Hi everyone! You're on the ToTheMoon channel. Tech insights from Silicon Valley. And today: Trump's executive order. Every new artificial intelligence model will be checked at the White House 30 days in advance. What will happen to the quality of AI if this keeps going? And which versions will end up reaching ordinary users? What is Anthropic really doing for the world, the real world? How every query can cost you 50 dollars, and why your work in ChatGPT 5 already costs 1,000 dollars a month, and for some, more than 100 thousand dollars a month. Why China no longer lets scientists leave the country, and what the Soviet Union has to do with it.
I think an episode ago we talked about the US introducing restrictions for all model operators, that they would have to clear their launches and releases with the government in advance. And now there's this story that Trump is pushing a bill so that absolutely all American models, 90 days before their release, give access to their update. Well, to the politicians, so they could check these models. That's the first big topic on the table. I don't know whether the US could really let something like that through, whether such a thing would work, because if it did work, it would be a huge blow overall. What does it even mean to hand a model over to the government 90 days before an update?
I mean, essentially sending the government, I think, not just the compiled program but practically the source code, right? Yes. So they can check how everything works, and only after that can it be released to the masses. A week ago we said this bill had been shelved, but in reality there's always some parallel story going on, because on June 2 the White House issued an executive order on advancing advanced... models in the field of artificial intelligence and security, and it describes a voluntary — importantly, voluntary — mechanism whereby you can voluntarily submit a model for review 30 days before release. But accordingly this raises a lot of questions on the market, with Sam Altman of OpenAI first taking the position that the launch of every strong model should not be turned into something requiring
prior government approval, which puzzles me a bit coming from OpenAI, because when Anthropic's conflict with the government started... We'll remember, yes, it started after the capture of the president of Venezuela. When it was said that during the capture of the president of Venezuela the security services, their equipment, were using Anthropic's models. And accordingly Anthropic said, we won't allow anything to be used this way. And then the entire top brass of the US came out and said that... So, Anthropic are some kind of villains, they should be banished from society altogether, practically banished from Silicon Valley, that's what they decided back then, I think.
And right away OpenAI, practically the very next weekend, signs a contract with the government, and Sam Altman, having signed the contract on Friday evening, writes on Monday, well, maybe we rushed it a bit there. And now he's saying that the data shouldn't be handed out in advance. Yes, I... the reason I started with this topic today is that it's very important. In what respect? That when it comes to the release of modern models, especially the most advanced models, I think we as users already don't get to see a lot of things. We used to see something good, right, now there are more and more restrictions. From what I hear, more and more restrictions for people... people hitting token limits, that whole story at Anthropic where people's access to tokens, the tokens ran out very quickly, resources ran out, compute ran out and so on. This example with the government shows that the system, of course, will be controlled more and more by the government, and the government will use these model systems for its own purposes. That is, they have essentially unlimited use, at least for the security of their own countries. For the US that's the US, for China that's China.
And we'll get less and less real access to models, and of course these restrictions are coming. They're clearly coming, yes, there will clearly be some kind of approvals. Maybe not for updates like an icon moving around in ChatGPT, right, but when it comes to the foundation model there will clearly be approvals. By the way, Ilnar, funny story, remember when we were filming the last episode? You wrote about the update. Was that last week or not?
Yes,
Yes, you wrote it.
It came out with an update,
that started rolling out.
but ChatGPT, by the way, never did come out with an update. So it came out, we've already talked about it, I think, or I mentioned in some mid-week videos that Claude 4.8 came out and there's Max, there's a million tokens. of tokens. But the most important thing, that they were promoting there, is that you can run simultaneously, in parallel, practically thousands of different processes that work with your repository and your system, doing different tasks in parallel. Honestly, I can say I haven't felt it at all. I mean, however much I tried asking that very system whether there's anything special about it, it tells me there's nothing special about it and then goes very seriously silent.
Among the important updates there's the dynamic workflow. That's most likely what you're talking about.
That's what I'm
When
talking about.
the model,
Of course, I'm talking about the dynamic
yes,
workflow. Yes.
yes, when the model itself decides which subtasks the task needs to be broken into, and along the way it can change the plan it drew up, and for each subtask it creates its own small prompt that sets the necessary context, and then agents, possibly small ones, possibly the model itself, other instances, a single one or something else, start working with these prompts. I mean, the example they like to give is: we need to rewrite a library from one language into another, a library, that's the thing all programmers love to talk about: we have some library, it's written in one language, and now we need to write it in another trendy, hip language.
The task is clear on the whole, it's clear how to test the end result, that is, it should work exactly the same as it did before, only with, maybe, a different speed, a different infrastructure or something else. And so here the system first analyzes what needs to be done, how one language differs from the other, then breaks it into small subtasks. For example, that all files with such-and-such an extension need to be processed in such-and-such a way, right, and it gives specific instructions. Like, what to do with these files? And with other files we do other things. And then, well...
In effect, a loop is launched that goes through all files of that type and calls an agent for them with instructions that the model itself wrote. And here, you know, it's as if we're going up one more level. If before you yourself, like an operator — there used to be computer operators, now it's chatbot operators — you say that with these particular parts of your project you need to do these particular things, you run some prompt, get the result, put it together with the other parts, do something else. Roughly speaking, let's take a more, well, not more, the opposite, a less programmer-ish example: we need to do some analytical research and put it together as a presentation, for example. So accordingly you have one part where you collect the data, a second part where you analyze the data, a third part where you put together a presentation in which these conclusions need to be shaped up to present to someone. And for each part you have your own prompt, your own request, or an agent run, or something like that.
And this is what a person used to do themselves, on their own. But when the dynamic workflow appears, it's as if an orchestrator model appears that does this itself, breaks it into subtasks, builds its own prompt for each subtask, and with these instructions the jobs are then carried out by other parts, they get checked, and only after that. Then we move on. Tokens just fly out like a river.
