Transcription
Well, friends, hello. Codex has been released, and GPT 54. I slept for 2 hours today. I fell asleep today at 4:00 AM, woke up, woke up at 6:00 AM, and started wipe-coding. Before that, I was wipe-coding on 54 in the CoEX App. And, honestly, I was surprised, because I usually do all this stuff inside-code. I'm a fan of code. For the last 6 months, probably, I haven't left code. And at some point, a month ago, I switched to the CLD-code and Codex Codex 5 and 3 combination. And, honestly, what came out yesterday, GPT 5.4, guys, it's a bomb. And this made me switch from clod-code. I haven't opened k-code for probably six, five, five years. It was my daily tool. Not a day without it, not a day without the terminal. And it seems, it seems, it has happened. - Codex. Today I will talk about it, I will talk about what I did with this thing, and why I am delighted, why you should try it too. Well, let's go. What's on the plan today? I want to talk about this thing. About Codex. The clod-code update will be released soon, and Seryozha will switch back to code-code. I hope, I really want to return to code-code, but this thing is too, too cool. And I had, for those who know me, you know that I have a wipe-coding club. And I had a long-standing idea to create my own LMS. In the sense that I want to have a website, and from the website I can go to Telegram. Telegram will redirect me to a bot. The bot will give me a link. Yes, that's how it will be. The bot will give me a link. I will open this link here. Or it doesn't work like that, or it doesn't work in incognito mode. Well, in general, and I will get here. And here I can watch. And here I will be able to watch lessons. I assembled this thing. I assembled this thing. What is it called? I assembled this thing in plotcode. Here's the pipeline that turns a simple video. A simple video turns into a page like this. I already had it ready, but I wanted to wrap it all in some kind of LMS. And I tried to do it with clod-code, but for the first time I managed to one-shot this thing. It worked for 3 hours. I made a very detailed PRD for it, with data schemas, with contracts, with a testing plan. And GPT 5.4 Pro wrote the PRD for me for 50 minutes. After that, I fed it into Codex. It tested it, it took a long time, it did something on its own for a long time. And in the end, and in the end, it did to-tests itself through Playwright, through Chrome, opened Telegram through Teletone. This is very cool. This is very cool. If you haven't tried Codex yet, I recommend downloading it. What will I show today about this whole journey, how and what I did. I will show GPT 54 Pro. Ah, here we have here, I have GPT54 Pro. This is a very cool thing. And probably, what we will compare, it turns out, it, yes, we will compare it with Codex. I will answer questions in the chat. To everyone who is not familiar, hello. I am Seryozha, a wipe-coder, and I have been wipe-coding hard for the last year, probably. We have a wipe-coders community. We already have almost 2,000 people. The link to the Telegram chat is in the description. And I help guys, non-technical specialists, wipe-code, to acquire a new profession. And I recommend, I recommend everyone to try, because now, to do this, you don't need anything at all. Honestly, I'm a little scared, because the barrier to entry has dropped to zero. This is incredible. Well, where did I start? We will test, probably, we won't test this thing for long, because, of course, my old friend, I'm very sorry. I haven't launched it for a long time. Let's see what it writes about GPT 5.4. The guys from OpenAI wrote that there are cool pictures, and I want to talk about them a little too. Well, first of all, it's clear that, as usual, as usual, they all say that this is now the best model in all worlds. And it is. Before that, I was using Opus 4.6, and it was a great partner. It had its downsides, we'll talk about them, but it was a super cool guy. You communicated with him like a friend. And at some point, Codex appeared, and I started combining them. That is, when you wipe-code, you ideally need to combine two entities. One entity that communicates at an abstract level. And you can plan, right? You make a plan. Let's, let's draw. We had clod-code with clod opus 4.6. We also had GPT Codex. 5.3. Yes, we also have the Codex application. We have the Codex app, which is probably one of the best. I think it's one of the best apps for all these things right now. We'll look at it too. I'll do a quick overview of Codex for those who haven't started. We'll also look at the cost. I want to talk about money, what costs what, what you can do. I also want to talk about it. And we'll test it. I think GPT will finish writing my PRD for me now, because the last PRD took him quite a long time. What is the cost of these guys? You can start with $20. For $20, you used to get the ability to do quite a lot. That is, a 3-hour session would be quite enough. Yes, if we talk about clod-code, there is also a $20 plan called Clot, what is it called, by the way? Clot Pro, if I'm not mistaken. Or maybe Pro Drive Pro. And this is a demo version. You could hardly do anything with it. You had enough requests for 10, then you hit the hourly limit and couldn't do anything. Here you can do a lot. And all this time, why did everyone like clod Clod-code with Opus? Because it had an awesome vibe. This guy had an amazing vibe. Awesome planning. It's relatively fast. When did it come out? 3 months ago, right? It feels like more than a year. Time flies incredibly fast. Relatively fast. I can't say it could one-shot. It had its own downsides too. That is, if we talk about the pros, then we also have the cons. What were the cons, guys? Tell me about the cons of Claude, Claude Opus 4.6. Among the cons, I can probably say it's the context. It's quite small, right? Because we launch, we launch Codex, we launch code, and immediately it consumes, how much does it consume? It immediately consumes 12k tokens. Just because, well, just because it's like with an iPhone. You buy an iPhone and you don't have 100 GB, but a little less. But, by the way, it's fine now too. And the small context, very bad auto-completion. And in general, it's fine, I set up my work with it. I set up my work. I've done a lot of cool things with it. For example, one of the super cool things is this blog, right? And I made it. I was interested in trying myself in SEO, AI SEO, how it would work. And we