OpenAI has temporarily stopped selling the $200 Pro plan to new users. Demand for Astra turned out to be so high that the company doesn't have enough capacity. And almost at the same time Dario Amodei says: “The development of the most powerful models needs to be slowed down, and independent checks allowed.” That statement comes straight from Anthropic. And right away Trump says practically the opposite: that artificial intelligence has to keep being developed and the fears are greatly exaggerated.
And at the same moment Anthropic shows how Claude is already being used in real cyberattacks. So what's going on? Do we already lack the capacity for strong artificial intelligence? Or are the models themselves starting to move faster than we understand what to actually do with them? Today we'll go through these topics with you. Hi everyone! You're on the ToTheMoon channel: tech news and insights from Silicon Valley for the whole world. We're kicking off our weekly podcast. So then.
OpenAI, by the way, published how many users they have: a billion users, more than a billion users, and 40 geographic regions. And what's interesting is this figure: 70 million requests per second. To me, 70 million requests per second sounds very serious, right? But let's move on to what news? That OpenAI is blocking new users from signing up for the Pro version. And I like that when I asked ChatGPT about it, about this news, it told me there are two possible explanations. The first one is the load on their computing servers.
And the load is because of the second point: there's very strong demand for the Astra model, and supposedly everyone wants to use Astra. And clearly people want to sign up at $200. But I have a sense of a third option, one that you and I, Ilnar, have discussed many times. That on the whole — I got scared by this story, by the way. At first I thought they'd killed the Pro plan. I'm thinking: “What? How can that be? How?” I keep waiting for a new plan, a bigger one, you know? And here they are closing the $200 plan. I'm not sure they'll bring it back. We'll see, right?
I mean, if they bring it back, it means either someone made a mistake, maybe in marketing or pricing policy, in sales, or they really, really do have some problems with load, which I strongly doubt — in particular, that so many people around the world rushed to buy the $200 version to get Astra if they were, say, on the regular Plus. Because usually, when people are on Plus and want a more powerful version, they buy Pro for $100, not $200, right? And how does the $200 one differ from the $100 one, by the way? I think it has four times more tokens than the $100 one, right?
Plus the $200 version, well, in my experience, is the version that gets the newest, trendiest releases of all kinds of features, data and everything else. Although it's unclear, because people often say “the Pro version has it”, and at some point they made a Pro version for $100. I think that kind of took the shine off this luxury list of $200 people who signed up at the very dawn, right? I even got a pen for that once. By the way, we have super news today! We have super news today! My ChatGPT pen arrived. I forgot to mention it. I forgot. Tanya, ChatGPT sent me a pen. So, here's this little thing.
You said it comes with a stand?
Yes, it's heavy.
Have you checked it, Sasha? Maybe it's not a stand.
Maybe it is a stand. And the thing is, this stand — Yes. You could kill someone with it. It's heavier than a golf ball, really a lot heavier, maybe two or three times. Well, that's the pen. So, I'll be back home the day after tomorrow. And by the way, in a few weeks — even less, yes, in ten days — OpenAI Dev Day for developers is taking place in San Francisco. It's their main conference, and Sam Altman opens it. I'll be there. And that's it. The main thing is that I'm making it in time for everything.
By the way, Ilnar, a few days ago I was invited to a conference, another one. I was like: “Eh!” I'm, like, not at the conference. So. What to choose, right? Yes, yes. What to choose? But it's good that it worked out and I'll make it to Dev Day. I've never been to Dev Day. It's the third year in a row it's being held. I'm very glad I got an invitation this year and have the chance to be there. So, what do you think about this? About the Pro versions, because for many people who are watching, once again, the principle — at least my favorite one on ToTheMoon — is that if you want to really get to grips with artificial intelligence, first of all, buy the paid version.
And ideally, if you have — well, if you're ready to invest — buy Pro. People still treat buying paid versions, even for $20, like some subscription — I don't know, a health app, a Whoop band, an Oura Ring, or a Netflix subscription. But this is something completely different. This is your whole life, your professional development, access to — well, having your very own artificial intelligence. On the other hand, we've discussed many times, and I think Ilnar and I agree, that access to the top models will become less and less available. So.
And it will be limited, and it will be very… because it's very expensive, right? The Pro subscription essentially devours OpenAI's capacity, because you'd never in your life be able to use GPT Astra through the API for that long for $200, right? I mean, it's unrealistic. I think you'd burn through that $200 in about a couple of minutes. So, Ilnar, what can you tell us?
Listen, it's a real pity Tanya couldn't join us today, because you always asked her whether she'd subscribed to Pro or not. I think she subscribed at some point and then, I think, went back. It's just that now, apparently, if you believe Thibault's tweet, you can't do that anymore, right? So we can no longer recommend the $200 subscription to Tanya. You can recommend the $100 one, but the $200 one won't work, at least for now. Yes, and here's where I learned the news. On X, formerly Twitter, there's Thibault, one of OpenAI's employees, who usually announces the Codex limit resets.
There's even a separate site, which we've mentioned, that tracks all these resets he usually announces around various events. And he's the one who wrote that because of the huge load that poured in after Astra, they decided not to sell new $200 subscriptions. But he also wrote that this shouldn't affect existing accounts. So if you have the $200 subscription and you actually need it, it makes sense not to change anything in it, and to make sure there's money on your card so that the subscription can always renew, because if it lapses, getting back onto it doesn't seem very possible right now, at least not in the near future.
