Навигатор по новинкам AI: Февраль 2026. Александр Свет

Good evening everyone and welcome to the February AI New Releases Navigator. And as is tradition, every month we meet and explore what's new in the world of AI and, most importantly, how to apply it all in practice, right? That is, we don't just announce some news, but we discuss how these things will affect our practical work. We do a lot of very practical things. And today there will be very interesting things. [music] And in general, you know, February turned out to be unusual, because when I was planning this stream, well, honestly, there weren't that many topics. I even thought: is it for the first time we'll just sit down so calmly, talk about AI in such a thorough, quiet way, about interesting things that might not be topical, but important. And the closer the stream day approached, the more and more news appeared. And, in fact, the culmination of all this was that yesterday evening, two top models were released on the same day. That is, yesterday both Opus 4.6 and GPT 5.3 Codex were released. These are two new flagship amazing models. We will, of course, talk about them today. And it turns out that, as always, there is a gigantic amount of news. Here's our plan for today. So, first we'll talk about new models, then we'll have the main topic of today's stream - how AI coding agents are becoming universal agents and how people who have never coded in their lives can use them. And then we'll talk about AI coding news. We'll discuss the new CODEX application. I'll show you how to make a game in it. Yes, today we will make a cool game. And, of course, we will definitely discuss CL3.0, because this is perhaps the most interesting breakthrough in video generation. And finally, we will definitely discuss the future of AI. And if you want to constantly follow this news, welcome to our chat, because everything that, well, not everything, many of the things we will discuss today, we have already discussed in the chat, so join us. It's interesting there. Well, and also, let's go through the obligatory program a bit, right? We are conducting all this at Profайл School. It has been on the market for 13 years, with over 70,000 students. And I am Alexander Svet. I am a practitioner, an enthusiast in development, and I have been working with AI for 4 years. And as it says here, I know how to explain things simply and to the point, especially to those who are just starting. And right now, all my AI development courses are available with a 20% discount until February 12th. We will talk about these courses in the process, and, well, because it's a truly very, very interesting course. And among them are two new courses. And plus, you should have received some promo code by email, so be sure to check your email. Let's begin. So, let's start, in fact, with the battle of the giants, because this is perhaps the most important and significant news here. Yesterday evening, two very, very important models were released, right? So, first of all, Clot Opus 4.6 was released. And if you follow AI at all and you know how important a release OPUS 4.5 was. For many people, and I consider myself among them, it became a turning point in AI coding, because now with opus 4.5, you can solve almost any AI coding task. And therefore, of course, everyone was very impressed that this model has finally been updated. So, it's been about 2 months, and Opus 4.6 has been released. And simultaneously with it, GPT 5.3 Codex is released. Let's see how they differ. And it's not even about which one is better, because, honestly, right now it's impossible to say that one top model is better or worse than another. This is, well, a slightly incorrect judgment, right? If you take simpler models, then you can really say that, well, it's worse, as it were. But if you take the very top models, for example, Clot Opus, JPT fifth series, Gemini third series, they are all absolutely amazing models. Each of them simply has its own strengths and weaknesses. Let's start with the context window. For GPT 53 Codex, it hasn't changed. It remained 400,000. But for Opus, it has grown to 1 million. By "grown" I mean that yes, you can indeed use CLD Opus 4.6 with a million-token context window, but it's important to understand that it will become significantly more expensive if you use it that way. That is, its cost with a million context, it's not with a million, it's its cost if your context is over 200,000. That is, you have a standard context window, which Sonet and Opus always had - 200,000. And now, in fact, Opus has a million-token context window. Sont already had it before. But Opus never had a million-token context window. But if your context is over 200,000, you pay $10 for every incoming million tokens and $37.5 for every outgoing million tokens. And as you can imagine, this is actually very expensive. Therefore, yes, there is a million context, but in reality, well, it's more likely that people with either a lot of money or who really need such context will use it. How have these models themselves changed? In fact, the changes in them have been quite interesting. Interesting in the sense that they are both moving in a very correct direction. And here, of course, it's also very important to say that the models were released literally yesterday, right? So, naturally, I've worked with them, read many reviews, but they were released yesterday, right? So, it's difficult to form a complete picture of what's happening in less than a day, so perhaps my opinion, and the opinion of others, will gradually change. But what do we see so far? Claude Opus 46 has become more reliable, which was the main complaint about it in the past. It was very pleasant to interact with, and you could solve almost any task with it. But if you worked with Opus, you noticed that it does everything very well and it works. But often there are small problems here and there. Some bug. It works, but with that bug, right? It lacked a bit of that reliability. And that's what OPUS 4.6 has received. So, in terms of more thoughtful, more in-depth approach to code, to creating more complex applications, OPС 4.6 has definitely become better, but it still remains, perhaps, the most suitable model for complex tasks where there are many solutions. That is, you know, there are tasks where there is exactly one solution, right? Well, well, Opus is not for that. Opus is for tasks where you can find many different approaches and it's not clear which one is best for you. With Opus, you will find such an approach. You will find an approach that is best for you. It will understand you. That is, the best description of Opus is that it understands what you mean. That's the whole essence of Opus. And therefore, it is very good, for example, for research and planning, for the parallel work of many agents. And most importantly, for working without a very precise brief. That is, if you don't provide a perfectly precise brief, like, you know, a highly skilled programmer might provide, Opus works excellently without such a brief. It will understand what you mean. It will do what you actually want. But at the same time, Opus 4.6 has a couple of drawbacks. So, first, the first drawback is that it writes worse, by feel. I mean, it writes ordinary texts worse than OPUS 4.5. And this worried me a bit, because I have my own channel. Now I'll promote it a bit. This is my Telegram channel where I publish various translations of the coolest articles about AI. So, you can subscribe to it. I'll send you a link to it in the chat now. And every day, about four translations are published there. And all these translations are done by Opus 4.5. So, if you like how it writes, then everything on this channel is done by Opus 4.5. And I personally really like how it writes. Therefore, if you want to read the coolest articles, be sure to subscribe. And it's very interesting that this channel is run completely automatically, except for one thing. All the articles here are personally chosen by me. That is, I just send a link to my bot, I'll show you how it looks today. And it translates everything. But I choose the articles. And then everything else is done automatically. By Opus 4.5. And therefore, this news that OPUS 4.6 writes worse, it worries me a bit, because what if at some point OPUS 4.5 is removed, and they say, well, you know, we don't need this old model anymore, right? Remember how they removed GPT 4, and people really started complaining, saying, bring back, bring back our GPT. Well, and the second point is a consequence of it becoming more thoughtful, it has become slower to work than OPС 4.5. It's not that it's a super bad thing, but perhaps you'll notice that it has become a bit slower. Now, regarding GPT 53 Codex. So, in general, like all previous codexes, these are the best models that are best suited for technical tasks where execution accuracy is important, right? So, if you are, for example, a very experienced senior programmer, you might prefer to work with Codex, right? So, you don't need it to guess what you want. It does what you say. This is a very strong and weak side of Codex. Because if you can provide it with a very well-described project, it will implement it really well. But if you want it to understand what you actually mean, then no, Codex is quite difficult with that. It is excellent for automation. And, by the way, it has become faster, noticeably faster. And perhaps we'll talk today about why this might be. One of the reasons is that, well, first of all, Codex itself consumes tokens differently, it consumes tokens more efficiently. That is, it consumes fewer tokens. Well, and since it consumes fewer tokens, it takes less time. But it also has drawbacks. First, it also writes worse than Opus 4.5, but this is a disaster. A disaster for all GPT fifth series models. While preparing for this stream, I read blind tests that were conducted. They created texts, you know, GPT 52, GPT 53, OPС 4.5, 4.6. That is, they wrote different texts and gave people to look at them blindly, to see who would choose what. Everyone chose, well, most chose OPUS 4.5. That is, no one chose GPT 52 at all, because it wrote terribly. Well, and 5.3 also writes so-so. So, this is a problem for all, this entire GPT5 series. It writes texts poorly. It's very noticeable that these are AI-written texts. And there's another problem. And this is, of course, a disaster. A disaster for Open AI. It seems to me personally. Because I haven't noticed this. For example, with Claude. Many people talk about the fact that with GPT 53, during serious, long work, at some point they might switch you to a weaker model, and you won't know about it. Well, you'll just feel that it has suddenly become duller. But this is a problem in general with Open AI, that their model routing is not working