallucinate strongly and where it is bad, I already, at the level of, I don't know, muscle memory, just don't lead it there. But if it's still necessary, I'll lead it by the hand, step by step. Uh-huh. That is, I don't perceive it, well, for example, for the sake of certainty, copilot, I don't perceive it as a replacement for myself and as a full-fledged gen, for example. I just know where it will cope well and where it will cope badly. And I understand that in many tasks it will do better than me, well, not better than me, but at least faster than me. Okay. And regarding hallucinations. Well, even if, by the way, even Anthropic themselves wrote about this recently, they released another document on how to use copilot. They perceive it themselves as a slot machine, that if it doesn't work out, try again, then maybe it will work out. But even if you look at it that way, just last time Anton was on my podcast on the same topic, he is very skeptical about it and says, "What's the point of all this if I have to check it and it doesn't always write good code for me?" The point is not that it always writes code for you, but if it writes good code for you at least 50% of the time, it will save time. Well, for me, it's even more. That is, well, you write a task, you send him a description of the task, maybe even simple code. In the sense, you have a working task, something simple needs to be done, but you need to go through the whole project, understand it. You send it to him, and in most cases it will do it well. That is, you just go through it, maybe make some minor edits, I don't even take them seriously. In 10% of cases, it will write complete nonsense, you know, well, in 10% of cases, you'll write it yourself or just guide it in the right direction. And, and in terms of edits, it's like, well, for example, I see right away that it's using the wrong framework, for example, the one we use. We have, for example, an internal one, and it uses some external one. I just tell it: "Use this." It says: "Okay, everything." And well, it's a matter of seconds. That's how it is. But I have, well, yes. Yes. Uh-huh. Well, in terms of some kind of full-fledged use, what you're saying, that it won't order the right thing, that's, in my opinion, OpenAI has such experiments in ChatGPT called operator, and it's available on the Pro subscription. It's probably available on Pro not because it's expensive and difficult, but because they don't want to release it to a large audience. I even, well, I have a Pro subscription now, I bought it for a month, the one that costs $200, just to try it. I don't even consider the operator. I'm afraid of such things. Like you said, I don't just allow it, I'm almost sure that it will do complete nonsense for me, order something wrong and call the wrong thing. And in general, I don't argue with that. Well, the question of applicability. Just the last case that greatly disappointed me, I just gave up on it for a while. I needed visualization for videos when, well, you have a buffer pool in memory, right, and you have files on disk. I wanted to do a visualization when, well, I couldn't find a ready-made one and it was too lazy to draw, when your buffer pool overflows, that some dirty pages can be, you know, pulled, flushed, and written to disk. Yes. I wanted to, I just described the prompt, and I wrote the prompt with another model, like, describe a prompt for me to do this. They can do these, like, almost all models can do these applet artifacts now. I don't know, well, like, I asked it to rewrite it 15 times and each time it was garbage, and I got so angry that I gave up. Yes, I drew it by hand on an iPad. Full-fledged whole coding, in my opinion, is still an illusion, because here with pictures, and with code, you can't always get the model to do it without delving into the code yourself. That is, I have Gosha Twist, for example, I have a utility for a video for a planner. I just don't understand the TUI, the terminal UI very well, and I didn't want to delve deeply into it, so the interface, this thing, well, copilot wrote it back then, there was no copilot yet, just through chat, but it doesn't matter. At some point, it, well, it did something similar to what I needed, yes, and then it started fine-tuning. It just says, like, change the font here, for example, change the window size here, and it, well, it did some nonsense, that is, it changed it randomly. I had to intervene. And yes, well, a guy who doesn't understand code at all, he would have his project fail on this, he just wouldn't be able to make the model fix something. And if you don't understand it yourself, then that's it, you can close the startup. But when you know how to do it yourself, well, damn, it's a 5-minute job. You look, well, it couldn't finish it, you finish it yourself, change the size, the pixel grid, or something else. Well, yes, in my experience, in reality, working with agents goes best when you yourself have described the project framework, you yourself have something similar to an existing module, you already have it, you wrote it yourself. When it is in an environment where, well, where there is something to rely on, then the results are much better. I've tried it in different ways. I tried to have it write something from scratch, and I tried when I wrote the framework myself, organized everything into folders, described the architecture nearby. Then it works much better. So here, you can use it as a full, well, a full-cycle tool, that is, write everything completely from scratch, but only once, that is, write it and then don't touch it anymore. And here there is also a paradoxical thing added that you have a complex task, and you entrust it to it, it fails, naturally. Then you try to break it down into steps. For example, as you said, write the framework yourself. But the funny thing is that it can even write the framework better than you if you give it such a task, like, make me a clean framework. And explain briefly, like, what parameters are important to you, like clean code, architecture, something like that. And the funny thing is that, in the end, you will still do it more effectively if you do it in pairs. Sometimes it gets to the point where, I don't know, you realize that you need a good prompt to do something complex, and you entrust this prompt to be created, perhaps by the same, the same model, and then you feed this prompt to it so that it performs according to the instructions. That is, you can say it will finish, what automates. Well, that is, it's clear that what's happening now, it's more or less the same. That is, it thinks for itself internally, m, and only then gives the result. But there's not enough of it yet. I often arrange consensus between models. It happens that, for example, Gemini 2.5 Pro, it is very good at detailed analysis. That is, if you ask it to tell you something, analyze something, it will tell you everything accurately and even more. Reading