And here the question is, actually, I don't know, at least within my own tests or the systems I'm building, not even tests, I haven't fully seen that exclusive thing yet. I mean, the direction itself is very good, we've discussed it many times, right, when you're not solving one single task, but when the system essentially... everything is in your context, that is, it understands what else is going on around it. So when you tell it to dig a hole to plant, say, a flower, it also understands there are other flowers around, right. Or when I make some decision in my life, it understands... Remember how it suggested I spend a whole weekend: going to the theater three times
To recover,
or some kind of recovery,
yes.
that is, without understanding that there's this whole other huge life of mine, a life, and not just some area of one specific isolated person, and it did that for me following some technique. So that's probably the direction of movement. Personally, I don't yet see any major shift in how the system works. Obviously the system still has a lot of errors, a lot of glitches, a lot of different details. And listen, I still haven't hit the limit in any system, even though I think I've got Claude running 24/7 now, both Claude and — you weren't here last time, we were just discussing it — Codex. And I still haven't hit it. Once I got some notification that I was approaching the limits. Maybe that's the American market, maybe. But I had bought
Yes, and actually
Well, well.
with ChatGPT it was strange, because in my version, as far as I understand, there were like 20x limits that week, right, and I basically couldn't hit them at all, no way. I even started turning off fast mode at night and, well, I kind of got scared. And then the next day I realized it was some glitch in the system, because I allowed it to keep spending my money on code, and it didn't spend a single ruble. But I had the follow... A classic story, I wonder who else has had this. We'll move on now. I realized that this whole token topic is being raised a lot. Obviously, for our audience, maybe, for many this isn't interesting at all.
But actually this topic is very important. We've told you many times that a time will come when access to proper, serious models will be terribly expensive, or essentially won't exist. Altman said there would be a situation where you could throw in 20 thousand dollars to solve a task. Now he's said there will be a situation where you'd have to throw in a billion dollars to solve a task. Billions. And look, here's a very interesting thing. The discussion now going on is this.
This: that it's not just a question of billions of dollars, and in their new document, maybe we'll do a special episode about it, maybe we'll cover some part of it today, they published, OpenAI put out a sort of description of the future of artificial intelligence, how it will be regulated within the lives of people and countries and everything else. Right now the top-tier version costs 200 dollars, practically across all the systems, but that's essentially free. You know, a lot of people keep telling me, Volchek, you don't live, you have a different attitude to money, for us 20 dollars makes a difference. But I'm not saying this from that position. I... I have close friends and acquaintances who buy bottles of wine for half a thousand dollars and stay in hotels for thousands of dollars, yet are too stingy to buy the 200-dollar-a-month version, which is... I mean, for them going to a restaurant and spending 200 dollars on lunch is nothing, but they don't buy it. It's their attitude, because obviously they don't understand, didn't understand what it is, right. So here's an interesting thing about tokens. What I saw: when I use the API, for example, for some post-processing, when I need to, I don't know, post-process an article, or a video, or audio, or some analytical report, I saw that all the systems are set up to not tell you at all how your tokens are being spent. And I had a situation the day before yesterday when analyzing one text
cost 50 dollars in tokens. I said once again to everyone that my work in ChatGPT costs not thousands of dollars but already about 100 thousand dollars a month. 100 thousand, no less. I mean, if I ran myself on tokens, I think it's not even 100 anymore, I don't know how much it would cost. Probably more. And there's a reason Jensen Huang has that line, that if you have a programmer working for you on a 500 thousand dollar salary, he should be spending no less than 500 thousand dollars on tokens. And there's a reason the head of Google said that we're interested in saving developers tokens and in being in this system at all. Here's what bothered me, Ilnar, maybe you'll understand this, that I spent a really big chunk of time on several of my projects, because I'm now building various very extensive infrastructures, including on my own, to keep the system from spending money, from spending money needlessly.
It's not that I'm against spending money, on the contrary, I urge everyone to use and test the latest models, so everyone keeps up with the times, so people don't use ChatGPT-4.0 for analysis, or like Zoom sends me summaries, and they don't match the conversation, it completely misrepresents everything I... it just lies. Why? Because Zoom uses a very cheap model for transcription. If Zoom used ChatGPT 5 Pro to transcribe your call, then at minimum an hour of call transcription and post-processing would cost 5 dollars. Of course that's unrealistic, the system doesn't work like that, nothing would survive. You just raised the token topic too, and I think it is incredibly relevant, yes, incredibly, and I also don't understand how the context windows fill up, I mean the million-token context window in Claude, it seems to me, just fills up,
for me personally it just fills up, a task is started and it's already taken up, and that bothers me all the time, I don't understand how that slider works, it's constantly full, then not full, full, then not full, then it's accumulated, I know there are auto-resets, but in ChatGPT, when I'm working in a big context window... Noise. That, by the way, is also one big minus, that the system doesn't tell you what your real context window is. And if they compress it, they don't tell you how they compressed it.
And honestly, my ChatGPT lately, viewers, please share, it's started glitching a lot in the sense that I have a huge number of chats, and I solve different tasks in different chats and projects. And it started merging information between different independent projects, between ones that must not be merged. And before, you could still sort of manage that. I remember six months ago I told you the case where I asked it a question, my mom asked me to ask about the garden, about plants.
I told GPT that I'm not asking the question for myself, I'm asking for my mom's garden, so it wouldn't mix up the locations, the climate, the details and so on. And it answered specifically for her. But here I can't explain endlessly every day that this project is about this, and that project is about that. And when you tell it "act strictly like this", it doesn't always do it. Claude 4.8 Max, easily, when you tell it "do an audit and write it up in the repository", it'll easily write into the repository. Then you ask it, "why did you write into the repository?" It says, "Sorry, I made a mistake." I don't know, do you have that, is it a problem for you or not?
A question for the viewers, for you personally, yes, but I think it's a big problem, a really serious problem for all future models and the robots to come, yes, and all the models.