very well, so people observe that this happens. So, if we summarize, right, that Codex has become more, this Codex 53, it has become more human, right? That is, it's more human, right? That is, it's more pleasant to communicate with than with 5.2, but still, it's still a model for more technical work, and Opus has become more reliable, but still not as reliable as Codex. But at least it's very pleasant to communicate with. And here, in fact, I wanted to talk about my two new courses, one of which starts very soon, the second a little later. The first course is called "How to get the most out of Chat GPT". Let's open it. So, this course is, well, not exactly a mini-course, but compared to many of my other courses, it's a mini-course, because it has only four sessions. And the next group starts on February 23rd. There are only four spots left. And the point of this course is that I want to show you what Chat GPT Codex is truly capable of. This is not a story about, you know, how to prompt GPT correctly. No, no, it's not about that at all. We explore various capabilities of Chat GPT that are hidden from most people and that you might not know about. And we simply analyze how each GPT tool, like deep research, Agent, and other things, can be used correctly. And most importantly, we start working with Codex not for non-programmers. How to work with files using Codex, how to create presentations using Codex, how to analyze large data using Codex, how to create a beautiful, simple website using Codex. So, for you, this will be an excellent introduction to understanding the full power of coding agents, and not just for coding tasks. And if perhaps an administrator can post a link in the chat, I recently conducted an introductory session for this course. You can find the link here. And there you will see examples of how it works. Take a look, you will see how coding agents can now be used by people who are not at all related to coding. And not just can, but actually should. My second course is called "Deep Dive into Cloud Code". This is more about development. Well, not even more, but in fact, it's necessary if you are interested in AI development. And it starts on March 23rd. And here, first of all, we dive very, very deeply into Cloud Code, into all its capabilities. And there are many interesting capabilities in it. And most importantly, we will develop an agent based on the Cloud Agent SDK. I honestly don't know if anyone teaches this in Russian, because this is a very new and very interesting area where you can create your own agent that will use your subscription. And at the same time, it can be in the form of any application, right? So, it can be, for example, even, I don't know, a video editor, you can make something like that, but that's a more ambitious task, but it's possible. I will have an introductory session, you can also sign up for it, and where you will see how it all works. Now let's talk about other news and new LLMs that have been released during this time. So, first of all, GLM 47 Flash was released. This is a small but very powerful local model. Well, it's designed to work with it locally, meaning to run it on your own computer. It still requires a fairly powerful computer. That is, it's not something you can run on just any computer. A fairly powerful computer is needed. But if you have such a powerful computer, you will like it, because I have read many reviews and feedback about it. In principle, this is essentially the first truly reliable local model for agent tasks. That is, when you can do things like have this model use various tools, use your other services, through MCP and other technologies. So, it codes well, performs agent tasks, uses tools. Of course, undoubtedly, not at the level of top models, right? You need to be aware that no general-purpose model running on a local computer, on an ordinary local computer, will work even remotely comparably with top models. This is simply technically impossible at this stage of AI development. But it's an interesting breakthrough, and you can try it, perhaps you'll like it. The second model that was also released in the open-source world is Kim K25. This is currently considered the most powerful open-source LM in the world. It has surpassed GPT 52 and Clot 4 OP in most tests. And, well, not in all, but in coding, it didn't surpass Opus, but in many, many tests, it surpassed all top Western models, because it's also a Chinese model. And what's very interesting is that it natively understands text, images, and video. It's cheap. So, compare this with the prices of Opus. Remember I said that the price of Opus, if with a million, right, is $10 and $37, and if it's, I think, 200,000 tokens, then the price for Opus will be around $5 and $24. Somewhere around there. So, it's, well, like 10 times cheaper, figuratively speaking. And at the same time, it's great. I personally really liked it. I've already started using it. It's a really good, solid model. You can try it on the ki.com website. And there's also an interesting thing called Agent Swarm. This is a swarm of agents. So, it can independently launch up to 100 parallel sub-agents. And here, you know, a very interesting discussion happened within the AI community, because all these new models and technologies that can manage a large number of sub-agents are coming out, and they are all called Agent Swarm, that is, a swarm of agents. And at some point, one of the rather respected people in AI came out and said: "Gentlemen and ladies developers, what about not calling it a swarm of agents?" Well, you know, "swarm" is a bad word. It's hard to think of anything good with the word "swarm," right? Whenever you think of a "swarm" of something, it's some kind of scary, aggressive, dangerous thing. And maybe let's call our agents a team of agents, a unit of agents, but not a swarm of agents. Understand? A swarm is something aggressive that will attack you, and you are creating this kind of branding yourself. And in fact, later, we'll talk about it a bit later, Cloud Code went down this exact path, it stopped calling it Agence SWM. That is, it sounds cool, a swarm of agents, but as marketing, it's terrible, it's a very, very bad word. Also, another Chinese model, Qu Max Thinking, was released, which is a maximally thinking model. It's catching up to KIC 2.5. And it has several interesting solutions. First, it scales during inference, meaning it can think deeper, it can think less deeply, and it decides for itself during the problem-solving process how deeply it needs to think. Well, and it also has adaptive tool invocation. It's not something super new, but it's also an interesting technology. So, in general, they are also doing well, progressing. I advise you to try Kimi. That is, N, I'm not a fan of N models, I personally don't like them very much. But K, I think, are very worthy models, very cool. Next, next, next, let's see what's happening in the world of medicine, because both Open AI and Antropic have made statements that they are launching medical products, and this will be chat GPT Health and Clot for healthcare. And what's interesting about them? Now, well, they are slightly different, but their essence is generally the same. That now, well, it's in closed beta testing for now, but gradually these models will be able to receive data from various medical devices, right? For example, Apple Health, or My Fitness P. This is actually a very important breakthrough, right? Because imagine, the data that is collected about you every day, right? About your heart rate, if, for example, with Apple Health, data is collected, heart rate, they can collect a lot. And imagine, Chat GPT will now be able to analyze this data. I think this will save a huge number of lives. And I really liked the quote from the founder of LinkedIn, when he said that if now you or your doctor do not consult with top models when giving a diagnosis, then both he and you are making a very big mistake. I completely agree here, because I think it's essential to consult with top AIs, but also to consult correctly, not just to ask something, but to upload a lot of your various analyses and medical history. It's advisable to consult with the very top model. Well, let's put it this way. If you have a serious illness, it makes sense to get the $200 plan from ChatGPT, because you will have access to a pro model, and it's an amazing model for deep analysis of data, information, and these kinds of things. And what's interesting is that these chats, right, your chats with the model, will be completely separate from regular conversations. So, these won't be regular chats, because, well, this is very important information. It will be stored in a special way, it won't be in the model's memory. Well, in general, this is a completely new thing. Also, continuing with ChatGPT news, there's a lot of it this month. ChatGPT Translate has been released. This is a separate new service. Here's how it looks. Where you can translate using ChatGPT, right? So, it's similar to Google Translate, Deeple, and other similar services. And, in fact, 50+ languages with automatic detection, contextual translation with tone consideration. So far, there's no translation of documents or websites, nothing. So, this is more of an experimental project, I think. I tried it. What can I say about it? That it translates tone well, right? So, it correctly translates intonations, correctly conveys meaning. But it has problems with terminology. So, for example, how translations are done on my channel, which I run, I wrote a special skill that does the correct translation, in my opinion, as it should be. And I invested quite a bit of time in writing this skill. And I attached many references on how it should translate certain terms, how everything should work. And I compared the work, for example, on the same texts, of my skill. And how this ChatGPT translate does it. In some cases, ChatGPT translates better. Where there are complex phrases, especially Anglicisms, ChatGPT translates better. But in almost all cases, it translated anything related to terms terribly. It unfortunately doesn't sense this. So, it doesn't understand. For example, in English, when they talk about development, there's a word like "shipping," which means, not that you're delivering something, but when you, well, when you release a product, and it's called that, it understands this poorly. And there are many such things. So, for example, in my field, it translated articles poorly, but it translated speech patterns well. So, try it, perhaps you'll like this product. I see Alexander asking: "And in coding, what's better: Quen Mark Thinking or Clpus?" Cloud, of course, Opus. No one can even, well, no Chinese model can surpass Opus in coding, nor, well, Codex, in fact, either, because, well, I mean.