gradually, gradually, gradually made this blog. There are useful articles. One of them is the GitHub project as memory for agents. It's a good article, and you can give yourself, give your agent a link to this article, and it will do pretty much the same thing. It will repeat this article, and it will set up a GitHub project that will eventually store, for example, all sessions in separate tasks. That is, yes, well, how could we solve this? By storing the task context somewhere else. That's how I solved it. And in the end, yes, for example, if I open the GitHub folder now, not the most successful folder, let me open my operating system, show it to you. If I open, tell me what tasks are currently relevant for me. High priority. I think it will understand. I think, I think it will understand. In the end, it looks at what I have. It shows what is currently relevant, yes, with high priority, everything that is there. For example, you can open it right from here, yes, another one here. Why isn't it working? Because it's different. Let's go here. And my Chrome crashed. This happens. This happens. I think I chose, I think I chose my dream Mac. It costs $5,000. It's a 14-inch with 128 GB. This is a cool thing. This whole story is completely managed by agents. That is, I have agents now, for example, this one wrote this, also wrote Codex. I initially made a large PRD, then broke it down into tasks, and it's doing all of it. If you are familiar, if you have seen what OpenAI is doing with its symphony now, then this is it. This is it. Why is this important? Because, for example, now I can save, yes, this large PRD here and just give this link to the agent and say: "Listen, buddy, do this thing." Yes, I can even do it in Copilot. By the way, Copilot is awesome, it costs $10. And it has almost everything you need. It has OPUS, it has 5.3 Codex, GPT is also there. For some reason, it hasn't appeared yet, but I saw it there today. And for $10, you get a lot. And you can do it from your phone. This is a cool story. And now, it turns out, you can, yes, do this, and in general, it was the solution. It was the solution for how to work with Opus's small context at all. Everything was fine. But Opus has, well, a certain history. You have to force it to refine, to finish it to the end. And in general, I can say that an experienced developer won't hurt in this matter. Because to understand what to ask, when to ask, how to formulate, yes, some guidelines, because in the end we have an agent, by agent I mean an LLM, which is a large language model. For those who are not familiar with what it is. A large language model that works in a loop and can perform actions. That is, it can create files, read files, call some tools. And the artificial intelligence, the large language model you have working, has a strong, strong, strong influence. But in general, the main idea is that it all works in a loop. And you can control this loop if you specify what will be the point when we exit this loop. And here, of course, developer experience was very useful. But at some point, I heard that Codex suddenly appeared, and it was, and it was generally not bad, only it was very stuffy. It was a classic developer, you know, a system administrator in the corner of the office, with three vertical monitors, and he, you know, in a t-shirt, some kind of group that only he knows. And, well, in general, well, in general, it was hard to communicate with him, he was stuffy, he spoke a language only he understood. But it had a cool feature. It could work for a very long time. It could work for a very long time. It could search. It could search by, by, oh God, I connected Teletone to it, for example, and it became very good. It became very helpful. You tell him: "Listen, find me, I, you know, I, you know, was doing, I need to run sync." I say: "Listen, I lost, I lost this and that file, please find it." I remember I wrote it somewhere, and it will, damn, it will search, it will rummage, and it will find it. And a new format of work appeared after that. It's GPT Codex plus Clod-code. For the last month, probably, this was my main format of work. We have code, we have Codex. Clod-code runs clod opus 4.6. It writes the plan. It writes the plan. This plan is decomposed into tasks in GitHub. And these tasks are sent to Codex. Codex does them and at some point reports back to Clod-code. I ended up setting up a secretary in my corporation based on Open Clow. And everything is crashing for me. I don't have enough. I urgently need 100 GB of RAM, 100 GB of RAM. Yes, in my company, there is also a manager based on Open Clow, when, for example, something crashes. And I, for example, can go in and see. It says there that some task was closed, yes, if it was open, if it was closed, and so on. Well, in general, they just write reports, nothing more, but in general, it's already useful, because I can add something to it, I can open it, I can send it to Codex from my phone, for example. That's how it worked for me. And it turns out that I paid for Codex, I paid $200 for it. It was extremely good. It is good in principle, but now there is a better tool. And I can say that now for $20 you get some kind of analog of a $100 version. It can do everything, but you just need to manage it so that it doesn't get lazy, so that you clearly indicate what this guy should do, what the result will be, how to write tests. A lot of things. And you need to connect to it, right? You need to connect Exa to it, and some external context storage, and I don't know what else. Tell me, by the way, what MCPs did you connect, and I'll open the comments and read them. And Papa, hello. Hello. Codex is updating soon, Codex or Codex's AI. Hello. Testing Codex - 4.6. Not such a big difference from 4.5. What do you think about the farming models? Hello everyone. Answering hello. Why isn't GPT accepted? Ah, listen, I haven't gotten to checking yet. I don't have a mat under the packaging next to it. Are you planning to make short summary videos for YouTube? Yes, I hope so. My PRD just finished writing. As soon as I wipe-code it, I will definitely do it. For now, there are so many projects that there isn't enough time. Let's see what I have in the plan. Plan for the broadcast. Yes, let's see what I wanted to do. I said, let's do this. GP Pro thought for 27 minutes and made a ready, like a full PRD package. Damn, it even made everything in separate files. We'll get to that now. Let me finish. The manager