it is impossible, because it writes such texts that you won't. But summarizing is very bad. That is, if you go to summarize, it will throw away a lot of useful information and leave a lot of unnecessary information. But it's good for generating a prompt or some kind of technical specification for another LLM, which you will feed. And it will already, for example, either do copilot, make a summary for you, or write code for you based on this prompt. Okay. And it happens that when I research some complex problem, I just copy between models, well, just in chats and say, like, tell me as much as possible about this. It tells, I send it to another, like, you tell me too, and then look, what did you think, what did Gemini say. And then I show it to Claude, look what Claude said. And then I show them, like, look what Gemini said about your previous result. And they start to communicate a little with your hands. That's how it is. That is, they agree with each other somewhere, and somewhere they don't. That is, Gemini won't say, "Damn, here, of course, he criticized me, but he's wrong, because I meant this." You show it again, it says, "Well, yes." At some point, they agree and find some consensus. Well, I also sometimes connect O3 or even O3 Pro to this. Well, it's a very tedious thing, takes a lot of time, but the results are cool, I like it. I do this in technical tasks, and even when, for example, I was just curious, I have two dogs now, we picked them up on the street as puppies, and it's interesting to see how they grow. That is, we measured their weight periodically. And I uploaded the information on weight and asked, "What do you think, up to what size will they grow?" And at first they diverged greatly. And it's also funny that it's like they were trolling each other, when I was cross-pollinating them. For example, Gemini and ChatGPT agreed, but Gemini bent a little under ChatGPT. He had a different opinion, and then when he saw the other's opinion, he sharply changed his own, like he re-educated himself. And Claude subtly noticed this, saying, like, pay attention, although he gave such a result, he thought completely differently at first. He agreed with GPT too quickly. And this thing is interesting, one of my interesting ideas is to automate all of this. That is, to write some kind of utility that will arrange this consensus itself. That is, three models sit in a chat and communicate, and each one on point. Moreover, you can even fine-tune it to understand at what moment which one will connect, so that Gemini first collects detailed information, then Claude compacts it, and O3 thinks about it, gives expert opinion, because each has its strengths. That's how it is. But you don't have this thing. Oh, I'm absolutely right. This really annoys me every time. Oh, I'm absolutely right. Depends on the model. Well, there is that, of course, yes. But I don't like that they don't have a backbone. They often bend. It seems to me that even if you constantly tell them incorrectly that no, you are wrong, I am right, at some point it will bend and say, "Oh, I'm absolutely right." In this regard, they say Gemini is more like it has a backbone. That is, it's often even difficult. But, by the way, about Gemini, I heard an opinion that it's difficult to convince him. That is, you tell him to redo it, and he says he won't, because it's correct, something like that. But in general, Kolya, you just called it a startup. I think you should do it. Make a thing for consensus. I don't think I'm the first to think of this. I had that thought too, but I haven't seen anyone do it yet. I think everyone thinks it's cool to do, but no one has done it. So let's go. And the second minus is that it costs money, damn it. I mean, you'll get A16, I don't know how much Oleg will give you. I mean, you'll have three models running these huge texts and coming to a consensus. And how many tokens will that consume? Probably a lot. Maybe more than those $1,000 a month that copilot eats. And big players, I've also heard rumors that when you pay for a subscription, you're kind of stealing from them. Oh, well, yes, there was a subscription model. They say in that sense that, um, their monetization goes into the plus because not everyone uses it to its full potential. That is, if you run tools on a subscription as much as the limits allow, then you go far beyond the cost, beyond the cost price. That is, for example, if you pay $200 a month for copilot, then you can earn $1,000. But the point is that not everyone runs it like that. And it's compensated by those who use it little. Or maybe they're just working at a loss to gain popularity? This is the valley. They print money, so to speak. How many years was Uber unprofitable? Well, I don't know, billions, for decades it was unprofitable. They don't care about money at all, it seems. They just print it like this and that's it. Well, I care. That's why it's hard for me to launch such a startup. And when you go to YC, they'll give you credits like crazy. So I don't know. At least do it as a proof of concept. Do it as a proof of concept. If it's cool, then you can do it so that users just bring their own keys. Well, well, as an option, yes, as an option. Well, okay, you convinced me, I'll launch the startup. Then you'll buy me a watch. You'll invite me to the network there. Okay, everything, agreed. Write in the chat who else to invite where. And I'll be the, what's his name, what's his name? Evangelist. I don't know. Well, in short, some kind of air-filled, air-filled. Angelion, yes, some kind of air-filled vacancy. No, really, Kolya can get in, honestly. Well, like, now it's such an environment that, like, try open source, not well, especially when you first make a tool for yourself, and then it turns out that many are interested in it. You felt the pain and conveyed it. Okay, accepted. So, okay, let's move on. We also had an article on the agenda that I also haven't read. It's, well, it's a bit outdated, but it was discussed a lot. An article from Apple where they criticized that this Illusion of thinking, that these LMs are not that smart. Well, Vlad, you said you read it in its entirety. And Darr said he summarized it through LLM. And I only read the title. Can you tell me? I've been talking too much. Ildar, you start, I'll finish. Okay. Well, the essence of the article, as far as I understand and understand ChatGPT, is that, well, one of the main conclusions is that models, they actually, even reasoning models, well, they don't think, they are just, well, a more advanced word generator, well, a sequence of words. Okay. They found this out by giving various tasks to various models, both ordinary and reasoning models. Okay. There were interesting, unexpected conclusions from the category that reasoning models cope better with more complex tasks, but they start to get stupid earlier as the context fills up, I think. Okay. And there was also another article that