But at the moment, especially if you use it very often and a lot in different areas, I'd probably just consider separate accounts. Especially if you're somewhere near the edge of the tokens available, you can just set one up, like, you know, some people buy different laptops for different jobs. And you just do some tasks from one laptop and others from the other. Well yes, that's the price of a solution like that. A black laptop for work, a silver one for personal life.
I
And
am not sure,
here's the thing, so that...
first of all, that it's legal, I'm not sure it's legal.
Oh, come on.
I don't recommend that, but that's for you
Just
to go
in
and do.
That's hard, too.
Keep in mind, I condemn it.
That's hard, too.
Wonderful,
No,
that's my thing.
it's a bit, you know. That's the case when you probably need to do a lot of audits and so on, independent ones. Yes, but for someone like me, who often works in parallel with the same context windows, and in general I need the memory, I'm still for overpaying so that I have one unified system. I understand what you mean, that if you're solving independent tasks, the same way I have a separate ChatGPT and Claude running. By the way, I've become a pretty heavy Anthropic user over the last two weeks.
Yes, and I don't know how much of the limits I've used up and all, but I haven't uploaded my ChatGPT history there, no. Although I talked about the case, literally in one of the special videos I mentioned that Maestra had a very cool case about ChatGPT, about downloading and exporting chats. Check it out, I talked about it in that video. But still I didn't upload it there, I mean I decided not to do it for some reason. For now. Maybe I'll do it in the future. But for now I decided not to. What I want to move on to, again on the question of different technologies, different countries.
I had a laugh this morning. A friend of mine was here, and about three weeks ago we were discussing robots, and he said Russia is way ahead in robots, that MIPT made a really cool robot that gets around. It has a super system of proper, specific legs. But I expressed some indignation and said that can't be, that they made something, given their lack of access to... cool models, cool infrastructure, chips, everything. He kind of even got on my case a bit, and I even thought, maybe I'm wrong. I even did some analysis online, studied a bit, looked, and indeed, they were promoting their technology as very good at standing on its feet. But today I saw a video from the St. Petersburg Economic Forum of MIPT robots playing football.
Yes, yes, indeed, indeed, indeed, indeed, absolutely.
And what was the video? The video is just, in Silicon Valley, the difference, and they showed Boston Dynamics. Boston Dynamics, let me remind you, that's Hyundai, they have a robot, and they have a whole project of teaching this robot to play football. And we'll post these videos, what Boston Dynamics playing football looks like and what an MIPT robot playing football looks like. This isn't to say that someone is bad or good. It's more about understanding the adequacy of the models and the difference between these models.
So I think what we're seeing now in artificial intelligence, especially when people use the free version, or versions of some open source models, or even the paid versions of entry-level models, ChatGPT Plus, right, that's roughly like this MIPT robot playing football. What the government is discussing in the US, in terms of, or the Chinese government, well, the Chinese government will simply never issue an executive order for anyone, asking permission to check models in advance, right?
They've been checking these models in advance for 15 years already. So this story is very important for you to understand the difference between models. I mean, this robot example is in no way... There are lots of smart people everywhere, including at MIPT, that's not the question here. The question here is about understanding reality. Not about arguing, about a real understanding of reality, of what this technology is. You're talking about quotas now; I was at a meeting recently, and a guy said you can code via the API.
And use different systems via the API, whatever, different models. Or connect them and code. I thought, how does he want to connect Codex and Claude Code to the API, because the whole point of these systems is that the system positions itself as a system that works with a highly branched architecture, you code inside it, you do everything entirely inside it, and in particular we can see the GitHub update from June 1: they stopped charging a single monthly rate, before you paid and coded with AI as much as you wanted, and now you have to pay for...
tokens, per actual use. I feel that, Ilnar, in Claude Code and in Codex, once they merge them. They will merge them, ChatGPT, OpenAI is creating a whole separate division, big Greg is supposed to run it, a division where the API, ChatGPT and Codex will be merged into a single system. We all still get confused about ChatGPT's models, but I think that'll never end, for the next 5-10 years we'll keep getting confused about them. But Codex has 5.5 High, and ChatGPT has 5.5 Pro. And you're like, and of course the difference is huge. I want to tell you that 5.5 Pro is a completely different model.
By the way, before we move on from Codex, I wanted to ask you, Sasha, and you too, Tanya, and viewers, write in the comments as well. It somehow happened that in recent months I've been using Codex more and more than the browser version of ChatGPT. Somehow I open the browser version less and less. Have you also moved to Codex, or to the apps, or are you still on the browser version?
Yes, well, I tried it once, so far I haven't
Was that long ago,
managed it.
Tanya? Try it now. Really, try it now, try doing something on the computer now. You need an idea. You need some idea, so it's not just a search, but so that, for example, there's some kind of system that does something, that reminds you of something every hour, or builds some system for you, or saves data. What's the big plus of Codex? ChatGPT's context window overflows. and data gets lost, it starts doing things in fragments. But in Codex you can follow the folders of your projects very precisely, strictly, and you can give it a task directly. All your projects are stored on the computer, you can give Codex a task: analyze all of my projects completely, put together complete analytical reports on them, organize all my files completely, and before organizing, tell me where the problems are, propose an architecture and storage, and organize it. And it'll do it, and propose, and you'll love this task. That, by the way, is your task, I want to say right away.
To say there's a whole lot of everything, that's my task. Not only that, it can even analyze additional images, find all sorts of discrepancies, sort all the projects by individual houses, or by some categories, or find patterns in the categories, where you had expensive, inexpensive, I don't know, projects, projects where this style or that style was used, clients of this or that kind, big or small and so on. You've actually had an incredible pile of projects, and you also had a lot of old projects. You're a well-known designer, actually, with features and
Yes.
in well-known magazines even, and everywhere, everywhere, everywhere. So it's a huge amount of stuff, and it'll help you gather it all. That's your first task; ChatGPT, alas, won't do it. It'll lose something, misplace something somewhere, get buried in some other task and so on. That's clearly your task, and for a lot of people it's probably a really cool task too, for everything. Ilnar, answering your question: I haven't moved to Codex at all, except that in Codex I started building things I'd never done before.