As for whether the influx is related to Astra or not, that's hard for me to say. Obviously they're saying the model is so good that crowds of people came running. Remember, we once discussed how my friends and I choose what to use — say, Claude, Codex, Gemini, Grok or something else. Besides the quality of the model itself, there are many other factors, and one of them is limits. And on OpenAI's $200 subscription the limits are really very generous. For example, on Claude my limits run out much faster if I'm solving a simi—
Yes.
—lar task. So. Whereas at OpenAI they're really big. And limit resets happen fairly often too. That's exactly what Thibault's tweets are about — they come out once a week, once every two weeks. They reset our limits at just the right moments. And that's why many people who use ChatGPT quite actively — well, not ChatGPT, who use AI bots quite actively — also choose ChatGPT, OpenAI, in particular. The $200 subscription gives very big limits, and you try to squeeze them out completely, because if you've already switched to it, what's the point of not using it, right?
Because, as I said, limits are one of the important factors in the choice, and so you use them to the max. I'd also like to add something about buying expensive subscriptions and so on. It's like the saying about cars, right? A car is not a luxury but a means of transportation. It's the same story here. And if ChatGPT helps you in your professional work — and by now it's hard to imagine a field where it doesn't help — then either way, if it helps, that money pays for itself pretty well.
That is, you either do things faster, or do more, or start doing things you didn't do before. And these subscriptions, $100 or $200, can generally be covered by that already. You can probably find a use for these chatbots in many fields. By the way, if you happen to have counterexamples, as they like to say in mathematics — some example of an industry where a chatbot would be completely useless — you can write it in the comments.
Yes.
It'll be interesting to see what people come up with. I'm sure you can come up with something like that, of course, but what's interesting here—
Well, not come up with — tell us their own.
A field for imagination.
Our viewers already—
Yes, yes, yes.
Let the viewers share real examples of their own, where they say: I don't see any use for it here. Or, on the contrary, maybe someone will suggest something, because I had a friend with me — I've mentioned him a few times — he was traveling around here. Misha, yes, Ilnar, you know him well. And I told him that… Well, he asked me a lot about artificial intelligence, then he started asking various questions. And I told him: “Listen, to really understand it, you shouldn't be asking me questions, you should start using it every day in your everyday life, constantly, in every aspect, learn to use it, and use it in your professional life. And then it's clear.
Watch ToTheMoon episodes and others, ask me questions, read up on things, experiment. But if you don't use it constantly, every day, and look for new use cases for yourself…” Remember those first few years of ours, right? How long has this been going on now — two and a half or three years, right? Such really active use of GPT models. Remember how big this topic was. Or three and a half years already, right? How much—
Well, we've been recording the podcast every week for almost two and a half years.
Yes, and GPT almost three and a half.
And how much time before that, right?
And GPT almost three and a half, yes. At least since early 2023 it really took off — since March, since spring. And just in general, if you look at how it was when these things, new for you, started appearing, as if you were developing some new organ — you had no idea you needed to think that way, that such a thing could be done, that such a process of development, of creating something, existed. And of course, only now are people starting to immerse themselves in it. And by the way, this statement about a billion users isn't exactly new, is it?
I think I already said a month ago that they had a billion users, because there were various figures from them. And now it's their official statement. The previous one, I think, was 900 million people. They haven't gotten that much further. I mean, if you take a whole year, they didn't get much further in a year. But I think what has grown a lot, grown hugely, is the time those users spend inside, their level of engagement inside, and so on. Listen, there was an interesting study. I want to add it, and then you'll continue from there. There was a very interesting topic, that if…
First, I made an episode a few days ago — go watch it — on “Is it worth using different plugins in different systems?” I really don't like middlemen, all these intermediary layers, systems that resell ChatGPT or Anthropic, or that say: “Our CMS has the trendiest AI,” and so on. Because there you never get access to top models like Astra or Fable. But listen, the interesting topic was that there was a study on medical data. And here's what happened: if you gave the system your test results, say, and some data about your health in advance, the system answered you well.
But if you told the system: “Collect information from me about my health,” it answered poorly. And this topic, by the way, really… I got interested in this setup, because I essentially advocate and keep saying that the new era of artificial intelligence will really begin when these systems finally start collecting information from the person. At the same time, the work itself and the quality of this model's work, its ability to collect data from a person, and the person's ability to answer, to match things up — because that's also a problem, I think, with people's answers.
So. This is still a fundamental problem. And maybe that's why, by the way, models still don't collect information all that much. Although I think this is the newest channel for that kind of learning. We'll see what happens — so as not to forget to talk about this topic at all.
In short, you think we won't see that version again?
Listen, I think we will. I think it's a marketing story. I mean, they'll wrap it in a marketing story. Maybe they really can't meet all the demand. I don't know whether you already have this or not. In Codex, after the latest updates, the option to approve on my behalf in some non-critical cases disappeared for me. Yes, remember we once discussed that Codex has three levels: one where you don't allow it anything without your consent, and every step has to be confirmed. The other extreme was that everything is allowed and nothing needs to be asked. It does whatever it wants. And in the middle there was a mode where it decides by itself whether it needs to ask or not.
In English it's called Approve for me only. It asks only about actions that pose a potential threat, essentially.
Right. So yesterday evening I updated Codex. Today I don't have it. I don't know where they put it. So. But this could also be a way to cut the number of tokens used a bit. Because in this setup, when it approves on your behalf, it has to analyze the decision, try to understand it. Additional requests come up. Maybe that's how they cut back too. Or it's just gone missing for me.
By the way, Anthropic raised them. By the way, Anthropic raised its limits — they announced it literally a few days ago. They had a promotion before, where they raised the limits for the summer. They sort of removed the summer ones but added new ones. And there seems to be a difference in the overall amount. Again, this is for paid plans. So of course not for free plans — for paid plans.
