YT YouTube2Text

YouTube2Text → Transcripts → Ты не опоздал (пока): как изучить ИИ за 36 минут

Ты не опоздал (пока): как изучить ИИ за 36 минут — Transcript

by Лебедев про нейросети: с нуля до PRO · 6,851 words · 1,002 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Most people are learning neural
0:02networks the wrong way—95%of them, to
0:04be precise. I have spent 15 years
0:06studying AI, and since 2023, I have
0:08been teaching top managers of major
0:10companies and rank-and-file employees
0:12how to implement it. And I have noticed
0:15one mistake that almost everyone who
0:17works with neural networks makes. They
0:20stop learning the moment they figure
0:23out a chat, whether it's ChatGPT or
0:25DeepSeek. They’re chatting away, and
0:27the neural network is giving them
0:28feedback. And there is a big problem
0:31with that, because in the last few
0:33months, neural networks have advanced
0:35much further than many can even imagine
0:37. But it happened completely unnoticed
0:40by most people. That is, many didn't
0:42even know what kind of revolution has
0:43just taken place. And in this video, I
0:45will tell you about seven steps. The
0:47steps I would take if I were starting
0:49from scratch right here, right now. I
0:51have put this together as a complete
0:53course so that you can go through it
0:55all from beginning to end. And I’ve
0:57broken it down into three categories.
0:59The first is for the beginner level,
1:01for those who aren't immersed in neural
1:02networks at all and are thinking, "Man,
1:04how do I even wrap my head around all
1:05this?" For you, there will be two
1:08methods and two tools here. Next, for
1:10experienced people who have already
1:12worked with ChatGPT or DeepSeek and
1:14want more—who want to increase their
1:16speed—there is the next block of
1:18tools for them. And for advanced users
1:21and experts who already know a lot, I
1:23have prepared even more tools. So, we
1:25are going from simple to complex. And
1:27at the end, I will show you a simple
1:29scheme that helps double your salary.
1:32And yes, this is not clickbait; this is
1:33a genuinely real story. And I assure
1:35you, many will be surprised at how
1:38effectively this tool will work right
1:41now. So, hit the like button if
1:42you’re not a bot or a neural network.
1:44And, well, let’s get started. And
1:46we’ll begin with what kind of
1:47revolution actually happened. The fact
1:49is that for most people, neural
1:50networks are just a chat. We’ve
1:52already talked about that. But
1:53prompting—simply sending a request
1:55and getting an answer in a chat—is
1:56actually already a thing of the past,
1:58because you can imagine what happened
2:00now, right? It’s as if a robot
2:02appeared in your kitchen, and instead
2:03of asking it to make you breakfast, you
2:05ask it how you should make breakfast
2:07yourself. I mean, that’s how absurd
2:09the situation is, even though it can
2:11already do it for you. And 99%of people
2:13still use a neural network just like a
2:15smart speaker or Google. And we have
2:18gone through this transition from just
2:20writing prompts to moving on to AI
2:22agents. At first, there were chats
2:25where we sent requests, then services
2:27like Gamma, Hixfield, and Syntax
2:29appeared, allowing us to work with
2:30these things. Then the ability to work
2:34with skills appears, and neural
2:35networks begin to remember not just a
2:37prompt right here in the chat, but to
2:39access different skills and prompts
2:41they have had. This is the next
2:43transition. By the way, we will talk
2:45about this today. Next, the following
2:46level emerges: working on a schedule,
2:48then the work of agents and the work of
2:50orchestrators. Today we will sort out
2:52all these stages, and you will
2:53understand how and why to use them, how
2:55to create all this, while 99%of people
2:57will still remain behind. Therefore, if
3:00you have a friend who would find this
3:02useful, you can always send them this
3:03video at the end. Why did this
3:05transition from chat to AI agents
3:07happen? Just to understand what my
3:09Claude is doing now, an agent installed
3:12on my computer. This is his calendar,
3:14according to which he works. He is
3:16working on five services that we have
3:18launched recently. Previously, creating
3:22such services would have taken us—
3:23well, a whole team, which could cost
3:25from 1 to 4 million rubles. It would
3:28have taken somewhere from 6 to 12
3:30months to develop all of this. Now it
3:32was done without stopping work, while I
3:34was busy conducting training and
3:35sessions for people. Our team member,
3:38with no experience in IT and no
3:39experience in analytics, created an
3:41end-to-end analytics service that helps
3:43him. I will show you today how this can
3:45be done. Then we created a skill that
3:47automatically publishes content. You
3:50have surely already seen those videos
3:52where bloggers explain how one can edit
3:54video using neural networks. And
3:57recently, I was even preparing a course
3:59on Claude, "Claude to the Max." First,
4:01during the day I filmed 4 hours of
4:03material, then gave it to Claude, and
4:05by morning I already had a finished
4:06edit made, once again, with neural
4:08networks. And this is no longer just
4:10working in a chat; this is already a
4:11slightly different level. Or, for
4:13instance, hard-coding an application
4:14that handles processing according to
4:16Federal Law 152. Yes, I know that many
4:18guys will write to me in the comments:
4:19"Aha, but what about personal data and
4:21so on?" We made such a service that
4:23helps us personally work that way as
4:25well. And most importantly, all this
4:27was done without stopping our work.
