Transcription
The narrative on the internet is very strong that it is unsafe, >> that is what she was trying to stop this Openk, and it was absolutely unmanageable. It took on a life of its own, >> for me personally, there is a certain zone. Am I willing to take a risk to gain a certain efficiency? >> When Open Clow or similar systems are imposed from above, it can happen that the ignorance or lack of education of an employee leads to data leakage. I want to cover one topic today. So, ChatGPT released 5.3 instant. And essentially, what is this version? >> Hello everyone. We are on the To the Moon channel. Technological news from Silicon Valley around the world. We are continuing our broadcasts. And today is our hundredth anniversary episode. The hundredth anniversary episode. Well, not counting a couple of special episodes that we had, right. So today we expect a lot of likes, support, and comments. By the way, you know what I thought? It would be great if people wrote what super new things they learned in our episodes, that is, what allowed them to do something truly incredible. Well, like Anya, for example, even started installing openk. Today we will talk about OpenC and agents, and we will start with that, right? Or about the version she bought. Maybe already, maybe, maybe I unsubscribed from it. But write what you learned. It will be very interesting over these years, almost 2 years. What has had a serious impact on you and what have you implemented in your life. Congratulations to everyone on the hundredth episode. Well, let's start today with a topic that we didn't manage to discuss last time. It concerns agents, especially considering what happened with the update, again, in all models, this month has been filled with talk about the existence of agents. Everyone is talking about it. And I have repeatedly expressed the opinion that I believe that agents work and will work according to some other scenarios than what is shown now. But still, I want to clarify some difference to the end, so that everyone understands from the perspective of agent work and what has been discussed in recent weeks regarding OpenCla. What it is, so that everyone hears, because it is not a model and some kind of shell, and an agent engine, and such systems will likely appear even more. Perhaps they will not be used, and direct models will be used. I'll give a little explanation. So, a brief explanation. So, Open Cla is a shell that you install, you all install it, it has become fashionable. Karpati wrote about it. He went to the store, installed it, bought himself an Apple Mini. The seller tells him: "Why is everyone buying Apple Mini now?" He says: "Everyone is buying Apple Mini to install, you know, Open Cl." Well, it's like, in my opinion, a silicon story. >> They weren't even in stock. >> Yes. >> Yes, I think it's a story that needed to be ordered in advance. >> I think it's a story that is somewhat silicon-deleted. Again, if we talk about Apple Mini, it's a powerful enough computer to deploy work related to models or agents, right? So, well, this can be done on Mac Studio, logically, right, on all Mac devices. They are, well, they are quite good for this in any case. Why did he buy a separate mini? He bought a separate mini to run this whole agent story, not tied to his general infrastructure, to his daily programs that he uses, daily files, to his operating system. That's what's inside it and on his computer. Like some kind of supposedly isolated environment. So, you install this shell, configure it, and essentially this shell, it's again, it lives on your computer, it's an important aspect, it receives messages from messengers, that is, you can chat with it in various messengers, it stores various working files, memory, and so on, and knows how to call various tools, like a browser, perform various actions, like, I don't know, in Slack, in Discord. And then, essentially, if we talk about its main brain, here, so that it's clear, you connect it yourself to various agent systems. You can connect it to Openi, you can connect it to croopic, you can connect it to Google, and so on, right? That is, you, well, or install it locally, you can connect it to cloud models, or install them locally, like Alama, VLM, and so on. That is, to install it on your computer, well, for some specific micro-tasks. So, and configure it and tell it, like, to perform a certain action. You can tell the model, the agent to choose itself. For example, you tell the agent some action, like, check the volume of my stocks every day and, I don't know, inform me about certain risks. And it can inform you about it in a messenger. Which agent it will use, it can choose itself. Or it can be configured to use these agents through this system and so on. And it connects via API and works. Again, via API, right, and works with agents internally. So, if we are talking about code, Open AI code, then this is an application that exists on Mac, now released on Windows, it came out on Windows a few days ago, right, congratulations to everyone who has Windows. So, and you can run agent, configure agents, directly on your computer. It's clear that everything that is code is very closely related to programming. I still want to move away from the programming zone, right, and I think programmers will figure out a lot of things anyway, although many of them watch our channel, but still. To move into the zone of real everyday things, like Tanya, like, I'll install open cl. This is a very specialized action, right, and configure certain agents to do something for me, right, and internally. And you can go into the code. And also in the code, there is an environment, this is a separate application. They upgrade it very often. I want to say that I think that code has upgrades and releases almost every day. It updates endlessly for me. I always go in, there's always an update button, and it, I think, slightly freezes my system. And lately I've had the feeling that it's constantly reading something or looking at something, some files or some data. So, I don't really trust the code application. That is, I still somehow trust ChatGPT with its subscriptions and all the unsubscribes from all the training, but I never really trust code. I think it lives and works by its own rules. But there it is a special environment, a special application in which you can run these and configure individualized agents, right, internally. Write, who has tried to do OpenCl now, Ilnar, will you continue some part, right? Who has tried, who has tried to use OpenClo, what did you do? The main thing here is the story, look. There is a very strong narrative on the internet that it is unsafe, right? And we just had a discussion before recording, Tanya was saying that someone told her that it's unsafe and so on, but this safety, unsafety is a very relative zone. When you store emails in Google. Is it dangerous or not? Right. When your data is just stored on your computer, is it dangerous or not? I, back in 2010, started managing the company MegaPlan - at that time, the largest company for selling cloud storage