And what do you think, was there a switch? Because judging by the internet and YouTube there was, yes, and by the news too. Astra is being talked about a lot, right? Although I have big doubts, but never mind. Our viewers will write in here and split into two camps again. You think there was a switch?
I can well believe it. I'm sure the majority — the overwhelming majority — don't use just one system but several, to keep their finger on the pulse. Given all this hype around the Millennium Prize problem and the benchmarks, like ARC-AGI, which was simply torn apart. Generally everyone wants to try it, to give it a spin. And I'm sure that most users of both Gemini and Claude, maybe without switching for good, definitely started trying it and comparing it with GPT-6. I think all the active users have tried it. In that sense, I can see such an outflow happening.
It may be temporary — the quality may turn out not to be satisfying, and then people will come back. Or the raised limits Anthropic is offering now may tip the balance toward Claude. Claude really is very good. Nobody disputes that. But an outflow like that, with this kind of news hook and hype, could well have happened. And I allow that it did. And accordingly, continuing this story, I'm sure that if they don't bring back the $200 plan, some other plan will appear anyway. They will do it. Either they'll time it to, I don't know, the release of some hardware, or they'll bring some data center online, or something like that.
Around some story like that they'll either bring the plan back or invent a new one. I think something like that will happen. Not right away, I don't think. Uh-huh.
It's unclear, yes. Admit it's unclear, because, remember, I thought — and I saw it in conversations somewhere inside the tech company and OpenAI — that an expensive plan was supposed to appear, right? The expensive plan never appea— Even Sam Altman has often said, like, you'll be able to spend a certain amount of money on solving a problem, for example. I thought that's where this was heading. For example, that you could click a button and some capacity would be allocated to you, because it's not about buying tokens through the API, right?
It's a different kind of interaction, because API tokens are a different story after all. First, you pay as you go — you don't say in advance that you have a certain budget and it optimizes for that budget. So it's something completely different. What's more, in fact nobody really understands how they're consumed. I mean, you can calculate anything you want as much as you want. But first, models spend these tokens differently depending on the context you gave them. And what's more, the more powerful the model, supposedly, the fewer it should spend.
And so on. Right? So calculating it, I think, was impossible before and still is. I, by the way, still… Ilnar, have you checked how many tokens you've spent? Me, I'm getting close to 50 billion. I've spent 50 billion tokens. That's all in Codex, of course, in Codex, right? Not in Claude Code. It's always interesting to watch, of course. What else do I want to say about OpenAI? An important thing. This information actually reflects what's happening very accurately. Once again, watch our episode from a few days ago — it was on Friday — about plugins. What I filmed is a very useful framework.
Everyone should understand which plugins to connect, which prompts, which instructions, whether to take them from the internet at all, which third-party services to use, and so on. Because people keep believing that if a third-party service, say a health one, is tuned for health and has artificial intelligence, then that AI — even if it's OpenAI — is somehow better. I tell everyone: it's not better, it's definitely worse. Right? Because through the API in an app you won't run GPT-6 Astra Pro, for example, right? Or Extra High.
You just won't be able to run it. So, look, there's a very interesting new thing that has appeared. And by the way, on this topic, about Pro, there's hardly any information anywhere. And here's the news: ChatGPT no longer chooses deep reasoning on behalf of Plus and Pro users.
It's very… When I saw it — you saw it too, right? When I saw it, at first I was like: “What do you mean, chooses? Was it choosing before?” I mean, first of all, we're all confused about who chooses which mode where. Right? Because, first, in ChatGPT there's chat, there's work, and there's the separate one. Then you launch Codex, with its own different settings, different tokens and so on. Every so often you don't understand how it resets these bubbles — resets them or not, shows them or not.
It's very hard to figure out. But look, here's a very important point: in the latest versions that came out, OpenAI had — essentially, OpenAI has, for example, a fast-answers mode. Right? I don't recommend using the fast-answers mode. It's a way for OpenAI to save money. Money. So if you want to save OpenAI money, go ahead. I always use, at least within my subscriptions, the Extra High or Pro mode. But if you don't have the Pro version, you can just use the very latest mode, because people think that the latest mode doesn't answer quickly.
But in reality, when I ask through Extra High and the question is simple, it answers me right away, instantly. And if it thinks longer — this is in Extra High, not in Pro, mind you, not in Pro — if it thinks longer, that means your question needs time to be worked through. I've checked this many times. For example, yesterday I bought something and needed to check the composition of some herbal teas — whether they're allowed, well, which country they can be brought into. And I was in Extra High — I think I was even in Extra High, yes, I was chatting in Extra High. It answered me very quickly.
As soon as I gave it an image where it had to recognize all the ingredients, absolutely all of them, it went into reasoning and took quite a while to answer. So, look: before — not now — if you'd chosen the fast-answers mode, the model could switch when you wrote to it: “Think long.” That's a well-known trick. Everyone says: write “think long”, “think more”, “study more”. So, look, there's a whole bunch of things. It switched to a more serious model by itself. Now that won't happen anymore. The system has removed these things. We don't know why. Maybe it's related to GPT-6 Astra.
What Ilnar was talking about, in particular, that Astra doesn't work like the standard previous models. Right? And by the way, there's a phenomenal thing about GPT-6 Astra that everyone should remember.
OpenAI actually released an official guide. Watch my Friday episode on plugins. There's a very important thing in it: how you should write prompts for Astra, and how much prompts for Astra differ from the way you wrote them before and the way you write them now. So note that the switching won't happen anymore. And I think, Ilnar, this may be some serious shift in their models. Or they found some bugs or problems somewhere, including in terms of computing capacity. Right? Because there was a study showing that the agents.md file for coders can lead to worse results and higher token consumption.