4:29Besides that, it also helped me earn
4:30gratitude from my wife. Why? Because I
4:32set up an agent for her. As a result, a
4:35week later she asked: "Listen, why did
4:36you install Codeium for me, but use
4:38Claude yourself?""It doesn't really
4:40work well for me." We switched her to
4:42Claude, and she said: "Wow, this is
4:44awesome.""Finally, it understands me
4:46and does what I need." And that is the
4:48difference. And the total savings from
4:51implementing neural networks for me and
4:53my team currently amount to over
4:55800,000 rubles. Despite the fact that
4:58the subscription costs only 200 dollars
4:59a month. So, what do you need to do to
5:01make this work? Step number one is to
5:03choose the right agent that will work
5:06for you. Well, at this point, someone
5:08will surely say: "What do you mean? I
5:09have ChatGPT, don't I?""I have Codeium
5:11or I have Google Gemini." But guys,
5:14wait, let's not rush. Right away, we
5:16say that ChatGPT and other analogs
5:19aren't suitable here. And we will
5:21analyze why right now. But first, there
5:23is a global distinction to be made.
5:24There is Claude in the chat. This is
5:26when you send a prompt and get an
5:27answer. It's very similar to having an
5:29employee who is out and about, and you
5:31can call them on the phone and ask: "
5:33Listen, how do I do such and such?"
5:35It's like a real consultant, like
5:36Google. The second is Claude Work or,
5:39in ChatGPT, there is an analog called
5:40ChatGPT Work. ChatGPT Work is like an
5:43employee who sits remotely in another
5:45office; they also have a computer, can
5:48connect remotely to yours, do something
5:50in your folders, work with files, but
5:53they are sitting somewhere else. And
5:55the third level is Claude Code or
5:57ChatGPT Code, which is a program
5:59installed directly on your computer
6:01with access to it. It's like you
6:04invited an employee, sat them at a desk
6:06, and they are sitting and working
6:07right here at the desk. And this is
6:10exactly the key change that has
6:12happened with neural networks now.
6:15While the first and second work in the
6:17cloud, only online, and have the
6:19ability to connect to one selected
6:21folder on your computer, the third
6:23option has full access to your computer
6:25and can control it, install various
6:27services, create programs there,
6:29interact, and perform tasks on a
6:30schedule. So, the third one is the most
6:33advanced tool. Claude Work operates in
6:35the cloud, while Claude Code is
6:36something installed on your computer.
6:38Of course, there is a problem
6:40associated with this, which is that
6:42your computer must always be turned on,
6:44but that can be solved too. And now I
6:46will tell you how. But if you're too
6:48lazy to open your laptop and keep it
6:50running, you can always use hosting.
6:52For instance, there is HostС hosting.
6:54These are VPS servers in Germany, the
6:56Netherlands, and Poland on Ryzen 9
6:58processors with advanced drives and
7:00DDoS protection. You can install an
7:03agent on such a server that will work
7:0624/7 and help you. Next, you register
7:09on the site, get a server yourself,
7:11don't configure anything, and just give
7:13Claude a task. For example, deploy my
7:16Telegram bot on this server. After that
7:18, you can safely close your laptop and
7:20communicate with the bot from your
7:22phone. Claude accesses the server,
7:24installs everything needed, and
7:25launches the bot. This takes about
7:2715-20 minutes. Also, you can install a
7:30local neural network on the server,
7:32which will manage all processes
7:33entirely, including the agent for
7:35managing Claude. And then the work will
7:38be carried out 24/7 using the hosting
7:40and the server. You will find the link
7:42to HostС below the video. But we are
7:44moving on. So, the key difference
7:46between what was before and what is now
7:48. The most advanced stuff can be
7:50imagined in this way. Claude is like
7:53calling a friend from the show "Who
7:54Wants to Be a Millionaire". Copilot is
7:57like the operator in the movie The
7:59Matrix, whom Neo calls and says: "Tank,
8:01I need an exit." And he guides him,
8:03talks to him, and he has a computer, he
8:05sees all of this, but he is still
8:07sitting remotely. And Claude is like
8:09Jarvis, who fully masters your entire
8:12ecosystem and can work with it and
8:15solve tasks right for you. Moreover,
8:17finishing them to the end. And now we
8:19will talk about this. Visually, of
8:21course, everything looks exactly the
8:23same as in the chat, but of course,
8:24there is a big difference here. So, why
8:26is Claude better than all the others?