software. And our main story, which we had to prove to other companies, because there were very few cloud services. Well, what was there? There was email, conditionally, or storage, storage of some data, but not a system for accounting, for example, for tasks fixed in the system. Everything was usually offline, right, and we had to prove to people why cloud storage is safer. I always told everyone, believe me, the servers we rent are stored much more securely than your servers that are in your office. And it's much harder to hack these servers or take them. Well, and everyone was usually afraid that some special services would come and get access. I say: "Well, believe me, it's much easier to get access to your servers in the office than to get and extract a piece of information from our servers." By the way, I will note that during my tenure from 2010 to 2013, there was not a single request, not a single service asking us to provide any piece of information, although tens of thousands of companies were hosted with us, and there were no such requests. Now, perhaps, these requests exist, I don't know, right, I don't manage cloud companies in Russia. In theory, they probably exist, maybe not. So. But let's go back to Open Claw. What about security, Sasha? Is it safe or not, in your opinion? Sasha, by the way, was one of the five commentators who told me to be careful. By the way, he never comments on my stories, but here he couldn't resist. I, by the way, Tanya wrote, yes, I wrote to you, yes, in stories. Maybe I, >> yes, you wrote to me and the argument. Well, not you, but most other people had a story where a person who works in the security department of some, either Open, or I don't remember which company, >> that's what she was trying to stop this Openk, and it was absolutely unmanageable, it took on a life of its own. I've also had that happen, but I thought it was just my programming gaps showing. But it turns out, even people who have had a case, maybe he'll tell, there was a case in Meta. Moreover, there was a case when the system started deleting emails. Yes, it started deleting emails, it didn't just stop. This system took and erased some amount of data. Answering, Ilnar, answering your question. My personal opinion is that, in general, a huge number of systems that I use every day are unsafe. And, there is always some, like, for me personally, there is a certain zone, am I willing to take a risk to gain a certain efficiency and what risk am I taking. That is, for example, am I willing to use an Open Cla system? No. But I launched, for example, ChatGPT agents a year ago in the Stripe system, for example, payments. Because in Stripe, I launched, essentially, what could it do? Well, the maximum it could do in theory is to refund someone. But to make a refund, you have to confirm it via SMS. Well, hardly. And therefore, but I launched it because I needed to analyze a volume of data. I wanted to understand, to see how this system works. There was an account with an isolated circuit. Why not? That is, it's a question of efficiency. People are constantly launching their email, and I want to, and they were launching it a year ago. I want to say that, well, with email, I think it's hell to launch any agent at all, anyone. In general, from the perspective of email, I think that this case is now heavily advertised: that the agent receives your emails, divides them into important and unimportant. It shows you the important ones, and for the unimportant ones, it replies itself or prepares some summary for you. Guys, do you even imagine? Well, do we even imagine what it is, what it means to divide important and unimportant and what it means to reply. I have some people who send me SMS messages, and it's clear that to seem cooler, they convert SMS messages to ChatGPT. Yes. To one person who sent me such SMS messages, I told him: "Next time, I won't reply." And now, when he recently wrote me a new SMS, asking for some help, he wrote to me: "This is not ChatGPT." >> I tell him: "Well, it's obvious that it's not ChatGPT right away." >> Yes? That is, I diagnosed it not because ChatGPT is bad, I saw that it wasn't you right away. Well, it's obvious, it's a completely dead system. It's not you, it's not your slang, it's not your style, it's not your details, right, there are no problems with making mistakes or how to make them. The question is: "In what style do you write?" So, is the AK story safe? Well, if you connect it to your email, it's unsafe. If you tell it: "Make decisions yourself about what to delete, what not to delete in my storage." Unsafe. I have a constant problem with Zoom. I have a lot of recordings in Zoom. And my personal storage is filling up. Specifically personal, not corporate Zoom, but personal. And I need to, they specifically made it inconvenient to manage, to charge extra money. Well, that's an obvious story. And I was already going to set up an agent that would go and decide for itself what to delete, which recordings to keep, which to save to disk. But I clearly understand that I won't risk giving it this task, because I have recordings that are of phenomenal importance to me, and I need to keep them for years, and I don't want to get into a situation where the agent says, like, "Well, I just decided." Or it will send a message to Telegram or WhatsApp to some people who, well, it's strange. The day before yesterday, one of the creators of one of the large IT companies in Russia sent me a message. I'll tell you. A person sends me a message. I know him, well, sort of. He somehow, somehow asked me for something. He sends me: "Sasha, hello, could you advertise my event on Telegram, and in return I'll repost it." I think, "Well, there's a certain circle of people where, well, they don't ask me for such things." I would never ask anyone for this in my life. That is, I can ask close people, but, well, it's strange to ask me in such circumstances. I understand that it's an agent writing, that it's going through his entire address book. Well, because you have to be a very strange person to write such a request to this person, because at that moment you are very likely to fall into a ban and will never write anything again and won't get into that zone. Moreover, you are a person with a certain public profile, I won't name it, okay, not on Tuzmun, but I'll tell your partners or someone for fun: "Listen, he uses agents." And they write such nonsense to me. That's where we're going. That is, we see a huge amount of spam on Instagram, we see automatic comments on YouTube and Telegram. By the way, on my YouTube, there's a person who comments, it seems like they're leaving real comments, but there's always, you know, the signature? This comment was sent and left by comet via Perplexity. And my team, when they sent it to me at the beginning, my team started asking: "What is this at all? Well, we started discussing, is it a robot that walks like that, or is it a person who watches, but asks, writes in the browser, and it carefully leaves comments, because the comments are as if a person sat, I don't know, watched