is a response to this Apple article, which largely criticizes their methodology. That is, it doesn't criticize the theses, but it criticizes the methodology. That is, for example, sometimes they gave models tasks that were impossible to solve in principle, at least there, and in this response article, an important point was that it was written almost entirely by Claude itself, right, something like that. Well, in conjunction with Claude or with another AI, yes, it was written. And who wrote it? Some company or Oh, I'll tell you now, they're called something like Lowson something Lowson and Co, I think, right? Oh, okay. Just not Anthropic itself, and not just some random guy. No, wait, there's like, wait, Alex Lowson is not a company, he's from Anthropic and somewhere else. That is, he is affiliated one way or another, but the original article itself is not about the Anthropic model, it's generally about LLMs, as far as I understand. Okay. But I'm interested in the thesis itself, that like models, well, don't think, but just generate text. This thesis is interesting to me because, well, we don't really understand what thinking is. Perhaps thinking is just generating a sequence of words or images. Therefore, we need to decide on this. And therefore, this thesis is, well, debatable, in my opinion. Uh-huh. Well, yes, I agree. Well, regarding the fact that, well, I often hear the opinion, and in some sense it is true to some extent, that LLMs are just like advanced T9 that generates text for you. Does anyone have anything to say in defense? In the sense of refuting this statement or supplementing it in some way? Well, we ourselves, in the sense, our brain also works in a similar way. There is such a hypothesis that, well, that thinking goes a little from language. Okay. And, accordingly, well, that is, people who speak different languages think a little differently. The approach to thinking itself can even differ. Okay. And, accordingly, well, perhaps the process of thinking is the process of generating a sequence of words in many ways. Okay. But this is a hypothesis, that is, we don't know how close it is to the truth. Well, we have something multimodal in our heads for sure, because there are pictures and sounds there too. Well, of course, yes. Well, here, modern models are also. They are actually multimodal, so But unfortunately, my context window is too large. I remember too much garbage. Yan LeCun, for example, says, Yan LeCun, for example, says that this is a dead end, a dead end branch in general for LLMs. Well, how right he is, no one knows. But Yan LeCun is a respected person. Okay. Do you have anything to add about this article in general? Well, the main criticism of this article, which is Illusion of thinking and Illusion of usionusion, the main criticism of this article is the metric they used. Their main metric was that they proved that it doesn't think by the fact that when they increase the complexity of tasks, they saw a dip, that is, yes, as the task becomes more complex, and at some point reasoning tokens, well, these internal ones, it cooks it itself, they grow, as they say, accordingly to the task, yes, they correlate, and then at some point, when the task becomes too complex, there is such a plateau, and they start to fall or even lie at the bottom. Well, that is, it seems to stop using reasoning tokens, stops thinking. And the criticism is that, well, is this really proof that it doesn't think, maybe it has already figured everything out, and it doesn't need so much reasoning, and plus, the prompt is not entirely clear initially, and well, that is, the fact that you use more reasoning tokens doesn't mean you think better. Okay, that's the main criticism of this paper. Well, yes, here I haven't read either article, but I immediately have something to say, that like, it seems our brain works like this, that when the task is too large, the context is too much, then we don't think straight anymore and we need to somehow either compact it or break it down into subtasks. That is, when you have something large, a lot of information, you try to summarize it first, and then work with these summaries or break it down into subtasks. So, more or less the same. Well, I think, here, here's what I see as the most important thing, that it's not the same, and that, I don't know, maybe a startup will launch, that, well, we still have a "cronjob" in our head, something built-in, I don't see that. There's no such thing as, I ask a question, yes, I, for example, I ask something, it tells me some nonsense, then I go away for a week, it thinks about it thoroughly, I come back to it, and it says, "By the way, I thought about it, so your question from a week ago turned out to be like this." Well, like, well, there's no such case anywhere yet. Well, I think we have something more like an agent inside, not just a pure LLM, if we're talking about an analogy. Well, maybe, but I honestly don't know anything about the structure of the brain, so I only know that my head hurts, that's all. That's how it is. Well, yes, in general, it's self-evident. Well, wow, they wrote 30 pages there, of text with a lot of, like, smart graphs, yes, basically, they gave them the Tower of Hanoi. Really, a task of moving the Tower of Hanoi disks. Well, in terms of whether it's an illusion or not, whether it thinks or not, it doesn't matter. What matters is how applicable it is. That is, like in physics, you can criticize Newton's laws and say that they are, well, they are incorrect, they are wrong, they don't account for some quantum moments. But here's the question of applicability. That is, if you throw a football, then Newton's laws are generally sufficient to describe it, oh, to describe its trajectory. But if you launch a proton, then, well, there will be questions. Or vice versa, if you have some large planet flying, then it's also very difficult to be guided by Newton's laws. So maybe it doesn't matter here that it doesn't cope there and there. What matters is that it copes here and with something specific. And what causes concern is that at first it was fun and cool, like, look, my copilot, oh, copilot wrote a short function for me, and wow, it understood what I was coding and even wrote the correct function. And then, a few years later, you launch an agent that writes your entire project for you. Well, maybe under supervision, but it can already do it with difficulty. By the way, I kept forgetting, I didn't have time, while you were talking, the thought was in my head. Last time on the podcast, I said, and I'll say it again here, about complex tasks and low, well, small context. I conducted an experiment with Iskander. Iskander Sharibov, who is also quite a well-known guy in the Go community. He's involved with games now. He has a game written in Beat Engine. A Go game engine. What's it called? Don't you remember? Ildar, what's the game called? Depends on which one. He has several, as far as I