Yes, I started creating big infrastructure projects, big ones, really serious tasks. Let me put it this way, honestly serious, because even the people I meet often say I'm doing startups. Well, every third person, when asked what they do. I understand that what I'm doing is already bigger than any startup, I just don't call it a startup. For me it's just some infrastructure project. Why do I do it this way, Ilnar? Because with Codex there are two things. First. I do understand that if I start moving all my ChatGPT requests into Codex, I'll have no limits left and I'll start paying for tokens. Yes, that's the first thing.
The second thing is how inconvenient data storage is there. And Codex... First of all, about 15 times in Codex a chat of mine disappeared. A fresh chat just vanished. My old chats get wiped, all the old chats get wiped. I don't know, there's probably some setting, I haven't got to it yet. Maybe I haven't been using it long enough, but my old chats disappear. Sometimes new chats disappear too. And that really bothers me a lot. Secondly, their menu is very badly made, and this structure. These confusing folders.
And in Claude you can work with several GitHub repositories, but in Codex you can't. A prime example, right? Sometimes I'm amazed and I don't understand, guys, why don't you just make a nice storage system, menu, conveniences, details right away? You position yourselves as systems for programming. Not for people like me who made something on their own, right? But for serious developers of everything. It all looks like kindergarten, to be honest. Just like the ChatGPT 5 menu, roughly, which is impossible to make sense of, millions of these chats, and it's impossible to find anything at all.
But still. The limits in ChatGPT aren't restricted, and plus there's the 5.5 Pro version. And even 5.5 Thinking, which by the way is still called Extended Thinking, my menu jumps around roughly every three days, it changes the name, and still these versions, I think, are way better than 5.5 High. Because Codex is, after all, a slightly different system. The way it works, I think, is you essentially go in, it's like 5.5 Pro did everything for you itself and handed you the result, whereas with Codex you go through the whole grind together, through the process of solving the task. And that's why I keep on, I keep on building the system. I don't call it vibe coding, I actually think vibe coding is an outdated word, because vibe coding is a word for working with code. When a person works in Codex and in Claude building a system, that's not vibe coding. I'm deeply convinced of this, convinced that this word will disappear and there'll be some
some other one. I like the story of creating, what I talk about in my videos, AI-native systems, and the fact that this system is created directly by artificial intelligence, that's not vibe coding. Because I don't do code at all, I've never once opened the code, I've never uploaded anything anywhere, I don't commit anything, I haven't done anything at all. I always position it as: use Codex and Claude without working with the source at all. And even that tab, the line running at the bottom, that says, like, changed, I don't even understand what that is, I mean, I don't even want to figure it out.
Lines of code. Lines of code
Yes, I,
that it either wrote or deleted.
technically I'm originally a programmer, so of course, when I don't understand something, I dig into it in detail all the way, but I don't open it, I'm not interested, because making sense of code that Codex or Claude wrote, first of all, is nonsense, why would I make sense of it, this code is written by AI for AI. and they're for me, and plus that's not vibe, that's not vibe coding. I mean, vibe coding, I think, is when, Ilnar, you started coding with the help of artificial intelligence, you have some things you code yourself, and some of it is vibe coding. And for some reason vibe coding came to mean that some guy sits down who's a total idiot at programming, who doesn't understand anything at all, and he's the one doing vibe coding.
Well, that's strange
Like, he gets
really.
Look, let's put it this way. OK, this term really is outdated. Karpathy coined it long ago, and it's more of a meme
Yes,
now
I think
than a specific
it's a term for programmers.
definition.
I think vibe coding is now done by real programmers. So, again, again, look, my, my point is why? Because you write part of the code yourself and deal with the code yourself, and that's it. And part of the code you write with the help of some other systems. and going further, somewhere the system gets merged, somewhere it's pure vibe coding, and somewhere it's purely your code, for example. Could be. But when I create, again, it's clear, or it's a play on words, it's
Yes
like a hierarchy.
yes, yes, everyone understands it their own way. For example, in my circle vibe coding more often refers to some analyst who can't write the code, or a product manager
Well of course,
or a project manager,
how would a real programmer
who...
say they're vibe coding?
who just sent a request, it spat something out, and he's like, I did it, and like, look, I've got
Although
something
look,
somewhere
usually
that works. No,
an analyst or a person who knows SQL. I remember, in the early days of programming, when I was, when I started programming,
An SQL analyst
we all, we
has to, yes.
all ourselves, we all wrote complex queries ourselves, complex joins, unions, a ton of stuff, all the tables, and we wrote these queries ourselves, did it. Well, new systems just appeared, and there are just people who aren't programmers, and and somewhere some query, but it's just, these are different approaches, so here, I think, it's very important not to confuse them. I just often hear all kinds of girls and boys saying they vibe code, and that's probably, Ilnar, maybe what it is. But they're just playing around. It's like putting up a landing page in a CMS, in a CMS system. For example, in Tilda, or earlier in Joomla, or in WordPress, putting up a site. Are you vibe coding there or writing a site? Or what's actually going on?
I remember the old Joomla, for example. Roughly like that. So that's that, right, or earlier making something on WordPress. You always had to be a programmer to put something up. Yes, it's just that later some modules appeared. So it depends on what kind of programmer you are, what you program, how you program, in what system you program, which pieces of code you program, right? It depends on what kind of programmer you are.
Yes, but someday, I hope, Karpathy will appear on our podcast, and we'll ask him to settle this term that he once came up with and tweeted.
Yes, I hope he'll appear when there are only AI systems, when there's no such thing as code written by a human by hand.