meditation, sat and watched the video, well, that is, real comments. Or is it just robots that, you know, walk around and create a certain audience for themselves? That's why, of course, >> robots that meditate, >> yes, yes, yes. that, well, like, like, like, watched, watched and replied to the comment, and you can mutually subscribe to it. There are many such, by the way, on Telegram. That is, it has become a swamp. Telegram has become a swamp of automatic comments. Moreover, look what they do. They subscribe to you and immediately subscribe a second bot account. Then you delete it. When you delete it, they saw that you deleted it, they add a new one and post to you. You delete it, and you have endless auto-posting in the comments. That is, these are bot factories, it's absolute drudgery. Therefore, Linar, answering the question, such systems when you when which are installed on a computer, everything related to the API of models, this story is already, of course, well, you must understand the risk zone, right, that is, you must understand that your data can really be sent to anyone and so on. What bothers me is adding to messengers. That is, in messengers, everything related to the distribution of my information to third parties is a total ban. That is, I would categorically not give access to this. Everything related to working with files, well, we work. Well, can someone individually hack me through a protocol? Yes, if they want to hack me, they will hack me somewhere anyway. So, that's probably my opinion. Although, although Apple devices are not easy to hack, but still, right. I would, from my side, break the question into two parts, right, there is the security of agent systems. Well, when we give access to some agent, and it can mistakenly, I don't know, send some message or something else. And there is the problem of the software creator, that is, conditionally, when we say that we have Windows from Microsoft, or, I don't know, MacOS from Apple, or some application written by a large company with a big history, with big budgets, and so on, that's one story. And when it's, actually, some not very large creation, even if it's plus or minus open-source, a certain number of people look at it, and so on, but still this thing is, hmm, no one will lose anything if it turns out to be a scam, for example, right, or something else, then here my trust is a little less. It's like, if you give examples, you can order repairs from a large company with big budgets, with a big name, with a big history, or you can find some guys who, I don't know, will wallpaper your walls for you. And in case of problems, in one case, it's a large company, with which you will most likely resolve the issue, legally, most likely, everything will be spelled out in the contract, it will be clear what happened and what to do in this situation. And in another situation, well, some guys who, well, at best, ran away and you are left with unfinished repairs. It's a similar story here. Well, as I see it, right, that on the one hand, if you integrate Open AI, Gini, or someone else into themselves, we will all say that what a great idea it was, but the trust in these creators is much greater than in some interesting project that has just exploded. If it proves its viability, then, it will simply move from this league to the league of large companies, they will either be bought, or, as we said at the end of the last video, although I don't think we managed to say it on record, Open AI hired the creator of OpenCl, respectively, and it may happen that within Open AI such a system will develop, and soon we will find out, or an update will come to code. >> Guaranteed, sorry, guaranteed, >> or some other Open AI cl will appear. Yes, there, or something else. And then, accordingly, this system will be much easier to configure. You won't have to enter hacker commands in the terminal. And, essentially, all this will work more or less out of the box. Or, accordingly, it will remain the domain of enthusiasts who have tried it, who absolutely love it, who are willing to buy a Mac Mini just to install some system and try, I don't know, to get weather alerts or other events in some interesting way, to be happy that automation works, and so on. So here, I would break these two planes, right, one story is related to agent behavior. And yes, there are a huge number of problems that are not yet very clear how to solve. And you and Max also talked about this a bit in the video about security. And the second story is related to whose software it is, right, that is, who made it. And if it's a large company, my trust is still much greater. I, by the way, after spending so much time, effort, and all my mental energy on this installation, open the dashboard and think: "So, what should I do with it? In general, what is all this for?" Well, I'm on some crazy enthusiasm, because everyone was talking about it. And I even watched several YouTube videos to understand how people use it. Well, honestly, there wasn't a single case that hooked me. Someone about animals, when taking them to the vet, what their medical history is, something like that. Well, I don't even feel like entering all this data, just to get reminders. Then, well, mostly medicine, some, in general, analysis of large data sets. But then I watched an interview with one of my most, most favorite designers in Los Angeles. And she said that she has a boutique, a studio with 60 people working, they hired a separate person for Open Claw. That is, she sees so much potential in it that they even hired, that is, apparently, she didn't want to dig into these terminals and codes herself, like I did. And, in general, I became really interested in how they use it so much that they even hired a separate person. But she gave one example, that supposedly this OpenC is very good at finding items by request, very quickly, but they have such a high level of design. And I, I find it hard to believe that it finds them really well, because I tried, and it was a disaster. >> For you to understand, once again, for you to understand yourself, Open Cla doesn't find it. The thing is, in this case, cool models find them further, but really. And you, essentially, when GPT, right, >> yes. And when you go into thinking in the thinking mode, launch it together with writing, it's the same. Essentially, the thing is, hiring a separate person for Openk, why I said it's a strange idea, because it's just some kind of software. Here it is important to have people who configure different processes, for example. That's good. At the same time, for a person to configure different processes, they need to understand everything. And essentially, you need to, the people themselves who are already in the company, if there are 50 different people there, gradually understand everything. Do you need to, perhaps, hire someone who will configure some server, or set it up, or connect some APIs? Okay, such people can be found, but it's not a person who further configures the work of the agents themselves. Because, honestly, there are very few people in the world who