writes the plan. GPT 5.4 has been released, which, according to them, finally contains all the best from Codex. It's just this Codex 5.3 plus that very humanity that you liked, that you liked in Opus. And how my work looks now, for the last day, let's say. GPT 5.4 writes me a plan, writes me a detailed PRD. So here's my advice. It does this very well. You need a model if you want to successfully wipe-code and bring your things to completion. Yes, I think this is advice for those who are definitely not from the development world, because there are slightly different stories starting there. But in general, I think it will be useful too. You need a model for abstraction, for dialogue, which can, talk in the language of strategy, for example, because Codex couldn't do that. This thing can do it. You ask it to do it. We ask it to make a plan. Yes, you can, you can ask it to make a plan first. You voice an idea, and this idea turns into a plan. An implementation plan. And here we turn something that is in your head into something understandable, a plan for development for a developer. We don't need to read this anymore, right? We have a model that can translate our abstractions into a language understandable to a developer. And what else can be done? You can ask it to do, for example, next time. How can you ask? You can say: "Let's not do a PRD." Include contracts, data schema. Tests. You describe the idea at the beginning. Well, ideally, a little more context. Ideally, add context plus context. Context is very important. What is context in this case? What are you doing? What is it for? Any important idea, it can, any, any information can help. I'll go back now, to this thing that I assembled, that I was just amazed that it did it. The LMS club. Let's, let's look. Now I'll reset its storage. It should then clean up. No, listen, it doesn't want to. Just curious why. Oh, it cleaned up. How did I make this thing? How did I one-shot this thing? Let's go through the full path to open this page. That is, it's a web application that contains, contains lessons. I have a bot that sends an entry link. When everything works out, I can enter and see and see the lesson and see the schedule and different techniques and everything else, right? Here. This is how it can look. You just can't imagine how happy I was that I did this today. I've been dreaming about this, I don't know, for a very, very long time. I dreamed of doing this. I asked it to make a chat, a pro PRD. I'll tell you how it works coolly now. At the beginning, at the beginning, I went into my favorite, here, yes, it was here. There is a file that includes and describes everything I do. Do I understand correctly? Let's do this, let's draw an ASCII diagram of the key files that describe what I'm doing, what I'm engaged in. I ended up assembling everything into one file. There was quite a lot of everything, key files, and so on. And it started doing it. I chose the pro version and the pro version, the pro version is incredibly good. I used O3 O3 Pro. And I remember then I paid for it for the first time, how much? $200. After that, I didn't use it, because I tried it once more and it didn't work for me, I didn't like it. So, yes, and here, yes, this is how it looks approximately here. This is some kind of, yes, files. And I also have, let's show another thing. For personal corp, yes, it's it. Show, show me an ASCII diagram of the services of digital departments. Wow. It did this to me. Show me an ASCII diagram of the services of digital departments. Do you know who it was? It was Vyspr. Vyspr sometimes has such a glitch. It's the mat I'm drinking. It's with taurine, I think. Yes, with some kind of invigorating herb. Red packaging. I'll send it to the chat later. The topic I'm passionate about. While it's thinking, let's go back to it. Wow, wow, wow, wow. It went great. Yes, this is, this is how the breakdown into tasks works for me and where it all goes. Thank you, thank you, my friend, for doing this for me. This is my context that I was talking about. That is, conditionally, I enable SEO and say: "Well, let's go, let's go coding." Because we are now, with you, wipe-coding. Wipe-coding. It has one drawback, it doesn't scale. To scale it, you need to switch to SEO mode. You don't need to approach a developer and sit at the same desk with him, and you don't need to, you tell him on the screen. Move this here. Yes, you need to plan and then managers should ideally pick up different things, maybe determine them themselves, yes, I have a product department, content, video pipeline, sales, analytics, research, each of them has something falling further. Well, there are also support services. What's there, research texts, knowledge graph, local transcription, normalization of terms that works. Well, here's what it could do, what it caught. In the end, yes, I have a few more Seryozhas who know me well. I fed all this to GPT and told him: "Come on, do this, buddy. This is what I want." I've dreamed of this for a long time. And what did it do for me? It made me this thing. D studentup mp access via Telegram. I told it everything, it gave me. Look, you need to do this. Everything is broken down. MVP scope scenarios, non-functional requirements, routes, contract schemas. Very, very detailed. And that's why this thing is good. It, by the way, currently supports, it has two modes. It has a standard mode, normal, but extended. This thing is only available on the $200 plan. And I can say that it's worth it, because in 3 hours I did what I couldn't do for a long time. It has a very interesting feature called steering. As soon as it starts thinking, you can give it different ideas, like I want this, and this, and this. And while it's thinking, it takes it into its context. Cursor has a similar feature, for those familiar, who have worked with it. And in the end, it told me, like: "Yes, okay, listen, here's your thing, everything is here, and I'll take it further, transfer it to Codex." I tell him: "Let's decompose, add everything, push to GitHub." It told me everything I have. What else? Let's quickly go through the overview. I didn't show. As soon as you download it, you need a plan to use it. I recommend starting with $20. It's extremely generous. It's an awesome application because it's not a clone of VS-code like all the others. And it's very simple. In principle, you don't even delve into the code. You don't see the code. I think by the end of the year, we won't