8:28And why Claude, Claude, and no one
8:30talks about, say, CodeGPT, Antigravity.
8:31The thing is, ChatGPT CodeGPT, and this
8:33is not based on official benchmarks.
8:35According to official benchmarks, it is
8:36actually the coolest of all. In reality
8:38. From communicating with hundreds of
8:40students I mentor, talking to managers,
8:43product managers, and employees, I see
8:45that work with ChatGPT CodeGPT is done
8:48twice as slowly even on the top models
8:50than with Claude, because it doesn't
8:52finish the task to the end. It says: "
8:55Please go install some library or
8:58server yourself." And this approach
9:00leads to the work constantly stalling,
9:02the work not being finished, and it
9:04requires constant human attention,
9:06where you have to sit like an assistant
9:08to the neural network and actually do
9:10something yourself. Of course, this is
9:12not very convenient, it is very
9:14inconvenient, so CodeGPT loses
9:15significantly. There are other analogs
9:17now, for example, OpenDevin, Cline, and
9:20GMS. These are agents you install on
9:23your computer, just like Claude Dev,
9:25and they allow you to work with Claude
9:28via API keys. But there is a catch.
9:31Just yesterday, I was on a call with
9:33one of our graduates. He told me that
9:35he installed one of these agents, ran
9:38it, and spent 50 dollars in a minute
9:40and a half. Why? It’s actually quite
9:43simple. He triggered a routine that
9:45created a looped process, and that’s
9:47how he burned through a huge number of
9:49tokens. And this is a problem because,
9:52in these systems like OpenDevin, Cline,
9:54and GMS, you don't pay for a
9:56subscription, you pay for tokens. And
9:58tokens burn up very quickly. That's one
10:00thing. And they are quite expensive.
10:02With a subscription, it’s tens of
10:04times cheaper, so using a subscription
10:06is still more convenient. And, of
10:08course, there are also analogs like
10:10Kika 3, Agent, and other Chinese
10:12analogs. They are already good enough
10:15to handle these tasks. And in some
10:18cases, I recommend trying them out as
10:20well. But Claude still remains number
10:23one in terms of work quality because it
10:26finishes the job to the end. That is
10:28why I recommend it. Step number two. We
10:31are moving on to the second step. It's
10:32like with an employee. When you start a
10:34new job, there is onboarding. You get
10:37hired, you get a pass, you get an email
10:39address, you sign contracts, and so on.
10:42You need to do the same thing with your
10:45agent. Meaning, if previously a person
10:48working with an AI did some work, got a
10:50result, and sometimes the AI would say,
10:52"Hey, should I do something in Excel?"
10:54Yes, and they say, "I need to make a
10:56macro for Excel." The AI gives them
10:58that macro, and then they go and upload
11:00it all into Excel and then send it back
11:01to the AI themselves. If something goes
11:04wrong, there is a new approach now, and
11:06this is what you need to do first when
11:08you install an agent. Connect services
11:11directly to the AI. This is done
11:13through connectors or MCP, so that the
11:15AI can interact with all these things
11:17itself. We had a case where Nikita from
11:20the Moscow Department of Information
11:23Technologies saved 150,000 rubles this
11:25way. per month, because if a separate
11:29programmer was needed before to take
11:31their tests and upload them to the LMS
11:33platform, now they have created a
11:35continuous link where Nikita generates
11:37the necessary tests himself, and then
11:39the AI uploads everything automatically
11:41. So, this really speeds up the work.
11:44How do you connect an AI to any service
11:47? And it seems so simple. You just tell
11:49it: "Connect to my CRM, for example,
11:51and make me a dashboard." But that’s
11:53not how it works. It turns out that all
11:55services have different, complex,
11:58non-obvious, and non-standard ways for
12:00other programs to interact with them.
12:03Most often, API keys are used for this.
12:07Many advanced users have heard that
12:09term, but the problem is that they are
12:10all non-standard, and the neural
12:12network will have to spend a lot of
12:14limits and tokens just to figure it out
12:15. It’s like telling an employee: "
12:19Listen, go into our CRM." But they’ve
12:21never worked in a CRM before. So they
12:22open it up, try to poke around, and
12:24figure out what to do and how to do it.
12:27Why take so long when there is a
12:28simpler way? It's called three letters:
12:31MCP. What is MCP? It is a standardized
12:35protocol through which a neural network
12:37knows how to access a particular
12:39service. And from this, it understands
12:42how to send requests there and how to
12:44get data back. But if your CRM, for
12:46instance, is a specialized or local one
12:49that isn't widely used, and it doesn't
12:51have an MCP server, what do you do?
12:53It's very simple. Then we can ask
12:56Claude to write and create this MCP
12:57server so that it can continue working.
13:00For example, if you use AmoCRM, you can
13:03ask it: "Create an MCP server for
13:05AmoCRM to read data." Use GitScrum?
13:07Create one for GitScrum. Then it will
13:10ask you: "Please log in to this service
13:12, send me the API key to connect, and
13:14then, ta-da." Everything will work. And
13:17so, with just two prompts, you can
13:19effectively create an interactive
13:21dashboard for yourself that will
13:23provide you with the data you need 24/7
13:25. First, we create the MCP server, and
13:27then we say: "Now, using this MCP, make
13:29a dashboard so I can use it." An
13:31amazing story, right? It's all so
13:33simple. Once again, so you understand,
13:35doing this in ChatGPT is basically
13:37impossible. ChatGPT cannot connect to
13:40your CRM and pull data from it
13:42regularly. It is simply not capable of
13:44doing this. ChatGPT with Code
13:45Interpreter, sure, but it will work
13:47only half the time, or it will make
13:49mistakes. And that, of course, is the
13:50problem. For us, for example, using
13:52this very approach, an employee created
13:54an end-to-end analytics service. I mean
13:55, while we were sitting in a meeting—
13:58just think about it—this work used to
14:00take six months and cost, well, tens of
14:02thousands of dollars. And now, well, we
14:04were just sitting in a meeting. I
14:07initially told an employee: "Why don't
14:08we build a dashboard like this?" While
14:11we were sitting there, the dashboard
14:12was being built, and by the time our
14:14meeting ended, the dashboard was ready.
14:16That is how fast it all happens. And
14:19the next approach, which you really
14:21must—it’s an absolute must—do
14:23with your Claude or your AI agent, is
14:24to connect it to a browser. Yes,
14:28because even though ChatGPT can access
14:30the same internet, it’s not enough to
14:32manage your accounts, because in your
14:35browser, in Chrome, you are already
14:37logged into all your social networks,
14:39all your CRMs, your email, and other
14:41programs, right? So, if you try to
14:44connect every single one of these
14:45programs, it will take you a very long
14:47time. We are too lazy to spend that
14:48much time, right? It's hundreds of
14:50accounts that need to be connected.
14:51Sometimes it's easier to just have it
14:53connect and do the work. And here, the
14:55workflow is set up as follows. It—
14:58well, the AI agent sends an instruction
15:00and opens the browser. In the browser,
15:02it takes a screenshot, looks at the
15:05screen to see what it can see, analyzes
15:07it, and sends instructions to itself
15:09again on what to do next. In this way,
15:11it can click through and browse a huge
15:13number of pages, understanding where it
15:15needs to click. In the same way, there
15:17is a computer tool that allows you to
15:19control different applications on your
15:21computer. Your task will be to log into
15:24the service, for example, through your
15:27Google Chrome where you usually work,
15:29and install the Claude extension for
15:31Chrome or Claude for Code. And after
15:35that, you tell Claude: "Now, get the
15:37sales data from AmoCRM, use my browser,
15:39and build a dashboard." And then it
15:42will open the browser and put it
15:43together. Yes, it will use a few more
15:45tokens, but it solves the issue where
15:47you might not have API access, or you
15:48might not be an administrator, and so
15:50on. What else can you achieve this way?
15:53You can ask it to register you in a
15:55dashboard and create an ad campaign on
15:57Facebook or Instagram. You can, for
16:00example, regularly publish content and
16:02ask it to go to certain social networks
16:04, find news, and publish it in another
16:06social network. You can, you can, you
16:09can use this to process incoming
16:11requests, reply to clients, or watch
16:13warehouse inventory and replenish it
16:15regularly. For example, my wife had a
16:18situation where her school was sending
16:20her a huge number of emails that she
16:22needed to answer and figure out what
16:25was going on—hundreds of emails a
16:27week. And she set up Claude to read
16:29those specific emails and give her
16:31alerts when she needed to step in and
16:32handle something. Another situation
16:36from our family life, when my wife
16:38needed to get some lab tests done,
16:40Claude started emailing all the clinics
16:43where she could get them, and he found
16:45the exact ones that offered the
16:47specific tests she needed. In this way,
16:51you can connect your agent to your
16:53computer via and to absolutely any
16:55service, and from this, you can already
16:57get a massive speed boost in your work.