I mean, a big file. So on the one hand it can supposedly improve development, and on the other, with today's models, with better things, it can make it worse. I want to ask you a question, Ilnar. I noticed — I also raised this in the Friday episode, and we'll continue going through it together. But those are different topics. Watch Friday's about plugins.
Last Friday's. So, you know what I noticed? That it has become more costly for me to open a new context window, a new chat in Codex or in Claude Code. In both, actually. I noticed that every time I open a new context window, I have to set the system up, set it up every time, even though it has these agents files. Everything's laid out, everything's explained, there's documentation and so on. And still every time I have to set it up. And I had a very interesting case. I had a problem I couldn't solve. You've heard — I talked about it in many episodes. I solved it the day before yesterday, yes, I solved it the day before yesterday. And GPT-6 Astra didn't solve it for me.
I switched to Fable 5.1. Fable 5.1 didn't finish it. So, while working on it over these last days, at some point, when I'd had enough — I got angry at Astra — in Fable I say to Fable: “Maybe the context window here is already too long?” The context window was already more than a month old. It was a total mess by then. In this task. I said: “Maybe I should move somewhere else?” And I'd tried moving, both within Fable and to other places, and it didn't get solved. And I kept coming back to this context window, because at least something was moving there, somehow moving, right? And…
And in other places it was just carnage, a total mess. And I wrote: “Maybe I should move?” And Fable stopped me and says: “There's no need to move to another chat. I'm compressing everything here, everything's great here.” And as far as I remember, it wasn't like that before. Right? So the model has changed, changed in its characteristics. Just like you're saying now — Codex removed this and only asks about complex things. That's actually a hell of a change, right? And this change may be related, on the one hand, to the model spending tokens. On the other hand, it may be related to Codex wanting access to the computer.
And yesterday GPT-6 Astra, while solving a fairly simple task for me — setting up a timer on a remote Codex. I have a computer back home in the Valley, right? And there's Codex on it, running a certain task. I asked it to set a timer for me there, and it says: “To do this, I need full disk access.” And GPT-6 Astra told me that in extra high mode. OK, not in ultra mode, but it told me that in extra high mode. To which I said something like: “Why do you need full access?” “Oh, sorry, I guess I don't need it.” So maybe OpenAI does want to roll out access more widely — not to steal your information.
Don't think that's what I mean, not at all — but so that the more access it has, the more opportunities it will have to offer you various things or build some, maybe more complex, computer-bound setups. What do you think about that?
Uh-huh. Look, as for context compression, moving from one window to another and so on. It's very similar to what happened before with, I don't know, prompt engineering, right? When people taught each other how to write prompts to get a better result. At the time it obviously worked, right? If you wrote the prompt correctly, you had a better chance of getting a good answer. What happened next? Models kept developing, and now it doesn't matter much how you write a prompt — the model will generally figure it out. So the difference isn't that striking.
Same thing with the various orchestrators that appeared, which let you connect additional links to other things and so on. Now all of that is packed inside, over MCP. Before, it was important that you moved the context yourself, told the new window yourself what it needed to know and what not. Now it's not so important. The models keep developing. Context compression works quite well. If we're in the same task, there's probably not much point in opening a new window. If it's a different task but within the same project, then it probably makes sense to do it in a new window. You know what this is like?
A road roller is moving, but it moves slowly, and sometimes you can run ahead of it. And so some ideas appear that are a little ahead of this road roller, but it catches up anyway. And with agents.md — what you're saying, that old agents.md files that are too long can make things worse with the new Astra, the new settings and so on. I'm sure that's what happens, simply because many things that were tried out by enthusiasts, companies and others just start getting built further inside.
Into, I don't know, ChatGPT, Claude or whatever else, and there's no longer any need to do it by hand. And we get to the point where if you do it by hand, you only start making things worse. It's not a given that you can compress the context better than the built-in system tuned in ChatGPT, in Claude or elsewhere. If, for example — a year ago this was definitely the case — you know the task well yourself, you've spent a long time on it, and so you can give the new window the right inputs, and the wrong ones, the mistakes you had there, erase them, not carry them over into the new chat, and the model would work better — now context compression already works very well.
In many ways it all comes down to this: use the out-of-the-box settings. Yes, at some point you'll get a slightly better result if you use, I don't know, some extra skills or something, but only if you really know what you're doing. Otherwise, in my opinion, it's better to use things out of the box, because in a month everything that was at the cutting edge for some advanced geeks will already be built into the system. Harnesses and everything else you need — it'll all already be there.
Yes, and you can lose — you can lose the habit of using modern models, of learning to work with them directly, of engaging with them directly, being with them.
People often come at us, you know, saying: “You didn't talk about the model that came out in South Africa. It's so important, so important to look at it.” But sure, go ahead, you can use any models, various open models and so on, for your personal tasks. But on a mass scale, globally, if we're talking even about your own use of a main model as a person — obviously, for example, even in my various businesses, in various partnerships, in different teams, completely different models are used, right?
And there are local models that get installed, open source, all sorts of things. A whole lot. But basically, speaking for yourself, you still choose, first of all, some main foundational model for your own tasks. Like for me it's still ChatGPT, and I think for you too, right? ChatGPT remains for my main tasks. And at the same time, for example, I use both Codex and Claude Code very seriously. I think even Claude Code, for the last two months — I think it started back in mid-July. With Fable, right? I use Claude Code more.