remember. Well, the first one, which he has already released. By the way, it froze. Yes. Yes. Roboden. The picture is something. Oh, it's alive now. In general, we discussed it with him too. He was also very skeptical, especially considering that Ebiton is a very niche thing, and of course, there is very little material about it. And practically, well, skeptics say that the model will only do what has already been written many times. That is, everything we do, we run routines, thousands of people have already done the same thing for us. Of course, it will cope. But try to write something new, that is, on Ebiton, that is, your game, most likely, it will be unique in terms of many aspects. And there isn't much documentation, not many examples. And out of curiosity, I just asked, like, give me a few examples of serious, complex tasks for copilot to do. And he left, well, he gave me several options. I chose the one that would be easier for me to test, so that he would have, like, strategy and these health bars at the top. He wanted me to attach them. They weren't there at all. That is, there was just a variable where they were stored, but there were no health bars. And I asked copilot to think. It took him about half an hour to do it himself. I just gave him permission to edit all the files as he wanted, and these utilities that search through files. And he fully implemented this feature. I'll also post a link to it. Then, after the podcast, I formalized it as a pull request on GitHub. And, well, he himself looked at it, he liked it. In general, he said that he would have spent roughly the same half an hour himself. But the fact that the model replaced him in this regard is already a little frightening. Uh-huh. And it's getting better and better. That is, version 37 was generally good, and now it's even better. That is, well, when Opus 4 came out. And indeed, it can take over a much wider range of tasks from me. That is, sometimes it's not a joke and I'm not exaggerating. I go into Jira, copy the task text, paste it into it and say, "Do it, even, well, just do it." The entire prompt can already connect the official MCP from Otvasin, in fact, without even copying anything. Especially, yes. By the way, I tried to connect something, but gave up. I think I connected it, but never used it to update the task status. But that's okay, I'll poke the status for it, move the card. I can do that for now. Okay, I'll read the comments, I got lost a lot. They write that the person sees three favorite YouTubers at the same time. Thank you, John Bee, for your pleasant feedback about all of us. Well, in general, I trust it to think. And in principle, when the task is not difficult, well, that is, many know that when we have things like DDD and so on, there is a lot of boilerplate. That is, when we transfer data between layers, like the service layer, transport, database, we need to convert it back and forth, write some mappers. And sometimes it's even difficult for me to keep it in my head. At some point, I forget where which handle was, on which layer it was, where to get the model. In general, my context, Vlad said that he has a large context, I have a small context, and I get lost at some point, I compact it and remember what I was just doing. And here this thing makes me better than me, because it quickly finds everything, quickly does everything, and the task is ready. I just have to drink coffee while I watch it. And where will this lead? Well, like, it won't lead to the fact that in the end, we as a layer between, say, a business analyst, or a product, and between the code, disappear. Okay. What do you think about this? Well, probably, we specifically, like, our generation. No, we, in the sense, programmers. Well, yes, I'm talking about our generation of programmers. Our generation of programmers, most likely not. And then? Well, yes, I think so in general. Well, here, it seems to me, we need to clarify who we mean by programmer, because, like, usually a programmer, if you are a good programmer and if everything is built correctly in the company, is a person who solves problems. Okay. And in principle, business doesn't care how this code and this information system was developed and launched. The main thing is that it works normally and can be developed normally in the future. Okay. And, at least, it seems to me, yes, indeed, well, maybe 15 years, maybe 10 years, it will still be okay, because it's still something that works under strict supervision. And there is, in reality, a lot of context, in a broader sense, that is difficult to express in words. That is, it's non-verbal, difficult to formalize, but which greatly helps us to work. And and this is, well, probably the main obstacle for AI to replace not text typists, but full-fledged developers. That is, for example, like, you need to make a feature, but I know that the team there usually does it in a tricky way in this place, but they don't document it, and here, this feature is being done in parallel, so I need to take this into account. And all of this is, like, not written in one place. You can't just go and get it through MCP or some other similar protocol. Well, look, this is purely a question of context and capabilities. Now, of course, it will be very expensive to automate all of this fully, but if you have, for example, an bottomless context and it costs almost nothing, then you can, because, well, you saw that, for example, in some DevOps chat, they already fed the context of this notebook LM. I haven't heard this news. They wrote at work that, for example, you can chat with an LLM, ask a question, like, based on the chat where DevOps communicates, and many questions will already be filtered out. I don't know, have you used it? Well, I haven't had any problems yet, so Well, yes, me neither. Well, that's what I mean, technically it seems possible, that is, to unleash this model so that it reads all the conversations in the chat, in Slack. Well, obviously, a huge context is needed to understand this. Unleash it on the code, then give it access to the pipelines, so that it writes code, submits a request, formats all the tests, runs the linter, and deploys to production. Then give it access to monitoring. There's also a lot of context needed. That is, maybe even migration isn't needed, it just needs Prometheus and logs to check that everything is deployed normally, and also give it, well, so that it writes tests itself, like integration tests or various unit tests and so on, and give it the ability to test it visually. That is, like Selenium. Now there's some other thing, I forgot the name too, what it's called. Magnet, something like that, came out. A new thing. That is, based on AI, something similar to Selenium, it tests only with AI, and they say it runs directly on your local machine. Give it something like that, and it seems like that's it. It replaced you. The only difficulty is that, for now, the probability is