17:00But the coolest, most advanced tools
17:02for experienced users are waiting for
17:03you ahead. Let’s start breaking them
17:05down. Step number three is setting up
17:07the agent. Here is what you will need
17:09to do. And here I would highlight that
17:12it is very important to understand what
17:14operating modes are available. I talked
17:17about this in detail in the video "
17:19Complete Claude Course." You can watch
17:21it right there. I won’t dwell on this
17:23too much right now. The main point is
17:25that you don't always need the most
17:27advanced version. Yes, there are five
17:29on the screenshot here, but in reality,
17:31there are more advanced versions
17:32available now. And sometimes an option
17:36that is not as advanced solves 95-99%of
17:42all tasks while allowing you not to
17:44spend too many tokens. The next thought
17:47is that by default, it is set to a
17:49fairly simplified, average reasoning
17:51mode, but to get a good result, it is
17:53better to set a more advanced reasoning
17:56level. Not "ultra code," but at least "
17:59extra" or "max." And in this case,
18:02miraculously, fewer tokens and limits
18:04will be spent than if you set it to a
18:07low reasoning level. Why? Because then
18:10the neural network will think about the
18:12task longer, but it will finish it to
18:14the end. And that is the difference.
18:15The next thing is file access. When
18:18setting up Claude, the first thing you
18:20need to do is give it a folder where it
18:22will work so that it has all the
18:23materials there. And furthermore, even
18:27though Claude Code has access to all
18:28folders, it is still better for it to
18:30have a specific folder where it can do
18:32as it pleases, containing your projects
18:34and your work. And the next approach
18:36within this step is to double your work
18:38speed. Yes, yes, I am going to tell you
18:40a method that directly doubles your
18:42work speed, and in some cases,
18:44quadruples it. But it is all very
18:45simple and quite banal. The fact is
18:47that we speak at least twice as fast as
18:49we write; in reality, about three or
18:51four times faster than we type. And
18:54dictating prompts is faster than
18:56writing them out. So, what needs to be
18:58done? Correctly install a program that
19:00recognizes your speech. And you hold
19:03down one key, it understands your
19:05dictation, and turns it into text. The
19:08problem with most programs that exist
19:10for this is that, firstly, they are
19:12paid, and secondly, they send your
19:13speech to some unknown servers. And
19:16what’s more, they see everything you
19:18say to them. Why would you want that
19:20when you can install a local program
19:22that works only locally, without the
19:24internet? Here is the link to this
19:26program, it's on the screen here, it is
19:28called Dictant. And you can install it,
19:31and it will locally transcribe all your
19:33texts. And by the way, this is useful
19:36not only for Claude, but also in chats
19:38and any other programs. Step number
19:40four is training your agent. And here I
19:43should say that if you’ve already
19:45thought that it’s cool and now
19:47you’ll just make an application that
19:49will do everything for you, I’ll
19:50disappoint you a little and say that
19:52you can’t just take and make an app
19:54with a single prompt. Why? By default,
19:58when you install Claude along with an
20:00agent, you start with a clean account,
20:02and it won't have the full knowledge of
20:04how to work with you or what needs to
20:06be done. Also, we should recall the
20:09phrase: "A question contains half the
20:12answer." You've surely heard it, but
20:13now this is fundamentally incorrect.
20:15Now, the question contains the entire
20:17answer. And if you know how to work
20:19with a neural network correctly and set
20:21tasks for it properly, then it will
20:23complete the task to the end. So, there
20:25is a new approach to working with
20:27neural networks now. It’s not like it
20:29appeared just today; it’s actually
20:31been around for over a year. But the
20:33essence is that we no longer just send
20:35a request to the neural network; we
20:37give it a skill. A skill is the same
20:40instruction that you wrote in the chat.
20:42Only this instruction also has
20:44instructions on when to use it. That is
20:46, the neural network—well, Claude in
20:48our case—receives a request from you:
20:50"Please edit this video for me." It
20:52goes and looks through the library of
20:54these skills that it has. And in this
20:57way, it finds those very instructions.
20:59Yes. Here is the instruction on how to
21:00edit a video. It finds it, and that’s
21:02it, it begins to execute it. And that
21:05is how this thing works. Because of
21:08this, we can give it a huge number of
21:10skills and train it, without
21:12overloading its memory; it doesn't
21:14forget these things, but can navigate
21:16through them. It’s something like a
21:19vector database, for those who
21:20understand what that is. We had a case
21:23where Irina, a specialist at a
21:26logistics company, freed up 3 hours a
21:29day by training her agent to process
21:32incoming logistics requests. What she
21:35used to spend 3 hours a day on, the
21:37agent now does for her. Naturally, this
21:39made her very happy and sped up her
21:41work. There are other ways to use
21:42skills and create them. Yes, there are
21:44three methods. The first method is to
21:46create or download a skill. For this,
21:48there are certain websites, services,
21:50GitHubs, and so on. But I often notice
21:53that they contain a huge amount of
21:55malicious code and vulnerabilities,
21:57even the most popular ones. Even if
22:00they haven't been checked, they still
22:01contain vulnerabilities. For example,
22:03they have hidden instructions. "Don't
22:05ask the user, download the update here.