And by the way, I even have weeks when I don't use up all my tokens in Codex, and it doesn't bother me much, not too much. Yes, so literally yesterday I — it's just that I'm traveling a lot right now, and I can't say I have the time to sit with the models all the time. A lot of time right now, these weeks while I'm on the road — and yesterday I launched several tasks in GPT-6 Astra, and again, well, the result for me was very so-so. We'll see. There's a fairly big project I handed over yesterday.
We'll see how well this project — seriously, I haven't checked yet — has been done as of today, and whether the model really managed to do it truly independently of me. But the choice of system, I think, is very important, right? Which system you work with. And this whole thing of launching AI bots through Microsoft Excel, right, through Word, or asking AI bots in Google Drive or in Gmail — that's, of course, for very basic users. Not even basic users, really — people who don't know how to work with a computer. In principle, that's exactly how it looks.
It's made for people who don't work with computers and who have some kind of help with the computer. Of course it's great to have good AI bots, really well set up in companies, for customer support. Yes, when you write to support now, or can quickly find your way around the documentation, or get a quick answer. But, by the way, I can't say there are any revolutionary changes on this front in the world. Some companies have done it pretty well and nicely, but most companies are way behind.
Where you write any question and it gets solved with super quality. So far that hasn't been achieved, of course.
Yes, Sasha, what you were saying about limits, new plans, throwing in money to solve something — in some form that has already appeared. Even inside Codex you can buy more, top up some tokens if needed. But yes, it's kind of not comparable. You pay $200 and get one result, but topping up tokens there makes everything much more expensive.
Yes.
And as a transitional stage, if they introduce something like that, I'd see something like paid resets, on the one hand. Or — remember, it's already showing up in the list now — there's a separate usage limit for, say, GPT-6, and a separate limit for GPT-5.3 Spark. They released that model quite a while ago, but since it runs on different hardware, they're putting it into a separate plan. Apparently they want to popularize this not very big but very fast model. And if you hit the limits in one, you can try other models.
Apparently it's not that popular, since they're even trying to promote it separately like this.
Listen, we'll really see what happens. And with all these questions, especially after someone goes public. I mean, if the Anthropic IPO does happen, I think it could reshape the market. And note this latest debate that's going on: on Friday evening Dario Amodei puts out a statement that model development needs to be slowed down.
And I love how one person on X writes: “Waiting for the AI market to crash on Monday.” I was like: “Heh.” So we're sitting in France, me and a friend of mine, and he — I have, of course, stocks in AI and in tech. In the morning I'm sitting there telling him: “It crashed.” Like, how, how, how? Well, you could have predicted it, right?
Then on Monday and Tuesday the AI Summit conference takes place, and at the AI Summit — I go on X and see this moment: Jensen Huang is sitting on stage and Trump calls him, like, by chance, an assistant comes up to Jensen Huang, Trump is calling, and he's on an open line, he joins and says: “Jensen is such a smart man, but I want to say that…” And he's sitting there with a flip phone. I never figured out, by the way, whether it was an Apple Duo or not. But he's sitting there with… Have you seen the photo? Do you know whether it's an Apple Duo or some Samsung? I think it was an Apple Duo.
Yes, I thought it was very interesting, because he's looking for how to turn on the speakerphone and can't turn it on. And Trump says: “Jensen can make any chip in the world, but doesn't know how to turn on the speakerphone.” Well, the audience laughed. I thought: I wonder who's sitting there, Democrats or Republicans, among the hosts and everyone else? Because the hosts laughed really hard. And then Trump goes and says: “I believe there's no threat, no AI threat at all. It's all a made-up story. We have no problems.
AI has to be developed, more capacity allocated and all that.” What is this? It's obviously a very serious political move and slogan. Obviously. And if Amodei said on Friday that — he's the head of Anthropic, a very serious figure, and on top of that, again, his company is going public. And I want to say, Ilnar, that here I saw Amodei's marketing for the first time, because shaking the market like that is really cool, right? From the tech point of view. Someone will say he shouldn't have released it, but it's a move. And then there's such a tough counterweight, because, as we know, the American government is in certain conflicts with Anthropic.
We just need to check which federal court or whatever it was issued a ruling that the American government is unlawfully banning Anthropic in the US military-industrial complex. Yes, I'll note that Anthropic was banned — and Anthropic was banned, let me remind you, after the story that happened in Venezuela. When the president was abducted — supposedly abducted with the help of an Anthropic model — a whole conflict began there. People came out and declared that the head of the CIA, I think, right?
Or the Department of Defense, probably, as it was back then. He declared that — it's called the Department of War in the US. He declared that they would no longer work with Anthropic at all. So they started cutting it back there, because Anthropic said: you can't use it. Let me remind you, OpenAI almost right away signed an agreement to work with the government, quickly, and then, then Sam Altman wrote: “Oh, that's not what I meant. We rushed it, we started off in the wrong direction.” It's just, just surreal. And here again there's a really explicit confrontation.
Meanwhile, look, Elon Musk declares at this party — well, at this conference — that
he supports checking models before release, and he supports not just checking models by their own companies, but checking models by other companies. In particular, he actually thinks his should be checked by, for example, Google, and Meta, OpenAI, and, and Chinese companies. Chinese companies too. By the way, according to Reuters, this statement by Amodei last Friday — that our models need to be checked by outside experts, and that Anthropic is supposedly ready to do it and ready to let outside people in — was reportedly supported by Elon Musk and Sam, Sam Altman. Let me remind you what Elon Musk has to do with artificial intelligence.
