Обзор курса «Разработка с AI: автоматизация и продуктивность». Александр Свет

Overview of the updated AI development course focusing on automation, loop engineering, and productivity improvements in 2026.

Key Takeaways

AI development is now largely agent-driven, reducing the need for traditional coding.
Loop engineering is a critical new skill for AI developers in 2026.
Fine-tuning LLMs is becoming an essential and accessible skill for all developers.
The course structure has shifted to more frequent, focused projects to enhance practical skills.
The curriculum reflects the latest AI tools and methodologies, ensuring relevance in a fast-changing field.

What the video covers

The course is an updated version of a year-old AI development curriculum at Profile school.
It focuses on AI development, automation, and productivity with a 100% AI-driven approach.
Significant curriculum changes reflect rapid advancements in AI, including agent-based development.
New major topics include loop engineering and fine-tuning large language models (LLMs).
The course offers more projects with a focus on practical, hands-on learning.
Integration of AI agents into products and advanced design generation tools are covered.
The course is suitable for those who completed the first level and for returning students with minimal repetition.
Discounts are currently available on this and other AI courses at Profile school.
The course emphasizes modern AI skills such as embedding AI agents and creating development loops.
Security topics like containerization and microVMs are also included.

Chapters

Who is this updated AI development course intended for?

The course is designed for those who have completed the first level of AI development at Profile school and for returning students seeking deeper knowledge with minimal repetition.

What are the major new topics introduced in this course update?

Major new topics include loop engineering, fine-tuning large language models, embedding AI agents into products, and advanced automation techniques.

How has the course structure changed compared to the previous version?

The course now features more frequent projects, typically one project per session, with a focus on practical applications and specific AI development skills.

Full Transcript — Download SRT & Markdown

Speaker A

Good evening, everyone, and welcome to the overview of the updated course on AI development, automation, and productivity. As you know, this course has been around at the Profile school for quite some time. I started conducting it almost a year ago. And

Speaker A

recently, well, a couple of months ago, I decided to take a fresh look at its curriculum. Maybe tweak a few things here and there, or swap out some topics . But when I started looking seriously and deeply into what needed to be

Speaker A

recently, well, a couple of months ago, I decided to take a fresh look at its curriculum. Maybe tweak a few things here and there, or swap out some topics. But when I started looking seriously and deeply into what needed to be

Speaker A

ago, for those who took this course, you remember that the concept of the second course was largely about diving deeper into more serious topics related to development—like talking more about databases, and things like that.

Speaker A

changed, I realized that everything had to be changed, because over this year, somehow imperceptibly, everything has changed completely. I mean, how did this course originally work? Right, it is the second level of the development with AI course. And originally a year

Speaker A

In principle, we have practically reached the point where the question of development is fully resolved. I mean that it is now almost entirely done by agents. For instance, the creator of Claude Code recently wrote that he essentially doesn't read code anymore.

Speaker A

ago, for those who took this course, you remember that the concept of the second course was largely about diving deeper into more serious topics related to development—like talking more about databases, and things like that.

Speaker A

everything. So. And with that in mind, I decided to completely rethink this course. And before we move on to the detailed breakdown, I just remembered that I forgot to say a very important thing that I am obligated to say. And

Speaker A

And here I am looking at all of this. A year has passed, it's October 2026 now.

Speaker A

automation, and productivity. Plus, you can get an additional bonus via email. Well, not just can, you will definitely get it, yes. So, definitely come join us. And that's it. I have said what I needed to say. Now, let's talk about

Speaker A

In principle, we have practically reached the point where the question of development is fully resolved. I mean that it is now almost entirely done by agents. For instance, the creator of Claude Code recently wrote that he essentially doesn't read code anymore.

Speaker A

of the course. That is, if you have recently finished the first level, then this course is definitely for you because we will be exploring all topics related to AI development in much greater depth here. And we will talk

Speaker A

And I looked at all of this and realized that, yes, some things we had in this course were certainly very reasonable. And, well, for their time, they were absolutely correct. It's just that everything has changed now and it's time, it's time to rethink

Speaker A

But at the same time, I believe that this updated version of the course is also great for those who have already taken the first version, because there will actually be very little repetition . I mean, in some ways we might overlap

Speaker A

everything. So. And with that in mind, I decided to completely rethink this course. And before we move on to the detailed breakdown, I just remembered that I forgot to say a very important thing that I am obligated to say. And

Speaker A

more human-led development, while here we are putting a 100%focus specifically on AI development. And so, if we answer the question of what has changed in the program, it's practically everything, and there are more projects. That is, if previously we had a system where we

Speaker A

that is, essentially, that there are currently discounts on all courses, including my own AI courses as well. So, be sure to sign up; for example, there is a good discount right now on this very course on development,

Speaker A

So, in general, what new global topics have appeared? Well, first of all, a major topic is loop engineering, that is, engineering cycles and creating development loops. I think this is the most important thing that exists right now, isn't it? You need to be able to

Speaker A

automation, and productivity. Plus, you can get an additional bonus via email. Well, not just can, you will definitely get it, yes. So, definitely come join us. And that's it. I have said what I needed to say. Now, let's talk about

Speaker A

And this is also, I think, a very big topic that has changed fundamentally since the last course. In the past, AI couldn't really create truly cool designs. Now, in the autumn of 2026, I think Opus 5.5, Fable 5.1, and Astra

Speaker A

who this course is actually for, what has really changed in it, how it changed, and in what direction. But let's start with the fact that this course is, on one hand, for those who have already completed the first level

Speaker A

become such a standard in working with AI, and creating your own skills, effective skills, has become a very important component. But at the same time, I decided to go further. And it seems to me, if the first level of the

Speaker A

of the course. That is, if you have recently finished the first level, then this course is definitely for you because we will be exploring all topics related to AI development in much greater depth here. And we will talk

Speaker A

create an AI, anything at all. That freedom, yes, of creation. then one of my goals for the second level of the course is to show that, in principle, the same thing has happened with fine-tuning LLMs for your tasks. You

Speaker A

about loop engineering, about skills, and about everything that is truly important right now in October 2026.