very high that, well, I sound a bit pessimistic now. In reality, I don't believe in this at all. Maybe in 50 years, I don't know, in some 10 years, this will happen. Now, the probability is just very high that it will stumble somewhere in the middle, or even earlier. Well, yes, I think the ratio in the market will just change. Well, that is, at some point website builders appeared, and they didn't destroy web studios and digital agencies as, in principle, as a class, but there became significantly fewer of them. Okay. Here, I think the same can be expected, but also, that is, well, there are still about 10 years left, so it's not too late to learn programming and so on. I think everything will be fine. Well, here, first of all, not only is it technically impossible yet, because the context is small, but even if you try to optimize it with the current million-context Gemini, it's a huge job. That is, to fit everything I wrote and be able to run it. You need to try very hard. And most likely, it will take much more time to automate this than for the automation to reduce your time, it will just be more expensive in the end. Yes. Yes. Automation will be more expensive than what it automates, than how much you save on it. Okay. But the main problem is that if there is no control, it will likely degenerate. Well, the project will just die, because, well, at these stages now, there was a loud case where some guy wrote so happily on Twitter that he launched some kind of startup. I didn't delve into it, you've probably heard about it. And then he started writing that he's closing the project because they messed up. Well, there are guys at the level who ask, "Can I somehow roll back what the LLM did?" Well, that is, guys don't even know Git. M, well, this comes down to the fact that in the end, yes, without control, it goes too far and too quickly in the wrong direction.
To that degree. That is, there can be complete bullshit, until you stop it. Uh, you lead it by the hand, it's wonderful, you let go a little, everything, everything is bad. And there were also discussions on the topic that maybe models don't need clean code, or any code at all. Maybe some optimized code for the models themselves can be invented, and let them run it a bit, well, maybe it's not zeros and ones, maybe something is optimized specifically for LLMs, but practice shows that it's nothing. They perceive markdown best, which we also perceive as humans, they perceive clean code better. And at work, in projects where the code is clean enough, they cope perfectly. But recently I dealt with a service that was written very badly, and it stumbles almost immediately. That is, I gave it the simplest task and left it to think, because, well, I hadn't yet assessed how bad that service was and thought that I would conditionally sit down, drink coffee, and wait. And it produced such the most furious nonsense that, well, um, it was impossible to use it, I just deleted it and started to delve into it myself and optimize some individual micro-moments with the help of an agent. And it would seem, write clean code, then everything will be fine. But it doesn't write clean code very well. If you tell it: "You wrote nonsense here, fix it, right?" It will say: "Yes, I know how to do it well. I'll do it now, it will be good." And it will do it well. But if there is no such person, a person who says that it is bad, uh, um, then it will be bad. That's it. And the models themselves show that they cope poorly. That is, I also had moments when this conditional cross-pollination with LLM consensus, I asked them to write complex code and drove it back and forth between them, asked them to make it cleaner, prettier, more optimal. And I didn't even delve into it much. But they made me so happy. For example, one wrote, the other sent it to O3, it optimized it, and everyone like Claude and Gemini admired it, saying, "Damn, O3 optimized it so well." Now I'll tweak it a bit more, you send it to him, and he says: "Damn, that's great, make it even better." And you're already sitting there, impatiently rubbing your palms, thinking: "Now they'll roll out a masterpiece for me, and this [ __ ] won't even run." And when you start to delve into it, it's complete nonsense written there. That is, what I'm getting at is that LLMs themselves can't do it either. That is, none of them will stop like a human. Uh, well, that's what I said, that an observer is needed who controls the process. And maybe someone will think that this observer can also replace it, but no, it can't. In my experience, it can't yet. For now, it will only make things worse, confuse everyone else even more. That's it. Well, and one more. Well, okay, for now, what do you think about this, I'll elaborate on this thought a bit more. Well, I have a banal thought that, well, if you take an ordinary error, right, well, let's say you have a 5% error on one event, and then two events, then three events, then four, then five, then then 20, then 30, then 40. But it just starts to multiply, and your error percentage grows and grows and grows. Well, and like, where is the plateau up to which it will grow exponentially, so that there's nothing adequate left. And the second interesting point for me is that all of us who are sitting here, we have worked all our lives with, well, like in our brave new world, with Lego companies, right? That is, our companies were founded before all this tooling became so popular. I don't know about Splunk, but I think it's right. You launched before everyone started using cloud code, cursor, or all that. Yes, yes, well, yes, yes. Just, I think, Chat GPT 3 with then was and something else. Well, yes. But I'm interested when companies appear that are like, well, their first development cycle and in general, they grew and matured, and they started with like LLMs, and they're bombing with LLMs. Well, like, how will their codebase look, how will they resolve tickets, escalations. I think that's how LLM-first companies will look, which started, how will all their affairs be. I don't think so, because sometimes you take some task from Snyk, right, and you see, well, from what a person wrote, it's hard to understand. You might think an LLM wrote it. Well, I don't know, maybe, maybe everything is really bad. Well, that's how it is, really bad. I've encountered services where, let's say, I guessed that a lot was written by a model and a person reviewed it poorly, and it's very bad. Well, yes. Well, so, so the fate of these LLM-first companies is interesting, which are like, they were initially optimized and like their main codebase is generated, they are written by people for us. And, well, yes. There was such an example when, uh, I was doing my task and accidentally fixed a conflict incorrectly, because of which, well, it was somehow missed. And after it was rolled out, everything exploded in the end. Well, not everything, but in that part of the system. And in the end, to