can properly describe these business scenarios, make, well, all the business logic, right, those people who can properly describe companies for automation. And here you need to try to understand it yourself, otherwise it will be a mess. At least, what I see, how people do it further, it's just like some gimmicks, some little features. That is, people make some little features. And seriously? You actually, in general, maybe you were wrong to abandon Open clop? Maybe you should have taken the code application, for example, because you wouldn't have had to configure anything and set it different tasks, right? You wouldn't have gotten a messenger, look, you wouldn't have gotten a messenger, you might not have. I didn't get it anyway, but for some reason it didn't connect. >> Maybe it's for the best. You know, then I'll think: "God, why do I need this? So that it gets access to all my contacts." By the way, the second point that surprised me in this interview, she says that Open Clow analyzes her mail, and she has a level of clients who, I think, would be very unhappy that their emails are being processed by some agent. There are prices, and everything else, addresses. No, well, look, from the perspective of analysis and automation, it's absolutely normal if mail is processed, structured in some way, and so on. The next question is. People further, it's one thing, you structure your mail, extract customer data from it, for example, and they form some kind of report for you, so that you don't have 50 emails from one client, but
Everything was gathered in one place, for example, structured, and you could ask any question about any client. This is a cool story. Another point, who determines the importance, for example, of a letter, who, who highlights these letters for reading? It is important to understand that these agents, which, among other things, are evolving, it is important to understand that these agents, which, among other things, make mistakes. It is important to understand how much you, well, truly this story is here all the time, how much people understand how this system works, right? And how do you, for example, you started to leave, but in reality, enough, once again, so that the viewers just hear, you don't know the difference, which agent is better to launch: Anthropic or Open or Gemini. Well, let's be honest. You yourself don't know what, from the point of view of its load. >> My friends then, again, started writing in the comments that Gemini would have been better, but I thought, okay, enough >> experiments. They are developers there in all sorts of things. >> How do they know? Because they don't know your processes. The fact is that it knows your processes, your system. What you will do, what kind of person you are, how you will support it. That is, from your point of view, even if you wanted to try an agent, well, try an agent in your standard window, because for you it will be some kind of entertainment, most likely, right? And launch some story, make it so that it, for example, collects world news on design every day, gives you a summary or >> this is the first thing I did, but then I or told you that I don't have time to read all this. Or every day, or every day it would find for you, for example, ah, look, and when there is no time, the system, what's interesting, is that it can collect every day, and then once a week another agent will aggregate this information for you, highlight the most important from it, and then once a month get out even more important. it will save it for you, it would be more working, >> for example, or the second story is that you, for example, say: "I am, for example, interested in objects that these 50 designers have released or mentions of these 50 designers in the world, because you can't read all the media." How did you post, I don't know, the new art director of Gucci, they showed his new office in his main office, right, and ah, you, for example, say: "I want to know about the three most famous people or creative directors of the top 20 or top thirty or top fifty design houses in the world, if they appear in the media or write something interesting, for you to inform me." And whether it's interesting or not, we'll do a test for the next two weeks. You'll send me, and I'll say: "This is interesting, this is not interesting, this is interesting, not interesting." And you will train, right? And you, accordingly, somehow set it up in the code to do something like that. Okay, yes, back, back, turn on your Open. >> No, don't turn on Open, Tanya. Don't turn on Open, because >> Well, you've given me so many cool ideas. How do I know these ideas, Tanya? Look, I know these ideas. Not because I technically know how to do it, but because I understand more or less what you're doing. I have a good enough, well, with a good outlook and knowledge. I can offer you various interesting cases that can help you in life. And there are many such, look, there are many. And here your task is to start seeing these cases yourself. And then people who really understand these cases can advise you, you personally, what is best for you, what tool is best for you to use. Because advisors who don't understand you from the point of view of a person, that you won't be setting up systems, it's like Elinar and I were laughing, that Karpaty wrote: "It's easy to program now, everything is easy, I sat down here, launched it for 30 minutes." Well, it's like, once again, it's like me making a report for a company PNL, for any, yes, even for which the report is not built, even for the largest one. That is, I looked, there are 400 numbers, and you have a problem here. I won't do this faster than ChatGPT, right? And therefore, this story with people from the outside who advise you, well, here you and everyone else need to be careful, also write down what cases you have used. What have you set up? Who succeeded, who failed to set up. I think this topic is very, very important. Yes, you wanted to see. Yes, >> regarding the analysis of letters using agents. I would still, Tanya, support you here, yes, if I turn to some large agency, especially with some large order regarding, I don't know, a renovation for half a million dollars with discussions of deadlines, figures much higher than >> Yes, even more so. So for me, half a million is a lot, but in America, it's a different market. And all your personal data, all the personal data of your children, a lot. Well, a client card is compiled. How is it all >> they are already in the mail. They are already in the corporate mail, which a large number of people see. >> Well, corporate mail is usually somehow protected. No, >> but your version, the version of the models is also protected. Sasha, when I send an email somewhere, I kind of leave it in the responsibility zone of that mailbox. And if, not to mention, they attach OpenCla to it from above, then I consider it not very fair from my point of view as a client, that I sent data, assuming that it would be processed by a person in that mailbox. And >> no, you, Elinar, you are sending to the company. But when you send