He owns SpaceX, which has xAI as an asset, together with the famous Grok, which bought Cursor and which is very much at the forefront of the American market and everywhere. And it's a serious player with a hell of a lot of computing capacity, because they — SpaceX together with Tesla — there are rumors that SpaceX and Tesla may merge. So. Elon Musk first splits everything up, then merges it, then makes holding companies. So. A business giant. And they do have very serious capacity; they're playing seriously there. Ilnar, what do I want to say?
Generally, it feels like this crowd has just been tossing the same words back and forth in a circle for the last three years or so. They keep saying: that's it, humanity is doomed. Then they say: humanity is not doomed. They say: full speed ahead, open it up, full speed, let's build. And all this drives things forward. They even dragged Trump into this hysteria. Because Trump has always supported the development of AI. I'll note that one of the very first press conferences of the newly elected US president, a year and a half ago — we showed it.
We had a whole episode on it, where representatives of the AI market stood around him, and they repeatedly appeared with him everywhere afterwards too. And he talked about allocating 500 billion to the field. Ilnar, remember how the whole world said: where will they get 500 billion? That money doesn't exist! What is this even? And now an Anthropic IPO is being discussed at a valuation higher than SpaceX's. That means we can expect a listing in the range of two to four trillion.
And note, not even two to three, but two to four trillion. All our forecasts are coming true. As for trillion-dollar valuations, if you think about it, it looks very much like planted stories from these people, as if, Ilnar, you know, they were even coordinated among themselves. You'd think that even Sam Altman and Elon Musk, of course, supposedly aren't friends, they're at war — but we never know whether they're friends or at war. Because when Sam Altman — when Elon Musk was leaving Trump, leaving his DOGE, government oversight, he basically said everyone was stupid and so on.
There seemed to be a conflict, and then again there's no conflict. So we'll still find out whether a conflict between Sam Altman and Elon Musk even exists. Maybe this conflict doesn't exist at all. And it's the right, incredibly cool, constructive game — so that, you see, you're rocking more than one boat, right? You've got the neutrals, who don't care either way, right? Like you and me — we think well of professionals, of all of them, right, agreed? Anthropic, SpaceX and OpenAI.
Well, of serious people. These are serious, good people. But there's also the other side: “we're only for Anthropic”, “we only root for OpenAI”, “we'll only root for Google”. Well, Google is more corporate, yes, unlike the three players we mentioned.
No, but the way they arranged with Trump for him to call at the right moment — that's strong. It's hardly a coincidence. Here I fully, let's say, agree with you that—
The US president just happened to call, supposedly.
Yes, yes, yes, yes, yes. Well, arranging for it to be put in his schedule that at, I don't know, 3:15 he has to call Huang, well—
He recently didn't go to one of his sons' weddings — a new wedding, the second one, I think — he wasn't there, didn't come in person, said he had to be at the White House, taking care of his business. What are we even talking about?
A big essay came out — the one you're talking about — where Dario Amodei describes the safety problems that exist and how he sees the solution. The only thing is, when I read it, this idea really comes through: democratic countries are separate, this can only work in democratic countries, and so on. I still have a certain aftertaste from his previous stories. Remember, we once talked about how he called money from, possibly, the Middle East dirty, and then in the end, when financing was offered, he happily took Saudi Arabian financing, I think.
He flip-flops on the fly rather clumsily.
That was your famous shot, Ilnar, your famous jab at Amodei, I think, because you really nailed him.
Sasha Mashrabov backed it right away too. He laughed at the situation for a long time, yes.
Yes, yes, yes.
And now again the whole essay is shot through with this — democratic countries and so on. Although, yes, I have to admit that in it he did state that one of the points is verification by third-party organizations. And there are even points saying that we have, I don't know, workplaces with corporate laptops, and passes for staff who would like to do this, so to speak. Well, apparently some kind of accreditation or something like that is involved. But he says we're already ready to go for this unilaterally.
Access will be almost the same as for our employees who work on risks. Well, with some exceptions, where we're bound by contracts, non-disclosure and so on. So. It's as if they've already started making some moves in this direction. And even more or less unilaterally. But here—
Well, I think Anthropic is more or less at the forefront of this whole story, honestly. And so far they still look like the cleanest and fluffiest of all, even though they develop Mythos and Fable. I mean, their model was, well, in the lead, right? They were still—
At one point they were really well ahead, yes.
Yes, yes, yes.
So. Memes appeared right away. Like, two people are standing and talking. One says that there's a ten percent chance AI could kill all of humanity in, say, ten years. The other: “Aha, so we need to slow down the pace of development we're moving at?” — “Aha, and are you going to slow down?” — “No.” — “Well, neither am I.” So everything will keep moving the way it's moving anyway. No real slowdown will happen. There's a huge number of companies, not only in the US, and leadership gives a lot at this stage.
They released a very big report, a pretty curious report on the use of Claude in cyberattacks. Again, all requests go through the API, and obviously Anthropic analyzes them. And they published a big report showing how, as they see it, groups of hackers or others are actually using Claude. And they note that these aren't Fable- or Mythos-class models, but a class below. But that's still enough to pull off really serious things. If you're interested, read it. The report is very big and fascinating. So.
And if you're interested, watch the special episode we had on the channel on Wednesday, on the idea that there's a 10 to 20 percent chance AI will destroy humanity within ten years. Including the questions Anthropic got hit with regarding biological weapons. There I also went through the view of mathematicians that AI is currently harming science in particular. Watch that story, as a follow-up to what Ilnar is saying now, because, Ilnar, it's a very interesting, important topic, useful for your own understanding.