understand what was happening, you needed to know that. Before that, the same thing happened several months ago, and it happened precisely for this reason. And this, precisely what I was talking about, that this informal context helped a lot. That is, probably an LLM, if it sat there, given time, and directed correctly, it would also come to this in the end, but it would take much longer to get there. And it would ultimately be more expensive in fact. And you don't have such a moment, Eldar, that you've seen, I've often seen this in tests, that it fixes one test, and for it to pass, it breaks others, like. That is, it tunes the results for it. That's also a painful point. Maybe with Snyk or something. I've encountered that. I first heard a lot about it. For me, it's almost become a classic. I thought, when will something like this happen to me? Recently, it happened that, well, it was my mistake in the early stages of working with Code Llama, when I hadn't yet figured out where it was good and where it was bad, and I gave it a too complex task, obviously, which it couldn't handle. And it started then to fix tests. And it started with, like, it thinks, thinks, thinks, then says: "Now I'll temporarily comment out this text, this test, uh, and then, well, let's see if it works or not." And it did that, runs it: "Oh, great, the tests are passing, let's fix the rest too." It comments out the rest too. And then it writes to me: "Happy, like, I fixed it, everything works." My favorite is when it creates a special environment variable so that when it's enabled, the function you're testing behaves in such a way as to pass the tests. Yes, I've heard about that. I've heard about that in other podcasts. But here's something else fresh. Well, it's a bit different, but it was also funny. I had a task where our contracts were updated, and some of them were deprecated, and it was necessary to update the gateway, which, in general, connects services, let's say. And it first thought that all these deprecated methods needed to be deleted. But I wasn't really watching closely, just with the corner of my eye. I just wanted to see what it would come up with first, and then intervene myself. I gave it too much time to think. It, in short, deleted all these methods, and then it turned out that it shouldn't have, because it breaks. That is, the interface and implementation are mandatory. It says: "Okay, I'll fix it now, but I won't undo the deletion, but I'll just make a stub for them." Well, like, it brought back these methods, rewrote them, and they just called a panic, like, not implemented, or something like that. Well, it's like it did a completely useless job, which only made things much worse. Yes. Ah, and regarding that, I'll expand this thought to the end, regarding what we will come to. There is also such a moment that there is an opinion that their development has already hit a limit, because they have studied all the information that is possible. That is, they first learned on GitHub, conditionally, then they started learning on commit history. That was also a great idea, a breakthrough idea, that you can learn not only from current code but also from how it evolved, which might even be more useful. They read all possible texts, articles, everything that could be. And then they come to the point where they do synthetic training, that is, they learn from each other. The model itself, either on itself or on other models, and, well, the quality starts to suffer. And to make it much better. That is, we are afraid that they are getting better, and maybe if they get even better, then that's it, like, we are not needed, but it seems like there's nowhere else to go. Possibly. But maybe the development will go differently, that LLMs specifically gave a big boost to this technology as a whole. Maybe, now, relatively speaking, I heard that these diffusion models, which used to generate images, can now generate text. That is, maybe Vlad can provide more context. No, that's completely beyond me, beyond me, yes. Well, in general, there are diffusion models that generate images for us. Uh, well, through noise, which works. Well, apparently, yes, and they are much faster for text generation. And they are currently being tried. Well, I'm not a specialist in all this, so I can't explain. There's also something, it was also called by Google. I also talked about it on the last podcast, that they invent algorithms using genetic algorithms. That is, well, to simplify, this thing is given a task for which an algorithm to solve it needs to be written, and it just randomly tries and attempts to solve it randomly. And, well, cleverly, using genetic algorithms, again, if you want more context, watch the previous podcast. Gleb explained in detail how it works, these genetic algorithms. Well, to simplify, it's not just random wandering, but a bit thought-out. And essentially, like this, out of nothing, by randomly trying, it invents an algorithm that is, well, a breakthrough, for example. Even a person wouldn't invent something like that, that it's very fast. Maybe it's very unclear, but it works very fast for some reason. Well, maybe it will go in this direction. And what I'm getting at is that even if we've hit a wall now, something might happen that will change the game in another direction, and it will work even better there. Well, Gleb also touched upon quantum computers last time, which can also give some boost. Well, we'll wait. But there will be a moment when LLMs will write better in Go. Kolya, do you know when? Guess? What? When you release a book, it should immediately be released with protection, like, AI, robot, something like that. So, maybe I'll write it with AI too? That would be synthetic training, essentially. Well, I think here, like, this limit regarding training data, it seems like it was reached a long time ago, a long time ago, in terms of AI development. That's several months ago. And I think the competition is not only in how and on what to train, but in the architectures of all these models. And, well, most likely, companies will compete in architectures, in how this model is structured. That is, perhaps LLMs are indeed a dead end, and they will invent, and all these transformers, and invent some new approach. That same diffusion might get its development specifically in text. And then, well, everything will continue with renewed vigor. Well, at least, well, there are no signs yet that all this is ending. On the contrary, it's accelerating. They're writing in the chat that strong developers don't want to use it. Apparently, LLMs for code, like, why? Well, apparently, we are not very strong developers. So be it, but maybe there are better ones. Is there anything else to add? Any, maybe, questions? That which we didn't manage to cover. Today, surprisingly, we went through it. In our podcast, there's usually a list of topics, and the conversation gets so scattered that we don't even cover half of it and that's it. And even the format stretches to 3 hours. Last podcast was such that we stretched it to 3 hours, maybe even a bit more, but we went through the whole list. Now we have about 2 hours, as I planned, and we've gone through everything I wanted to discuss exactly according to the timings. And that's wonderful. There's probably some experience. I'm getting my hand in, how to host a podcast. Maybe next time it will go just as well. Does anyone else have anything to add? Il. Well, uh, the only thing I brought up was the topic that you can look at the experience of other industries, what's happening there, and try to apply it and understand what will happen to our industry. And in general, in all industries, at least where people's work is mainly intellectual, the trends are roughly the same. Tools are also developing there, maybe not as actively, but, for example, there's a lot of movement in music right now. That is, even Daniel Ek, who is the founder and CEO of Spotify, said that, well, soon we will all come to AI and AI will write all the music. And against this background, they even cut payments for not very popular musicians. Well, and it seems like on the one hand, yes, that is, if we talk about purely functional music, about music for elevators, for YouTube background, for advertising, and so on, then people who do such music are at great risk. And if, then, well, it's probably not functional music, but commercial, it would be more correct to say. Well, and if we talk about artists who tour, whose music is still sold on vinyl, and so on, then I have big doubts, because, well, when fans follow some artist, they follow the personality and go to concerts also for some personality, and they find it more pleasant to perceive that there is a person behind this music. Well, and I think here you can draw a parallel with IT as well. It's clear that there's less creativity here and more engineering, craft, but still, it's easier for us to work with people anyway, and also, that is, well, people who just sit on fl.ru or on Upwork and do some simple things, well, I think they are under threat now, under great threat. Well, and those who solve business problems very effectively, well, that is, in relation to the money invested in them, or create some cool engineering products, technologies, and so on, well, they will probably be fine. Well, observing neighboring industries, these are roughly my thoughts. Everything is roughly the same there. Vlad, do you have anything to summarize? M, yes, thank you for inviting me. It was great to sit here. Sorry if I sometimes talked a lot about gossip. No, you actually talked too little in the first half. So, so what I wanted to say? Well, look, there are things in life that depend on you, and there are things that don't depend on you. Well, and like, the development of God from the machine doesn't depend on you. And like, if you sit and think about it and worry and do nothing, then, well, nothing will change in your life. And there are things that depend on you, and maybe something will change in your life. I don't know, you'll start learning something, studying, maybe software development, maybe you'll do something else. like, you are responsible only for what is directly under your influence, and, like, God from the machine is not under your influence. Therefore, if you think about it a lot, nothing will really change in your life. So do what is within your power. If you want to learn, if you want to, I don't know, become a developer, or something else, learn, become one, if you don't want to, don't become one. That's about it. Well, yes, I agree. While I was listening to you, I wrote down a couple of theses to summarize myself. Regarding, regarding whether they will replace us or not, I look at it absolutely positively. Well, because they won't replace us, but complement us, let's say, because, as Vlad said at some point, the smart word is exponents, exponential complexity. That is, when I layered all this, one on top of the other, and said that you need to write code, you need to predict slack, everything, something else, it's like, the complexity grows exponentially. That is, it's not just adding the complexity of analyzing slack and the complexity of writing code. It grows non-linearly, in my opinion. And the more of these components you layer one on top of the other, and you want to deploy your startup FPO at some point, it will be such an incredible complexity for implementing this entire automated thing that, well, not in our lifetime. And as an example, you can also say that, well, Vlad disappeared somewhere. I don't know what's happening. I see myself on the stream, but my browsers have frozen. Vlad, can you just try to re-enter through the link I sent you, and I think everything will be fine. Hello. Uh-huh. So, can you hear me, sorry, I don't know what happened, but I managed to finish my thesis. Yes, yes, yes. Okay. You just froze for me, that's why I don't know what happened. Sorry. No, you finished your thought, I went on to tell you. And now you'll have to. Don't disconnect from here later, because Versai needs to load your data, well, probably that recording it won't load anymore, my browser died. Well, now we'll check, we'll see. Okay. So, in general, uh, as an example, also when, I don't know, everyone thought that finally high-level programming has appeared, that you can build such complex systems that it takes your breath away, and some complex games. And now there's 4K gaming and all that. And the complexity of developing games, for example, has increased so much that, um, even large companies can't afford to finish a game, just bring it to completion, for example, with the same Cyberpunks and so on, right, which had some Napoleonic plans, but even, I don't know, they threw out half of what they planned, and it still didn't work out, it still turned out badly. Well, in the end, they polished it, of course, but I'm saying that we often underestimate the complexity when there is some exponential growth, and all these components are interconnected and interact with each other. And I'll also remind you that once, I don't know, maybe someone remembers links, proofs, I can't provide them, there was some committee of scientists who wanted to create artificial intelligence many years ago, and they set themselves an ambitious goal. Everyone was sure that they, well, I don't know, it would take a maximum of a week to develop a full-fledged artificial intelligence. Well, I think they agreed that they had 10 years, and in 10 years they would build something. Well, initially they had an ambitious goal that within this session, they gathered that they would invent something during this time. As far as I understood. Well, yes, even 10 years turned out to be not enough, apparently. And it will be like this here too, but I look at it absolutely positively, because what we see, first of all, it speeds up a