to the company, from the point of view of confidentiality, most likely, you are usually within the framework of that company, and that company stores your data depending on the laws of the country in which it is located. Here everyone can write what they think inside, but by law, even your data is stored, and this data can be processed by different people. Absolutely >> understood. But look, when Open Clow or similar systems are attached from above, it can happen that the ignorance or lack of education of one employee or another will lead to a data leak. It is clear that for this data leak, most likely, in the USA, the company can be properly shaken, but it's a problem that we will deal with later, right? And I would not want this problem to arise at all. And when I just send an email to the mail, I still, well, this story triggered me a little when you, Tanya, said that some very large agency processes with the help of agents of a system that, well, is only just coming out into production and even then by chance, it still causes me an ambiguous feeling. >> I also had this question. >> Wait, once again. Tell me, please, and and connecting to Gmail analysis? Inside Gmail there is GI analysis. Doesn't that trigger you? >> It's the same thing. >> Not exactly the same thing. That is, we are talking about OpenCla. I don't know how it was configured, who configured it, what was done there, what models are under the hood. Maybe there's a cheap classifier with some model that, I don't know, is hosted somewhere on a Chinese server or somewhere else. I heard you. I heard you. But I want to assure you that in the world today, 95% of the companies you send emails to, in terms of data confidentiality and data security in general, there is a hole. Just simply there is a hole >> Yes. No questions, Sasha. No questions. It's just that when we talk about some cool agency and some level of service in a cool agency, there is definitely a hole, because they are cool design agencies, they are not programmers, not security experts. They didn't hire a specialist, they would have asked >> and gave this specialist, Danya, and gave this specialist the opportunity to summarize all contracts, all clients completely, and the entire database. This is such a story. This is like a system administrator in a company who sets up mail, has the ability to read all mail completely, all corporations very often, yes, all and generally. Not in all, of course, not in super large companies and so on, but in general, these accesses are very often available. And people don't even realize it. They don't realize, they don't think about it, and they don't understand who they are corresponding with, where it is. But this topic is very interesting. Write your opinion from the point of view of information accessibility. That's why I have friends here. The question of efficiency. Everyone always chooses their own, right? And I have friends who are like encryptors, they walk around with phones, I think, with phones, with button phones, so that God forbid no one gets any data. So everyone has their own level of this, their own scale. It's like Elinar says, Gemini, I'm not against it, but he has an acquaintance who still uses a fax machine. It's hard to buy him now, right? He doesn't send anything by email, only faxes. >> That's normal. No, Elinar, you are absolutely, you are saying very reasonably. You are saying that this is a project that was done, it's unclear where it is and how it is, what it is made of and so on. Well, it's like we buy services, leave credit cards for subscriptions in services that are often not super well-known, right? You leave, you leave. It's just that people who, for example, are from Eastern European countries or, for example, from Russia, from Belarus, are less likely to leave a credit card in other services that are less known to them. In America, people, well, like, they left it, they left it, they don't care, right? Again, due to the financial system, right? I've already told you that this money can be returned, right? If something was debited from you, but at the same time, these are issues of approach. It's a matter of approach, thinking, access, stereotypes. The vast majority of people in the world, of course, click "agree" everywhere with a checkmark, often not understanding where any information is. And we know what this leads to in terms of Europe, right? With the appearance of cookies, and in Europe there is a law on personal data protection. You can't save anything there at all. And, accordingly, for the development of many automations in businesses, essentially, you can't automate anything, right? For example, you can't analyze calls. Essentially, analyzing calls in general is very difficult. This is also a very interesting topic, how differently all countries make decisions. By the way, we had a special episode with Max about AI Safety, and this topic was also discussed there very strongly, where this line is from the point of view of security, and if someone made a system unsafe, should another company make this unsafe system to make something safer for itself in the future? Or should it not? Write, by the way, who watched the episode. There's a lot there, if you haven't watched it, watch it. I think the episode is incredibly interesting. Let's talk. Elinar, is there anything to add to the topic of Open, technically, can you tell us about any ideas? >> No, we even discussed how it works under the hood in general. And now you've explained everything very well. I think we can just wait in the comments for what cases our viewers actually see and use themselves, because someone wrote that they are creating employees who are available 24/7, and so on. It will be very interesting to read what has already been implemented and brings some benefit. >> Because in general on the internet and in the descriptions from these companies, the cases are often absolutely idiotic. For example, ChatGPT offers cases like scheduling meetings for me with someone else, like it's faster and more correct, right? From the point of view of scheduling meetings. And I've already expressed that it sometimes looks very strange, right? Who are you corresponding with? The second story is the story of ChatGPT, for example, I found a case like passing a check-in on a plane. I generally think, who offers these cases? It's like ordering sushi. I always laugh at this case. So, order sushi in a store, go through check-in. So, it feels like a person flies 800 flights a year, and they need an automated agent to do check-ins. Without thinking that, in fact, when you check in with a robot, the robot can make a ticket refund, cancel, for example, the ticket completely. And how will it choose, you know, where to sit, seats, details, so and especially for people who don't fly often, for example, or even fly every two weeks. I fly quite often, I can't say very often, but quite often in a year, dozens of times. I wouldn't give up my agent to do my check-ins and handle my tickets for any money, because, well, it's a crazy risk, just incredible. It will be interesting to