That's true. About what you're saying, you started talking about startups, remember, I think a year or a year and a half ago people started saying that soon there'd be a unicorn with one employee, where one person creates a company and it becomes a unicorn. And it seems we're, maybe a bit slower than expected back then, but still moving in that direction, where a person who can create some business doesn't run into the bottleneck of technical specops... specialists, you know.
Yes,
Now the comments will pile in that models can't write enterprise code, there'll be lots of bugs, everything of yours will leak, and so on, but at least an MVP can definitely be built now.
Wait, but the startups that write code in YC and create
they're
a company,
all now
they
all, yes
they
just
make real enterprise
generated, yes.
stuff, yes, they make real enterprise code.
Yes, yes, yes, yes, yes, and they raise decent money at the same time. And about what you also said, writing code via the API, there's a joke on that topic, there's a meme. Programmers, when they're doing some project, send so-called pull requests to GitHub. What does that mean? The piece of code you wrote, you send it to your repository, and other people look at it, and if a few people have looked at it and said you wrote it fine, nothing will break, then you add it to the main project. So a meme appeared: you made a pull request with ChatGPT, sent it, then Claude Code added some comments of its own, then Gemini redid it all, then more, and so a human isn't involved at all, this poor little piece of code just gets sent to different systems, gets worked over from all sides, and then it all gets sent off.
I
then
now
copy
do specs like that
that.
often.
Right.
gives you some edits, you throw it back and say, this needs fixing. That's roughly how the tossing back and forth goes from one place
But
to
there was
another.
actually a study on this, that having one model check another is a mistake, and you can't trust it. Tanya and I talked about that in the last episode.
Ilnar, tell us about employees not being allowed to leave China. I think that's a pretty serious story in this modern year of '26.
Yes, several months ago, maybe even six months ago, they introduced a restriction on employees of state companies leaving the country. That is, if it's not a private company, a state company, say some researcher or someone else, and they plan to travel abroad, they can only leave with written permission. You write an application saying I'd like to be abroad for some time, then I'll definitely come back, your request is reviewed and you either get permission or you don't. So for specialists in industries especially important to the country, there's this procedure. But it only applied to employees of state companies, research institutes, some large companies, and so on.
And literally a few weeks ago this was extended to commercial companies, companies that are supposedly independent, but if some important special... works inside them, an important specialist, they may not be allowed to leave the country either. Older people may well get flashbacks to the Soviet Union here, when you could only leave by being on a list. So we're moving in that direction, which once again confirms that people, scientists, specialists are hugely important for entire countries not to fall behind in one industry or another, not just companies, it's already a question...
already a strategic question for countries. And China, for example, is closing its borders from the inside so that especially important specialists don't end up abroad and then get headhunted there, or anything like that happening.
Well, wait, there was also that story when they sold, what was that company called, the one that made the hype, that Chinese company that made the hype that they'd made the coolest agents in the world, because nobody had heard about agents anywhere.
Ma-something, Manus?
Manus, yes, Manus. And that was the story when they sold to Meta, they moved the assets to Singapore and sold to Meta, and they weren't let out. And now, by the way, there's the question of how to return that money, they want to do some kind of restructuring so Meta can get it back and so on, to sort it out. They don't understand how to do it, because as far as I understood, they didn't leave and won't leave. Yes, and they won't. We really live in a very difficult time in terms of...
Yes, I think right now, in terms of the movement of people, the laws on people's movement in general, government restrictions, you just compared it to the Soviet Union, I think what's happening in the world now, the Soviet Union just doesn't hold a candle to it, in every country, I think, and it's in every country, in every country there's some kind of chaos with these restrictions, with documents, with registrations, with companies, with everything, I mean, it seemed we were heading toward a world, The world was heading toward a world that's open, right, reliable, broad in what gets adopted and so on. And now there's just a colossal number of restrictions everywhere, and some people they don't touch. If you're unlucky and you end up in the wrong element, the wrong time, the wrong place, it just gets surreal, really, really everywhere, really everywhere.
Well, that's essentially what you started with, the order that Trump signed, and initially, as we discussed, the wording was completely different, that companies must be obliged to provide their models for testing, and only after that a decision is made. But in the end Silicon Valley managed to, well, push through, or sell, or edit in its own version, the order became voluntary, and the order itself even says, of course, that it can't be interpreted as licensing of models and that it's done on a voluntary basis. But they strongly recommend providing access via the API, and the government creates its own benchmarks by which it will determine how valuable a given asset, a given model is.
And only after that, and 30 days in advance, as you already said, you can let the government use it. This is all... A lot of this actually grows out of Claude Mythos, the Mythos one, and Project Glasswing, when they were able to discover a huge number of critical vulnerabilities in programs and systems, and Anthropic then initially gave access to 50 companies, and now there was news that they're expanding to 200 companies that get to use this system to look for vulnerabilities in their own systems.
And here the government, apparently, also makes it onto that list and 30 days before release will be able to... can use the systems, apply them at home, possibly close the holes that come up because of this, and only after that will the models be released to the public. If they're released at all, of course.
And next week, literally in a few days, on Tuesday or Wednesday, we'll release an episode about professions. Watch it, about which future professions will be in demand and where you can make money on that. But the point is, Ilnar, when I read this news, I thought, could they really have built the functionality they wanted to make a year or a year and a half ago, when there was a rumor on the market that ChatGPT was releasing a killer for the recruiting market, for vacancies, for resumes, for everything, a global competitor in the American market to LinkedIn, Indeed, in short, all these job portals? And now they've released it, I started looking into what they ended up making, and it says they supposedly made some special feature that can now search for vacancies and resumes across their sites. I thought, how backward is everything going on, even in OpenAI's PR department, just so backward. I think all of them, these people, and they earn hundreds of thousands of dollars there, should just be swept out with a broom out of OpenAI, I'd chase them out with a broom, such backwardness. Never mind that your regular ChatGPT, especially the Pro version and the Plus version, can do really great search. And GPT,
I think, works better than Claude, in particular at going through various resources online and checking those resources. And it can easily go through a specific resource, find vacancies, resumes and everything you need. That's the first thing. The second thing. You can now write yourself in Codex... a system that will crawl all day, or once a day, and it won't spend tokens. You can write a system that will run independently as a service for you, go and search through all the job portals for what matches, for example, your resume.