Speaker A

either a programmer or you weren't a programmer, and that's how it feels now , as if you either know how to fine-tune or develop LLMs, or you don't . And it's like a really, really complex skill. And my task, including,

Speaker A

But at the same time, I believe that this updated version of the course is also great for those who have already taken the first version, because there will actually be very little repetition. I mean, in some ways we might overlap

Speaker A

principle, even for free on, well, like , hobby scales. It's fast and it's very simple, right? That is, you just tell the codebase that I want to fine-tune the LLM like this, and it will fine-tune everything for you. You just

Speaker A

a little, maybe 10-20 percent, but for the most part, you'll see the course curriculum yourself and realize it has changed quite significantly. And most importantly, the topics have changed; because in the previous course, everything was somehow centered around

Speaker A

very simple. It's not perfect, of course, there are still some complexities, but any person, absolutely any person, can now train their own LLM. I also became interested in embedding an AI agent directly into a product, because if you remember, in

Speaker A

more human-led development, while here we are putting a 100% focus specifically on AI development. And so, if we answer the question of what has changed in the program, it's practically everything, and there are more projects. That is, if previously we had a system where we

Speaker A

right in the first course. Well, it has basically become such an elementary thing. But Pie is interesting. Pie is interesting because it offers more engineering solutions, gives you more to think about, and it's interesting to work with. As you remember, in the

Speaker A

had one, one project per two sessions, here we have almost always—with the exception of a few sessions—one session per one project, because the projects will also be slightly different, with a more specific focus.

Speaker A

thing, that even at the first level of the course, we can integrate Codex into a browser extension without any issues.

Speaker A

So, in general, what new global topics have appeared? Well, first of all, a major topic is loop engineering, that is, engineering cycles and creating development loops. I think this is the most important thing that exists right now, isn't it? You need to be able to

Speaker A

interesting in that it's like a construction set, like LEGO, and it's fun to work with. I am currently integrating it into many of my projects , and it's interesting because it's so simple, so lightweight, and you can assemble it to fit your specific task,

Speaker A

create cycles and run an agent so that it does everything itself, develops everything itself, and so on. The second point is design with AI agents.

Speaker A

; you can build very interesting things with Pie. And, again, the API subscription works with it too, so there are no problems. And finally, I would also like to talk separately about the security of AI projects. This is something that we will, in fact,

Speaker A

And this is also, I think, a very big topic that has changed fundamentally since the last course. In the past, AI couldn't really create truly cool designs. Now, in the autumn of 2026, I think Opus 5.5, Fable 5.1, and Astra

Speaker A

happening. But let's see what has changed. Well, in fact, firstly, almost all the topics have changed. And we start right with the first lesson— that is loop engineering, or loop engineering. And what will we study here? Well, firstly, we will look at

Speaker A

GPT6 Astra. All of these models can actually create very, very cool designs that don't look like AI designs. Also, of course, another very big topic is creating skills, because, well, you definitely don't need to explain to anyone what skills are anymore. It has

Speaker A

about that, too, what that's going to be like. And also, how to budget everything correctly so it doesn't consume a gigantic number of tokens.

Speaker A

become such a standard in working with AI, and creating your own skills, effective skills, has become a very important component. But at the same time, I decided to go further. And it seems to me, if the first level of the

Speaker A

Codeium: "Make me a game"—that's not a problem at all. The problem is making it play great, right? So that it is, for example, moderately difficult. Uh ...And so that it is, for example, interestingly balanced, yes. So how do

Speaker A

AI development course, well, for us, for all of us, including myself at one time, it kind of opened up such a opened up a new world of development, right, the fact that you can just take and create something like that, just

Speaker A

not stop until it achieves the desired results. And I think it's interesting to look at it through the example of a game, because a game is something that is, on one hand, very visual, and where we can immediately feel whether it's

Speaker A

create an AI, anything at all. That freedom, yes, of creation. then one of my goals for the second level of the course is to show that, in principle, the same thing has happened with fine-tuning LLMs for your tasks. You

Speaker A

design with AI agents. And here I see a different goal for myself. I want you to learn how to create websites that don't look like sites created with AI, yes—so they aren't, uh, these cookie-cutter, AI-made sites, but rather we will try to create truly

Speaker A

know, this is a topic that is such that it always, even now, when we have AI, it seems very complex, some kind of engineering, very technical, right, just like programming used to feel, actually, yes. meaning if you were

Speaker A

and how to use them all so that the output is something that is truly non-templated AI design. And, essentially, the second project you will make is a site that doesn't look like it was created by AI. So that's

Speaker A

either a programmer or you weren't a programmer, and that's how it feels now, as if you either know how to fine-tune or develop LLMs, or you don't. And it's like a really, really complex skill. And my task, including,

Speaker A

the entire class, all two hours, we will study how skills are structured from the inside, and generally, when you should choose skills versus something else. And, for example, how to do skill evaluations correctly, and how to build a self-improvement cycle,

Speaker A

by the way, this very session, because we will be doing this today as well, is to show you that, in fact, fine-tuning an LLM for your task is not difficult, really. I mean, right now in 2026, it can be done almost for free. Well, in

Speaker A

implement a local model, yes. So, this will be a similar story to what we had in the, uh, first version of this course, right? We—well, because I believe it’s very important to start working with local models, but this

Speaker A

principle, even for free on, well, like, hobby scales. It's fast and it's very simple, right? That is, you just tell the codebase that I want to fine-tune the LLM like this, and it will fine-tune everything for you. You just

Speaker A

fact, local models have also advanced very significantly. And, well, you’ve probably seen that, say, Qwen-3-827 B came out, which catches up to Opus 4.5 in a number of tests. It’s just something incredible, right? Well, obviously, it’s a fairly old model—

Speaker A

need, again, you just need to understand what key points to give it, right, like where it will get, for example, a GPU, that is, video cards for training. So that is the question, but the process itself, honestly, is

Speaker A

essentially, we will integrate them into our tools. Then, a session—uh, we will have session five. And that one is very important because we will be fine-tuning them for your specific task . And this is—I'll repeat again— this, I think, is the most important

Speaker A

very simple. It's not perfect, of course, there are still some complexities, but any person, absolutely any person, can now train their own LLM. I also became interested in embedding an AI agent directly into a product, because if you remember, in

Speaker A

thing. Well, I mean, simple in the sense that we have the code for it, in that sense. And, well, we will also, of course, figure out where you can get a GPU for free, because there are different ways to do that. And what to

Speaker A

the first version of this course, we were embedding, actually, the codebase. And honestly, this has completely changed as well, because now we integrate Codex in the very first course. It has become so simple that we now build a browser with built-in Codex