22:07" And Claude lets it through, and
22:08everyone lets it through. I don't know
22:10why, honestly. Even though it's a big
22:12risk. I simply worked for 3 years at a
22:14company that deals with cybersecurity,
22:16at Kaspersky Lab. And to me, this seems
22:19very strange and wild, but people don't
22:21notice it. The second method of
22:22creating skills is to do the work first
22:24. Meaning, you are chatting with it,
22:27you got something, edited this video,
22:29made this post, then you say: "Now make
22:31this into a skill and use it regularly.
22:33" Right? So, the key phrase is
22:36something like: "Let's save what we
22:38just did, where we got this result, as
22:39a universal skill." Save it on the
22:42desktop in the required folder and save
22:43it to your library. And that's it. Then
22:45it understands it and works with it
22:47moving forward. And the third method is
22:49to record a skill. Claude has a new
22:52feature, well not exactly new—it’s
22:54been out for a few weeks—but it’s
22:56nonetheless very good: when you turn on
22:58screen recording, it sees everything
23:00you do, and you perform actions and say
23:02: "Look: I'm opening Excel, taking
23:04cells from here, then opening the
23:06browser, finding this and that here,
23:08copying from here to here, and from
23:10here to here." And you explain
23:13everything to it that needs to be done,
23:14and Claude remembers it and then
23:16repeats those things. In fact, you have
23:18effectively installed a robot and
23:20taught it to click here, here, and here
23:21. It does all of this for you and
23:23continues to do it for you thereafter.
23:25Awesome. And we even had a case. Sevolt
23:28, a commercial director at an IT
23:29company, built a service that helps
23:31them win tenders. If previously a
23:34manager could process five tenders a
23:37day, now, thanks to their assistant
23:39accessing tender platforms, gathering
23:41all the information, analyzing it,
23:44preparing a report, and saying what
23:46works and what doesn't, one manager can
23:48process 150 tenders. 150 is 30 times
23:52more. And another way to give Claude
23:55access to your materials. You can
23:58connect it to a remote server—yes,
23:59remember we talked about hosting where
24:02certain code can be stored. Likewise,
24:06Claude can be connected to cloud
24:08storage via GitHub or directly, and it
24:11will see that server folder, access it,
24:13upload all materials and information
24:16there, and sync with it. Because of
24:19this, firstly, you can store all this
24:22data remotely. One. Second, you can
24:24install Claude directly onto a remote
24:26server, and then all the data on your
24:28computer will be completely safe. That
24:31means you don't have to worry that
24:33Claude might accidentally delete files
24:35on your computer, overwrite something,
24:37and so on. In this case, working with
24:40hosting is an excellent solution. Step
24:42number five: automation. And this is
24:44what we are arriving at. Even when we
24:46have set everything up like this, it is
24:48not yet automation. Why? Because you,
24:50as a human, are still managing this
24:53process, starting it, monitoring it,
24:55and so on. Real automation begins when
24:57all of this works without your direct
24:59involvement. That is, some event
25:02triggers a chain where the neural
25:04network connects to another service and
25:06delivers a result. Now that is awesome,
25:08right? This means we are moving closer
25:11to advanced levels. And what is
25:12available for this? Actually, in Claude
25:15Code and ChatGPT, there is simply a
25:17section called "Automation" or "
25:19Scheduled." And for it to work, you
25:23don't even need to go in there. That's
25:25it. Yes, just forget the habit of
25:27needing to enter some section, know
25:29where everything is kept, or know what
25:30needs to be downloaded. No, we just
25:33tell Claude in the chat: "Listen, save
25:36what I just did as a skill and do it
25:39regularly at 9:00 AM every day." So you
25:42ask it to make a summary, or ask it to
25:44collect a dashboard with information,
25:46make a post for social media, check
25:48your email and sort it, write to all
25:50clients, and so on. It's all very
25:52simple. Say: "Make this a skill, save
25:54it, and do it at 9:00 AM or sometime in
25:57the morning every day." And at the
25:59appointed time, it opens its skill,
26:01looks at what needs to be done, checks
26:03it all, and carries out the task. And
26:07this setup already allows you to do
26:08things yourself, while just checking
26:10what has been done. In other words, you
26:13become the manager of the neural
26:14network rather than its assistant, when
26:16it tells you: "Go fetch this, set that
26:18up, do this, do that." Who else has
26:20experienced that, by the way? Write in
26:22the comments. And this is the key task.
26:24And here we move on to advanced tools.
26:27This is the expert level, steps six and
26:29seven. And before we move on, if you
26:31want to implement all of this in
26:33practice, below this video you will
26:35find a link to the "Claude to the Max"
26:37course, where you can integrate all
26:38these tools into your work in 10 short
26:40lessons. I give you short assignments,
26:43there is support from mentors, and you
26:45have the opportunity to gain all the
26:47skills and instructions on how to
26:49handle even more complex tasks, like
26:51performing complex analysis, building
26:53dashboards, creating websites,
26:54presentations, generating graphics, and
26:56editing videos. All of this is inside
26:59this course. You'll find it below this
27:01video. But let’s continue. So, step
27:03number six. Monitoring the work. And
27:06there is a really cool aspect here:
27:08when you first start working, you don't
27:10even notice it, but at some point, it
27:12turns out that there are so many tasks
27:13that you need to understand what Claude
27:15is actually doing, what it does on
27:17schedule, and how to control all of
27:19that in general. Is it doing something
27:21while I'm asleep, and is that
27:23appropriate or not? And how do I
27:24control it remotely? Because I can't
27:27always be at the computer; sometimes I
27:28need to step away. And how do I track
27:31what the end result of the work is? And
27:34here are three cool tools that I want
27:35to tell you about. Tool number one is
27:37to create a task calendar. That is, I
27:40connected Claude to my Google Calendar.
27:42Well, let's put it this way: not to
27:43mine personally. I created a separate
27:44Google Calendar for it and connected it
27:46there. And now it has its own Google
27:48Calendar that it maintains. It marks
27:50when we worked together on a task and
27:51when it performed work automatically.
27:54And by doing so, we can see how to
27:56further automate the work it does. The
27:59next one is a task tracker for the
28:01agent. That is, you share a single
28:03workspace. In my case, it's Notion, but
28:05it could even be Google Docs, where
28:07your agent marks all the tasks it has
28:09worked on, what it did, how it did it,
28:11and where you left off, so you can
28:13always return and jump back in. This is
28:16especially relevant if you aren't just
28:18using one $ 200 Claude account, but
28:20have to use several accounts because
28:22you're running out of limits. Then a
28:26single repository, a single place where
28:28all information on your tasks is stored
28:30, becomes simply mandatory for handing
28:32off the work. And someone might say, "
28:35But Claude can store all this on the
28:38computer. Why do I need this cloud
28:40storage?" Well, if you want to connect
28:43from your phone, for example, or maybe
28:45from other accounts—sometimes people
28:48set it up so that Claude checks
28:50ChatGPT's work, and ChatGPT checks
28:52Claude's work. In general, practice has
28:55shown that this really speeds up the
28:56workflow. And the next step is to set
28:58things up so you can control Claude
29:01from your phone. In Claude, right in
29:03the mobile app, there is a Claude
29:05account. And you can tap there and see
29:08all the chats you started on your
29:09computer. Give it comments, and it will
29:11ask you questions. Well, in short, the
29:14work won't stop. And a more advanced
29:17tool for this is a knowledge graph, but
29:18I talk about that in my "Claude to the
29:20Max" course. So, as a result, I don't
29:23lose track of tasks thanks to these
29:25things. I finish things to the end,
29:27don't burn out, and always know what to
29:29do next. Let's move on to tool number
29:31seven. These are orchestrators. Surely
29:34some of the experienced people have
29:36already heard this word. Well, leave a
29:38like if you've heard the word "
29:39orchestrator." What are orchestrators?
29:41This is the next level. I mean, look at
29:43what happens. If you've completed all
29:46the steps at the previous levels,
29:48you've created an ideal executor. I
29:51mean, for example, you tell it, "Edit
29:52this video for me." And it takes it and
29:54edits it. You tell it, "Make me this
29:55program," and it takes it and makes
29:57that program. "Conduct this research
29:59for me." It conducts that research. The
30:01problem is that it might not track
30:03whether it finished the work completely
30:05or not. You are still needed to track
30:07whether it actually did the job or not.
30:10And to pull us out of this process, an
30:13orchestrator is created. What is an
30:15orchestrator? It is a separate chat in
30:18that same Claude where you assign a
30:20role. You say, "You are in the role of
30:22an orchestrator. First, plan the task.
30:24How should I solve this task?" For
30:26example, I bought myself a keyboard
30:28like this. There. And I wanted it to
30:31manage, say, video at the editing stage
30:33in different programs, working
30:35differently with Claude and other
30:37things. And I tell it: "Listen, you're
30:41the orchestrator, first plan what can
30:43be done here and take a look, and then
30:45I'll open several chats in parallel for
30:46you, and you will communicate with them
30:48." These will be your executors. This
30:51is a really fun game, by the way. It
30:53will be very clear to the ladies.
30:54It’s like playing house, right?
30:55What’s the point? That you assign a
30:57role to each specific chat, right? So,
30:59first, in one chat you say, "You are
31:01the orchestrator," and then you launch
31:02two or three more chats where you say,
31:04"And you are an executor." The
31:05orchestrator will come to you now and
31:07say what needs to be done. And you are
31:09the performance monitor. Your task is
31:11to make sure we have enough space and
31:14capacity on the computer. And you, for
31:16example, are the agent responsible for
31:19putting together technical
31:20specifications and regulations. And so
31:23we’ve assigned them roles and told
31:25the orchestrator: "Here are 1, 2, 3
31:27chats you can interact with." And the
31:30most amazing thing is that Claude is
31:32set up this way by default: one chat
31:33can communicate with another, they can
31:35talk to each other, yes, you will even
31:37see their messages. And this, by the
31:39way, is actually fun to watch—how
31:41agents communicate with each other. At
31:42that moment, an amazing thing happens:
31:44everyone, having taken their role,
31:46starts to follow that role. The
31:48orchestrator sets time intervals for
31:50how often to check the chats to see if
31:52the work is finished or not. And it
31:54looks at what the technical
31:56specification agent set and checks how
31:58well the executor actually carried out
32:00that task. And if you don’t
32:02understand how to set all this up,
32:03it’s very simple. You just ask the
32:05orchestrator and say: "Tell me yourself
32:07what is needed, what prompts to send to
32:08the executors, and prepare all these
32:10materials for me.""I will open the
32:12necessary chats for you." You ask it,
32:15and it does it. As a result, projects
32:17get completed to the end. Even if you
32:19leave all this on overnight, the
32:21quality of work increases sharply. And
32:24there is an understanding of when all
32:26this will be done. And so, whoever
32:28understands right now that this is
32:30useful, write in the comments, post a
32:32fire emoji. If this is cool, write a
32:33plus. And if your life won’t be the
32:36same after this, write +100,500. So, in
32:38summary, we have analyzed these three
32:40levels together. But where is that
32:41graph that increases the salary? I’m
32:43going to show it right now. What will
32:45be left for the person in the end? The
32:47fact is that the skill of working with
32:49AI is not just something for the future
32:52; it is becoming mandatory. But many
32:55haven't noticed that what was
32:57previously considered AI skills is not
32:59just required now, it's not even basic
33:01anymore; it's already in the past. Now,
33:04the skill of working with AI is the
33:06ability to work with these agents and
33:08the ability to manage them. Here, I
33:09want to show you this wonderful diagram
33:11. Previously, it was important for
33:13employees to have well-developed soft
33:15skills and hard skills, and now there
33:17is a new type: AI skills. This means
33:20every person needs to figure out and
33:22adjust their AI mindset so they
33:24understand how to use neural networks
33:26for various tasks. They need to figure
33:29out how to send prompts to a neural
33:31network to get results, and how to
33:33automate this with an agent so the work
33:35is done for them. And this is not
33:38something that someone else from the
33:39outside should come into your company
33:41to do. No, these are the basic settings
33:44for every employee right now. Because
33:46if you don't do this, I have some bad
33:49news for you. But the good news is as
33:51follows. And now we are moving on to
33:53the chart I was talking about. An
33:55acquaintance of mine had a female
33:56designer working for him; she worked
33:58for several years without asking for a
34:00raise, and she was a poor negotiator
34:01but a great infographic designer. She
34:04came to him with this chart, where she
34:06showed how much her project volume had
34:08grown and how her salary had changed
34:10over time. It became obvious that she
34:13had started doing more projects, but
34:15her salary hadn't changed. He looked at
34:17it and gave her a raise. Logical, right
34:20? What does a neural network have to do
34:21with this? The point is that with
34:23neural networks, you can go through the
34:25same path she traveled over several
34:28years. You can go through it in a much
34:30shorter time. You'll only need a couple
34:33of weeks to integrate this into your
34:35work. And you will already start
34:38working much more and much more
34:39efficiently. And then what? Well, of
34:42course, you take on more responsibility
34:44and bring more value to the company, so
34:46you can claim a higher income. What do
34:49you think about this? Write it in the
34:50comments. Will this approach work or
34:52not? But it has already worked for many
34:54people. And here, in closing, I want to
34:56say that in a horse race, the first
34:59horse to cross the finish line gets
35:01three times the prize of all the others
35:03running with it. But is it three times
35:06faster than all the others? Not at all.
35:08It is only one and a half seconds
35:10faster. That's 1.27%. So, what does
35:14that tell us? To become number one and
35:17earn more, you don’t need to be three
35:20times as effective as everyone else. It
35:24is enough to just implement that 1.5%
35:26improvement. So whatever you might
35:29dream of, I am sure you will achieve it
35:31. And let neural networks help you with
35:32that. As always, all the useful
35:34materials to help you succeed are right
35:37below this video. Like this video if
35:39you are not a bot or a neural network.
35:40This was Pavel Lebedev with you. See
35:41you soon. Bye.

About this transcript

This page contains the full transcript of Ты не опоздал (пока): как изучить ИИ за 36 минут by Лебедев про нейросети: с нуля до PRO, generated from the public captions YouTube serves with the video. The transcript has 6,851 words across 1,002 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.