It's an absolutely workable thing. In ChatGPT you can prepare it as a technical spec for the project, upload your data there, about you, about your resume. And then ask it to write code for Codex, which Codex will then spin up, where the system will match whether a vacancy suits you or not, and it'll go through all of it fully.
Yes, and moreover, Sash, you can sync it with your calendar, mark the slots when you can do an interview, and ask it to schedule interviews for you
Well,
in those
if
slots
needed,
with
yes,
the companies
if they respond,
that fit, yes.
yes. You mean if they also respond to you, you can also read the emails related to it, in a specific folder in your mail or by a specific subject, yes. Yes, right there on your computer, just an ordinary computer, nothing extraordinary needed at all, and it'll run on its own, separately, and again you don't need tokens for it, tokens to write it, but you just need Codex, the basic version, I think a regular subscription is enough to write something like that. It'll also do the syncing, you can also connect a bot, WhatsApp or Telegram, that will notify you about it, and remind you, and ask your permission whether to put this or that date in the calendar. Obviously that's a system of a more complex order, but even the ordinary thing, ChatGPT search, works for vacancies and resumes, and the release of news like this is astonishing, of course. Obviously, there'll be a lot of videos on YouTube saying ChatGPT is entering this market, how it's all going to work, some very strange fake, and I ask you to pay attention to this, because it's a very peculiar thing, it's unclear why it's even needed. Needed.
Here, Sash, another interesting thing, you were talking about how soon we'll start paying for tokens. Yes, it's interesting when that will happen. And in my view, one of the moments when it might happen is after the API, after the API, the API.
Separate.
I've got total professional deformation already: after these companies' IPOs. So, for example, well, you understand professions much better here, but it's obvious, in my view, that for OpenAI, market share matters a great deal, so accordingly these models are handed out for free, essentially for free. When you get a huge number of tokens for a subscription, and accordingly your number of users keeps growing, but the company doesn't start squeezing money out of its customers, doesn't move them to tokens and so on.
Possibly after the IPO, when the shares are sold, when it becomes necessary, on the contrary, to earn, to earn money rather than grab market share, the situation may change, and at that moment, instead of subsidizing users, we'll move to another option, where accordingly you'll have to pay a lot of money to use these models. And the same with Anthropic; not Google, because Google isn't going public, but at least it's in this race.
80 billion, exactly 80.
80 billion, and that's a serious sum, considering that it's in the same race, because essentially SpaceX, including xAI, Anthropic and OpenAI, when they go public, of course, they'll be raising... a large amount of money gets raised. So why go public? Why go public at all? And that's
Very good
one thing:
question, yes.
you go public so that the valuation isn't a trillion, but so that at some first point the valuation becomes 10 trillion. That's the first thing, yes. And someone recently wrote somewhere in the comments: "Alexander, you're wrong, Anthropic is already worth more than OpenAI." Just read the news on how the deals are going. Obviously OpenAI is worth more than Anthropic. It's just a question of reading the news: here they raised at this valuation. A crisis at that valuation. You can't buy any of these companies right now even at a 2 trillion dollar valuation.
You can't, no one will let you. Even at 3, at 4, at 5 they won't let you. They won't let anyone in at all, they won't let you in. They'll let you in for a certain piece once the IPO has passed, you'll really be able to buy some shares. Whether you should buy them or not, that'll be your choice. And of course these companies will strive to be worth everything, and they will be worth more than 10 trillion dollars. The question, well, the skeptics might strangle me, or someone else will say something, but OK. So look, here's the situation: on one hand, you need to do the IPO to get a higher valuation, that's how they get money. When you do a public offering, you get a lot of close scrutiny and far more requirements than now, and whether to do it or not is a big question. Why else is it needed? Why else is it needed, other than to get a big valuation? It's a peculiar topic. So they should all already be worth more than two trillion. When you say they'll go public and their task will be to show profit, this is where the market differs a lot.
Because if, for example, I buy their stock and see that their purchasing power and the size of their audience are falling, and I personally believe that whoever has the bigger audience wins, that is, whoever sits on a billion-plus audience wins. It doesn't matter how they do it. This doesn't mean that Anthropic now, for example,
has less than a billion audience, unlike OpenAI or Google. Anthropic, I believe, is betting not on... not on coding. I believe they're betting on making an artificial intelligence that will create all the software the whole world will use. And honestly, I only understood this in the last month. Before that I didn't see it that way. But now I see that it's not a bet on coding. It's not a bet on programmers at all. And that's why they have such a valuation, because if I start creating AI-native software with Claude Code, then people will use that software. That's the point: people will use not Anthropic and not OpenAI, but the software they created, their system created. If Claude Code created a new CRM system used by 100 million people, or wrote a new messenger used by a billion people, then that's Anthropic, yes, and it's like Microsoft.
It becomes their story. I mean, how many people use Microsoft? Billions. Why? Because there's Windows, there's Word, there's Excel, that's it. What difference does it make how many use it? Tomorrow they... all these tools of theirs now have AI built in. And as long as people use these solutions, people will use the AI inside these Microsoft solutions. This is very important. And at some point I migrate, like I said, I don't use Excel, because I don't understand why I'd use it now. ChatGPT creates my Excel, ChatGPT analyzes my Excel. I stopped using it. That's it, it practically doesn't exist anymore. I have PDF, I have HTML, there's the display inside already. And Codex, and ChatGPT, Anthropic, they all build.