Speaker A

dataset for training. That is the real challenge. And I recently watched an interview with the creator of Jeff, remember this new technology, the one that makes decisions really well and quickly, a decision model of sorts. I watched an interview with him, and he

Speaker A

right in the first course. Well, it has basically become such an elementary thing. But Pie is interesting. Pie is interesting because it offers more engineering solutions, gives you more to think about, and it's interesting to work with. As you remember, in the

Speaker A

are all, honestly, weaker than Jeff. There is even one that recently came out from Cloudflare. It is also smart, but it costs twice as much, for example . So, how did Jeff manage to be, on one hand, very cheap and very fast, while

Speaker A

first level of the course, or rather, as you remember, in the previous version of this course, we integrated Codex into our product. Now we do this at the first level of the course. It has become so simple, such a basic

Speaker A

there be Jeff?" Can Jeff actually be integrated into a lot of places here? I haven't fully decided where exactly yet , but I just haven't, I haven't quite figured out where it would fit best.

Speaker A

thing, that even at the first level of the course, we can integrate Codex into a browser extension without any issues.

Speaker A

I think maybe we will integrate it into one of these topics too. But Jeff is now just with us, well, Jeff, the hypothetical Jeff, right? So, we might, by the way, we might try, for instance, to run a local model, not an LLM, but

Speaker A

And it happens on the first try, without any problems whatsoever. Uh, it's just an elementary thing. And at the same time, at the same time, I personally find it much more interesting to work with PieAgent for these purposes, because Pie is

Speaker A

worse than Jeff. I mean, they are, well , let's put it this way, for example, where Jeff makes mistakes in 20%of cases, these alternatives make mistakes in 50%of cases. I mean, they're just that bad. And so I'm thinking, for this

Speaker A

interesting in that it's like a construction set, like LEGO, and it's fun to work with. I am currently integrating it into many of my projects, and it's interesting because it's so simple,

Speaker A

to create datasets properly, how to do, well, that's exactly what this model fine-tuning process will be. And by the way, we'll talk about when fine-tuning isn't actually necessary, and a good, proper prompt is enough; that's also a common situation. And, in fact, your

Speaker A

fifth project will be to fine-tune a model. Uh, well, you could use this lesson, or maybe we'll use a new one.

Speaker A

You see, when it comes to AI, to be honest, it's become very difficult to plan everything, because take the same Jeff, for instance, who knew it would be released? I mean, seriously, who could have predicted that Jeff would be

Speaker A

released tomorrow? Exactly. But it was released, and it changed everything significantly. So that is roughly our tentative plan. And for the sixth lesson, where we'll be embedding AI, I was thinking about what to embed it into, and the idea of making our own

Speaker A

site generator, like, you know, Relume, or something like that, or even V0, just to make something of our own, seemed interesting. Yeah. I think it's a very interesting idea because it's engaging. On one hand, it's actually a very technical task. I mean, there's no

Speaker A

problem at all in doing it. In principle, the problem is doing it in a way that it works truly efficiently and is also secure. So these two lessons are very closely linked, because as soon as you give third-party users who

Speaker A

come along a literal coding agent that runs on your device, it's dangerous. So , you really need to think everything through very carefully here. Exactly.

Speaker A

And so, in this lesson, we will be creating our own AI website generator, the kind that other people can actually use, right? Meaning they come to it, enter what they want to generate, and it generates it for them. Maybe not

Speaker A

necessarily websites; we could think of a presentation generator, or a generator for something interesting and visual. We'll see how that evolves, but I think we will definitely do this with Python and see what else—I have a few ideas on what interesting things can be

Speaker A

built in. Well, most importantly, there are all sorts of isolation tools, because that overlaps quite a bit with security. And we will discuss more here , for example, what the difference is between a standard container and a microVM. We will look at what indirect

Speaker A

injections are and what the "deadly triad" is, right? So, in general, we will be studying various things related to security. And accordingly, our final project will be to conduct a full security audit of your own project.

Speaker A

Finally, we will talk about AI as a productivity tool, just from a somewhat new angle, because something like Jev, for instance, opens up completely new horizons for that. For example, if previously you needed to monitor something, say, competitors 'websites

Speaker A

or just keep track of certain things. And, honestly, in the past, it was, well, yes, you could do it fairly inexpensively. But you still would have had to pay for it, right? I mean, it was still something you had to pay for.

Speaker A

And with Jev, well, it becomes practically free. I mean, maybe, I don't know, a dollar a month, if you are analyzing a whole lot of things, for example. So, that is roughly what the updated plan looks like. And now

Speaker A

let's see in practice how these things work. Let's start with what I want to do with an arcade game. And this is a rough concept of what I would like to make, and how I see, how I see its

Speaker A

creation, right? So, first of all, I noticed that right now AI has certainly advanced a lot in beautiful graphics. I mean, the graphics—it can actually create this nice pixel art, and it can truly look beautiful. And I want to

Speaker A

combine this in this project. I mean, creating loops is one thing, but it’s something completely different to try to learn how to make your agent create such a really cool, beautiful image that we can actually work with further

Speaker A

in an interesting way. The second project is a website that doesn't look like it was created by AI. And here I want to show you the Magic Path tool, because it appeared recently. It is now , like, it can now be integrated into

Speaker A

uh into ChatGPT and accordingly, let me just show you how it works, because I think things like this are starting to change everything. So, for example, let's, well, for example, ask it to make so, I'm just in Codexe here. And

Speaker A

let's use Magic Path, yes, it's a plugin now. By the way, notice that plugins in Codexe and ChatGPT have become much more than they used to be.

Speaker A

I mean, if before plugins were just a kind of packaged, uh, packaged skill, a packaged MCP, well, some kind of connector to a service, or packaged hooks, now plugins have become something else entirely. Plugins have now become, well, not exactly for

Speaker A

everything, but if a company wants to, it can essentially embed its product into the chat or Codexe, just like Magic Path did. And, for instance, let's ask for something like this. Uh assemble for me a few examples of

Speaker A

beautiful, uh, buttons, minimalistic ones, but ones that will press and toggle very nicely. Just various beautiful ones like that. Make some UI elements that I want to take a look at.