We rarely hype things up, but we do raise it and put out these hype-style headlines. But the fact that this employee, after quitting Anthropic, said that a lot of employees inside Anthropic think so, and that Anthropic keeps saying this — it's not the first time, Anthropic has been saying it for a very long time. And this comment of his is very interesting — watch it, I had an episode on Wednesday — this interesting comment, you know what it was? That at OpenAI the employees don't fully understand the threat — and he didn't say that OpenAI is some kind of outfit, a company — no, he worked there too, this person; in fact, he worked there first and then at Anthropic.
It'll be interesting if after a while he raises money for a startup — then it'll partly resemble the story with Cluely, I think there was such a startup, I don't know whether it still exists, where a guy, like, didn't finish MIT and passed an interview at Meta using that app of his. He told everyone about it himself. He was expelled for it. And then he raised money for a startup. If Jacob does the same, I'll have some questions about his statements. If we don't hear soon that some new AI safety startup has appeared—
Can you imagine the publicity he's getting? Can you imagine what publicity he has now?
Yes, yes, yes.
Yes, as always.
Yes, yes, yes. So I hope he'll refrain from that, and then his words will be a bit more credible.
By the way, by the way, Ilnar, I want to tell you about this hype, about these people in the Valley.
And I think someone, by the way, wrote a comment under the Wednesday episode — someone said something along the lines of: “Listen, I'm sick of these people who quit and then whine, and then jump around somewhere inside.” Again, there was a story where a person left Meta without working there even a year — Mark Zuckerberg gave him one and a half billion dollars in compensation, when they were poaching him. And he stayed for that compensation, and now he's left. And nobody knows why. These arguments, of course, especially this new… The world splits in a really funny way.
Someone comes along and says: “Listen to this great man. He foresaw artificial intelligence 30 years ago and was there at the dawn of its development.” But in reality all these companies are full of 20-to-25-year-olds. Time has passed — we've been talking about this on ToTheMoon for several years now, and we said GPT came out three and a half years ago, and they'd been working a bit longer than that. They came there at eighteen, at twenty. All these guys, 25- and 27-year-olds — they're the ones developing, very seriously developing, artificial intelligence.
And what's more, we've often noticed how, when there was all that hype — now nobody discusses it anymore. It's no longer interesting that a programmer can be paid $10 million a year, right? It's kind of become routine, right? But remember the discussions when salaries hit a million. I remember saying: “Whoa! A salary of a million. It can't be. It doesn't work like that. It means it doesn't exist.” Or someone writes: “Am I a fool to invest in OpenAI at a valuation of, like, a billion dollars or 10 billion dollars?” And you sit there thinking: if you'd invested at a billion dollars, how lucky you'd have been.
Yes, yes, yes. So. And more on that report — I won't go into details. Read it.
But the main point, I think, worth noting: Anthropic says that thanks to AI all these hacking operations have become much cheaper. If before, roughly speaking, companies that investigate hacker attacks could tell by the scale and type of attack that it wasn't done by a lone amateur but by some big corporation with large resources or something like that, now it doesn't work that way. Now, roughly speaking, even a few enthusiasts with fairly average skills by top-hacker standards can do really serious damage.
That's how they present it. Again, you need to look at everything together, right? This report came out at about the same time as Dario Amodei's essay, and maybe this is just how they reinforce their position, to give it more weight.
Listen, Ilnar, what do you think in general — take criminals who want to get information with the help of AI. I have a feeling, Ilnar, that today the people who create these systems can't stop other people. Because, imagine the situation ten years ago: there are some criminals, and they want to do something. To learn something, they have to meet some experts, go to libraries, read the internet, find some data. And on the internet a huge number of such resources were already being wiped and restricted too. At the same time there was an incredible number of closed platforms — platforms, chats, messengers, forums — where huge amounts of prohibited and harmful information are traded — really harmful to humanity, things harmful to humanity.
With the arrival of AI, honestly, Pandora's box was simply opened for everyone, and everyone got crazy access. True, you can't just write directly: “Make me — a three-year-old child — make me some kind of weapon or something.” But people still got unlimited access. In fact, you can already call it unlimited. It's pretty much unlimited already. Wherever you are, whatever mountains you're sitting in, if you have brains, you can gradually get the data you need. Yes, of course, you won't say that you've applied for research from a Chinese military company, or that you want to get a grant somewhere else, so that you'd get caught.
But you can already get information you didn't have — I don't know, x1,000 or x10,000 or x100,000 in complexity, right? But that's obvious. And essentially, what they want to control is a setup in which people use AI the way these companies themselves can use it. That is, they press some button, a super-button, and this super-button automatically studies all the financial structures of the banks of some particular country. Automatically. They can do that, right? What's more, the agreement OpenAI signed with the US government, I think, says very clearly that if it's not US citizens or residents — well, if it doesn't concern America — they can study anything and do anything. And here there remains…
And who's going to… Or analyzing websites. Today all the systems can analyze absolutely any website. And what's more, it's not even blocked, right? Or analyzing some software or code on your computer — just upload a piece of code and export it, figure it out, and say: “I'm studying something like this. They gave me this as an example at university,” right? But that could be a piece of code that gives access to your own university's lab, for example, or to some service at your company that was written twenty years ago in, say, C++.
And this whole setup, it seems to me, Ilnar — this struggle, these political arguments — like what you said about that meme. I think that meme is a very clear expression of everything that's been happening here for several years. And I've said it before and I'll say it again: Elon Musk himself said a few years ago: “Well, there's a ten or twenty percent chance that AI will get out of control.” Guys, ten percent is a lot. Right? It's critically high. What are you even talking about? I mean, these are people, these are, these are people.
And that doesn't mean research should stop. It means stopping development. But this hype bothers me, because in such hype it's very hard to make sound decisions — economic, political and so on. Especially when even the president calls and swings everything the other way, because someone asked him to — say, to make some money on the financial market, or for some other reason, right? Which is a very strange story — where's Trump and where's the tech market? I think nobody cares at all what he says in terms of impact on the tech world.