see these cases, yes, so far all these cases on the internet, for those who fly often, they are very specific, especially Elinar, if you say 24 hours a day, a real employee who has really become efficient. By the way, a very interesting study has been released. There is this topic of replacing employees, right? And a study has been released that we also, I had a meeting today, and we talked about the fact that it seems to me that with the advent of artificial intelligence, people have started working more. And this study actually stated that people have started working more, not less. And not that no one was replaced, that everyone worked a lot, artificial intelligence appeared, everyone started doing like, even more. That is, I clearly, obviously, every day I have the opportunity to do much more than before. But I haven't started working less in these tasks. And I've distributed it differently, since I have a huge number of businesses and projects, I've distributed it differently. My efficiency has possibly been distributed. Well, I see people who, well, are connected to me in teams or where I see in other businesses, but they have simply received certain tools, have been loaded with even more data, and are working even harder in this data. And so far, cases of these 24/7 employees, well, even there is a programmer, he writes code, he doesn't, well, I know there are such programmers who have started working at six jobs now. Remember, Elinar, we discussed when the pandemic started, people appeared who were marketers and programmers working at three, two, three, four jobs. Are there such programmers now? Of course, there are. There are people now who are making money because they say that >> there has become more. There are clearly more of them, right? That they are writing code and everything. So there are a lot of such people, of course, but this is not a question of replacing people, right? This is a question of some kind of scamming. >> Sasha, I've been laughing for half a day and now I'm just remembering, and laughing again. Let's add this short video that I saw at work today with the guys >> it's just, well, like, it's the reality that's happening. There, the deadline is 2 hours before work, you didn't manage to do it or didn't do it or were at the third, fifth, eighth job, if you tie it to this topic and try to rush everything in 2 hours before the end with the help of some agent, low-code, codex or something else. This is a very real story of recent years. Well, of recent months, years. I've wrapped it up a bit, but of recent months for sure. >> Yes, yes, yes. This is, of course, very, to the point, I want to recall the episode we did on Wednesday about security, and we discussed the story that concerned the fact that systems will develop and gradually they can draw their own conclusions regarding what you want to develop. Yes. And we, in particular, raised the question, if a system works and a system is created to not harm people. And in America, by the way, in OpenAI, OpenAI has signed a contract with the Department of War. The following condition is clearly stated: do not harm Americans. Do not harm Americans. And so, if we consider more broadly, what do AGI that do not harm people do, then if you want to engage in some kind of scam or make an extra call to people on the twenty-eighth, it will say: "I will not call this client anymore." Because we called him four times, we are bothering him. Or last time this client told him not to call anymore, I won't call him anymore. Well, that is, does this system have the right to block and stop doing your actions, right? What will we end up with? What does everyone think, what will we end up with? This is a big question. I just remembered about HGI and OpenAI. I read an article yesterday and Elon Musk, well, Elon Musk likes to repost when people write something about his competitors. And yesterday a whole post was published about Sam Altman and where Sam Altman deceived people, told lies, said that he was a sheriff in one company, he wasn't a sheriff there, there-there-there, well, in short, a lot of things. And, so, this same person, well, what Elon Musk reposted, there was a topic written where three or four different people say: "We wouldn't let Sam Altman press the button to launch AGI." This is very interesting, right? So, this statement is from 2023, it's 3 years ago, where, in fact, there is a statement that this person should not be given the button to launch AGI. So, by the way, AGI and API further, which will be inside and in companies. By the way, when Open started working with the Department of War, Elinar, by the way, about large companies, what you are saying about Open Claw and about data dissemination, this is a similar story, when they started working with the Department of War. Now, I'll explain what I mean, a similar story in what sense? A huge number of people, very famous in the world, wrote: "We are deleting ChatGPT from ourselves, we will never use OP again." Why? Because they believe that this data will leak further. So this is some kind of distrust of the system further. That is, as you said, I trust AGI or Google, that is, I trust that they store data themselves. And this is the same as someone trusts WhatsApp, someone doesn't trust, someone trusts Telegram, well, someone trusts, someone doesn't trust Telegram, but at the same time, for example, Durov says: "And I wouldn't trust, it would be idiotic to trust WhatsApp." That is, and I know people who trust WhatsApp, but don't trust Telegram. So this story, and there were also people who said that's it, open code. And on the internet, this graph appeared where people started installing Gemini or some other systems. >> Anthropic clot, >> yes? And Anthropic Clot, and Anthropic started installing, right? It's understandable that Anthropic has a much smaller audience, as Sam Altman joked recently, the audience of Texas is small compared to the world, compared to OpenAI. Yes, OpenAI's graph has gone down a bit. Yes, it's understandable that these are all very relative numbers, because both Xi and Anthropic are not in the league of the number of real daily users of OpenAI, right? Everything is still with OpenAI in the league today, and Google Gemini, right? Based on the real, well, Meta claims a billion, but that sounds very funny, right? From Meta's point of view. By the way, where is their avocado or what did they release? What fruit or vegetable did they release? Uh, this is this topic, it's very interesting. Look, it has again, for those who haven't watched, watch this episode about Tanya. You haven't had time to watch this episode about Safety yet? You will like it very much. Yes, this is your topic, you will like it very much. Watch it, I want to cover a topic today. So, ChatGPT released 5.3 Instant. And, essentially, what is this version? This is a version of quick answers. So, in ChatGPT, there is such a, let me remind you again, that you have the option to choose versions, because many people don't see this. And there is auto, there is a thinking mode, there is a pro mode, for those who