Speaker A

Make five different button elements there, uh, buttons, radio buttons, and all that kind of stuff. And make their toggling look aesthetically very pleasing. I'll try Astra medium, because Astra medium is good for these kinds of tasks. And look at what

Speaker A

happens next, right? I mean, this is how we activate this plugin, which activates within Codexe. And, uh, it's now, it's also, uh, Magic Path will read the skills now. And what is Magic Path anyway? Magic Path is, essentially

Speaker A

, a service for design generation that, well, in essence, has built-in skills for how to work and what good design principles are. I mean, relatively speaking, you take a good model. Well, in our case, it's GPT6 Astra, and it's

Speaker A

given, well, you could say, a more proper wrapper. And, uh, Magic Path is paid, but I think you can get about 100 requests a day—oh, wait, a month— for free. So, for a hobby project, that’s enough to try it out. And it

Speaker A

just draws everything right here. You can also edit everything here, add things, or change pictures if needed.

Speaker A

And this is exactly that plugin that’s now available, because you can see exactly what it’s about to do.

Speaker A

But, of course, I have to say that it might take some time. And while it’s working, let me show you an example it made for me while I was practicing. I asked it to make a toggle switch; I

Speaker A

just said, "Make me a beautiful toggle button." And to be honest, I didn't even expect it to produce such a result . Look, when I say a design that doesn't look like it was made by AI, this is what I mean. Just look at it. I

Speaker A

mean, could you really have imagined, say, a year ago—if you were in the first version of this course—that AI would be making designs like this, transitions like this, and buttons like this? I think this is, of course, a

Speaker A

very significant leap forward. We are literally witnessing progress in design right before our eyes. And that’s what I want to highlight in this course , because it’s a very interesting thing, and I don't know where it will lead us, but this, you know,

Speaker A

cookie-cutter AI design will go away; it will disappear very soon. It will more likely be a sign that a person doesn't know how to work with AI. So, if you see some template-based, obviously AI-generated, "sloppy" design , it will just suggest that, well, yeah

Speaker A

, the person doesn't know how to work with it, while good AI can do things like this. And just pay attention to how many subtleties there are, how the shadows are distributed, and how it all works. Let’s see what it has done for

Speaker A

us here. It’s not very fast, even though I have the fast mode enabled right now. But, by the way, I hope to upgrade to the Ultra Fast plan by the time we start this course. I'm just not switching to it, because I currently

Speaker A

have a very favorable subscription; because until October 30, these $ 200 subscriptions provide a 20x usage limit , but starting October 30, they will become 10x. And so, while they're offering this, I'm not switching. I'm not switching to Ultrafast—oops, not

Speaker A

Ultrafast, but that $ 500 plan that includes Ultrafast—but when I do switch to it, I'll be using Ultrafast for my classes. I'm really looking forward to that. Uh. Let me answer a question while it's processing. So.

Speaker A

Marina asks: "What if I took the first level a year ago, do I need to retake the first level?" Well, look, the first level has changed too, but I wouldn't say it has changed fundamentally. Just a second. It's finished now. I want...

Speaker A

no, it's actually working well. But I just want...here, right here, I recognize Astra. There's an Astra style here, so I'll tell it: "Make it full minimalism." Uh. A minimalist, almost black-and-white image with small color accents. And use the most current

Speaker A

design trends for October 2026. Do some research and make it really trendy. I mean, honestly, this is already a very good design. I mean, if we look, well, look, for example, now. Now. And why isn't it moving? See, yes, if you look,

Speaker A

these are very beautiful button transitions, and this animation is really very, very pleasant. But I want —I want it to be unrecognizable, because here I recognize it, and here —I don't. Here I feel that it's done really coolly, but over there it was

Speaker A

recognizable, and I want to get to the point where it isn't. So, yes, answering Marina's question: if the first level of the course was a year ago, do you need to retake the first level? Well, look, in general, the

Speaker A

first level has changed too, but not that critically. Well, it has changed. But overall, if you completed the first course and haven't abandoned it since then, and have just been developing things yourself, then I don't think you need to. I mean, if you just finished

Speaker A

it a year ago and haven't worked much since then, then perhaps it might be worth it. Mm. I mean, well, if you’ve already gone through it and kept your knowledge up, then I don’t think you need to. He’s mentioning Linear, by

Speaker A

the way. Interesting, right? He might just make it like Linear now. Moving on . Oh, Andrey asks: "Is Magic Path only for Mac?" I don’t think so, because look, it’s a plugin, it’s a plugin.

Speaker A

Well, I mean, there is a separate service, but generally, it’s a plugin . Right now, I’m using this plugin in Codeium and ChatGPT. Since it’s an official plugin in ChatGPT, it shouldn't just work with Mac; it should work with everything. It would be just

Speaker A

strange if it were a plugin that only worked for Mac. I don’t know for sure , but it seems like it would be very strange. I mean, it seems to me that OpenAI themselves would have to say, "

Speaker A

Listen, if you want to be in our plugin section, you have to be for everyone, not just Mac." Uh Right, Bogdan asks: " Everyone is running around with' Hey, we've moved on. 'Good old AutoGPT." A question regarding the upcoming agents

Speaker A

course. Is it still in the ranks, or is it hopelessly outdated? Maybe it’s not worth spending time learning it anymore? I think you asked a very similar question on the last stream you attended. No, just look at the

Speaker A

situation. Uh AutoGPT is just a way to learn how agents work. I mean, what’s the problem, for example, with, say, Grok or bots, right? What is their actual problem? Globally speaking. The problem is that they’ve hidden everything from you completely.

Speaker A

Everything. They give you cool functionality, but you have absolutely no idea how it works. And at that moment, if, for instance, say, OpenAI, or Grok, or now ChatGPT has GPTs— which are also agents—if they block your account, you won’t be able to

Speaker A

get anything out of it. They will take it away from you; it’s not yours, really. And AutoGPT, or not just AutoGPT, but LangChain and others, they are yours. You fully understand how it’s built, how it works, where it

Speaker A

stores data, and you can switch to another provider at any moment. And so this OpenClow course, it will never go out of date in the sense that perhaps at some point we won't be studying OpenCl, but we will definitely be

Speaker A

studying the agents that you control. That is, if something fundamentally better than OpenClow comes out, and it's as open as OpenCl, we will most likely switch to it in the course immediately. But switching, for example , to Grok-bots—personally, I really

Speaker A

like Grok-bot. I use it myself every day. It's wonderful. But I just don't trust Zuckerberg to the point that I don't want to give him even a kilobyte of my data. And so, I don't use it. But , for example, once they slowly roll

Speaker A

out GPTs to everyone in ChatGPT, I will gladly use them, but at the same time these are just completely different things, right? I mean, I would never build my own agent entirely on Grok, or on GPTs, or on Muse, because they could

Speaker A

take it away or change it, and you wouldn't be able to do anything about it. You don't even know where its personality is stored, where its data is, or where anything about it is kept at all. So, that is the fundamental

Speaker A

difference. So, meanwhile, it has done —oh, well, look, this already looks more interesting. There. And look at the transition as well. Can you feel how interesting it is? So, what about here? We also have, yes, a beautiful button there. So, here, this doesn't

Speaker A

feel as much like it was made with that . That's interesting in that regard.