even need to look at the code. Honestly, most people will probably forget about it. What's interesting here? You can probably change the language right away. Yes, there is Russian language, and there are some translation quirks, so I'll probably show the main settings that are extremely important. Speed. You can choose the accelerated mode by default. It's 1.5 times faster, but it consumes twice as many tokens. In general, try it, it's a little faster, not super fast. And here is the steering, when the model is thinking and you can add your thoughts to its context, and it changes the plan as it goes. This is a very cool thing. Works very coolly. Probably, probably, probably, that's all. The only thing else you can, I think in the config, there is a history where you can set up y mod. Yes, here it is, config. I think in the config you can enable it, I think. Yes, yes, these things. Confctomal, here what's interesting, probably, is Sandbox mode, where you allow it to access the internet. My personality is set to pragmatic. There is, there is like friendly. If it's friendly, it will tell you, like, a cool guy, everything is great, and pragmatic will be like, I really like it. Awesome. Useful MCPs. Chrome dev tools is a must-have. I think it's the best right now. You can turn it off. Yes, Dev Tools is the most awesome MCP for testing, for browser automation. The only thing I recommend you install is, let's do it right here, as a custom. Per session, I see, in this case, it launches the Dev Tools version with my profile. And in this case, I can, for example, say: "Go to Google Search Console, see what's happening there." Well, in short, to manage, to manage.
Your browser uses your specific profile. And it works like this. I asked CloudCode to set this up back when I was still around. Tell me how this Alias works. Tell me how this Alias works. Well, it actually launches A, yes, and it launches a version of Chrome, on a specific port, and does so with a attached profile. There. And this allows you to use those services and automate them. What else is here? Here we select the GPT-4 model. If you have Fast mode enabled, it will have an icon here. Extra high mode is default. You can offload tasks to the cloud. Here you can choose the cloud type. Either you work on Work 3, if, uh, you need to upload tasks in parallel. The beauty of this is that you can work in parallel. To start working, you select a folder here, add a new project. Let me begin. What I wanted to do is I need a new folder. By default, I work in the GitHub folder, where all projects are. I'll name this folder, hmm, I don't know yet how, I'll rename it later if needed, but the main idea is that it's part of my content production, content production that will generate short videos. I described the idea that I had in my head. I described it to ChatGPT, and it made a rewrite for me. We'll look at it now. Video. M. Let's see, I don't know, it could have been named, uh, I'll call it TikToker, because I want to make good short, uh, short-form content, and have clear titles. There. Once you've done that, this appears here. On the left, yes, you can switch it to chronological order, the last, last things you did, or, uh, in this order. GPT-4 is selected for me. And now we say what we are doing, respectively. I'll open my GPT, which wrote it for me. This was in one of the plans. Okay. M, it's also cool with presentations, by the way. I mean, not presentations, but those PDFs. PDFs were also made well. I liked it. So, I told it the idea of what I want to do. The idea was as follows, which I will personally use in the computer. I told it by voice. It will be an electronic application. And here I wanted to say, of course, electronic. That is, that it's an application. I streamed live on YouTube. And, in general, I upload a video to it, it puts the necessary titles. So, it needs to be rewritten for agent development, finds key moments, and does everything based on the internal system. Here, Wiper is meant. You can actually log in there with CloudCode and use the subscription for analysis. How I use CloudCode now, I used it as a transcriber for transcription based on, uh, based on five sub-agents. And I have a post about it on Sergey's Tech blog. He talks about how I did it. So. But in general, it understood, yes, that an electron app is needed, and these things need to be collected. I know what that is. Not just that I memorized it, but simply because I tried to do it, learned it, and, uh, God, and now I remember what it's called. So it's not that I know a lot, I don't know anything at all. It fixed the stack: desktop first, local first. And here we have a prompt. Let's see what it did for us. I haven't even looked at it yet. And here I have Z open for convenient reading. Life to Reels desktop, AA coding agent plus a person who will review the result live. And this thing, you remember my prompt, it was extremely, extremely bad, the initial one. Let's see what the Pro version did with this idea. Context and problems. Now the workflow will be fragmented. Long videos separately, transcript with a separate tool. Yes, I want everything. I'm losing all of this. Product goal. Personal desktop app that turns long content into ready or almost ready short videos in one go. That's what's needed. Product KPI. North star with a scope of VP app, local video transcript, import local transcription via Whisper. That's what I wanted. The only thing it didn't add, of course, because I didn't look at it, is that it misunderstood the idea with CloudCode. Who uses Pcel, yes, you can log into different applications using CloudCode, and it will use the power of CloudCode or Codex LLMs for work. What do I have now? A week ago, I paid for Max's subscription to CloudCode and now I don't use it because there's this beast, which is very cool. Well, in general, it's clear. We wouldn't be bypcoders if we read all these rewrites to the end, who is even interested in this. We now upload it here. And oh, it's cool that it tells me: "Let's use linear for this prompt." That is, it immediately suggests me, in fact, to decompose it into tasks. and upload it all to a GitHub project. Therefore, and it also suggests me to use plan mode, which is also awesome. I say: "Yes, let's, let's try." I'll tell it: "Next with this prompt." What do I do next? I further decompose it into tasks. Decompose it into tasks and write these tasks into GitHub Issues. And create a separate board for all of this. Don't just write and create a separate board for all of this. Okay, let's go. Extra high mode. And that's it. At this point, you can switch to the next thing. Anyone who has worked with AntiGravity knows that there is an agent manager. This is approximately it. AntiGravity has one major drawback. They also left an editor there for some reason, it's not needed there. Uh, yes, I created a GitHub Sla Ops skill, which, uh, in general, just went through GitHub's documentation and learned to use all these things. But in general, you just need to upload this article to the agent and it will work the same way. This is like a part of personal corporation, when you can scale and, uh, there at that moment, when you have these tasks, tasks, yes, described in detail, each of them, uh, you can, by doing cycles, you can do them in parallel through Worktris. Each of these tasks will speed up the process. Also, by the way, there is an important thing here, which I didn't show, the last thing I didn't show, is the limits. Look. The limit is quite generous. I have 5:92. And today, for today, I have used 23% of the limit, uh, of the $200 plan. It's weekly, but I've been wipecoding for 12 hours, yes. So I, well, there, okay, I also, I had 3 hours of calls. Well, the remaining 9 hours, uh, I, well, yes, yes, I'm in ipod. So, where to prepare GitHub Issues for life to Reels desktop? New Tiktoker, Leave to Reels desktop. Existing content factory, which, by the way, exists. Yes, here, damn. Ah, I clicked through. It asks questions, sometimes asks for clarification on some things. This works in planning mode, when in planning mode, before it starts doing something, it will ask you a question, how detailed to break down the prompt into issues? It told me 20-30 issues. Well, it's recommended, so, okay. It will think now. Also, in GPT, there is a thing called Coax Park. This is based on, based on SBRC. Codex with a small context and fast inference. Here I have a context window of 258,000 tokens. I haven't enabled the million, million tokens because, uh, in general, I think it's not relevant for me yet, I don't need it. I really like how context compression works. Uh, well, I think I'll try it. It's enabled in the settings. You need to go to config.toml. And how is config enabled? A million tokens in GPT 5.4 within Codex App. Config is enabled within Codex. By the way, I've heard stories that people in the chat share that they use, they change subscriptions for $20. two or three subscriptions and you can just switch within the application. Well, of course, the only thing is that it's a bit slow, but the guy is reliable. His strength is that he can work. You tell him, for example, now we will break this down into issues, and I will tell him: "Please start with the first one." It will write each of them in detail now, and there will be evaluation criteria in each of them. And I will tell him: "Please complete each of them and don't stop until you've completed everything." I'll also tell him that he checked everything with Playwright. Uh, by the way, I customized my Telegram bot through Teleon. And for some reason, I didn't think of this before, but now it did it itself because I had Tele installed. I was like: "Damn, of course, you can test Telegram applications using Teleon, using an agent. This is just top." And if you don't test your frontend stuff using Playwright or, by the way, Skill, it has an awesome skill. So, let's read the plan, uh, that it made. Here it describes the tasks, collects the electron block safely. So, SQL schema for core entities. Local first storage. Implement project CR system, ground job runner. Okay, okay, okay. This task is for a couple of hours. Okay, maybe 40 minutes. And I tell it: "Let's do it." Let's do it. It will first create all these things, write them down, then probably stop, and then I'll ping it again, like, friend, let's do it, report on your work. This is the first stage - planning. If you want to scale, and move away from wipecoding, when you can do it, yes, my best engineer friends can simultaneously manage, well, I don't know, four, probably, agents, because, well, it's hard to keep more in your head, very hard. If you want more, you need to package all of this into some kind of project, project history, and team lead, enable SEO mode. Now it will do it for me, uh, it will do the project for me. Yes, the project is already done. You can go to it. Here is the project it made in the desktop. It's empty for now, but it will be filled with new issues. And, accordingly, some different things. I do it like this. The advantage is that you can connect different agents, yes. I now have CloudCode, I have a subscription to VS Code, which I manage. What else do I have? I have GitHub Actions pipelines. You can also run them through GitHub. Honestly, I believe that GitHub is probably among the top five most important infrastructure applications of this year, because it's a center where your code is, and you can build anything on it. You can build awesome automations on it. And you don't necessarily have to do it, yes, from the terminal. Like, now Codex is here. Also, Codex is available to you in the cloud and so on. And what else? What else do I want to tell you? Ideally, in my scheme, yes, in the end, I have GitHub. This is the agent's memory. And Codex, Codex is now the executor. And it turns out that in my case, the architect is, uh, GPT-4 Pro. Where was the diagram? Here it is, yes, it turns out. Let me fix this. This guy, this is my coder, but he can be more than just a coder, actually. He can also write texts, right? He can be a writer. Why not? Why not? In short, the executor. I have an architect and I have memory. Well, I also have OpenCL. as a secretary, as a secretary for all of this, who just sends me pushes to Telegram, and something happens, and I can also tell him something like, listen, guy, let's do it differently here. Or assign an agent directly through him. Tasks. The architect is GPT-4 Pro. It makes a prompt for you, and it researches, researches, and connects different parts. This is a very important thing that cannot be missed, cannot be missed from the plan. At what stage are we? Still. It's still thinking. Well, okay, while it's thinking, you can read what could have been started earlier. And I would also switch to a poll. So, if there are any questions, let's chat now and see how it breaks down these tasks, and I'll send it to do them, to execute them. M, where