They build tables, and you're like, why do I need that? There are external platforms that display everything for me. And this is where Microsoft starts losing millions and tens of millions of people, then a billion, just like search engine optimization. We recently had an episode about the new search, where I said search won't exist at all, it won't exist, there won't be search engines, there'll be something else. They may have the same names, like Google, but they won't be search engines, and here it's the same as with Excel. It doesn't mean Microsoft will disappear, but Excel just won't exist in its current form. Accordingly, if users during the IPO, after the IPO, fall, Ilnar, and the number of people falls for the sake of making a profit, that is, if tomorrow they cut off the tokens. Look, Codex has only 5 million MAU, 5 million monthly active users.
Guys, what is 5 million? Even Tanya is among them, she tried it. Tanya tried it. But I'm in no way, about Tanya. Tanya is an incredibly smart, broad, multifaceted person. But she's not about programming. We made her try this Codex, right? And now she'll try it again, and put a project together. My sister tried it. What kind of programmers are they? My sister, Tanya, right? There are tens of millions of technical specialists in the world, engineers, people in analytics, in finance, accountants, people working in sales. These are people who desperately need these systems. They haven't even...
And they haven't even tried them, never mind being active users. Active users. And, Ilnar, imagine they make tokens paid in Codex. Well, how many users will there be? You and me, and the two of us, right? That's roughly how it'll be. I'm waiting, by the way. And I'd be very glad, I'll tell you honestly, here's my statement. I'd be very glad if ChatGPT Pro started costing a thousand dollars tomorrow.
Why?
I'd be glad.
What for? Why? I mean, what's the motivation?
My motivation is
Yes,
why... that
you
so that
want it to be like that?
then I'll know I'll have the wider access that OpenAI is hiding. Because for 200 dollars... It's not that I... I mean, I... You all know I'm always for people, for people to... I made this channel for that, for people to grow. I have other channels just about people, about a person's personal growth, about their spiritual growth. But for a person to grow here too, together with artificial intelligence, keeping up with the times, just moving along with artificial intelligence. But I understand that these systems will restrict access to the models. I'd definitely like to have that access.
And the more expensive the access, if there's an expensive plan, they'll try to stuff something into that plan. Or... remember when ChatGPT 5 Pro appeared, at first they threw in everything they had. It just seemed like there was nothing there, right? So only enthusiasts bought it. But now it's a completely different story. Now those in the know, those who sit inside it, understand the difference. Well, for one, the tokens don't run out, for example, right? And everyone around is shouting that they're running out. And you're like, what are you doing on your computer? You can't be pressing that many keys.
You couldn't have. What, 40 parallel sessions open in Codex so you ate up the tokens? What are you even doing? And accordingly, the IPO, this whole thing, the IPO, it's a very peculiar story. They'll have to grow even more. They'll have to, with these... a hundred billion dollars or 200 that they'll raise. I think they'll all raise more than a hundred, maybe more than 200, maybe more than 300. They'll have to, with that money... But in theory, if there's an offering and the valuation is 2-plus trillion, they'll all raise somewhere around 300-plus. They'll probably have to give up 15 percent, a range of 10 to 15. So they'll have to spend that money. Or on... And what do they spend on? It'll go to compute, to people.
They'll spend it on the audience, they'll spend it to reach every single person and really work. Look, here's what a cool situation it is now for you and me, and for Tanya, for our viewers. There's Claude and there's Codex. But there's Gemini too, granted, still coming with its huge infrastructure. And now xAI has arrived with code too. Now everyone's going to fight, but Claude and Codex are direct competitors, very, very close ones, because Gemini is a slightly different story, xAI is a different story. They'll be fighting precisely in terms of competition over tokens.
And their task will be to engage the person so their tokens don't run out. Because when you say, Ilnar, since Claude Code has more tokens, I'll give it the architecture to work out, what did I do? Why did I do it? I read online that Claude, Opus, is better at doing architecture. I started analyzing all this. By the way, I still don't have an answer. I think it's a marketing move. Because judging
Even
by
by the
how
benchmarks.
ChatGPT, Codex, Claude Code and Opus, and Claude, Claude itself, react. And they often spout such nonsense, and at the same time trash each other, then don't trash each other, then do one thing, then another. I'm not sure whose architecture is better, but in general, to get an independent assessment, you know how you can get an independent assessment? Just open a new ChatGPT, a new window, upload the archive there and ask it the question again. And for some reason it starts telling you something completely different, yes, totally different, that, it turns out, these plants...
they die when you... I recently bought a vacuum that automatically cleans the bottom of my pool. And I ask it, I say: my pool isn't level, and it says: that's no problem, it'll all fit, it'll all fit. I bought it, threw it in the water, and it died on me right away on that level pool. I took it out, thinking, how do I return it, I've already put it in the pool. Well, in America you return everything within 30 days, right. Recently a guy returned a TV to Costco that he'd bought 23 years ago, a big one, bought for a thousand, probably, 10-15 dollars. And I go back to ChatGPT and tell it, it didn't work after all. It says, well yes, I underestimated one thing there.
I made a mistake, it says. And you're like, well, and we all understand what fundamental mistakes there are in people's health and so on. It's like doctors making mistakes. It's like when my dad died, I call the doctor and say, my father died, he had a stroke, and the doctor says, well yes, it happens. And it was a clear mistake by that group of people I was calling. I was present at that meeting when they made the decision to operate on my father. I'm by no means shifting the blame onto them, but they acted unprofessionally. They said we need to do a new examination, because at the other hospital they don't know what they're doing.
And at the other hospital they told me, absolutely do not do a new examination, because you'll disturb the blood vessels and the person will die, which is exactly what happened. But the doctor didn't react. For him it's just a kitten. Or a rat? Or, as Max Grigoriev says about us as guinea pigs in ChatGPT, right? That we're test subjects. So, Ilnar, that's my view on the IPO. Well, we'll see. We'll see what happens. I'd recommend to everyone, maybe we'll do a separate episode or some segment, or we'll add things during the week.