Speaker A

Look, also quite a nice animation. Cool . Great. Anyway, Magic Path is an interesting tool, give it a try. And you can also use it to create full designs, not just some design elements.

Speaker A

Ah, and so, let me also answer a question right away. Bogdan asks: "Is it possible to implement the function of an art director today?" That is, someone with good taste who would hold the whole picture in their head? It's

Speaker A

clear that you can make very beautiful buttons and transitions, but if a person without taste or a good eye does it, they'll just make a cookie-cutter mess. The elements are fine individually, but they don't go together as a whole. That's an

Speaker A

interesting question. And what do I see here? I am curious about what will happen sooner: will models evolve to the point where they become that very art director you are writing about?

Speaker A

That is the first path, and it is a very possible one, actually. I mean, it doesn't exist yet, but maybe by winter it will, who knows, right? So, maybe in a few months we will already have such a model. But there is also a second

Speaker A

path. And the second path is that, well , many people around the world are currently asking themselves this same question. And I fully admit that quite soon we will get, uh, we will get some kind of service like this. To be honest

Speaker A

, it is in the air, it is very much in the air. And it feels like a couple of puzzle pieces are missing for this service. And such services are already appearing. It is just that, well, the moment hasn't fully ripened yet, it

Speaker A

hasn't arrived, but it will come. The service is such that imagine you have, say, tens of thousands of real art directors, tens of thousands of, mm, other professions where this human judgment of taste is needed. And imagine that you literally give your

Speaker A

agent an MCP, an MCP access to this service. And your agent, it generates an image, sends it to this service.

Speaker A

There, this image is shown to some completely real art director, who just, it starts going, their phone goes: boo-boom, boo-boom. You know, like it does for taxi drivers, right, that, poo-poom, poo-poom, poo-poom. A new request for direction has arrived,

Speaker A

would you like to evaluate it? And the person is like: "Well, yeah, I have five minutes free, let's take a look." And they open this image, they might just say: "Listen, this is bad, this part is great, this is cool, redo this.

Speaker A

" That is it, they get their credits for it, and your agent gets instant feedback. And then the agent comes back to you and says: "Listen, I evaluated this with a real art director." They said that this part needs to be changed

Speaker A

. I will take care of it now. I think the question is, what will appear faster? I do not know, to be honest. I mean, maybe the models, or maybe not, because, you know, the visual aspect is still a major hurdle for models. And if

Speaker A

we, for example, let's say, if in the next six months or so the models don't develop that much, I'm almost certain that in six months such a service will really take off. And it will be a very necessary service. And it will be

Speaker A

needed not just for, say, an art director, but also for checking texts and so on and so forth. And I'm sure that many people will be happy to do this, because it will actually be a great way for everyone to earn money

Speaker A

too. So that’s roughly my vision on this matter. But what the path will be, I don’t know. I mean, nobody here knows that. Maybe tomorrow some, not Jeff, but I don't know, a dev or some design entity will come out and just

Speaker A

say: "Oh, that’s it, we’ve solved this problem.""Everything, we’ve now made a new model, we’ve trained it on a new principle.""And it’s trained on all the best designs in the world and boom, that’s it, it makes stunning designs right away." Who knows, maybe.

Speaker A

Ah, yes, that is, accordingly, what concerns our second project, the site that doesn't look like it was created by AI. And the third one is our project , which is skills. Well, here we’ll just focus more on the skills

Speaker A

themselves and all such things, right? I mean, we will look at how skills work , how to actually create them professionally, how to evolve them, right? I mean, just one big session dedicated purely to skills. And I, I

Speaker A

currently only have skills covered in Claude, but not the whole session is dedicated to skills there. There, skills are, like, an important part of one of the sessions. But here I want to do a full 2 hours just on skills. And

Speaker A

then we’ll have a project, essentially a command line for your product, where we’ll embed some local model. And the more I think about it, the more I would like this local model to be some kind of decision model. I

Speaker A

think that will be really, really interesting. And most importantly, we will fine-tune the model on our own data. And how will we do this? First of all, to start, let’s see what can be done. I, when I was thinking about what

Speaker A

example to show you clearly, right, how to show you this clearly, right? So, how do you show that a model can't or doesn't know how to do something, and then suddenly it can? Here. I decided to make this example for you. Let's

Speaker A

take Notion, right? There it is. Well, Notion, I think you know what that is, right? It's a program that's used, among other things, for managing projects like this. So we have "to-do," "in progress," and "review" here. It's a standard dashboard, right? And let's

Speaker A

try taking, for example, a small local model that almost anyone can run on, well, a more or less decent simple computer. And let's see if it can control Notion, right? So, can it actually do that? And, well, I actually

Speaker A

did this because I’ll tell you right now that these things aren't done live anymore. I mean, we will be doing them in class because we'll have two hours for it, but in general, many things still need to be prepared in advance.

Speaker A

Or, or if I have UltraFast, we can do everything in class, but that requires UltraFast because it works eight times faster. Ah, and so I prepared, I prepared the following thing. This is, in essence, a control panel. I can even

Speaker A

make it smaller like this. This is a control panel that manages this Notion of mine. Well, it actually does it via MCP. So just like any other MCP works there. It's just connected to MCP. And there is a small model connected here

Speaker A

right now called Mini CPM 2B. So it's 2 billion parameters. Such a tiny little model. It will run on almost any computer. Well, I mean, if you have at least some amount of video memory, yeah , it will work, or it will probably

Speaker A

even work on just RAM. So it will work. Right, let's check our internet connection now. Yeah, seems okay. And let's see, right, how it works. So I made three examples. This is just this model with a basic prompt, well, a

Speaker A

decent one, just basic. And this is the same model with a heavily improved prompt. And this model is fine-tuned to work specifically well with Notion, so it can control Notion. And, well, let's , for example, take a simple task.