was GPT? GPT-4. Our latest model. Ah, yes, very useful. The link is here in the documentation section, where the benefits of GPT are described. There are also several guides on how to enable a million contexts. Here's another interesting thing that appeared: it has a built-in computer vision capability based on their new API, which is called, I think, if, uh, if I, if I'm not mistaken. And this is something I will test. And I hope that I can, uh, on one of the streams, guys, together with you, make it play Heroes of Might and Magic III, and watch, because, uh, all of this is still done through screenshots, but it's my dream that, uh, for example, I don't know, Codex fights with another Codex to test all the models in Heroes. God, how I love Heroes. The game is just childhood. At what stage is this, by the way, there's also a guide, prompt guidance, and latest model guide. Always check, see how they work. Using GPT-4. Here there are always interesting, uh, interesting details on how the model works. Regarding the prompt, yes, I need to look. I'll need to look further. By the way, they have good. Yes, I think it's in the cookbook, I think it's somewhere in the cookbook. I saw it yesterday. I can't find it now. Okay, I'll remember later. So, Codex. Codex, my friend. Project fields. Massively. Okay, it's updating little by little. Updates are being filled in. Here they are set to to-do for some reason. Oh, it sets priorities. What do we do first, what do we do next? How many did it make in total? 23. No, even more. Not 23. Ideally, the more time you spend on planning, the better. Seriously, this is the thing. Never start just by hammering out a prompt to turn your idea into something. You need to validate it, test it. And now it's checking what's opening there for me, looking at all of this. But I have another session running here. Let's check. Here, yes, they appeared. These are the prompts, yes, it broke them down into tasks. Each of them has a thing called context. Understand electron shell, connect React UI shells, design a secure PC layer. And there are acceptance criteria, meaning it won't move forward until it completes this thing. It also has a good skill that I recommend installing. This is a must-have skill that you should install. There is a thing called Skills here. They are located in a separate tab. Also, definitely check out Automations, play with them, install something for yourself, just for fun. It works interestingly too. There are Image Generation skills here. Someone asked how to generate images. This image generation skill allows you to generate images directly, for money. It works through the open API. Playwright. Playwright is what allows you to test. For some reason, I don't see the skill here called, this is my skill that I created. Playwright interactive. This is the skill that allows you to test electron apps. And this is a cool thing. Uh, so I recommend installing it too. It's available by default too. This is probably one of the most important skills. And I'll show you how it works now. I learned about it in the GPT-4 demo, in this article, they showed it. Uh, yes, it's clear that the best is everywhere, as usual, everything always is. I hate these benchmarks. They mean absolutely nothing, absolutely nothing ever, because as soon as a benchmark becomes available, it becomes a source for training the model. And, well, it's like, and it makes really cool presentations. I tried it a couple of times, I liked it, I made them, uh, yes, five-four and made them in Pro. It works well. Uh, what else was interesting there? There were a couple of interesting pictures. It's clear that it's also, it's really fast, I liked how fast it works with tools and tool management. Uh, yes, this impressed me a lot. Guys, have you seen these games? This is, of course, this is the joy. Or an RPG game. This, this is what was this game called? On Nintendo, on Switch, damn, what is it called? The last one, or the penultimate one, where you interact with your students as a teacher, all that cool stuff. I love this series, I forgot what it's called. Well, and some kind of T-shirt. Ah, yes, here is this thing called, watch this video, I won't play it, because I played this video once, and there was some sound. And then I had to edit it. The cursors said it's the best model. I completely agree with them, but the best model now is better than Opus, simply because it can work for a long time. With Opus, you need to be very careful. You need to sit with it and watch how it works. With the correct setup, your GPT will work for a long time. With a good plan, it will do everything. If you also specify tests for everything, it will independently check everything in the browser. And you'll even get a ready app. That's what I got with what I asked it to do, uh, an elmask for myself, uh, in conjunction with a Telegram bot. So, let's fight. This is about what I was talking about, when you launch the model, it starts thinking. And at first, I thought it wasn't a super important thing, but it's a very important thing because it thinks for a long time, and at the moment when it starts thinking, you can give it. Actually, no, I want it differently. Or give it additional context. This is especially relevant with how Pro works, because it thinks for about an hour. So, pricing and values and all that. Well, we can start. Friend is still thinking. I'm setting up another useful connection. I'm bringing a new board to the server tracker repository so that the log is visible from the repo context. Also okay, good. Well, well, well, that's how it is. And doing the final machine verification of account project item field values. It just rechecks everything for this. For this, it was a good code. Because it rechecks everything several times. Uh, Gosha, hello to you. You're unlikely to be watching me, but Gosha is my friend, top from Yandex, and we had an eternal beef. I was a fan of Opus, he criticized Opus. But now, Gosha, I'm on your side. I'm on your side. Uh, yes, yes. This is how this thing works. This is how our Chrome works. So, everything is excellent. It did it. It did it. Uh, awesome. Awesome. Well, it's suggesting me to fix something related to layouts, if I need it. Let's, let's see. Okay, here are all these tasks broken down. Layouts are also here. You can look at them by layouts, break them all down. There's also a cool roadmap. You can also set time, uh, like set deadlines. Well, this is if, for