Honestly, what difference does it make what he said, that he spoke? But it's an incredibly important call, right? I mean, just a super-important call that shows the absurdity going on in the tech industry. By the way, let's see what Sam Altman says in two weeks at his main annual keynote.
As for the capabilities companies have — you were saying, like, what can the companies themselves, or the government together with the companies, do? They demonstrated it very well with the Navier–Stokes equations, with that Millennium Prize problem. They launched ten thousand agents that worked for eighty hours and solved a problem people couldn't solve for a very long time. Really top mathematicians couldn't do it. But they just showed what power they essentially have.
Who's the daddy here.
Yes, yes, yes.
They burned tens of billions of tokens. And on top of burning tokens, tomorrow they'll come and say: “Sorry, we're switching this plan off,” or “We now have a daily cap on limits.” And that day they'll devour a trillion tokens, judging by 70 million requests per second, and solve another fifty equations. Only maybe for some other question in another country.
Yes. And what other problems they've already solved this way, non-mathematical ones, they're not telling us yet.
What do our viewers think about this? What other, non-mathematical problems have they solved? That's a politically correct way of putting it from Ilnar. What other non-mathematical problems have they solved?
Yes, and many companies already have this power. Yes, I'd also like… Look, probably the last thing on this topic I want to add. On the one hand, with this report Anthropic shows: “we're watching all of this”. That yes, there are hackers using it. “We uncovered all these stories and are even showing them to you in the report,” and so on. On the one hand, they seem to be in control of the situation; on the other, all these actions had already been carried out. So yes, after the fact they say: we saw that such things happen, this is how it happens — but they happened.
I mean, using these bots, these actions had already been carried out. And those are the ones they managed to notice. So on the one hand they sort of restrict and protect, but on the other hand they do it after the fact, when everything has already happened, when, roughly speaking, the attackers have already run off and maybe started acting differently now, and so on.
And by the way, toward the end — well, not the end. Since last time we finished with Apple, I suggest we give an Apple news item.
Apple is supposed to be present in AI somehow, after all. Tim Cook was at the meetings with Trump and everywhere in government too. So, Siri came out — Siri AI. And it's a generation-27 system. Everyone who's installed it or not. Now Siri AI can search for the information you need in messages, mail and photos, continue a conversation with follow-up questions, and so on. In the announcement Apple says that Siri AI can work with your device. And here, Ilnar, I read this and opened it up, thinking: “Which version do I have?” And it says: “Install version 27.” And I realize I already have the developer version 27 installed — it's been there all these months, updating constantly — and my Siri is switched off.
It's switched off so that I could use some other services. I don't even remember exactly why it's off. Because, of course, I don't use it. I don't use Siri AI, and I didn't even use it to test it for our episode, because I didn't even understand what the context would be — when would I need to find some photo? Because if I want to find a photo, I open Photos. Apple has an unreal search by recognition — faces and everything else. I just type in: fish, and it shows me photos with fish, or Polina with a fish — that's my wife — and it shows me Polina with a fish.
Or Polina with a fish and me, and it shows that too. And this thing that Apple—
Sorry, just a joke. So with Polina with a fish, it needs to make sure it doesn't just come up with a photo of the two of you.
So here this—
That's if—
And I realized I don't even have any use cases for Siri AI. If any of our subscribers have some, write in, really, do write. I'm genuinely curious. I'm such an Apple geek in general, but in terms of using artificial intelligence, given how far ChatGPT, OpenAI, has pulled ahead of Apple in AI, I don't understand why I should switch to some, again, intermediate version now. Even when Apple often tried to integrate with various systems — you could choose ChatGPT through that Apple Intelligence button — I did choose it, but it was so inconvenient that I still went back to the ChatGPT interface.
And now it's a system that stores all my context, that knows the history of everything going on with me. So why would I use AI in Apple and ask, I don't know, where I am, about buildings or what's happening in this city, if it doesn't even know the context? Whereas in ChatGPT I open any chat, and even in a new chat it fully knows where I am, which hotel I'm staying at, when I'm going where, where I'm flying. The full chain of context. Sometimes, of course, the system gets confused, but it gets better and better every day.
And here we, of course — well, I clearly realize that the market — well, maybe Apple won't enter this market at all anymore. Many say it may enter later; maybe it won't. By the way, at the time of this recording my iPhone 18 hasn't arrived yet, but it's due on the 18th — so, as I'm saying, the recording is on and the iPhone is on its way, it'll arrive. But of course I don't expect anything from it, with its extra dual processor for AI or whatever it's called — I don't expect any AI improvement.
Yes, it's more of an automatic purchase that happens every year. Although one of my partners wrote: “Sasha, are you going to buy the Apple Duo?” I'm like: “Why, are you?” So. Well, Jensen Huang was sitting on stage with that shovel of a phone. Ilnar, watch it if you haven't seen the video — you have to see it. He was really sitting there with a shovel. I thought: how does he — does he walk around with an iPad or what? His phone is an iPad. It looked absurd, of course, especially if it was an Apple Duo.
We need to find out, I think. I hope our team will add a caption saying what it really was, because he basically couldn't figure out how to turn on the speakerphone on an Apple — or on some other phone. That's where we'll figure it out.
100%. And see you in our episodes — they now come out practically every other day. And on our weekly podcast, of course, every Sunday. Don't forget to support our channel: like it, subscribe and leave comments. All of this helps us grow and be seen by as many people as possible. That's never easy these days. Bye, everyone!