For those who follow this and are even interested in investing: with projects like these, always look not at the IPO but at who they'll be spending the money on. That's my very real, honest recommendation. Believe me, everyone will be trying to get into this IPO, but you need to get into those who will be building all this for them. And... They'll have to spend the money on someone, yes. And even if we look at the latest contract between xAI and Anthropic, it's only a billion dollars a month. And when they spend 300 billion there or... That's the big story, especially since they're already earning decent revenue. Anthropic, by the way, showed some slice of profit, decent in some respects. I don't know how reliable that profit is.
Well, they showed some, some slice of profit, they really showed it, a slice of profit. But that's probably not what to look at right now.
But shares, that's... that's air. Or... imagine, you're sitting at home, and you need to sell not a suit or a table, but some piece of paper, you know, just some piece of paper. They sold some piece of paper, they sold shares, shares that a huge number of people hold, employees and investors, and everyone got diluted. Many got diluted, yes. There are shares that get diluted. Well, you sold it, and there's a certain problem in that, right? When you sell a big chunk of shares, the effectiveness of each share goes down, so the shares themselves get cheaper. So they can't do that. And Google... Google, obviously, is very... There used to be a problem, look, a very well-known problem, Apple's cash problem.
Apple, 15 years ago, already had around 100 billion dollars in cash. Companies back then, by the way, weren't worth trillions of dollars. And they had no idea what to do with that cash. That problem got solved in Silicon Valley. Everyone says there are overvaluations here. I was talking with a friend yesterday, he's a financier in the investment sector, and so on. But he doesn't like these markets. He keeps saying they're all overvalued, they're not real, and so on. But that's the whole point: before, their cash just sat there, they did nothing with it. Now where are they spending it? On compute for artificial intelligence.
They spend it on capacity. They buy servers, boards, memory, flash storage, electricity, in short, the entire infrastructure, and invest in it. By the way, an interesting project, a very cool project, I think, a very good way to wrap up. There's a startup, they're doing it here now, together with NVIDIA. I don't know whether it'll happen or not.
You'll be able to get, at home, for 150 dollars a month, actually for free, equipment that costs 200 thousand dollars; they install special computing equipment at your place, they connect to your electricity, pay all your monthly bills, as far as I understand, no matter how much money you spend, and my electricity bill is roughly one to one and a half thousand dollars a month, and you'll pay, they'll use your internet. You'll pay a fixed 150 dollars a month, and they'll pay for all the rest for you, everything entirely, all the electricity, and you'll have this station running like servers, rather. By the way, that seemed like a really cool idea to me, because in America right now one of the big problems is the construction of all these plants, they pop up across from people's homes, many states are banning them by decree. As for these spaces appearing at all, obviously these spaces don't bring jobs, they bring money, but they ruin the environment, they ruin the infrastructure, they're noisy. And there are a lot of cases in America where people come out to protest, because they're noisy.
And here they do pay attention to what bothers people; there are countries where they don't pay attention, but here they do. So, for what bothers people, maybe this will be the solution. It's a 200 thousand dollar installation.
It's a bit like, you know, about a year ago, I think, the Liberman brothers announced a project, Gonka, I think it's called, to do distributed computing, that you don't need to build a data center, like with crypto, so that it's all computed on different hardware and then all comes together. And here it turns out, well, regardless of how I feel about the project I just mentioned, when NVIDIA is doing
That's what you should
something
have said.
similar. It's
When people ask me, I'll have to tell you how it's done.
Yes, yes, yes. So. But here they themselves install their own hardware at people's homes,
My politician
yes, so not on
will never
your computer.
talk like that.
They don't do it on your hardware, they scatter their own hardware across the whole country like this and then collect from there. So. But it's cool.
Yes, they do go further there
Just a parallel like that.
and talk about a kind of distributed artificial intelligence in terms of decision-making, details, all of that is still not... These are slightly different things, I think, Ilnar, because all I was saying is there are big plants being built, and I'm just talking about a small plant being built, and I think it's not new, like in crypto, they didn't invent it. No, I think the biggest example of that was blockchain, when back in the early days a friend of mine, actually a mutual friend of ours, set up on his balcony in Moscow, set up servers, bought graphics cards, and they mined bitcoin for him back when you could still make money on it, so it's probably a similar concept, but they're doing it, you see, together with NVIDIA, I don't know what the economics are, that's the question.
A question of economics: how much easier is it for you to buy and set up a big plant, or set up this thing. And I think in this day and age, I'd say the project would be pretty risky right now, because... the thing is, everything becomes obsolete and changes very quickly, and you need to redistribute all this infrastructure fast. It was working on one thing, and you need to change it. And so big plants are easy to re... re... rebuild, and repurpose. This, I think, is one of the most fundamental problems there are: repurposing this. So personally I wouldn't invest in this project, because it looks very cool and neat, but I think some new modern boards or new code may appear that will be more efficient in some way, and it'll get applied. That doesn't mean these plants will die, the big infrastructure that Elon Musk, for example, is building, or these gigawatt ones, giga, mega, tera, or Zuckerberg, who's building near New York, or the size of New York, or rather Manhattan, right? So I don't think they'll disappear. Tanya, next time you'll definitely tell us how Codex set everything up for you, structured it, sorted everyone out, counted all your paintings, tables, chairs, lamps, kids, cats, dogs, fish, whales, plants, everything else. The idea, by the way, is very cool. The main thing, don't overthink it, tell it you're not doing anything, so it does all of it. And when you've spent, when you've spent all the tokens and all the money on your credit card, then you come back.
Then you'll have to deal with my husband.
Yes, then you'll... No, no, it wasn't me... Yes, say the wizard advised it. The ToTheMoon channel advised it. See you in this show exactly one week from now, and almost every two days we've got new episodes coming out. Watch us, don't forget. Bye, everyone.