example, something is time-sensitive for you, also a useful thing, planning sprints. And I strongly believe in this. And this is a kind of personal corporation. If you're interested, subscribe to the channel. I'm starting an experimental stream soon about how to go from wipecoders to SEO. A kind of pro track for those who have already mastered wipecoding and want to learn to manage agents. I believe that the future is in this. So, well, in general, everything is ready to launch it. So, well, friend, we are ready to start developing this product of ours. I hope we'll earn a million with it, so start executing. Everything from the first, from the first task. Don't stop until you do the last one. And don't dare to make a single mistake. Not a single mistake. So, in general, in general, this will be enough, of course, for it to launch. It's just important to launch it. We don't care about the context at all. Let's also enable the context. Why not? Why not? So, where to enable it? Ah, oh, cool. What else do you specify? I need links. Where is it? This is it. Oh, cool that it started searching for nuance app.to in history. Let's try to enable it. Ah, yes, I'll just enable it now, you know, I'll enable it in a new thing. I'll copy all of this to it. Enable yourself to tokens and check that this is current info from sources. Let me enable Codex Park here so it does it quickly. And here I'll enable rising on medium, because for this task you don't need to be a super smart guy. If you also enable Exa here. Well, it's searching quite well in general now, but Exa, Exa would be faster. Plans and prompts in your context are the same thing. Let's see. So, so, so, so. Questions, questions, questions, questions. I'm answering. Don't you group your chats by projects, like Inbox Zero, so that chats aren't neglected? Listen, I just haven't used ChatGPT until now. So, in general, in general, probably not. I think I'm more likely to delete unnecessary chats, or I assign specific tags to them, but in general, no. Can you explain for beginners, what's the difference between plan, spec, prompt? In general, they are different names for the same thing. So. Because everyone calls something, I don't know, a sub-prompt. Well, in general, it's just project documentation, yes, a technical specification, you could say in Russian. Uh, human experience is still better. Yes, possibly. So, GPT Pro is good for developing a project from scratch, and for further development, do you need the project context? That's the beauty of it. That, listen, it did it in the end, I recommend the minimal config users scope. Yes, okay, let's do it. So, how relevant is Super Power now? I just tried to make a bot with it, but it turned out to be quite weak, even though it thought for a long time and asked a lot of questions. Uh, listen, it depends heavily on how you, uh, how you guide it. It's cool because it helps you write a plan. That's what I showed, how Pro generates, uh, it generates a detailed plan. This is a similar story. Super Pers. How will I do this? Yes, I have brainstorming. And here, and I don't know how to earn a million. The main thing is not to make a single mistake. It will start asking you questions, and at the end you'll say: "Well, let's, uh, use brainstorming for idea development." That is, it asks you additional questions. You can also, actually, okay. A million in what currency and for what period? A million Pesos, of course, a million pesos. So, well, by the way, it's okay. Let's say a million rubles a month. A realistic goal. So, uh, have you tried Super Power? Yes, Super Powers, there's an article on Sergey's Tech blog about Super Pers. And a couple of streams were about Super Pers too. Uh, it's a good thing. I love Nasa the most. Now let's look at the end, approximately current monthly revenue. Uh, I don't know, I'm a bum. Uh, now I'm just taking out loans. What's happening with it? Damn, it's getting stuck here. Codex Park is getting stuck here. So, it's done, idiot. Idiot. So, and now revenue is zero, income only from loans, so you need to go from zero to 10k. And you already have a working school, content pipeline, one-on-one mentoring, audience community. So, at some point, it tells me, and you say to it, like, let's write a plan. This is how you tell it "fating plans" and it writes a detailed plan. This is approximately what we did with Codex, where it broke everything down into tasks. This is unl superpers, it's just stored in a file. In general, in general, it's a cool thing. Uh, okay, okay, theme. If, if I think you're on CloudCode, must-have. Where do you get information from? Is there a list of resources or subscriptions? I have a wipecoders chat, and guys there constantly bring good news. But in general, I can say that you shouldn't chase news. FOMO is such a thing, it's better to just try, to implement your current pipelines from Cloud to Codex. How are they transferred? They support pretty much everything the same. For example, in Codex settings, there's a thing called configs. Yes, it's in the config. Look, it pulls, by default, yes, migrate user settings, cloud settings into users resers codx config tomal, meaning it does everything itself by default. And it transfers skills and agents md, it takes from code md. So, in this regard, compatibility is top-notch. Yes. But you need to select. But in general, it also pulls by default if you start working in an old folder. So, do you need to restart? The check is underway. Parameters. Open a new session. Code-code. Let's try a new session. And you can't choose a million here. Let's check what it did. It set the model window limit to 25,000. It's unclear if this is a lot or a little. Well, in the sense, it seems like it's a bit too little, because, uh, I think you can do more. I need to check. I think the limit for Opus is now 100,000 tokens per go. Well, let's try. I have a board. The board is my context. And here it is. I'll copy the previous prompt and see if a million tokens will work here. Not a single mistake. So, yes, a million tokens are enabled. Here it is. Uh, look, there's a nuance. that when it reaches 300,000 tokens, everything above is calculated with 2X, 2X token consumption. This is an unpleasant story. So, so keep it in mind, if you're going to use it, look, I waited. Again, you can look. Pay attention to these things. They are generally very good. For $20, you get, I think, analogously to a hundred on Cloud, maybe a little less, but in general, it's