programmer's work many times over. Well, in my opinion, even the most pessimistic views say 20%, but I'll say many times over, and, well, in my experience, that's how it is. And another important point is that it eliminates some burnout. Well, this is purely my experience. Maybe someone else has noticed it too, that at some point it becomes so awful to work, because you spend, I don't know, almost 90% of your time, even if it's 50%, it's still a lot, on some boilerplate, some routine things, and now it's perfectly taken over by it in my case, and you're left with only something interesting. And that's very cool. Burnout seems to be pushed back to a very distant future. And I experience some kind of childlike delight sometimes, even in pet projects, I have several pet projects that I've been postponing for years, and now I've returned to them and started to finish something, because it became more fun. That is, you don't have to remember for a long time or figure out how this library works. You just ask the LLM specifically, check that it's right, and that's it. And the process goes much faster. Much less friction, and the feedback loop has become shorter. Yes, that's true. Yes, you don't have to Google much anymore. And regarding replacing, well, they recently released a cool thing that the artifacts that, that is, conditionally, it can generate an application for you in JavaScript, you can run it directly from the chat. It used to be like that, but now you can also share this artifact with someone, and they can use it with their tokens, if they are needed at all. That is, now you can do it this way too. They are actively pushing this direction. Somewhere it will become easier. That is, some simple freelance tasks, as they correctly said, LLMs will perfectly replace them. That is, conditionally, some owner, as I'm thinking of right now, the owner of a shawarma kiosk, will be able to write some simple things for himself, like a website, a mobile app, anything. Here, in principle, he won't have time to hit the point where everything goes out of control. He will get some result. But more complex projects, no, but still, the trends are still positive for me, because it will be much easier to get into it. That is, if when I got into IT, I even studied, well, I wrote posts about it, well, it was very difficult, that is, I simply didn't have materials, I didn't have the internet, and it was very hard. And now, everything is there, there is the internet, there are many books. And most importantly, you have AI, which will support you, help you, tell you everything. And learning has become much easier. I learn a lot, not only related to programming, I study Japanese, other things. Now I'm deeply delving into cynology, dog training, how it all works. And it's not that it takes a lot of time, because I have AI, which gives me proofs, with links to articles, tells me something. Well, something like that. And getting into IT, accordingly, I think it will be much easier, because, well, as Eldar already said, the feedback loop is shortening, that when you are a young programmer, it is very important for you, you want to write something and see the result, and you can do it perfectly in pair with an LLM, and it will motivate you even more. Yes, and you don't need to be afraid that someone will replace you, you just need to follow the trends, because this train is accelerating very strongly, and you need to keep up with it, because, in my opinion, people who say, "I don't use it, I write much better without LLMs," it's not that they will be replaced, but they will simply become less competitive than those who follow the trends. That is, you will simply be pushed out by them, and programmers who follow LLMs. Something like that. And it's quite fun to follow this now, so why not. That's all for me. The teacher writes: "Well, yes, LLMs are messed up." Well, yes, you need to know how to use it. You just need to get your hand in. That is, if you use it every day, and validate it, not just perceive it as a given, like a monkey, just accepting everything it writes for you, then at some point you start to understand. That is, yes, I think the problem with accepting all these tools, especially among people who say, "Well, I write with LLMs, without LLMs, rather, I write much better and so on," is that many, when it became hyped, thought that this is a thing where you come, and it immediately gives you the perfect result or what you wanted. But in reality, it's, well, the ability to use these things is another skill that, well, you need to train. And therefore, yes, and it will be difficult at first, won't it? Well, it will be bad, it will be less effective than if you did it yourself. But then, when you get your hand in, your efficiency and productivity will soar, practically. Well, yes, I've gotten my hand in quite well at this, and I just don't think about it immediately. I have a router that directs it to the LLM, this one can go there, and this one I'd better do myself. And I don't really encounter hallucinations. That is, I know where it will hallucinate, and I just use it a little differently there. And when you get some information, you also need to understand that on this topic, there are likely many articles written. Here, it will most likely be right, and it's easy to validate. But here it's better not to ask, or to tell it to go through scientific articles specifically, or somehow, well, combine all this. For example, with dog training, you study the methodology, you ask it not just to read some yellow websites, but to go through research specifically, and then also analyze it for you, like, summarize it, provide some easily readable summary of all these scientific studies specifically. And here you start to understand that the answer was correct, because it matches the research, and this one is incorrect, because it was read on some yellow website, for example, something like that. Well, and again, I don't take everything they write as pure truth. I take it as something very probable. Most likely, it is so. And I wouldn't risk training my dog the way ChatGPT wrote it. I might adjust it a bit here and there. And a dog trainer is still in the first place, a live dog trainer. It doesn't replace a dog trainer for me. Here's the key thought. It complements the dog trainer. Something like that. Well, thank you all for. Thank you to everyone who came to watch us. We had a stable plus or minus 100 people online. Thank you Vlad, Eldar for coming. We had a great conversation. I really enjoyed it. Thank you for organizing. Thank you for inviting me. Often in podcasts, the conversation flows so much that time just flies by unnoticed. We've been chatting for almost 2 hours now. Okay, well, I'm ending the broadcast now. Yes, well, yes, regarding the recording, it will be available. The recording will be available in a couple of days in better quality and as usual. That's all, everyone, bye. Bye-bye.