Speaker A

Return it. Wait, let me just switch to like this. Move the landing page card from "Done" to "Review." I'm not at all sure it will do it now, because, again, it struggles a lot with this. Why is it

Speaker A

giving me an error now? Hold on. Hm, let's reset everything. Uh, you've started acting up on me. Unexpected.

Speaker A

Just a second, we'll ask Codex what's going on. Everything was working before this. Tell me, why is it writing this?

Speaker A

Why is my Notion management app showing me this now? Okay. I literally tested everything right before the class, and it was all fine. Ah, well, let's see now. It’s interesting that these modes all work with this specific one.

Speaker A

Something happened to this mode. Let's write it down. And for some reason, the cards aren't showing up in simple prompt mode. Now we’ll write this down, too. And like this. And it writes like this. Yes. But let's try using an

Speaker A

improved prompt, perhaps. Maybe it will work now with an improved prompt. Let's take the improved prompt. And, for example, let's do: move the landing page from "Done" to "Review." The landing page is this card here, right?

Speaker A

So, we have the landing page card. And we need it to move it from "Done" to " Review" using a command. Let's hit execute. It's started. It has started executing now. And, uh, with an improved prompt, in some cases it might

Speaker A

actually work. On the course, we will analyze when you need to improve the prompt and when, uh, well, when you actually need to fine-tune the model.

Speaker A

Okay, let's see. It’s started thinking. Another thing is that when a model is poorly trained for certain actions, it will take a very long time to perform them. Meaning, it might eventually do it, but the action will be very slow. It could just, well, it

Speaker A

could just take a very long time to execute. Let's see. Maybe it won't do it at all. So far, it hasn't done it yet. We're watching this landing card.

Speaker A

It should, like, move over here to " Review." If it does it, then the model handled it. If not, then no. But again, let me remind you, this is a very small model. Why this specific model? Because , for example, I compared it with

Speaker A

Gemini and, oh, wait, not Gemini, Gemma . And it turned out to be somewhat worse. So, for some reason, Gemma handled this worse. Right, looks like it won't handle this either. It seems like it, like it went somewhere off

Speaker A

there. Somehow it, somehow it went somewhere. Right, now we'll ask it now. Ah, check the logs. Oh, look at the logs. I have an improved prompt now, it's executing something, but is it doing it or not? I feel like the model

Speaker A

is working, because you, you can hear my computer has started to hum, the computer is humming, because I think the little model is chugging away there . We'll see, maybe it will do it now. I I am asking Codex right now, but it's

Speaker A

like, it's normal, because the model is very silly and small, and it needs to control Notion, you know. I mean, imagine, a tiny 2B model is going to control Notion now. That's definitely something, but we'll see now. Maybe it

Speaker A

will manage in the end. We just need to wait a while. Now we will now we'll find out for sure, or if it has already frozen. Let's see. Oh, the model returned an incorrect plan. The changes were not applied. So, that means it

Speaker A

spent a total of 183 seconds thinking. And, accordingly, that's how much it spent on it, yeah. And now now let's try the same command, yeah, for this same model, only the one that was fine-tuned to control Notion. Like this

Speaker A

. After training, we have the same thing here, the landing page. I say, move the landing page from Done to review. It also doesn't happen instantly, because Well, all done. How long did that take? It took, so, how

Speaker A

much in total? Generation. Well, in general, roughly five or six about 7 seconds. It took 7 seconds. So, just look at that. The model without, this model without training, even with a good, correct prompt, a well-thought-out prompt, it couldn't do

Speaker A

it. This one did. And now let's, for example, let's see what else, for example, to ask. It still can't do everything, but let's, for example, try , uh, for instance, move the announcement and the newsletter to Done . Let's see if it handles it or not now

Speaker A

. I haven't tried this myself yet. Right, which card do you mean now? Let's check. Ah, maybe it's, maybe it's only trained on one. Let's try now, move announcement to Done. Let's watch.

Speaker A

Yes, everything, it works well with one . Maybe it needed more training so it could perform multiple actions. And now , for example, let's say, for example, let's make an instruction for the card.

Speaker A

Set priority to high and make Anna work on it instead of Olga. That’s a complex one. I don't know if it can handle such, you know, a comprehensive, yes, a comprehensive task. Let's see.

Speaker A

Look, what we just did took some time. It took 5 seconds: 3 seconds for generation, and the rest for loading.

Speaker A

Well, in short, it’s, well, you saw it, it’s happening literally before your eyes. So, once again, a standard 2B tiny model. We trained it, I trained it, and I'll tell you right now how I trained it. You can do this too, and

Speaker A

after the course you’ll know how to do exactly this. It is trained to control Notion. Just imagine, a tiny, little model that will be able to control Notion right on your computer.

Speaker A

You write to it, and it manages it. I think that's really, really cool. So, I think let's take a break now, and after the break, we’ll talk about how this was done. I will show you how to do it,

Speaker A

where you can do it, and, naturally, we’ll finish discussing our course plan. So, be sure to come back in 10 minutes. Yes. Good evening once again, everyone. We are back from the break.

Speaker A

And let me first answer a question. Victoria asks: "What is the difference between simply tasking Claude-code with tracking tasks, statuses, and queues and connecting Notion? What are the real advantages of using Notion to manage these processes?""Well, look, let me just say right away that the

Speaker A

example I’m showing with Notion is just an illustrative example, as we just saw, right in front of you, how a small 2B model can control Notion. And it's really just done to show that it can control anything at all in the same

Speaker A

way. It can control some other program that you need to manage. For example, it could control some corporate system.

Speaker A

Just imagine, for instance, that you, let's say, if you approach integration professionally, yes, just imagine the company has some internal system, or maybe not internal, perhaps one of the popular systems. Or maybe they use it right in the company, who knows, right?