super generous. In the app, in this thing, you get 2x more limits, if they give it. It constantly writes here that it doesn't write anymore. Well, in short, it often writes here that you should go to the Codex app and do your things there. So, I'll translate A1 to In Progress on the board and immediately after that I'll start setting up the working skeleton of the project. And I need to add, you know what? I feel like Julia Vysotskaya when I start talking to the camera in a familiar way, but I think it's a normal format. If I, who watched, this woman who cooks terrible food on TV, I can wipecode terrible projects live and say: "Oh, listen, hello. You know what we're going to wipecode today? We're going to wipecode this application today, which will turn our, you know, >> What else do I want to do? I'm setting up OpenCL through Cloud, so I need to connect another thread here, which will monitor what's happening on this board. So, connect this thread, uh, to our chemist. Connect this thread, please. In this project, I have something like, uh, look, yes. I think I have SSH passwords here, but I'm not sure. I'm not sure. Ah, yes, this guy, my Cloud lives on a server, on a VPS. So it will definitely show the password now, I'm sure, it constantly shows me, it constantly shows me passwords. So, before making changes, I compare the structure with the Salem Electron White to avoid building the framework on incorrect entry points. How clever. And, uh, here it has already started using MCP. It also has one of its features - Tool Search. It's good at selecting the right tools. If, for example, you were on an MCP diet before in CloudCode and you didn't install many, many MCPs, then now you can not worry, you can install anything you want here, and it will load them itself, if needed, in context. In context. The model is very cool. It really feels new. It's super, super smart by default. No, by the way, it seems, it seems it hasn't revealed anything yet. Let me check. Yes, I think I connected SSH. Look what it does next. I have a Cloudic file here. It just shows how everything works. And it connected. Now I have a new thread here. Lif to Reels. And, accordingly, now, as soon as the reels change, it will upload them to me. This is a group. This is a group with a bot. Ah, yes, there are all sorts of useful things here, like the chief editor, Naval Ravikant, who gives advice, uh, well, and so on. So, what else is interesting to tell? Uh, I don't know, I think that's it. I think that's it. I'll answer questions. Yes. So, this is the result. This is the result. Let's see. Opus conversational prompt architecture plan for the project. This is my setup, yes, memory for the agent, decomposition, handshake between roles, uh, GPT, generalization, connection search, re-checking, system architect, hardcore coder. I am the cycle manager. I monitor the context, I monitor the evaluation criteria, well, and the quality. Uh, after release, yes, my GPT thinks and does coding, long running, autonomy, yes, it can really work for a long time. Very cool. Uh, Codex surface SE and de cloud, one command center, yes, together with windows. So, in general, you can launch different things from your phone too. Workce automations. So, it's clear. It just collected it for me. This is a presentation that GPT made for me. What did it say interestingly? Social facts. You can close the plus, guys, a lot. Well, okay, let's read. Ah, by the way, okay, we've done all this. So, I'll go and set up my plan further and test this app that I made. God, I'm so happy I made it. Everything is here. And I need to send it to the guys. My March circle started yesterday. And I can't wait, I can't wait to send the lecture, the lecture, the lecture from yesterday and ask questions. So, well, oh, Gosha Gashan, hello. Damn, I'm so glad to see you. Is it still worth using? Is there anything better? Context. Yes, in general. In general. Oh, listen, look what's happening. What's going on? It's already started. Damn, it disappeared. It launched, but it disappeared already. So. Why not make the architect in the same Codex App A-4? Because Codex App A-4 doesn't have Pro, and you need the smartest model. The smartest model in the world right now, the most awesome, which has the longest, uh, how to say, uh, budget, budget for tokens, for thinking, and a good cycle. This is GPT-4 Pro, yes? This is a separate model that I showed. Yes, follow the hand. Here it is. GPT Pro-4. This is it. It also threw me a diagram here. Yes, also contracts. I didn't do this, by the way. I should have uploaded them. I could have uploaded them there too. Add them. Ah, this is an important thing. This is an important thing. I'm honestly going to add them right now. Uh, oh, guests have arrived. So, okay, yes, it will understand everything. I'll upload this now. This will break it, actually, now. Well, I don't know, I don't know, let's see. So, chat further. Ah, add this to config.toml. And one context appeared. Yes. Is Ax 7 worth it? In general, I don't know, honestly, I haven't used it. I installed Exa for myself and now, in general, with Codex, I don't know. You can do without Exa, you can generally have normal search. GitHub Skills up to 400 tokens, 54 with cache, but I don't think it's 54 Pro, I think. 54 Pro is nowhere, because it's a super expensive thing. It's, I think, 50 times more expensive than regular 54. This is purely so that people buy Pro. If I have more than 300k tokens, and Fast mode is enabled, which also consumes tokens, then how much will it consume in total? You need to check, you need to see. But in general, $20 won't get you far. But for $200, why not try? Why do you drink tea through a straw? Damn, because drinking tea through a straw is cool. This is not tea, this is mate. This is mate. I'm in Argentina, so hello everyone. It's 6 PM here. Leave the recording, please, I'll watch it tomorrow. No problem, bro. Is GPT-4 Pro available from the cloud? No, it's not available. What if I buy it for GPT for one day? Codex will be enough. I don't know for what. It's unclear, it's unclear. Maybe it will be enough, maybe not. So, well, guys. So, well, guys. M, I wonder what this guy will do, and I'll write in the chat later. So, so, so, so, subscribe to the channel, this one and that one. Read, try, experiment. If you need company, come to the circle, uh, to the March circle, you can still join. Write in private messages. Well, so, bye everyone. See you all again. Moi.