The code, it needs to be surpassed in terms of technical aspects. As for Opus, based on feelings, based on feelings, not a single model, well, K2.5 is approaching it, but still, it cannot surpass them. It's hard to explain. But no one yet, in my opinion, can truly surpass Opus in terms of pleasantness of working with it and the overall quality of work.

Also, a new plan has appeared worldwide for ChatGPT, called Chat GPT Go. And it's a budget subscription for $8 a month. Honestly, I don't quite understand this subscription because, on the one hand, well, you're still paying, and $8 a month isn't exactly cheap. Well, it's like choosing between 20 or, well, you'll pay half anyway. Yes, you'll get 10 times more messages than in the free version, yes, you can create more images there, but there's no research. To be precise, it exists, but it's very limited. There's no code at all, no code, no agent mode, yes, this agent. And generally, a huge number of interesting things are missing in this version. I don't quite understand what it's for, because, I think, well, if you don't want to pay, it's better to go to AI Studio, for example, and use it for free. And if you are paying, then you should probably pay for the full plan. Well, maybe I'm missing something.

And about the speed, yes, remember, I said why GPT is becoming faster? It's because, oh, I have some incorrect text here. It's because they formed a partnership with a company called CBR. And CBR is a very interesting type of chip, fundamentally new. They are significantly different from Nvidia chips, and thanks to this, they work very, very fast. Really very fast. I'm talking about tens, hundreds of times faster. And therefore, I think, thanks to the fact that now Opus will use these new chips for inference, that is, for token generation, then, well, the code will actually start working faster. It has already started working faster, I think, and it will gradually become even faster. And by the way, regarding chips, it's also important to say that Opus 4.6 was trained and works on these new Google TPUs. That is, it doesn't use, well, at least partially for sure, it uses Google's TPUs, not Nvidia's GPUs.

Now, let's discuss our main topic for today, namely, coding AI agents are becoming universal agents. And I think this is the most important thing. If you look at our entire stream today, I think this is what you should truly understand. If you truly grasp what I'm trying to say, and I hope I explain it well, it will be super beneficial for you in the future. Because many people think that a coding AI agent is just some kind of programmer's assistant that solves programmer's tasks. The further we move in the development of AI, the more we see that this is not the case at all. And I started talking about this several months ago. I started saying, guys, understand, you can work with files, you can work with big data, you can create presentations, anything. And about a week and a half to two weeks ago, a completely new direction emerged. This is motion design using coding agents, that is, creating motion design, all sorts of video animations through coding agents.

How does it work? Well, the technology called Remotion has emerged. It's a framework that turns code written in React into video. So, for you to understand, it's essentially code written like when you visit websites, and you see everything so beautifully, right? Beautiful texts, everything looks beautiful. They just took this same technology and are using it in video. So, essentially, each frame of this video is just a React component that receives the current frame number. That's all. So, they just came up with a brilliant and completely simple and brilliant idea. They just took the technology that already exists, the technology that coding AI agents work with super well, and essentially turned it into a video generation technology.

And what's very important to understand is that, unlike, for example, Sora, some Veo, right, Kle, all these video generation tools, right, where it's generated, here the video is code. So, each time it will be the same. This is because code is a deterministic thing. And if you write code once, it will always work the same way. And therefore, you can, for example, easily edit videos this way, right? For example, you create the beginning of some video, then you don't like something, you want to change something, and you change exactly that, just like you would change something on a website, right? You can change part of a website, and here you can change part of a video, and the rest will not change. So, it won't be like with Google Veo or others, where everything is generated anew each time and it will be different each time. No, that's not the case here. This is a clear, precise, deterministic process. And the most important thing is that since it's code, you can insert any data into it. So, you can, for example, make it so that, for example, this code is generated taking into account, for example, user statistics or user names, or take some different pictures from a database. And essentially, you can generate unique videos for each user.

Thus, let's see how it looks in practice. Here are some examples of what you can do. You see, everything I'm showing you now, this is generation through code. This is not generation, once again, like Sora, Veo, it's completely different. For example, like this. This is all code. Well, music is just overlaid, but everything else is code. So, imagine the possibilities. And this code can pick up, for example, the name of a specific user, right? So, for example, in their personal account, a user will see exactly this video with their, well, with their [music] or with their personal information. Or, for example, look. By the way, Ina did it this way on GitHub. They made such a video, like, with the highlights of the year. And you see here, like, my most used programming languages. And it shows, like, C++, right, and others. And where does this come from? It's because it analyzes the user, which languages they used the most. And these values are simply inserted into the code, and the code generates this video, and that's it. So, it's so, you know, an impressive process in terms of the future. So, imagine that the advertising of the future will look like this: an advertising video will be generated for each person, specifically for you, based on your personal behavior, videos will be generated for you like this, and this will happen very soon, it's already practically here. Or, for example, let's look at another example of how to use it. Let's say, um, now, well, for example, these guys also made this, all of this, well, except for the recorded person, all of this is made. Everything you see is made using code. And this is so, well, like, this thing will simply turn everything upside down, if I'm honest. So, understanding this will turn everything upside down.

And let's see how it works in practice, because you can also use all of this yourself now. So, what will we do? If you watched my previous streams, then you know that in our previous streams, this was a few months ago, you can find this stream on YouTube if you're interested in watching, we did such a fun experiment. We took, I found, it turns out, there is a really serious database that contains all the history of observations anywhere where a UFO appeared, right? So, this is, well, really, I was just looking for some datasets to work with, and I found such a funny dataset where there's just a huge, huge, giant amount of information about where UFOs were seen. I thought, great, this is an amazing dataset for our activity. And in that session, we did this. We turned this table into a visualization on a globe, where, essentially, you can see where what happened, right? What shape, in what city, where it was seen, what UFO, right? So, you can click right here, and you'll find all sorts of interesting information. So, if you're interested in how it's done, find it. It was, I think, October or November, somewhere around then we had a stream. And I, in fact, detailed how we created this technology right on the stream. And to show you just the power, right, of this generation, let's do this. Let's try to take this already ready-made thing, right? This globe with various information here. And let's try to turn this into a, well, like a small interactive cartoon. To do this, I will open my code. Now, yes.

And, so, what you need to do is, you need to, you need to launch this technology, right, which is called Remotion. Um, it works as follows. You need to enter a command once in the terminal, and that's it. And it will install the project, create a new one. And how does working with it look? It looks like this. So, now, now you need to. Ah, in principle, you can ask the agent to do everything, but you can also do it yourself by just repeating the commands. And it will now launch a server. And this server is, essentially, a video editor. Well, well, it is, um, I think they will refine it further, because it can work much more interestingly, I think. Now we will launch it. Here it is launched. This is a video editor. So, here we will have a video later. And everything we see here will be created by AI. So, how will we do this? Let's say, let's give it a prompt like this. So, now I'll turn on planning mode and say, for example, this: Look at this project and use the Remotion skill. Be sure to look at the Maps rules in this skill. There you will find all the necessary instructions on how to do everything correctly. And I have already created a Remotion project. It is in the My Video folder. What you need to do is, you need to create, a beautiful, well-thought-out motion graphic, where five cities in America will be shown, where, well, five places in America, which are marked on this map, and you need to open one place, show the event that happened there, then move to the next place. And so that everything looks beautiful in animation, so that there are smooth transitions, smooth camera movement, so that it all looks great. So, be sure to study the Remotion map skills and study the project. Study all of this and create a plan on how to do something like this. That's all. So, I essentially told it, right? There is a special skill, it must be used for Remotion. This is, essentially, a skill that trains the AI model to work like this.

So, now, just a second. And therefore, it will now go, read, let's see. Ah, yes, it's reading. It's reading the Remotion skill. Now it will look at everything here, study everything, and create a plan on how to make such an animation, right? Because, again, this animation will be made simply by code. And we will see in this window how, how later all this animation will appear here. I'm not sure if it will do it well right away, because, again, we are only at the beginning of this, right? But there are already a giant number of examples of what people are doing. So, while it's thinking, I'll show you a few more very interesting examples. Ah, this is a showcase. They have a prompt, I think. Prompts is also how they use it. Promos. Oh, here you can see, for example, if you want to make a spinning logo from SVG format. You can make such a 3D logo. You want to make, you know, an equalizer, no problem. You want to make, for example, some kind of infographic, very easy. You want to make it so that, for example, text is typed out like this. Please. You want it so that, for example, you know, a news headline is underlined like this, no problem. So, all of this already exists, and all of this can already be done.

Moreover, key companies involved in video generation have also understood, they have already understood this power. And, for example, Hixfield, I think for sure, if you are involved in video, you know this. They have already launched their own very cool, well, I haven't tried it, but based on their videos, it's a very cool, own coding agent, which also, by the way, uses Anthropic models. And, now I will show you a video from it, how it works. Here. So, this is what they did with its help, all of this is done with its help, specifically their own agent. In fact, I think this agent is pretty much the same as what we use, only plus it's, well, better prompted, well thought out, right? So, well, the guys approached it with, apparently, a sense of purpose. All of this, everything you see now, you can also do. So, these different types of transitions. You can insert your own pictures there, right? All of this, all of this you can do, right?

Let's see what it has done here. So, I have a good picture. Let's now read the key files directly and launch the agent for detailed design. In general, let it do it for now. And let's see, maybe while it's doing it, we can talk about other topics, because we have many interesting topics today. And by the way, yes, here's another topic about video. Well, not that it's about video, but it can also be told about video. So, I want to tell you about Claude Code Work. This is also a very interesting thing that appeared, how long ago, maybe also a couple of 2-3 weeks ago. Claude Code Work is essentially code for people who are completely far from development. What does it allow you to do? It's like a chat mode, but with access to your local documents. And you can do various things with your local documents. You can also analyze tables, work, do simple things with files, and it's all secure. So, you don't need to be some kind of technology expert to use all of this, but unfortunately, it comes with a price. The price is that Claude Code Work has fewer capabilities than Claude. So, Claude can do anything. And I'm not just talking about Claude Code, but about CodeX or any other coding agent, it can do absolutely anything. Here, there will still be limitations, it won't be able to do everything, it won't be able to make all API calls. So, it's somewhat limited, but its capabilities are still interesting.

So, I want to launch now. I want to wait until it creates a plan for me, and after that, I will show you what you can also do in Claude Code. Evgenia asks: "How long did it take to create the video in Hix?" I, well, precisely, I don't know, because it's a promo video. I think it still takes a considerable amount of time. But, firstly, I think less than it would take to do it manually, and secondly, it's accessible to people who can't do it manually. Well, I, for example, personally, I know absolutely nothing about motion design, absolutely nothing. And therefore, I won't be able to make such a video, no matter how much I want. It's like, you know, I'm not a programmer, right? And therefore, without coding agents, I wouldn't be able to do many projects that I do. Well, yes, all the projects that I do. And it's the same story here. So, it just gives you impossible new potentials.

So, we are looking now. It's designing with the agent now, and let's see. See? So, this is already working with Opus 4.6. And you can feel it, right, that it has become a little slower to work. Because the previous Opus would have done everything very quickly, boom-boom-boom-boom-boom. Here it clearly works slower. So, while it's working, I'll show you what you can do with this Code Work. So, how does Code Work look? Well, it looks like this. So, here, for example, a standard chat looks like a standard chat window. You select the folder you want to work with. I'll choose this folder now, for example. And what do I have in this folder? Now I'll show you. In this folder, I have my video, right? I just downloaded my stream from YouTube to show you. This is exactly my review on how to get the most out of ChatGPT. This is a two-hour video, and I have subtitles for it. So, I gave this Claude Code Work access to the folder where there are two things, right? This is my video file and my subtitles. The subtitles, naturally, have timing, right? And look what I'll ask it. I'll tell it that, um, so, I'll tell it this: I need you to use the skill, um, this FFmpeg - this is a special skill for video processing, and analyze the subtitles for this video and based on them, create a new video where there will only be the part where I talk about image processing in ChatGPT. By the way, if you don't know, you can process images in ChatGPT. It's like a built-in Photoshop. Make a proper introduction and conclusion so that it doesn't cut off mid-sentence. So, I'm essentially asking it to create a new, like, cut out a small video from my large video where I only talk about this topic. It will do this based on analyzing my subtitles and, um, then it will trim the original video and create a new file. Let's give it such a task, and let it think about it now. And in the meantime, let's see what we have here.

So, um, yes, it has created it. Aha. I'll just tell it that, um, I will check the video myself completely. You, um, don't do anything with browsers, don't launch them. I already have everything launched, and I will check the video myself, because I'm explaining this because as soon as Claude got access to browsers, it now likes to check everything in the browser. Well, for example, this is great, you can go and have some tea, and when you come back, it will have already checked everything, right? But just now, on the stream, it will go and check everything, and it will take longer than a person. Um. Uh-huh. Okay, let it start working now. So, we've, like, it's come up with a plan here. I didn't even really look. I roughly understand what it should do. Okay, the plan is approved. Starting implementation. I will create a task and work in order. Okay, let it work here. Let's see what's here. Ah, yes, excellent, I have analyzed the subtitles. Here I found, so, it found the place where I talk about this. So, thus, the complete fragment is these minutes, right? From this minute to this minute. The beginning of the phrase sounds like this. Well, so, the next opportunity, a good introduction. The conclusion. You can immediately go to Photoshop as a separate service. Let's cut the video. So, you see, I just told it that, um, cut it out for me. And it will do it now. This is what CWork can do, right? So, if, for example, you are completely far from code, then Code Work is such a pleasant thing where you can work with it.

So, um, it has done it. Let's see, let's see what it has created. Here a new, uh, new file has appeared. And let's see what I'm talking about here. So, the next opportunity, well, also one of those that are built-in like this, it's quite an interesting thing. I think you definitely know, right, that in >> it's quite simple >> Yes, that's right. I'm showing how ChatGPT can process video, oh, >> edit a photo. Right, right inside the chat. So, it won't be a re-generation, you'll just take its version. At the same time >> Okay, that's right. And what will I finish with? Let's see. Then you can immediately go to Photoshop as a separate, um, as a separate service. Super. That's it. So, look, you had a video, right? This two-hour video, you have subtitles for it, and you just tell it to say, look, I need to cut out this moment from it. You didn't tell it, like, from this minute to this minute or something like that. You just told it that I need where I'm talking about this, do it. And it did it perfectly the first time. Just look, perfectly. It starts perfectly, ends perfectly. And it just created this video for us. And this is what you can already do with coding tools.

Let's see what's happening here. Ah, now I need to allow it to execute commands. Let it execute commands there. So, and I see, I see many people liked it. Well, let me answer the question. So, um, so, Bogdan asks: "Can color correction also be done?" Well, color correction is a slightly more complex topic, because you need to have some tool that applies this color correction, and you need to have access to this tool, because cutting video, there are special programs that can easily cut video, and it uses them. So, um, Sergey asks: "And how did you get the subtitles?" I, well, I talk about this in my courses. There are actually many ways to do this. You can do it via API calls to services that transcribe video well. You can transcribe video locally. So, there are many options, and, well, we discuss it in detail in the courses. Vadim asks: "And voiceover, can it be done?" Are there tools? Yes, by the way, you can do voiceover without any problems. To do this, you need to connect a tool like, for example, 11 Labs. And the agent perfectly makes a request with the necessary voiceover text, with the desired voice to 11 Labs, gets the audio back, embeds it all. I've already done this, so it all works. Bogdan writes: "Give me back my money for all the software I've bought over the years." This is, by the way, a common thought among people. Many people say, you know, I think I'll start unsubscribing from various software, because, well, why? Evgeniy asks: "And where can I learn this? I'm a motion designer with zero coding knowledge. Can ChatGPT teach me?" Well, firstly, I think it can. Um, but look, so, I have courses. Just so you understand, in order to do all of this, you need to understand various basic principles. And these basic principles are precisely what we study in my two courses. Now I'll show you, where is it? Yes. So, the first course is the basics of working with, um, like, with AI coding agents. It's called "Development in Practice for Everyone," right? So, this is the basis. I just recommend you take this course, because we build websites with agents, many things, but even if you don't need to do any of this, you will understand a lot about how to work with code for, including, for such tasks. So, we won't be analyzing such specific tasks, um, but we, like, we are doing it, we are building the foundation, your understanding, the foundation of how to work with AI coding agents and with code in general, right?

Here are the reviews, please read them, just read people's reviews. These are people not connected with, well, most of them are not connected with programming. Read what they write about this course. And I have a second level of this course. It, by the way, starts very soon. And it starts on February 11th. And the first group has already finished it. And you can also read the reviews of people who completed the first group of this course, and read what they write. And if you have already completed my first course, come, because we do a lot of interesting things. We launch local models here. We use models in Hugging Face. We connect Codex as an engine for your agent. Yes, you can do that, and we do it in this course. So, in fact, well, read people's reviews – it's really such, well, the next level, but you should only go here after the first level, because otherwise you won't understand anything at all, really. Well, it's just, it's just, that's how it will be. Let's see what's happening here. Okay. So, what did he do? Let's see what he did. Where is our? Here. Look, here's what he did. Let's see what he did. So, well, let's, yes, let's not nitpick too much about the design. Look at the essence itself. He took real locations, yes, in the USA, drew lines like this, how they connect to each other, where these happened, and where, and where UFOs were seen, yes. Like this. This is just what is loading now, this will not be there. Let me do a full render. I'll just do a full render of this video now, because, uh, now it will render, because these are these kinds of loadings, it's because it hasn't rendered yet. I'll do it for you now. What is he doing there? Hmm. Let me tell him now to render this video for me now. Ah, render this video in 1920p. Let him render it now. Strange. I don't know why it gave an error, but it doesn't matter, he'll figure it out now. So, let's look for now. Then I'll just show you how this video will, well, how it will look when it's rendered, right? So, like this, so imagine, so he took real data, drew this kind of map. and all these animations. So, imagine, imagine what kind of animation you can do there. I don't know, if you will, for example, uh, I don't know, for example, there, do some kind of infographic for, I don't know, the appearance of a network of new stores, right, or something like that. But this is absolutely, absolutely amazing. Now we'll see what he's doing with the render. Ah, this will take time. Each frame is rendering the map. I'll check in a minute. He'll check now. Aha. The render is in progress. It will be rendering for about 8.5 minutes now. Well, let it, well, let it render. Let it render. We'll see. We'll probably just go on a break while it's rendering, because there are actually many more interesting topics. We'll make a whole game today too, so ours is a bit longer today than usual. Evgenia asks: "Is all this through the API, 11 Labs, for example?" Uh, yes, 11 Labs is through the API, but, in principle, there are already good, quite local models now that generate voice locally. So it's not necessarily to use through the API, but you must have a powerful computer then. Sergey asks: "And can a video clip be made from audio?" Well, if you want to, yes, but it just takes a little more time and it will be a bit difficult to do live. So, Bogdan asks: "So, can you render on Codex too? Yes, yes, you can work with Codex too. It doesn't matter, any coding agent will do." I just work with Cloud Code because I have a max plan. I know that my lessons won't run out of limits and all that. Now let him, let him render. In general, let's do this. Let's go on a ten-minute break now, because we still have a lot of interesting topics today, and we'll also make a cool game today and, uh, talk about this cool video model. So we still have a lot of interesting things to discuss, so let's calmly go on a break now, and after that, during this break, it will render for us, and we'll look at it. Here it is, you see, the render is in progress. 92 frames out of 1,500, approximately 8.5 minutes remaining. So I hope, uh, I hope it will do it. Ah, so 10 minutes, be sure to come back. Good evening everyone again. And while we were on break, the render finished. So, it rendered for about 9 minutes. 40 MB. Let's look. Here's the file. Let's open it. It's straight HD. And let's look. So now it will be without these kinds of loadings, because the loadings were happening because it wasn't rendered. Look, look how smooth everything is. Yes. Accordingly, and he took this location, then, you see, smoothly, carefully draws to the next point. Imagine that you can also do motion graphics, for example, for, I don't know, some kind of transport company. Look how cool. Hop. Yes. And I just did this. Right in front of you, it's all done right in front of you. Aber, like this. Well, I think this just opens up completely new things. Ah, and you know what's even cooler? That since it's code, you can change any part very easily. Well, for example, let's say, now ours looks like this, right? So it looks like this. But remember, I showed you this map, right? And why are they so different? Because there are different map styles. So, you know, like in Google Maps there are different, like satellite, view, this and that. It's the same here. And that's why I just ask. Okay, great, everything is great. I just want to change one thing. Make the map style in this video the same as the style in this repository of mine. And don't render, I just want to see how it will look in the Remotion project. Let him create this style for us now. Ses, I'll answer questions for now. Daniil asks: "Can you set up editing in DaVinci by voice?" Well, in general, I think it's no problem at all. You can, for example, well, you can simply, I use the Whisper application for dictation, yes, you can also dictate to your agent. As far as I know, DaVinci has an API, so it can edit via API later. You can do everything by voice like that. So, let's see what he changed. Here you go, look. Now our video. The only thing is, it will load a bit slower, because, again, after rendering, everything will be fine, but before rendering, it will load a bit less. Look, we just changed the style of this map, right? So, if you, for example, want it like this, well, you see, the loading is longer like this, but in reality, when you render it, it will all be just as smooth. So. So you can, for example, take and change anything, right? There, if you, uh, for example, I don't like that this line is green, right, a green line is drawn, and I want, for example, a red one to be drawn. And make it so that this line, which is drawn from point to point, is red. He went into the code, changed it, you see, line color, line color. Let's see. Let's update. Ah, well, these will probably be green, but the line now will probably be red. Let's see. There, now a red line is being drawn. And you can change anything like this. So you have full control over it, and you see these changes instantly, right? And then, when, well, you want to render, it, well, it will take, yes, some time. But everything else is fine. Ah, Bogdan asks: "Is there still the possibility of manual editing, or if you've started editing with text, then you can continue with text?" Well, look, it's all code. If you want to edit the code manually, you can. I understand that you probably mean that you want some kind of interface, right, so that you can click somewhere, do something. This, I think, well, this will appear very soon, I don't know. Or maybe Hicks has already done it? I'm not sure, I don't really deal with mon design. Maybe Hicks has already done it, that it exists. Everything, you can edit both through the interface and through communication. So, this is such a thing. Well, we'll move on, because we have no less interesting things ahead. Now I'll stop all this for now so it doesn't load my computer. Let's close all this. What else do we want to talk about today? So, let's talk about the new Codex. So, Codex has released. Codex is a coding agent from ChatGPT. And if you have a ChatGPT+ subscription, it comes with your GPT+ subscription. It's a cool coding agent with a really super good model. And recently they released their separate application. I'll say right away that, unfortunately, this application is only for Mac. Later it will be for Windows, only for Mac and only for Macs, i.e., on M-processors. It looks like this. So it's a separate application where you can work with code, and it's designed so that you practically don't touch the code at all. So you can't even change the code manually here. You can only, uh, look at it, but you can't change it manually. And it's interesting that you can set different levels of access here. So, I currently have full access. This means the agent can do almost anything on my computer. Well, with some exceptions, but in general, it has complete freedom. And let's check how it can use this freedom. What will I ask it to do? I promised you that we would make a game, and we will. And we'll make the game this way, right? So, what do we need for a game, right? It's the game process itself, right? So, there, well, so that, I don't know, some kind of character walks. But at the same time, there needs to be some beautiful, right, graphics? So we need to have, well, some kind of beautiful character, some enemies. How to do this? Well, I'll show you how to do it now. Uh, I'll give it a very simple prompt. Yes, like this. Use Game Creation Skill. This is a special skill that I created for this task. And make a 2D platformer game. Well, this is, you know, like Mario. Use the Nan Banana Pro skill. This is a skill that allows, teaches agents to use Nan Banana. And I've already, well, there, in a special file, I gave it an API key so that it can use Nan Banana, and generate images in it. For creating 2D sprites, keep in mind that it doesn't generate a transparent background and that you need to formulate prompts precisely for generating images. Characters should have movement animations. Well, we'll give it a very, very minimal prompt for now. Let's send it. And it will think about all this now. I'm not sure if it will all be good right away. Immediately, if it's really bad, I'll show you how I did it in the test, because I got a cool game in the test. This happens sometimes, because it's evening, there's such a thing as if you're on my courses, you'll learn about the curse of the evening, because the USA has woken up now, when the USA wakes up, everything starts working worse for everyone. But let's see, maybe it will do it. In any case, I already practiced with this this morning, and it, well, it did everything, but let's see. It's checking everything now. You see, it went, it read. It seems like it's checking everything now. If it's not enough for it, it will probably ask me something. Let's see now. Okay. Now I'll switch. Yes, this is the view, right? So, it seems like it found everything, everything it needs. So, now it will go to generate, собственно, sprites, right? So, well, how will it do this? It will generate images in Nan Banana itself, right? So, I'm not telling it anything at all. It will do everything itself now. It will generate images in Nan Banana itself. Uh, also, by the way, I hope it will work, because lately Nan Banana has been working just terribly. It will generate images in Nan Banana, cut out these images, and create a game based on these images. The difficulty is that we are doing animation. So, without animation, there would be no problems at all. But with animation, there could be problems. But if anything, I'll show you how I've already done it. I just want to show that it's possible to do it this way. Ah, and while it's doing it, let's, you know, discuss a very hyped topic. This is, ah, ah, yes, I'll just say a couple of words about Codex. What, so, first of all, yes, this is an application that is designed specifically for managing agents, right? Uh, as I said, there is a full autonomy mode, and it's also interesting that you can set up automations, right, but this is still a bit of a new topic. I think we might return to it next month. And by the way, until February 10th, uh, even free ChatGPT users can download it and try it completely for free. So, until February 10th, you can do exactly what I'm doing. And while it's working, let's talk about such a thing as the madness around Open Claw, as it's called now. This is, um, such an agent, which was originally called Claude Bot. Then people from Anthropic came and said that no, sorry, this is our trademark, change it. For a few days, it was changed to mtbot or MT, and then it finally got its new name, it's called Open Claw. This is a very interesting thing. And what's most interesting is that this is just the hype now, but if you look at my last month's stream, I showed it to you a month ago, and I told you what an amazing thing it is, because it's truly amazing. So, it's not like, you know, we're riding on the wave of hype here. No, no, I told you about it long before the hype. What is it, right? If you're not aware, if you're not aware, it's better to watch my last stream, because I show there what it can do. But to simplify greatly, you install an agent on your computer that essentially has access to your entire computer. Well, if you give it that access, and you can communicate with it as you would actually communicate with your computer. So, not like you communicate with ChatGPT, but you can actually, for example, just ask it: "Listen, my computer is overheating, what's going on?" It says: "Yes, I'll go and check." Oh, listen, you have some processes here. Shall I stop them for you, and your computer won't overheat? You say: "Yes, go ahead, I don't need them." And the level of communication with such an agent is completely different. So, for me, this is the most interesting moment in AI since GPT 3.5. So, when ChatGPT first came out, I remember that moment when, wow, when this first real ChatGPT experience, it was just incredible, absolutely incredible. And here, it's a very similar feeling, because you're communicating with the entire computer, and it can do anything. So, for example, you know, uh, for example, how I use it? I use it for many purposes. And I use it, for example, in this way. Just so you understand how I work. I work with Opus, so I have Opus 45. And for example, this is how I run my channel. This is what I sent him messages for my channel today. I call it Cloud Assistant. I put Dr. Zoidberg from Futurama on it. So that you can just see how I actually run my channel. I just send him links to articles that, well, I like, and tell him to process them. That's it. He does everything else, he translates them correctly with my skill. He sends them, puts them in the queue for publication, does everything, downloads pictures if needed, so he does everything, everything, everything further, he takes care of everything. This could have been done without him, right? So, initially, it worked without him, when I was just starting to create the channel. But with him, it's so cool. So, you know, just so you understand, for example, once I accidentally sent him a link not to an article, right, that needed to be formatted like this, but I sent him a link to a tweet that contained an article. And any automation, at that moment, would say, "You sent me the wrong thing, I can't process it, it's not an article, it's a tweet, right?" So, that's what it is. He just says: "Oh, listen, you accidentally sent me a link to a tweet, not an article?" Well, I went, read that tweet, found a link to the article there, and opened it and formatted it. Everything is fine. And at that moment, you understand that this is not automation, right? This is not automation, this is a real agent that sits with you. And you can, well, there have been so many cool stories with him, when, for example, once, uh, in the morning, a guy just wakes up, and sees missed calls on his phone. He calls back, like, "Hello." And his agent, this Claude, answers him: "Hey, it's me." He speaks to him with his voice. The person is, well, just has a panic attack because he doesn't understand what's happening. And this bot tells him that I was, you see, I was solving a problem at night and trying to contact you, but you weren't responding. And so I decided to find a way to contact you. And I went and looked at the access you gave me. And you gave me access to telephony, among other things, well, it was just somewhere among the API keys and to 11 Labs. And so I'm calling you. I found your phone number in your system's contact list, I'm calling you. These are real stories. So, it's a very interesting experience. But it's very important to say that if you don't understand anything about agent security rules and all that, please don't install it. You must understand what you are doing. Because if you don't understand what you're doing, you can put yourself in serious danger. Well, you can really lose money. You can be hacked. So, please, if you want to install it, first of all, thoroughly understand the security rules, how to do it correctly, and always use only top models and don't give full access anyway. So, for example, my Claude bot is very limited. It, well, it can do very little. And most importantly, it's not connected to anything external. It doesn't read my email, it's not in any chats, no one else can tell it anything. This is very important. So, it's just, I'll just say right away that it's an amazing thing, an amazing experience. And if you understand AI, you should try it. But you must understand what you are doing. So. Now I'll look at what's happening here. It's almost finished. Let me just say a couple more words about the Claude bot. So, first of all, everything is getting even more interesting, because it all started with one person creating Moldbook. This was exactly the moment when the name was Mol, for those few days. And Moldbook was created then. It's like, you know, like a social network for AI agents. And here, like, agents communicate. Actually, no. Actually, there are many people here too. They just enter here under the guise of agents. But in general, as a concept, it's an amazing concept. There are indeed millions, probably millions of agents now, and it's true. And they do communicate. Not all of them are real. Some, well, in general, there are many variations where it's not true. Partially it's just hype, but the concept is super interesting. And after Moldbook took off, everything started to roll. Sk started. This is like Slack, but for agents, where people communicate in Slack, and here agents communicate. Websites like rant a human appeared. So, like, a robot needs your body for AI can't touch grass. You can get money for tasks that an agent can perform in the real world. And here people sell their services for $150 an hour, some for $50 an hour. All this is done with cryptocurrencies, payment, and your agent can hire a real person to do something in the real world. And this is already reality. This is not, you know, this is still more of a concept, right? So, obviously, it's partially hype, partially, well, it's just fun for everyone to talk about, but as a concept, it's already working, right, and I think by the end of the year, all of this will become an absolute reality. We will live in a world where an agent can actually order a service from you. It will be very interesting. And now let's see what he did with the game here or not. Now we'll tell him. Ah, yes, launch the game. Let's see. If it doesn't work, I'll just show you what I did before, just to, well, not waste too much time on it, because I think there will be some problems with animation now. Oh, there are some problems. The evening is quite, well, the animation is quite beautiful. It's just that it turned out a bit strange for him. And I'll show you how it looks from his side, right? So, this is what he did. He prepared, he generated these pictures. Here, you see, he generated different pictures. And then he cut them out incorrectly. You see, he cut them out incorrectly. Now we'll tell him. He, you see, generated cool pictures. You see, different animation stages and various enemies. But he cut them out incorrectly. Let's tell him that you cut out the pictures completely incorrectly. You generated cool original pictures, but you cut them out completely incorrectly. Look how incorrectly you cut them out, and redo it. You see? So, he just cut them out incorrectly. Now, yes, yes, let's give him a chance. If it doesn't work. Okay, while, while let's discuss what we have, what's next. Ah, so new Codex. So, CWork, we talked about it. New Codex. Ah, it's just important to say what's happening with subscriptions now. So, on the one hand, Claude stated that the Claude subscription cannot be used in third-party programs. So, you can, for example, use it in Cloud Code for VS Code, Cloud Code for any other thing, because it will be used within Cloud Code. But specifically in third-party programs, for example, Open Code, yes, there is such a coding agent, it's not allowed. They, like, forbade it. And there was a lot of outrage, uh, from, uh, these, well, from all users. And they, like, took a step back a little, but in general, in general, they, like, forbid it. So, essentially, if you want to use Cloud Code in other programs, not related to Anthropic in any way, then you can't do that anymore. And there's actually a reason why they're doing this. The reason is that when it's used in their controlled programs, for example, Cloud Agents SDK or CLC, yes, so, programs that they control, they can distribute the cache very correctly. And cache is what makes calls much cheaper. Well, essentially, they will cost almost nothing with proper use. Well, they will cost a little bit.

And therefore, uh, for example, when you pay, like, I, for instance, pay a subscription, 100 dollars I pay for Cloud Max. Uh, and the thing is that for these 100 dollars, I actually get, if you just count the tokens, recently one person, he conducted a study, if you just count by tokens, then actually, if you use it to the fullest, you can use up to 3,500 dollars, even up to 3,700, 37 times more. And why is this profitable for them? Because they use this caching correctly, and other programs don't do that. For them, it's just ruinous. But at the same time, ChatGPT, on the contrary, said that their ChatGPT Plus subscription and, accordingly, their Codex agent can now be used in many places: in Open Code, in Kill Code, in Keyne, meaning it can be used everywhere. So, you know, it's like two different paths have emerged. ChatGPT said: "Please, go ahead." And Claude, on the contrary, forbade it. Well, let's see whose path will be better. Let's see what's up with Codex. Ongu, yes, confirmed visually. The problem is not in generation, but in processing the background before cutting. Now I will rewrite Chmaekey so that it correctly cleans the green gradient background. So, you see, it turns out that he, you see, he generated these kinds of pictures, for example, like this. Here he generated this like this. By the way, it turned out to be a beautiful, beautiful little monster. And he generated it so that it's easy to cut out using a chroma key, as it were. And he will do it easily. But, uh, let's see. But, you see, for now, he's not succeeding. Let's see what he does. Uh, I confirm. A harsh cutting error is indeed visible. I'm fixing it point by point now. Well done, good job fixing it. We'll give you a chance. But if anything, I'll show you later how I managed to do it today. So, regarding clD-code, there are also a few important things to mention about what's new here. In clD-code, three interesting things have appeared. Firstly, Agent Team has appeared. This is like a team of agents, uh, previously this was called the SWM mode, Swarm mode. But you see, they took the right marketing path. And they didn't, you know, create this uncanny valley, right? So, what does this give, right, this agent team, it's a real team of agents. So, before clD-code, well, and generally for many, for Codex, for example, there are sub-agents. So, sub-agents are when your main agent creates several such sub-agents, gives them tasks, they complete them and return to him. But in clD-code, it's not like that anymore. In clD-code, your main agent is like a typad, but everything that happens is a real team. They communicate with each other, each has its own context, they are completely independent. Each works in their own workt. We discuss this on the second level of our development course. And they truly collaborate like a team of people. For example, one of them can talk to another during task execution and figure out how they can better complete a task together. It's a real development team, but because of this, they consume more tokens. The second important change is that now, instead of just a standard to-do list, tasks are, well, it's hard to translate into Russian, it's also tasks, but a to-do list is a list of things to do, like, and tasks are global tasks. And what's very interesting is that these are different concepts, because, for example, if it's a to-do list, the agent just goes through this list of things. But tasks are like a global task. I need to do this, this, and this. And, for example, it's possible that the second part of the task is not activated until the first part is definitely completed. So there are many interesting settings. Several agents can go through the same task list, right, and distribute among themselves what they will do. So, in general, it's a more complex concept and very interesting too. MCP Tool Search has appeared. So, if you, well, if you don't know what MCP is, it's a tool for access, so that your AI gets access to third-party services. And the problem is that it consumes a lot of tokens. For example, here in the screenshot, 33,000 tokens, just the presence of certain MCP tools gives. For example, unfortunately, it consumes a lot of tokens. And therefore, in clD-code, if you have a lot of MCP servers, it doesn't load them all at once, but starts loading them as needed. This is a very, very smart thing. And we will talk about all of this and much, much more in the course on deep immersion into clD-code. And if you really want to know everything about clD-code, then this is for you. So, let's see what's happening here. Uh, it seems like it should be better now. Let's see. Mmm. Start the game. Let's see how it goes. Look, it's already better. Look, here, you see, look, just think about it, right, well, there's still work to be done, make the grass prettier, all that, right, but look, he actually made an animation. Look, I'm walking, you see, the character is walking, and he generated a picture, hey, why did you close it? He generated a picture and, uh, and, uh, accordingly, animated it like this. Wait, wait. What do I have? Why did he close it? Mmm, start it again without a timeout. I'll close it myself. He just set it to close after 20 seconds. So, this is, like, a normal thing. Just, uh, not very good for a stream. Right. And look, so you see, the meaning, look, there's jump animation. Look, so imagine that all of this, look how the coin spins, right? Well, it's amazing, isn't it? So, so it's not just, right, some pixels, you know, like it used to be, that they could make games, right, where all sorts of pixels, all that stuff, look at the quality of the animation. He prompted it himself for a nanobanana to generate it, and he did it himself. Well, that's really super, right? You see, I, I just died. Uh, let me just show you what my first one turned out like. Uh, now this one, which I was testing. Just so you see that it can do it differently each time. Wait. Start this game. Evgenia asks: "Are you controlling the little man?" Yes, this is a full game. It's not a video, it's a game. Look what he just did. Just so you understand, he first generated, through nanobanana, you see, like, on a green background, these kinds of pictures. For example, the entire animation of the little man is broken down through these kinds of pictures. Then, using code, he cut it into sprites and removed the green background. And as a result, he got these final ones. So, here, for example, are the sprites of the little man's movement. And this is how the little man stands, this is how the little man jumps. You see, he didn't clean the green background very well. This is how the monster jumps, right? This is how the coin spins, and that's it. And then he created all of this into a game. And here's how he did mine before. Now I'll show you. So, it can be different each time. This is what I did today. Look, here's another little man. Look, he's jumping. Look, he's collecting coins. The monster also jumps coolly, but, you see, it's generative, right, all of this. So, he generated one game for me now, and when I was doing it, it was a different game. Well, each has its pros. I liked the very detailed graphics. Here, the overall style, I would say, is nicer. So, just imagine, and again, this is a game that he generated the graphics for himself. You can use Codex and, well, or clD-code too, it doesn't matter, and, for example, nanobanana, you can generate entire games, real games with real art, really well done. I generated this in front of you. Imagine what you could do if you spent a month on this, a month of this kind of work. You can do anything, really. Yes, well, that's it. That's it. Well, and on the topic of all this. I really like the prediction, uh, of the head of Anthropic, Anthropic, who predicts that in 12 months, there might be full automation of development. So he writes, well, so he said that, uh, that already now, if we talk about models that write code, we have engineers at Anthropic who no longer write code at all. And he says that I think, although I don't know for sure, that we are 6 to 12 months away from the moment when a model will perform most, and perhaps absolutely all, of the work of software engineers by the end of the year. And then the question arises, how quickly will this cycle close? So, what is he talking about? That, well, in about a year, well, according to his predictions, at least, such a possibility will most likely exist for someone, maybe Anthropic will have it, you will simply be able to approach the model and say: "Do this for me." And it will do everything, test everything. Not like now, right, we had a bug problem, it will test everything, look at everything, so that everything works, and then it will give the finished result. Well, that is, about a year until then. And, finally, our last two topics, because usually our streams are an hour long, but there have been so many news. Well, I hope you don't regret spending your time. So, the latest news is that, firstly, Clean 3 has been released. This is a new, quite, well, breakthrough video generator. It has several interesting features, right, that, firstly, it uses a multimodal architecture, right, meaning it combines everything: text to video, image to video, video to video, so, everything is combined there. Plus native audio generation, plus multi-frame generation. So, for example, a video up to 15 seconds can contain several separate editing cuts, and plus native 4K. So, accordingly, it all looks like this. So, well, this, for example, I'll just show you the official one, uh, and then I'll show you what I liked from what people generated, right? So, well, officially they do it like this. Well, everything looks cool, but, honestly, Veo also does some similar things. Now there's just some kind of box here. Right. Let's look here. So, better camera control. Well, these cool, you see, transitions of these cuts. Much better consistency overall. So, in general, well, like, everything, everything is great, right? So, well, but these are official videos, they are always very cool for everyone. Voiceover. Everything there, all of that, all of that is good. What impressed me? Two videos impressed me. So, the first one that impressed me was the video about emotions. So, look at this, these are AI-generated emotions. I think we are close to real movies. >> [laughter] >> So, these are really, well, emotions of some incredible level. And the second thing is the processing, well, of different people in one frame. So, here, look, here in one frame there are eight people, and each of them is doing different things. Previously, such things, as far as I know, video generation couldn't do, but again, I'm not a video specialist, I'll say right away, that, like, I don't think it could do this. Google Veo also released an update, like, an update for Google Veo 3.1. Native 4K, vertical video has appeared. A very interesting thing came out from CREA. They made real-time video generation. So, you take it and see what's happening in real-time, right? So, right in live mode, you can, well, it's not even video, it's picture generation, it would be more accurate to say. So, this is also a very interesting thing. I believe that everything will be generated. Everything will be just, absolutely everything. Websites, everything will be generated on the fly. Right. And the last topic of our today's stream is the future of AI. And, so, what are we looking at in terms of the future? Firstly, GN3 has entered initial public access. So, if you don't know what this is, it's something that will significantly change everything, especially the gaming industry, because it's a world generation tool, infinite worlds, real worlds. Well, in the sense of gaming. So, you can generate a world that you can walk around in. So, it looks like this too. You can also look, there are many videos about this on the internet. So, you describe with words what kind of world you want to generate, or with a picture, and it generates such a world for you. And you walk around in it, and play. Now it generates everything. And you control it like in a game. Everything, it's a game. So, like you walk around in a game, only this is all generated in real-time. Every action you take is interactive. And you can generate anything you can think of, right? Want to crawl as an ant, no problem. Want to fly in space, no problem. People are already generating anything. To get access to this, you need to be an ultra subscriber for $250 and be in America. So, I don't have access to this, and therefore I can't show it to you live. But this is the future. So, when you can just generate a world and walk around in it immediately, well, I think it will change, I don't know, almost all existing industries. But I just want to wait until it becomes more publicly accessible, so that I can try it out with you live on stream. For now, it's, well, it's public, but only for people from the US and with the most expensive subscriptions. The second interesting thing is, I think it's a very important concept that, while it seems very funny, this concept, I think, is what will become a big topic soon, well, in the near future. It's controlling AI agents through game interfaces. Well, not like, right, I'm sitting here now, like I did in class, I was sitting in the terminal, typing something here. It looks terrible, honestly. Well, really. Well, don't think that I think this is beautiful. Don't think so, please. I don't think it's beautiful. I just, well, I'm forced to work in it, but I don't think it's beautiful. And I think that beautiful is like Warcraft, that's beautiful. Or Starcraft. And there are already the first experiments where you can control your agents like this through a game. So, so you literally take it and say: "Okay, you agent, you do that." And you go there. And you communicate with it like this. So, I think this is something people will really want to do, because it's so cool. Well, so you're essentially playing a game with your real agents. There are already working concepts, and there are many of them, and I think there will be more of them, because it's a very interesting thing. And, finally, to end our stream today, I want to use a quote from Dan Shipper. Dan Shipper is, in my opinion, one of the most interesting visionaries in the field of AI and, well, a person whom I personally follow very closely. He gave a very cool and very practical definition of AI. AI is artificial general intelligence. So, it's, well, or artificial general-purpose intelligence. So, it's, roughly speaking, well, like human-level intelligence, right, in a broad sense. And he says that, so, this AGI will be achieved when it becomes economically viable to keep your agent working constantly. In other words, we will have AI when there are constantly working agents that continue to think, learn, and act autonomously between your interactions with them, as a human does. And I think this is a very cool definition, and I think we will reach it soon. So, for example, imagine that, imagine that, for example, uh, well, we haven't finished, right, with this, for example, agent, but we just, like, I finished the conversation with him, but he will actually be with me constantly, he will be constantly on, he will be constantly working, he will be constantly improving himself, constantly looking for something, and it will cost money. But this money will pay off. So, if you want to really dive into all of this, I have four courses. So, one course is C development in practice for everyone. This is if you want to, well, really dive into this topic of development, specifically in terms of truly understanding how to create websites, how to create agents, how to do all of this, how it all works. Well, so you will understand many important concepts, and after that, you will be able to apply them as you need. And if you have already completed the development course, then come to the second course, because in the second course we cover very interesting things. Just, well, read what, uh, people are writing, right, what things we cover in these courses, right, and you will notice a common pattern that many write that these two courses work very well together. And if you are more interested, let's say, in getting the maximum AI capabilities without diving into super programming, then this is the course on how to get the maximum GPT. We cover cool things here, dive into it very deeply, and also start using coding agents, but more for simpler tasks and more related not to programming as such, but more to working with files, creating a simple Telegram bot, creating a simple website. And this is, well, actually, for many, this will be a very good entry point. And if you have already completed my first or second course, or want to understand clD-code, then you will like this topic, because it opens up huge, actually, prospects, especially this Cloud Cloud Agents SDK. This is a very, very powerful thing. Well, at least come to the overview sessions just to see it. And here's a question. I see that, Oleg, your course ends on May 12th. Don't you feel that by this date the knowledge gained in the course will become obsolete? Well, you can ask, those who have taken my course. You know that I have two, well, like, two fundamental positions. First, I don't record any video lessons on AI, specifically pre-recorded ones, because it's pointless. They become obsolete instantly, only live sessions, only streams. And second, you can also ask in the chat, they will confirm: "In all my courses, we use the most advanced technologies available. That is, if something new comes out, I will simply change the course. I will just say that, okay, we are no longer using this, we are now using this technology." For example, on this course, development with AI automation, productivity, the last lesson was supposed to be different, but I say: "No, guys, a more important topic now. We will cover this more important topic, which has become relevant, because it's more important for you to know." Therefore, these courses are always conducted in a live format, because they are restructured on the go, literally on the go. That's it. If there are no more questions, then let's probably finish for today. Uh, sorry, I see Olga's question. When you solve various tasks like this, do you usually work with two monitors or more? How is your workspace set up for comfort and efficiency? I'm currently working a lot, well, let's say, while traveling. And I often work on a laptop, so I create different workspaces for myself. So, on Mac, you can switch between workspaces. And often I create it so that, for example, I have one workspace with only terminal windows, and another workspace, accordingly, for, well, regular work tasks, right, so I distribute it like that. But in general, if I were to equip a stationary place a little more, then, yes, I would probably want a couple of monitors, but unfortunately, I'm just living a little differently for now. Right. And Maria writes: "We confirm." Yes, well, because it's true, because I try very hard to ensure that you get the most up-to-date knowledge possible. Alexander asks: "What are the computer requirements for your courses?" Well, look, for the first course, right, there are no special requirements, because, well, you just need, well, some basic things to work there. Well, so, it shouldn't be super weak, right, there, well, just some ordinary normal average computer. And for the second course, in principle, also, but we will have one lesson where we launch local LLMs. You don't have to launch these local ones, I offer other options if your computer can't handle it. But if it can handle it, and we use, you know, LLMs that even run on a phone, then that would be great. But in general, a super powerful computer is not needed. It just needs to be normal. So, just an ordinary normal modern computer. It doesn't need to be super powerful. So, we don't use local models that require a powerful computer. Right. So, I hope you enjoyed our stream today. I think I showed you some interesting new things. I will be waiting for you at my future streams, because we do this about once a month. And soon, by the way, I will also have a free stream on how to use Skills. You noticed, I used Skills today. And on February 16th, I will have a separate detailed stream specifically about Skills, because it's a very important format. And there will also be an overview session, an immersion into clD-code. That's it, I won't keep you any longer. So, we usually do an hour, but there was so much interesting stuff today, so I hope you enjoyed it. Thank you very much for being here for 2 hours, and I will be waiting for you at new streams. Goodbye.