Speaker A

And they need, for instance, let's say they have a meeting, like a call, and as a result of this call, a transcript is produced. And they need this transcript to be turned into tasks, right, for example, in their in task

Speaker A

management system, or to have tasks updated within their system. So, what are the options? For example, the first option is—well, let's imagine they want all of this to happen inside the company. Meaning, for example, this company is a large one, maybe a bank or

Speaker A

a medical company, or something, and they don't want any data leaving their premises at all. So what are their options? For example, buying hardware for a lot of money, say, I don't know, 20 or 30 thousand dollars, and running

Speaker A

a powerful LLM on that hardware that will handle this task. Another option is to buy weaker hardware and hire you.

Speaker A

You train a model for them that will do this task well, and on weaker hardware, they can run it much cheaper at the exact same quality of performance, right? So that’s one example. It’s not necessarily just about saving money

Speaker A

, though. Perhaps they might buy a large hardware setup, but they’ll run your model and maybe another one as well for other needs of theirs. So, when we start talking about serious AI application, the option of just having

Speaker A

it do something brings up the question: can we send this data to Claude, for instance? I mean, can Claude, for example, see all of this? So, this is exactly what this course is for, right?

Speaker A

It’s about these next-level things. People often ask me if I’m going to make a course on how to integrate AI into companies. Well, probably not yet, but things like this are partially about that, right? So that when a

Speaker A

company asks you," Can you train a model for us? "By the way, let me tell you right now, the vast majority of people are 100%convinced that all models can be fine-tuned. And that is just, well, it’s a bit astounding. On

Speaker A

every call with a client, the client asks quite seriously:" Well, you'll fine-tune the model, right? "and they might be talking, I don't know, about something like GPT-3. And I'm like:" Uh-huh, sure, right away, let's go fine-tune GPT-3. "Yeah. I mean, for

Speaker A

some reason, everyone thinks that you absolutely have to fine-tune models. And it's funny that the situation is now reaching a point where, in principle, anyone can actually fine-tune them. Well, we're talking about small models, of course.

Speaker A

Fine-tuning large models is much more difficult. Now, let's actually talk about how this was fine-tuned, okay? So what do you need for this, what do you need for this? Well, you need the following. Essentially, to fine-tune any model, you need two things. You

Speaker A

need a dataset of data that the model will be fine-tuned on. I mean, for example, in the case of our task here, what is that? You just need to create a large dataset of correct actions. That is, for example, if this incoming

Speaker A

command comes in, these specific actions should be performed. So the model should say to perform such and such actions. And we just take, say, 10-20,000 such commands that are generated synthetically. These are commands that, say, Codex would generate for you without any problems,

Speaker A

even thousands of them. And this dataset is taken. And with this dataset , we get some GPUs. And, well, let's put it this way. If you have a powerful computer at home, you can train at home . If you have a weak computer at home,

Speaker A

then you can rent a GPU. And, for example, there is a service called modal.com. And they actually, well, I don't know how it is now, but at least for me, it still works. I don't know, maybe something will change, but they,

Speaker A

for example, give 30 dollars a month for free. Just to everyone who creates an account there. You can just create an account, and every month you'll be given 30 dollars. They don't accumulate , meaning they expire every month, but

Speaker A

30 dollars, so you understand, is basically enough to train what we did just now. So, to train this model, 30 dollars will be enough, because training with a good dataset like this, well, if you already have a good

Speaker A

dataset and training takes about 50 minutes on it, you can just rent a GPU for those 50 minutes. And just so you understand, renting a GPU costs about ...I'll show you now how much. Right now. Well, we can go to pricing; for

Speaker A

this kind of training, well, actually, plenty of things will be enough. Like an H100, for example, an H200, right? I mean, it'll definitely be enough 100%, even something smaller would work, it would just take longer. It’s just a

Speaker A

question of what kind of speed you want . And accordingly, well, let's calculate the price per hour. Look, see ? So, these are the prices we have, and for 30 dollars, you can actually train for quite a decent amount of time. The

Speaker A

thing is, it probably won't work out on the first try. So, why do I say that 30 dollars will be enough? It's because, for example, well, most likely, there will be several stages where you train, then improve the dataset, then train

Speaker A

again, and eventually you will reach a good result, but that will cost you, roughly speaking, those 30 dollars you get for free here. Plus, in the course, I will also talk about what other free options there are. So, in principle,

Speaker A

there are options. But again, if you have a reasonably powerful computer at home, you can train at home, it just won't be 40 minutes, maybe a few days, right? So that is also a perfectly viable option. And, for instance,

Speaker A

specifically here, right now I'll show you, I even made a little artifact to show you how it was trained, right? So again, how was it done, right? Well, right now so, the code generated 20,000 examples, right, and for completely

Speaker A

different actions, so we took various actions, like transferring, for example . Well, you saw it, I did it right in front of you, changing the priority, changing the name, so there are many, many different actions. And he shows

Speaker A

20,000 examples. He created different actions. So, it's just a dataset of 20,000 examples like that. After that, he did what sort of he took those 20,000 examples and sent them for fine-tuning. So, we took the MiniCPM-2B model. And actually, I didn't just take

Speaker A

that one, I also took that Gemma one I mentioned, I think it was 2 B4, I think , if I'm not mistaken. Well, it's also a very small one. Or 2 B-it, honestly, I don't remember exactly. And I trained

Speaker A

both, but it turned out that this one is better, just better for such training, specifically for this task.

Speaker A

And that's it, it trained. It was trained using the LoRA principle. I'll also tell you what LoRA is, because, well, there are different approaches.

Speaker A

And it was trained on one B200 machine. So, this means it cost me, if one B200 machine is, well, let's say 6.25 per hour. So, in principle, well, again, 30 dollars is definitely enough for you, well, even with a few failed attempts,

Speaker A

to cover this training. And accordingly , it trained. Moreover, I did this with different variants, also different things. Well, in the end, everything took about 40 minutes. it was training.