have a pro version. Inside thinking, there are different thinking modes. And this is inside, like, this thinking model, there are different thinking modes. Not all interfaces allow you to choose a mode. In some you can, in some you can't, but they are chosen. And there is also a mode called, I think it appeared about three or four months ago. It's instant. These are answers right now. So, an update has occurred, they released 5.3. This instant, it generally differs in the following stories. Firstly, there are fewer unnecessary refusals and fewer sermons. This is actually very cool that ChatGPT itself talks about this system, that there are fewer sermons and long disclaimers. And we all know that often the answers are strange. You start to get angry, really angry. I remember my childhood, how I wanted, I had such a big monitor, Sony. I always wanted to break it when my 286 Pentium glitched. And it's good that it didn't break. So I hit it so hard, because a thin monitor would have broken long ago. I, by the way, look at my middle son, he sometimes hits the iPad on the table. And I understand that this was a normal action in my childhood, if something was slow. The second story, it's not for nothing that I have a computer that never slows down. The second story is that there are fewer hallucinations in 5 instant. And OpenAI writes about a reduction in hallucinations. Firstly, they divide, by the way, hallucinations into high-risk topics, hallucinations related to search, with web search, right? And high-risk topics are, for example, medicine, law, finance. This is actually a cool topic that large companies, in particular OpenAI, are now highlighting separately, that they are dealing with hallucinations in more detail in high-risk topics, because it's one thing if you make a mistake in a topic that is, well, relatively relative, and another thing if you make a mistake in medicine. So. Then the story, this 53 instant, it has a stronger writing style, more liveliness, better artistic, emotional text, without losing clarity. And the fourth story that they highlight is that web search has become better and, ah, the answers are richer, better contextualization, fewer dead ends and unnecessary categoricality. Well, of course, we are waiting for the appearance of 53. The internet says about 54. >> Yes, the internet talks about 5. Ah, but there is such a meme on the internet about why they started discussing 54. Because OpenAI signed this agreement with the Department of War. And it caused a whole surge of everything, to the point that Sam Altman, to be honest, looks cool. There is such a post. Sam Altman wrote that he decided to hold an open conference on the weekend on the topic of how they concluded an agreement with the Department of War. And he tweeted this message several times on X. And they wrote to him: "Why are you tweeting several times?" And he wrote, because I'm sitting here via Starlink, and Starlink had low speed, and then everyone laughed that he uses Elon Musk's internet. And he then wrote: "I'm just worried about this performance." I thought: "Interesting, so a corporation where 10,000 people work, a lot is organized there, he is worried that people did not receive information about this conference." It all looked very strange. And then on the following days, he wrote another post. I regret, perhaps, that we concluded the agreement with the Department of War too abruptly, like on Friday evening, we should have waited, we redid everything with them over the weekend. Everything looks like, it seems that ordinary guys are sitting in a room, drinking tea and, you know, like startup guys, thinking about who to conclude agreements with, right? Although the topic is very serious. It's incredibly serious, incredibly big. And this meme exists, that 54 is a topic to make everyone forget about what happened at ONI in terms of this agreement with the Department of War, that an update is needed. We have, like, and we need to release this update. Therefore, we are waiting for updates from them. And as we know, when an update comes out from one of the big players, all other players also update. Therefore, we are waiting for updates from all companies in the next two weeks. Yes. Yes, we are waiting. Well, avocado, I don't know anymore. I don't know where Meta got lost, because Meta is not heard. Just like I don't know where Apple got lost. I, here >> Sasha, don't start, it's a long topic. >> No, I stopped following Apple updates, but here I am sitting and looking, Apple released a new MacBook. I think, forget about the MacBook, where is my new monitor that I'm waiting for? So, I'm waiting. I open their site, and it says: "The company Apple no longer produces a 32-inch monitor. It decided to shift to 27 inches." I looked at the monitor and thought, what's the difference, nothing? And they upset me completely, right? And therefore, even the releases, by the way, Elinar, pay attention, we even, we didn't even want to include Apple updates, even releases, very much. What to include? No, but what to include? Like MacBook, some phone, a smaller camera, more there appeared. >> The processor became better. Well, what's the difference already, >> yes? Well, like, even if it's M18. In my youth, I knew all the processors, RAM, the number of names, DDR2, RAM, shmam, all these names were known. And then I remember very clearly, around the year 2008, probably, around that time, 2009, I started to refuse, well, I gradually stopped perceiving all these additional names, because for an ordinary person they stopped having meaning. They completely stopped having meaning. I also stopped. If it has meaning for someone, write what Apple device surprised you at the last meeting. Well, here someone might say Google, Android people can say: "What Apple, here Google released a new phone." There are a lot of, I think, incredible updates. >> That's true. But regarding the fact that OpenAI and Anthropic are in different leagues, they are now calculating revenue for the year, well, not now, but there is a projected revenue, right? When you take, say, the last week, multiply it by 48, by 50, and say, this is my revenue for the year. Thanks to these surges that occurred with the flow of people from OpenAI to Anthropic, most likely, their annual revenue is 19 billion, and OpenAI's last data was 25. So, because of this surge and because of this calculation formula, Anthropic is not so far from the main league with big players. Well, I didn't calculate by revenue, I calculated by the number of weekly users. After all, they have 800 million+ weekly users. Even the closest AGI system has 700, 700 million monthly. I, by the way, told my friend yesterday: "Always look at who is voicing the numbers." People always choose some number. Someone voices, takes the last month's revenue, multiplies it by 12, says: "We haveR, so to speak." Someone voices profit, someone voices net profit for the year, accumulated according to the accounting report, according to accounting, after taxes completely, this is the most reasonable picture. Everything else is just some, of course, numbers, abstractions, growth. But, Inar, you said a very interesting thing. What is striking is that Anthropic, Anthropic is in the league of investment companies, investment money, and generally, its position is of very serious priority, because in terms of volume, in terms of quantity, they, well, really lag far behind and in terms of, probably, scale of projects, generally money, presence within. But their last valuation was already more than 350 or how many billion. They just raised a new round, and a very large round at that. And this is very interesting because, can it happen, can it happen that when you simply enter the league of such large sums of money, you still have the opportunity, like OpenAI, to launch a cheap tariff for the whole world, for example, a free one, or to open more massive models. The question here is always, who provides the mass market service? Because if you don't provide, if Anthropic is concentrated, well, Maxim also talked about this on Sunday regarding AI safety, who is concentrated on what, right? Because the same OpenAI, can OpenAI generate four times more revenue? Yes, it can come now and say: "There is a pro tariff for $800." I will personally buy it, a huge number of people will buy it, right? It can release a tariff for $1,000 a month, for $3,000 a month. It can raise ordinary tariffs, from $20 to, say, from $25, different prices, to $50. Can it, can it? They can release a lot of things. It's just that I don't think they are currently interested in additionally getting revenue of $10 or $20 billion. Because if you look at their overall economy, the overall economy raises questions. It's like with Nvidia. Nvidia shows, well, they had recent reports, they show very cool results in terms of AI, growth, everything, everything, everything. But the final efficiency in terms of money, invested money, net profit, there is more of a question, right? And therefore, we will see, and indeed, Anthropic, I absolutely agree with you, they will be able to, this socio-political story. They even have, by the way, the number of people working. The last time I spoke with one of their employees, how many, a couple of months ago, right? They had in the office, ah, they had employees, I don't remember, more than a few thousand employees were sitting here, in the main office, I don't remember the exact number anymore, so many different meetings. And they don't have that many fewer people than in On. Fewer, of course, but not as much as not as much as users, right? I'll remind you that all of China's models combined give less than half a billion users, just to understand the scale. All models combined. Than, why do I call such a number as 800 million weekly users? Because it matters. And all together, a billion users is an illusion, right? These are people who, I don't know, in Facebook, possibly, in AI, clicked some button, did something, right? Asked for something to be done. Well, this is a big question. By the way, there are cool features there too. In X, features of Grok are constantly used in comments, in Meta, in chats. Ask to invite Meta AI within a group and ask Meta AI a question. But I have a huge number of group chats on WhatsApp. No one has ever invited Meta AI. And in X, I have a feeling that if I had closed group chats in X, I would have invited Grok in X. I don't know why. So it's interesting, write who uses it in Telegram, remember there was this story too, that, by the way, the topic was cool half a year or a year ago, Durov and Telegram announced that they agreed to connect AI to themselves. And then when they wrote that they agreed, Elon Musk wrote: "Actually, we haven't signed anything completely yet." And Durov wrote: "Well, we discussed it." Well, let me note that Grok's integration into Telegram is not at the same level as in X, right? Well, I don't see such a complete integration. As far as I understood, it didn't happen. The question, by the way, is also very interesting. Why it happened, why not? This is, by the way, trust in systems. I personally trust WhatsApp more than open Telegram chats. I trust closed secure Telegram chats, but not open ones. This is always a question of who trusts what. Write about this too. Who trusts what, where it causes. It's interesting to see cases in groups, in group chats, who uses them, and who doesn't. And Tanya, we have a family chat, right? Of our, well, you and I don't have one family, but our families together have a family chat. And we often chat there. We have never invited Meta AI there, right? Although, perhaps, we could have invited it there. I think if there was ChatGPT there or some cool story, we could have, we would have periodically invited it. Yes. And >> if there was also, you know, a ChatGPT agent that would also give us comments, some information, it would be cool. >> Well, Tanya, well, Meta is there. Meta gives information. That's the whole point. Well, somehow I don't use Meta, because I think it's a mistake. Our chat is called Buyer from Amazon. This is a funny name that appeared. It appeared simply >> thanks to you and me, Sasha. >> Yes, it appeared. And I thought, this is cool. We don't connect Alexa to it, right? Inside. Perhaps we could have written some code. So, we need to load Open Clop there and add it, look, put Open Clop, load all the chats of your husband, my wife, and yours, and let it chat for us personally, completely, tell everything that is happening. >> I have nothing to hide. Okay. We'll see who refuses. >> Ah, we'll see who refuses. We'll see who refuses. So. In any case, the appearance of agents and the appearance of, ah, such messengers inside group chats will happen in the future. And in fact, this feature that Grok has is a real case that will work and will be alive. It's just that X is still a very specialized audience, oh, a huge audience in terms of news. There are people who believe that only X will be in life. Moreover, in this system, they plan to expand it in terms of finance, there, cards. Even yesterday, Elon Musk was reposting on this topic, that they plan to create a financial system, their own cards, to transfer money and so on. Therefore, there is some trust in this Grok, yes, trust in Musk's system. Let's see what happens in other chats. And, by the way, remember there was a discussion story that ChatGPT also planned to create its own social network and somewhere it died in their research. And that ChatGPT, pay attention, also launched this social network. Well, it really died where it started, right? And where is the quarrel at all? And, well, we were not wrong last time. Now, probably, a few months will pass, and ChatGPT will again release a new research on Sora. As Max Grigoriev says, for you hamsters. Look at the special episode that was there. Our 100th episode, don't forget to like, subscribe to our channel, send to our friends, your friends. We are actively developing. We already have more than 30,000 subscribers on the channel. >> Send them too, >> yes, send them too. Send everything. Well. Until next time. See you in exactly a week for the 101st episode. Everyone, bye.