Speaker A

Ah, and that's all. And accordingly, based on tests, for example, on test actions, on average, out of ten requests, for example, only one request succeeded before training, and it took 40 seconds on average. Well, you saw, we waited 108 seconds there, and it

Speaker A

still didn't pass. With an improved prompt, it, ah, passes on average. Well , that is, if you just take such an untrained model and write a good, high-quality, correct prompt right for this specific task. And then five out

Speaker A

of ten actions are performed and it's also much faster, like 12-14 seconds. But if we fine-tune the model, we have 10 out of ten test actions performed, and the whole request usually takes like 5 seconds. Well, you saw it, right

Speaker A

? So it all happens very quickly. And accordingly, that's how it works. So you just take it and, in practice, what does it look like, right? So you don't think it's some kind of process or anything like that. It's literally just

Speaker A

you telling the code. So, well, we will be doing this in class. We just take and install the model CLI, and say:" Look, we have this specific task, we want to train for it, let's think about what kind of dataset we can create for

Speaker A

this training. "And perhaps, in some cases, you can use third-party datasets , because, in general, there are many datasets. For example, one of the datasets that you can, well, use relatively freely, although there are some nuances, is Kaggle, or they also

Speaker A

have different, they have different datasets. You can download them. Some datasets can be used for commercial purposes, while others only for, well, non-commercial ones. And there are a bunch of datasets here, like the price history of Bitcoin, or the results of

Speaker A

all football matches since 1872. Well, just a huge number of datasets. So, you can train your model on the dataset you need. And that is exactly what we will be doing in class. Training, or fine-tuning, to be more precise, of

Speaker A

models will be an important component of this course. Apart from that, we will be creating our own AI website generator. And here, well, I would like to emphasize that by creating such a generator, you, through creating such an AI website generator, will learn to

Speaker A

think a bit more broadly in terms of both security and potential functionality issues. And, for example, what if, let's say, you have built, let's say, a website generator, and a user wants to generate some illegal websites, right? How do you get around

Speaker A

that? So, I want to use this project not just to embed an agent, but to look more at such practical nuances that arise here. And Evgeny writes:" In the July one, seems like in the model training navigator, there was some very

Speaker A

special program on GitHub, and there are such open-source ones too, right? " Yes, yes, there was one. I think it was for Mac, if I'm not mistaken, but for serious, well, serious work, you definitely need a GPU, because

Speaker A

fine-tuning models on your own computer —it's, well, you can, but it needs to be really powerful. But otherwise, it's easiest to rent a GPU. Again, from a number of sources, you can get them for free. This is, of course, not as simple

Speaker A

as it seems. I mean, there are certain nuances to it, but overall it’s a genuinely interesting thing. And most importantly, once you realize that it’s actually quite straightforward, you will be able to use it for various tasks of your own. So, what else do we

Speaker A

have left to discuss today? Well, first off, which coding agents we will be using. Honestly, I think we will be doing most things in Cursor, especially if I have the ultrafast mode on, then we'll do the vast majority in Cursor.

Speaker A

But perhaps I will use Claude for visual tasks, because overall I prefer how especially Opus 3.5 works with visual stuff right now. It works really , really great for that. But we will be doing most of the work in Cursor, as

Speaker A

Cursor is a very good workhorse for all of this. As for additional expenses, look, I expect that the free tier of Modal will be enough for you. However, of course, when you are building your AI website generator, you will need

Speaker A

some kind of API, like OpenRouter or something similar. It is, of course, better not to use free models. But honestly, this course is not a beginner-level course, and it mostly makes sense for those who see that certain AI expenses are definitely

Speaker A

justified. So, you will likely need to top up your OpenRouter account a little bit. Ah, Vladislav asks:" Did Cursor lower its limits? "Yes, they did, but again, generally, we are moving towards the idea that while you might be able

Speaker A

to do it with a $ 20 subscription, we are more focused on potentially switching to higher-tier subscriptions, because it really is the next level, right? So, if you feel and understand that you can use AI for more serious

Speaker A

tasks, then, accordingly, you just have to pay more for it. There we go. We will also talk about using AI coding for productivity. And this is where I really want to focus; yes, I'd like to focus more on things like various

Speaker A

automations: monitoring automation, file operation automation, email automation, and automating everything you possibly can, especially with models like Claude, which make it all so much cheaper. Well, they make it all hundreds of times cheaper now. And Vladimir writes:" I'm on the 200 plan,

Speaker A

this affected me quite a bit, well, look, theoretically, the 200 plans haven't been touched yet. "I mean, I'm also on the 200 plan and, theoretically , they haven't been affected yet. Maybe they have already been touched, but officially they will be affected

Speaker A

starting from the 30th. So. Most likely , starting on the 30th, you will really feel that things are changing. But by the way, in that regard, I feel like Claude allows you to do so much right now for 200. Opus 3.5—I have this

Speaker A

setup now where the 200-dollar plan is enough for me to have Opus 3.5 working almost non-stop for about 3-4 days every week. Well, mostly working on one or two projects in parallel. That wasn't possible before. It's really cool. Uh, yes. If you have any

Speaker A

questions, go ahead and ask. But overall, let's summarize. I decided to create—well, it's a completely reworked previous course, a level two course that focuses on loop engineering , designing with AI agents, creating skills, training for your specific tasks, and a more mature, serious

Speaker A

approach to embedding agents into your products. And also security. AI and products. So, if these topics resonate with you, if you feel that this is something you are also interested in, then I will be glad to see you in the

Speaker A

course. And accordingly, our next group starts very soon, I think it's in less than—how long until it starts? There are only a few spots left. It starts on October 14th. The next group, rest assured, will only be in January. So,

Speaker A

either join this group or wait until next year. So if you're interested, be sure to come. And I'll be very happy to see you. I see a lot of familiar names.

Speaker A

That is very pleasant. Oleg asks:" What about those who have already taken this course before? "" There seemed to be information about a fifty percent discount for taking the updated course.

Speaker A

"Listen, I don't know for sure, but theoretically, it seems to me that it should work. It's just that, I think, it's done through—you need to go into your personal, hold on, I'll show you now. Right now you need to go into your

Speaker A

personal account. Here is the personal account. This is automation and productivity, course retake. And, accordingly, if it's a retake here, then theoretically it should give you a 50%discount. But if anything, please write to the administrators. There. Yes . And administrator Mikhail is writing

Speaker A

right now that it's in the previous course's resource, yes. So, if you have already taken it, you can come to this updated course with a very good discount. And I think it will be great, yes. So I hope you also learned

Speaker A

something interesting today, at least about Magic Path and that models can be fine-tuned this way. And I will be glad to see you all on this course, which starts in less than 2 weeks. So thank you very much for coming, and see you

Speaker A

later, take care.

Topics:AI developmentautomationloop engineeringLLM fine-tuningAI productivityAI agentsProfile schoolClaude CodeAI course 2026AI project-based learning

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries and speaker detection. 30 minutes free.

App Store Google Play

Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools