OpenAI has temporarily stopped selling Pro for 200 dollars to new users. Demand for Astra turned out to be so great that the company does not have enough capacity. Almost at the same time Dario Amodei says the development of the strongest models needs to be slowed down and independent checks allowed. That statement comes directly from Anthropic. And right away Trump answers practically the opposite: artificial intelligence must be developed further, and the fears are greatly exaggerated.
And at that very moment Anthropic shows how Claude is already being used in real cyberattacks. So what is going on? Do we already lack 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 will go through these topics with you. Hello everyone! We are on the ToTheMoon channel — technology news, insights from Silicon Valley for the whole world. We are starting our weekly podcast. Well, OpenAI, by the way, released information about how many users they have: a billion users, more than a billion users, 40 geographic regions.
And what is interesting, there is this figure — 70 million requests per second. It seems to me that 70 million requests per second sounds very serious. But let us move on to what information? That OpenAI is blocking registration of new users for the Pro version. And I like that when I asked ChatGPT 5 about this news, it tells me: so, there are two options here. The first option is the load on their computing servers. And the load — because the second point is that there is very strong demand for the Astra model.
And supposedly everyone wants to use the Astra model. And clearly people want to connect at 200 dollars. But I have a feeling there is a third option, which, Ilnar, you and I have discussed many times. That on the whole... I got scared by this story, by the way. At first I thought they meant a tariff in roubles. I think: surely not, how can this be, how can this be. I keep waiting for a new tariff — a bigger one, so there is more. And here they are closing the 200-dollar tariff.
I am not sure they will bring it back. We will see. That is, if they bring it back, it means that... Either someone made a mistake, maybe in marketing policy or in pricing policy in sales. Or they really, really do have some problems with load. Which I strongly doubt. In particular, that so many people in the world ran to buy the 200-dollar version to connect to Astra if they were, for example, on the ordinary Plus. Because usually, when people on Plus want to switch on a cooler version, they buy Pro for 100 dollars, not 200.
And the 200-dollar one, I will note, by the way, differs from the 100 in that there, I think, there are four times more tokens than at 100. And plus the 200-dollar version, well, by my feeling, is the version that gets a more fashionable new release of all sorts of features, data and everything else. Although it is unclear, because they often say 'the Pro version has this', and they once made a Pro version for 100 dollars. It seems to me that by doing so they slightly diminished this luxury list of 200-dollar people who once connected at the dawn, yes.
I even got a pen for that once. By the way, we have super news today. We have super news today. A ChatGPT pen arrived for me. I forgot to mention it, I will forget it. Tanya, ChatGPT sent me a pen — a little thing like this.
You even have it with a stand.
Yes, it is heavy.
Have you checked it, Sasha? Maybe it is there — well, it is a stand.
Maybe it is the stand that writes, because it — it is the stand that writes, you could kill someone with it. It weighs more than a golf ball, well, quite a lot more — maybe two or three times. So, the pen exists. The day after tomorrow I will be back home. And, by the way, in a few weeks — even less already, less, yes, in ten days, it turns out — OpenAI Dev Day for developers will take place in San Francisco. Their main conference, which kicks off, opened by Altman. I will be there.
And the main thing is that I am managing to fit everything in. I was invited to a conference a few days ago, by the way, to one, and I am like — I am sort of not at the conference.
What to choose, right?
Yes, yes, what to choose. Well, it is good that it coincided so that I make it to Dev Day. I have not been to Dev Day yet. It is being held for the third year in a row. Very glad that this year I received an invitation and there is a chance to be there. So, what do you think on this topic — about the versions? Because for many people who may be watching, once again, what is the construction — at least my favourite one on ToTheMoon: if you want, first of all, to really get to grips with artificial intelligence, buy the paid version.
And preferably, if you have — well, if you are ready to invest — buy Pro; people still treat buying paid versions even for 20 dollars like some subscription, I do not know, a health offer, a Whoop band, an Oura ring, or like a Netflix subscription, yet this is something completely different, and it is your whole life, professional development, access to — well, having your own artificial intelligence with you. On the other hand, we have discussed many times, and here, I think, Ilnar and I agree, that access to cool models will be less and less accessible, and it will be limited, and it will be very — because it is very expensive, yes; the Pro subscription essentially devours OpenAI's capacity, because you will never in your life be able to use GPT Astra via the API for 200 dollars for that long, yes, that is unrealistic.
You would eat up those 200 dollars, I think, in about a couple of minutes. What will you tell us, Ilnar?
Listen, I am very sorry that Tanya could not join us today, because you always asked her whether she had subscribed to Pro or not — I think she subscribed at some point, then I think went back. It is just that now, apparently, if Thibault's tweet is to be believed, you can no longer do that — that is, we will no longer be able to advise Tanya a 200-dollar subscription. You will be able to advise 100 dollars, but 200, for now at least, will not work. Yes, so where did I learn this news?
On X, formerly Twitter, there is Thibault, one of OpenAI's employees, who usually always announces limit resets for Codex. There is even a separate site we have mentioned, where all these resets he usually announces around some events are tracked. And he is the one who wrote that because of the very heavy load that poured onto them after Astra, they decided, essentially, not to sell new 200-dollar subscriptions. But at the same time he wrote that existing accounts should apparently not be affected.
Therefore, if you have a 200-dollar subscription and you actually need it, it makes sense not to change anything in your subscription and to make sure there is money on the card so the subscription can always be activated. Because if it lapses, getting back onto it, at least right now, does not seem very possible, at least in the near future. As for the influx — whether it is connected with Astra or not — that is hard for me to say, of course. Obviously they say that, like, the model is so good that a crowd of people rushed to them.
Here — remember, we discussed at some point how I and my acquaintances choose what to use. Say, Claude, Codex, Gemini, Grok or something else. Apart from the quality of the model itself, there are many additional factors, and one of them is limits. And at OpenAI, on the 200-dollar subscription, the limits are really very generous. For example, on Claude the limits fly away much faster if I solve a similar task there. And at OpenAI they are really very large. At the same time, limit resets happen quite often.
Those tweets come out about once a week, once every two weeks. They reset the limits for us very timely. And so many people who use ChatGPT quite actively — not ChatGPT, who use bots quite actively — in particular also choose ChatGPT, OpenAI. The 200-dollar subscription gives very large limits, and you try to squeeze them out fully, because if you have already moved onto it, well, there is no point in not using it. Because, as I say, limits are one of the important factors of choice.
Well, and you use them to the maximum. I want to add, regarding buying expensive subscriptions and so on: it is like the saying about cars — a car is not a luxury but a means of transport. Here it is the same story. And if ChatGPT helps you in your professional activity — well, it is already hard to imagine an area where it would not help, but one way or another — if it helps, then this money pays back quite well. That is, you either do things faster, or do more, or start doing things you did not do before.
And these 100-dollar, 200-dollar subscriptions can on the whole already be covered by that. Probably in many areas one can think of a point of application for these chatbots. By the way, if you happen to have counterexamples, as they like to say in mathematics — some example of an industry in which a chatbot would be absolutely useless — you can write in the comments, it will be interesting to see what people come up with. I am sure one can, of course, think of such a thing, but here it is interesting.
Well, not think one up — give your own. Our viewers have already given real examples of their own, where, someone says, I do not see an application. Or, on the contrary, maybe someone will suggest one, because I... There was a friend with me — I mentioned it several times — he was travelling. Misha, yes, Ilnar, you know him well. And I told him that... Well, he asked me a lot about artificial intelligence, various questions, then he began asking. And I told him: listen, to really get to grips with it, you should not ask me questions, you should start using it every day in everyday life constantly, in all aspects, and learn to use it, and use it in professional life.
And after that, clearly — watch ToTheMoon episodes, some others, ask me questions, read some data, experiment. But simply, if you do not use it constantly every day and look for new use cases for yourself — remember those first few years of ours; this has been repeating for how long, two and a half or three years, yes, such really active use of GPT models — remember how big this topic was, or three and a half years already, how much...
Well, we have been filming the podcast every week for almost two and a half years, and before that how much more time, yes.
And GPT almost three and a half, yes, at least from the beginning of 2023 it went really actively, from March, from the spring. And just on the whole, if you look at how the story was when these new things appeared for you — as if you were developing some new organ — you did not understand at all that you had to think that way, that such a thing could be done, that there was such a process of developing, of creating something. And of course only now are people beginning to immerse themselves in it — this statement about a billion users, by the way, is not exactly new, yes; I think a month ago I already said they had a billion users, because there were various data from them. And here it is their official statement, so to speak; the previous one, I think, was 900 million people — they have not run that far ahead, that is, if you take a whole year, over the year they have not run far ahead, but I think it has increased greatly, the time these users spend inside has increased greatly, the amount of engagement they have inside, and so on. Listen, there was an interesting study I want to add, and then you continue exactly with this. There was another interesting topic here, that if...
First, I filmed an episode — watch it — a few days ago on whether it is worthwhile to use various plugins in various systems. I very much love the middle layers, all sorts of intermediaries, those systems that resell ChatGPT 5 or Anthropic, or that say 'in our CMS we have the most fashionable AI' and so on, yes, because there you never have access to cool models like Astra or Fable. But listen, there was an interesting topic: studies were done on medical data, and what turned out is that if you handed the system your test results in advance, for example, and some data about your health, then the system answered you with quality, but if you told the system 'collect information about my health from me', then it answered poorly. Although this topic, by the way, it...
It really... I got interested in this construction, because I essentially advocate and keep saying that the new era of artificial intelligence will begin in earnest when these systems finally start collecting information from a person, yes. At the same time, the very work and the quality of this model's work, its ability to collect data from a person and a person's ability to answer — to match them up — because that is also a problem, I think, of people's answers — this still remains a fundamental problem.
And perhaps, by the way, that is why models do not collect information from a person all that strongly, although it seems to me this is the newest channel for that kind of learning. Let us see what happens, so as not to forget to talk about this topic at all.
In short, the Pro version — do you think we will not see it again?
Listen, I think we will. I think it is a marketing story. Well, I mean, they will wrap it in a marketing story. Possibly they really are not managing to satisfy all the demand. I do not know whether you already have this or not. In Codex, after the latest updates, the option to agree on my behalf in some non-critical cases has disappeared for me. Yes, remember we discussed at some point that there are three levels in Codex: when you do not allow it anything without your consent, every step has to be confirmed.
The other extreme was that everything is allowed, nothing needs asking, it does whatever it wants. And in the middle there was the moment when it decides for itself whether it is needed or not.
Yes, in English it is called 'approve for me' — it only asks about actions that essentially represent a potential threat.
So, yesterday evening, it turns out, I updated Codex, and today I do not have it. I do not know where they put it. Well, but this could also be a consequence of the same thing — to slightly reduce the number of tokens used, because in this situation, when it approves for you, it, it turns out, analyses that decision, tries to understand, and some additional requests arise. Perhaps they reduced it this way too. Or it only got lost somewhere for me.
By the way, Anthropic increased. By the way, Anthropic increased the limits. And literally a few days ago they announced it. They had a promo where before that they raised the summer ones. They, like, removed the summer ones but added new ones. Well, and there is a difference there. Sort of by the overall difference.
That is, this is of course not for free tariffs but for paid tariffs.
And what do you think — was there a switch? Because judging by the internet and by YouTube, there was a switch. And by the news. And Astra is being discussed far too much. Although I have a big question. Well, never mind. Our viewers will write here and again split into two teams. And do you think there was a switch?
I fully allow for it. I am sure that the majority, the overwhelming majority, do not use only one system but use several, to keep their finger on the pulse all the time. And given all this hype connected with the Millennium Prize problem, with benchmarks like ARC-AGI, which was simply torn apart, on the whole everyone wants to try, to try running it. I am sure that maybe they did not switch permanently, but they definitely started trying, comparing with GPT-6. Most users of both Gemini and Claude.
It seems to me the active users have all tried it. In that sense I allow that such an outflow may exist. It may be temporary. It may turn out that the quality does not satisfy them, and then people will come back. Or the raising of limits that Anthropic is now offering may tip things towards Claude. Claude really is very good — nobody argues with that. But such an outflow with such a news hook, hype — it could well have happened, and I allow that it did. And accordingly, continuing this story, I am sure that some other tariff — if they do not bring back the 200 dollars, then some other tariff — will appear anyway.
Whether they do it, or time it, I do not know, to the release of some hardware, or connecting some data centre, or something else like that. But under some such story they will either bring the tariff back or invent a new one. Here, I think, that will happen. I do not think it will be very soon, right?
It is unclear, yes — admit it is unclear, because I still, remember, it seemed to me — and from conversations somewhere inside the technology company OpenAI I saw that an expensive tariff should appear, right? The expensive tariff has still not appeared. Even Sam Altman has often said, like, you will be able to spend a certain amount of money, for example, on solving a task. I thought this topic was going to go that way — for example, that you could click a button and some capacity would be allocated to you, because after all this is not a question of buying tokens via the API, right?
It is some other construction of interaction, because API tokens are a different topic after all. You, first of all, spend sort of after the fact; you do not say in advance that you have some budget and it optimises that budget for this. That is, it is completely different. Moreover, essentially, in fact, nobody understands how they are consumed. That is, you can count as much as you like there, whatever you like. But first of all, models spend these tokens differently depending on the context you gave.
And moreover, the cooler the model seems to be, the fewer it should spend. And so on, right? So calculating this, I think, was not possible and still is not. I, by the way, still... Ilnar, have you looked at how many tokens you have spent?
I have...
Oh, I am approaching 50 billion. There — I have 50 billion tokens spent.
That is, of course, always in Codex, in Codex, right?
Not in Codex.
That is always interesting to observe, of course.
What else do I want to say about OpenAI? An important thing. Very... This information actually reflects the reality of what is happening very precisely. Once again, watch our episode. A few days ago it was — on Friday. On plugins — what I filmed is a very useful construction. For everyone to 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 continue to think that if a third-party service — I do not know, on health — is tuned for health, and it has artificial intelligence, then that artificial intelligence, even if it is OpenAI's, is somehow cooler.
I tell everyone: it is not cooler, it is definitely worse, right? Because via the API in an application you will not run GPT-6 Astra Pro, for example, right? Or Extra High. That is, you will not be able to run it. So, look, there is a very interesting construction that has appeared. And by the way, about this topic, about Pro — there is basically no information anywhere about it. And here is the following info: that ChatGPT no longer chooses deep reasoning on behalf of the user on Plus and Pro.
This is very... It got me... When I saw this in general... You saw it, right? When I saw it, at first I was like: what does 'chooses' mean?
And what — was it choosing?
That is, first of all, we have all got confused about who chooses which mode, where. Right? Because there is also, first of all, ChatGPT-8, there is Work, and then separately you launch Codex. And these have their own distinct settings, different tokens and so on. You periodically do not understand how it resets these bubbles — resets, does not reset, shows, does not show. Very hard to understand. But look, here is a very important construction: in the latest versions that have come out, OpenAI had... Essentially OpenAI, for example, has a fast-answers mode, right?
I do not recommend you use the fast-answers mode. It is a kind of money-saving for OpenAI. That is, if you want to save OpenAI money, please do. I always use the mode — at least within my subscriptions — Extra High or Pro, but you can use, if you do not have the Pro version, simply the very latest mode. Because this latest mode — actually, people think the latest mode does not answer quickly. But in reality, when I ask through Extra High, if the question is simple, it answers me instantly.
Instantly. And if it thinks longer — this is in Extra High, not in Pro, look, not in Pro. If it thinks more, it means your question requires time to work through. I have checked this many times already. That is, if I uploaded — there, yesterday I bought something, I needed to check whether the composition of herbal teas passes, which country they can be brought into. And I, for example, was talking in Extra High, and it answered me very quickly. As soon as I passed it an image and pictures where all the ingredients had to be recognised, absolutely all of them, it went into reasoning and answered me for quite a long time.
So look — earlier, not now. Your model, if you chose the fast-answers mode, could switch over when you wrote to it 'think long'. It is a well-known topic: everyone says write 'think long', 'think more', 'study more', 'look' — there it is cooler, a heap of everything. It switched itself to a more serious model. Now this will no longer happen. That is, the system has now removed these things — we do not know why. Perhaps it is connected with GPT-6 Astra, with what Ilnar was telling — in particular, that Astra does not work like the standard previous models.
OpenAI released an official instruction outright. See my Friday episode on plugins. So, there is a very important thing there that they talk about: how to write requests in Astra and how much requests in Astra differ in terms of how you wrote them before and how you write them now. Accordingly, note that the switching will no longer happen. And it seems to me, Ilnar, that this is possibly some serious shift in their models, or they found some errors or problems somewhere, including in terms of computing capacity.
Because there was a study here that the Agents.md file for coders may lead to a worse result and increased token consumption. That is, a big file. That is, on the one hand it can seemingly improve development, and on the other hand, now, with modern models, with better things already, it can make things worse. I want to ask you a question, Ilnar. I noticed — I raised this in the Friday episode too, and you and I will continue taking it apart as well.
But these are different topics. On Friday, watch the one about plugins. In the past one. So, you know what I noticed? That it has become more expensive for me to open a new context window, a new chat in Codex or in Claude Code. Moreover, both there and there. That is, I noticed that every time I open a new context window, I need to configure the system. Configure it every time, regardless of the fact that it has these Agents files there, everything is laid out, everything is explained, there is some documentation and so on.
That is, every time I have to be configured. And I had a very interesting case. I had a task I could not solve. You 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 did not solve it for me; I switched to Fable 5.1, and Fable 5.1 did not finish solving it. So, in the process of solving, in these last days, at some point, when everything — I was pouring over Astra, Fable — and I say to Fable: maybe the context window here is already very long.
The context window was more than a month old; it was already a total mess there, in this task. I said, maybe I should move to some other place — and I did try moving Fable to other places, and it did not get solved. I kept coming back to this context window, because at least something moved there, it moved somehow. And in other places it was just carnage, well, a mess. And I wrote: maybe I should move — and Fable stopped me and tells me: there is no need to move to another chat.
So, I am compressing everything here, everything here is great. And from my memory, earlier it was not like this. That is, the model has changed — changed in terms of characteristics. Like you are now saying: Codex removed this, it only confirms complex things. That is actually, damn, what a change. And this change may be connected, on the one hand, with the model spending tokens. On the other hand, it may be connected with Codex wanting to get access to the computer. And yesterday GPT-6 Astra, while solving a fairly simple task — there, the settings of one timer — on a remote Codex — that is, I have a computer standing at home in the Valley.
And there is Codex on it, on which certain tasks run. I asked it to set a timer for me there. It tells me: to do this, I need full access to the disk. And GPT-6 Astra told me this in Extra High mode. Fine, not in Ultra mode, but in Extra High mode. It told me this. To which I said, like, why do you need full access? Oh, sorry, so I do not need it. There. Maybe OpenAI does after all want to roll out access more widely — not in order to steal your information. By no means think that I am talking about that.
But so that the more access it has, the more opportunity it has to offer you various things. Or to create some, perhaps, more complex constructions tied to the computer. What do you think about this, damn it?
Look, as for the compression of context there, transferring from one window to another and so on. This is very similar to what happened earlier, I do not know, with prompt engineering, when people taught each other how to write a prompt to get a better result. In the moment it obviously worked. If you wrote the prompt correctly, you had a high probability of getting a good answer. What happened next? Models developed further. And now, well, it is not very important how you write the prompt — the model will figure it out on the whole.
That is, not such a striking difference. The same thing — various orchestrators appeared that let you connect some additional links with other things and so on. Now all this is packed inside via MCP. Earlier it was important that you somehow transferred the context yourself, told it yourself in the new window what it needs to know here or not. Now it is not so important. Models continue developing further. Context compression works quite well. And if we are in the same task, there is probably no great sense in opening a new window.
If the task is, well, different, but within this project, it probably makes sense to do it in a new window. You know what this resembles? A road roller is driving — well, it drives slowly, but sometimes you can run ahead of it. And it turns out that some ideas appear that are a little bit ahead of this road roller, but it catches up with everything. And Agents.md — what you are saying, that some old Agents.md files that are too long can make things worse with the new Astra, with the new settings and so on.
I am sure this happens simply because many things that were tried out by enthusiasts, companies, whoever else, just start getting built in further inside. Into, I do not know, ChatGPT, Claude and whatever else. And there is no longer any need to do it by hand. And we reach a situation where when you do it by hand, you only start making the situation worse. That is, it is not a given that you can compress the context better than the system inside will, the one configured in ChatGPT, in Claude or wherever else.
If, for example — I do not know, a year ago this was definitely so — you knew this task well yourself, you had spent a long time on it, and you could accordingly give the correct inputs in the new window, and the incorrect ones, the mistakes you had, erase them, not carry them into the new chat, and the model would work better — then now context compression already works very well. In many ways it all comes down to: well, use the settings out of the box. Yes, at some point you will get a slightly better result if you, I do not know, additionally use some skills or something else, but only if you really understand it well.
In other situations, using it out of the box is better, in my opinion, because, well, in a month everything that was at the cutting edge for some advanced geeks will already be built into the system — the harness and everything needed, all of that will already be...
Yes, you can lose — you can lose this habit of using the modern models after all, learning to work with them directly, being with them directly; we are often confronted with, well, you know, people say: but you told me about a model that came out in South Africa, that is so important,
and it is so important to look at it — but yes, please, you can use any models and various open models and so on for some personal tasks of yours; but en masse, globally, if we talk even about one's own use as a person of a main model, clearly, for example, even in my various businesses, in various partnerships, in various teams completely different models are used, yes, and there are models that are local, open source ones get installed for all sorts of things, there is a lot of everything — but basically, speaking of yourself, you still choose for yourself, first of all, some main fundamental model for your tasks — like ChatGPT remains that for me, and I think for you too, yes, for your main tasks ChatGPT remains — while, for example, I use both Codex and Claude Code very seriously; it seems to me even Claude Code more over the last two months, it seems to me this started from mid-July, with Fable, yes, I use Claude Code more, and by the way I even — I have weeks when I do not use up all the tokens in Codex, and this does not bother me much, not overly, yes; that is, just yesterday — it is just that I have a lot of travelling right now, and I cannot say I have the time, yes, to sit with the models for a great deal of time constantly, these weeks while I am travelling — and yesterday I launched several tasks in GPT-6 Astra, and again, well, the result for me was very so-so; we will see, there is a fairly big project there,
I handed one over yesterday, we will see how seriously this project — I have not checked yet — has been completed as of today, and whether the model really managed to make it truly independent of me. But the choice of system — that, it seems to me, is very important, yes, which system you work with; and this topic of the kind 'launch artificial intelligence bots through Microsoft Excel', yes, artificial intelligence bots through Word, asking artificial intelligence bots in Google Drive or in Google mail — this is, of course, a topic for, well, very simple users; that is, not even for simple users, for people who cannot work with a computer, in theory — that is, this is exactly how it looks, that is, it is made for people who do not work with computers and who have some support with the computer.
Clearly it is great to have good artificial intelligence bots, really well-tuned support in companies, yes, when you write to support now, or can quickly sort out the documentation, or get a quick answer — but by the way, I cannot say that revolutionary changes in the world on this topic are visible; there are companies where this is done quite well and cool, but most companies are very much in the back line — so that you write any question there and it gets solved super-well for you — that is, for now this is, of course, not implemented.
Yes, Sash, and what you are saying about limits, new tariffs, throwing in money to solve something — in some form this has already appeared, that is, even inside Codex you can buy more, throw in a few more tokens if needed, but yes, it is incomparable: that is, you pay 200 dollars, you get one result, and there, throwing in tokens, it all becomes much more expensive. And as a transitional stage, yes, if they introduce something like that, I would see something like paid resets on the one hand, or — remember, right now a separate limit is already appearing in the list for using, say, GPT-6, and a separate limit for GPT-5.3 Spark. They released that model quite a while ago, but given that it runs on different hardware, they are moving it into a separate tariff — apparently they still want to popularise this not very big but very fast model, and in case you hit the limits on one, you can try other models. Apparently it is not particularly popular, since they are even trying to promote it separately like that. Listen, let us see
in the evening, Dario Amodei puts out information that the development of models needs to be slowed down. So, I like it — one person on X writes: we await a market crash on Monday in artificial intelligence. I am like — here we are sitting in France with a friend of mine, and he has, I have, clearly, stocks in artificial intelligence, and in tech; I sit there in the morning telling him: it crashed. So, well, it was to be expected — and here on Monday, Tuesday the AI Summit conference takes place, and at the AI Summit I go onto X and see, in the moment, Jensen Huang sitting on stage, and he gets a call from
Trump — as if by chance an assistant comes up to Jensen Huang, Trump is calling, and he joins on an open line and says: Jensen is such a smart man, but I want to say that — and he is sitting there with a flip phone, and I never did understand, by the way, whether it was an Apple Duo or not. And he is sitting with — did you see the photo? You do not know: an Apple Duo, or was it some Samsung? I think it was an Apple Duo, yes; I think it was very interesting because he is looking for how to turn on speakerphone and cannot turn on speakerphone, and Trump says: Jensen can make any chip in the world, but he does not know how to turn on speakerphone. Well, the hall laughed; I thought, interesting, who is sitting there — Democrats, Republicans — among the hosts, among everyone, because the hosts laughed hard, and then Trump goes and says: I believe there is no threat from artificial intelligence, no threat at all, it is all an invented story, we have no problems, artificial intelligence must be developed, more capacity must be allocated, all of that. What is this — this is of course a very serious political move and slogan, obviously, and if Amodei on Friday said that — this is the head of Anthropic, yes, a very serious figure, and moreover, once again, his company is going public — I want to say here, Ilnar, that for the first time I saw model marketing, because shaking the market like that is very cool from the point of view of — someone will say he should not have
knocked it down, but it is a move, and here is such a harsh counterweight, because as we know the American government is in certain conflicts with Anthropic; by the way, Anthropic through the courts — our editors will pull up this news, the exact name — a federal court or whichever
issued a ruling outright that the American government is unlawfully banning Anthropic in the military-industrial complex of the USA, yes; I will note that Anthropic was banned — and Anthropic was banned, I remind you, after the story that happened in Venezuela, when the president was abducted, that he was supposedly abducted with the help of an Anthropic model; a whole conflict began there, people came out and declared — the head of the CIA, I think, yes, or the Ministry of Defence, probably, as it was then — he declared that — it is called the Department of War now in the USA — he declared that they would no longer work with Anthropic at all, so they started cutting it, because Anthropic said 'you cannot use this' — and I remind you, OpenAI practically right there signed an agreement that it would work with the government, quickly, and then — and then Sam Altman wrote: oh, I did not mean that, we rushed, we started off in the wrong direction — just such surrealism; and here again there is this very explicit confrontation — and at the same time, look: Elon Musk declares at this party — well, declares at this conference — that I support checking models before release, and I support not just companies checking their models
themselves, but I support models being checked by other companies — in particular he generally thinks his should be checked by, for example, Google, Meta, OpenAI and — and Chinese companies, that is, Chinese companies too; by the way, according to Reuters, this statement by Amodei, where he said last Friday that our models need to be checked by outside experts — supposedly Anthropic is ready to do this and ready to let outside people in — was supposedly supported by Elon Musk and Sam Altman; I remind you what relation Elon Musk has to artificial intelligence: he is the owner of SpaceX, which has the asset xAI together with the famous Grok, which bought Cursor, and which are very seriously at the forefront of the American market and everywhere — and this is a serious operator with hellishly big computing capacity, because they, SpaceX together with Tesla — there are rumours that SpaceX may merge with Tesla — Elon Musk first splits everything apart, then merges, then makes holdings; a giant of business, and they have very serious capacity, they play seriously there. What I want to say is that generally there is a feeling that this crowd has simply been tossing the same words around in a circle for about the last three years, all the time
saying that humanity is doomed, then saying humanity is not doomed, saying let us race, let us open up, let us race, let us not — and all of this drives things forward; they even dragged Trump into this hysteria, because Trump has always supported the development of artificial intelligence; I will note that one of the very first press conferences of the newly elected US president two and a half years ago — we showed it, we had a whole episode on it — where representatives of the AI market stood around him, who then repeatedly appeared with him everywhere, and he spoke about allocating 500 billion to the field — remember how the whole world said: where will they get 500 billion, there is no such money, what is this at all — and here, an IPO of Anthropic is being discussed at a valuation more significant than SpaceX's, which means we can expect a range of 2–4 trillion before exit — and I note, not even 2–3 but 2–4 trillion; all our forecasts coincide. So, as for trillion-dollar valuations, this looks, if you think about it, very much like these plants from these people, as if they were, you know, even somewhere coordinated among themselves; it would seem that even Sam Altman and Elon Musk — of course they supposedly do not get along, they are at war — but we never know whether they are friends or at war, because when Sam Altman — when Elon Musk was leaving Trump, leaving his DOGE, government control, he generally said everyone was stupid and so on, there seemed to be a conflict, and then again there is no conflict, and
so we will yet find out whether a conflict between Sam Altman and Elon Musk exists at all; maybe this conflict does not exist at all — and that is right, an incredibly cool constructive game so that, you understand, you rock not just one boat — you have, like, neutrals for whom it makes no difference, yes; you and I, well, treat professionals well, all of them, yes — agree, Anthropic, and SpaceX, and OpenAI — well, serious people; serious people are good; while I am still a counterweight, like 'we are only for Anthropic', 'we are only for OpenAI', 'we are only for Google' — there is no anti-stance, but Google is more corporate, corporate, yes, unlike these three structures, so to speak, but how
did they arrange with Trump for him to call at the right moment — that is strong; that is, it is unlikely to have coincided — here I fully agree, let us say, that yes, yes — but to arrange for it to be put in his schedule, that at, I do not know, 3:15 he has to dial Huang — he was recently at the wedding of one of
his sons — at a new wedding, the second, I think; he did not attend, did not come in person, said 'I need to be at the White House dealing with my affairs' — we say, fine, but apparently his interest in this is quite strong.
this story really runs through it — that democratic countries are separate, that this can work only in democratic countries, and so on; I still have this residue from his previous stories — remember we once talked about how he called money possibly from the Middle East dirty, and then in the end gladly took advantage of Saudi Arabian financing, it seems, when financing was offered — your famous 'call', I think, and Sasha immediately supported it clearly too, he laughed at the situation for a long time — and now again all this runs through it, about democratic countries and so on, although yes, it must be admitted that in it he already indicated that one of the points is checking by third-party organisations, and there are even points that 'we have workplaces with corporate laptops and passes for employees who would like to do this', well, and accreditation for it; what is happening is that he says 'we are already ready to come to this unilaterally; access will be almost the same as for our employees who deal with risks, with some exceptions where we are bound by contracts, non-disclosure and so on' — that is, as if they have even already begun making some moves in this direction, and even, more or less, unilaterally.
It seems to me that Anthropic is more or less at the forefront in this whole story, honestly, and they still look the cleanest and fluffiest of all, although they develop — well, they had their own model, but in leadership, yes, they were really well ahead in the moment, yes yes yes yes, there
memes appeared right away — I do not know, two people stand talking, one says, like, 'with a probability of ten percent it may kill all of humanity in ten years' — 'aha, so we need to reduce the speed of development we are moving at' — 'aha, are you going to reduce it?' — 'no' — 'well, neither am I' — like, everything will still move the way it moves; no real slowdown will happen, there is a huge number of companies, not only in the USA, and leadership gives a lot at this stage, and so I doubt it. The second story, also from Anthropic, which was fuel for the same fire: they released a very big report, quite
a curious report on the use of Claude in cyberattacks — again, all requests go via the API, and obviously Anthropic analyses them — and they published a big report showing how, as it seems to them, hacker groups or whoever else are actually using Claude, and they note that these are not Fable- or Mythos-class models but a class below, yet that is still enough to pull off very serious things; whoever is interested, read the report, it is very big and fascinating.
And whoever is interested, watch the special episode that came out on Wednesday on our channel, about there being a probability of 10–20 percent that artificial intelligence will destroy humanity within 10 years — including the questions Anthropic caught regarding biological weapons; there I also took apart the question of mathematicians believing that artificial intelligence is harming science in particular today — watch that story, in continuation of what Ilnar is saying now, because Ilnar — this is a very interesting, important topic, needed for your own perception; we rarely go for hype, and we raise it and talk about these hyped headlines, but the fact that this employee said, on leaving Anthropic, that many employees inside Anthropic think so, and that Anthropic constantly declares this — not for the first time; after all, Anthropic has been declaring this for a very long time — and that comment is very interesting; watch Wednesday's episode, I have this interesting comment there, you know which: that at OpenAI the employees do not fully understand the threat — he did not say that OpenAI is some sort of shop, a company; this person also worked there, in fact worked there first, then worked at Anthropic — but he did notice the fact of what is happening, and he said that at OpenAI they discuss it too anyway. It will be interesting if after some
time he raises money for a startup — then this will partly resemble the story with Cluely, it seems there was such a startup, I do not know whether it still exists, where a person, like, did not finish — got hired by Meta after passing an interview with the help of that app of his, told everyone about it, was expelled for it, and then raised money for a startup; if Jacob does the same, I will have questions about his statements — if in the near future we do not hear that some new startup in the field of safety has appeared, as always with him — so I hope he refrains from that, and then his words will deserve a bit more trust. By the way, I want to say about these hypes, these people
in the Valley — and it seems to me, by the way, someone wrote in the comments to Wednesday's episode, someone said something like: listen, we are sick of these people who resign and then whine and then jump around somewhere inside; again, there was the story that a person left Meta without working a year, to whom Mark Zuckerberg gave one and a half billion dollars in compensation — he was being poached, and he stayed for the compensation, and now he has left, and it is unclear why; these reasonings, of course — especially since this new world in general divides very amusingly: someone comes and says 'listen to this great man, thirty years ago he foresaw artificial intelligence and stood at the dawn of its development', whereas in fact in all these companies there is a huge number of simply 20–25-year-old guys — time has passed now, we have been telling this on ToTheMoon for several years already, and we said three and a half years ago GPT-5 came out — and they had worked a bit longer, they came at 18, at 20, all these guys, these 27-year-olds — they are the ones developing, but very seriously developing, artificial intelligence; and moreover we noticed a lot how, when there was this hype — now nobody discusses it any more, it has become
uninteresting — that, like, a programmer can get 10 million dollars, it has already become sort of commonplace, yes; remember when there were these discussions when salaries reached a million — I remember saying: you, so, a million salary — no, maybe, so, it does not work like that, so it does not exist — or someone writes 'am I a fool to invest in OpenAI at a valuation of a billion dollars or 10 billion dollars' — you sit there now: if you had invested at 3 billion dollars, you were lucky.
Yes, yes, yes. Right. And one more thing about this report — I am not going into details there, read it — but the main point that, it seems to me, needs noting: Anthropic notes that thanks to AI all these hacking operations have become much cheaper.
If earlier, say, companies that investigate hacker attacks could tell by the scale, by the type of attack, that it was not a lone amateur but a large corporation with big resources or something like that, now it does not work like that. Now, say, even a few enthusiasts of fairly average qualification by the standards of, say, top hackers can cause very serious trouble. That is how they show it. Again, everything has to be looked at together, yes. This report came out at roughly the same time as Dario Amodei's, and perhaps this way they are simply driving their position home, to add weight to it.
Listen, and do you think, Ilnar, in general — if we take criminals who want to obtain information with the help of artificial intelligence? I have the feeling, Ilnar, that today the people who create these systems are not able to stop other people, because imagine the situation 10 years ago, yes: there are some criminals, they want to do something. To study something, they have to meet some experts, they have to go to libraries, they have to read the internet, find some data.
And on the internet, too, a huge number of such resources were being erased, restricted. At the same time there existed an incredible number of closed platforms — chat platforms, messengers, forums — where a huge amount of forbidden information and things harmful to humanity was traded. With the arrival of artificial intelligence, honestly, we all simply opened Pandora's box, and everyone got insane access. True, you cannot write directly here 'make me a child of 3', 'make me some weapon' or something else, but still people received unlimited access.
In fact it can already be called unlimited. It is already fairly unlimited. Wherever you are, whatever mountains you are sitting in, if you have brains, you can gradually obtain the data you need. Yes, indeed, you will not say 'I applied for research from a Chinese military company or somewhere else, I want to get a grant', so that you get caught — but you can already obtain information you did not have, I do not know, of x1000 or x10000 complexity. But that is obvious. And essentially what they want to control — they want to control the construction in which people use artificial intelligence the way these companies themselves can use it. That is, they will press some button, a super-button, and this super-button will automatically study all the financial structures of the banks of some particular country. Automatically. They can do that.
Moreover, in the signed agreement between the US government and OpenAI, I think, it is said very clearly that if these are not US citizens and not residents, if it does not concern America, then they can study anything and do anything. And here it remains — and who will, or analysing websites. Well, today all systems can analyse absolutely any website. And moreover, this is not even blocked. Or analyse some software or some code on your own computer — just upload some piece of code and export it, work it out, say 'I am studying something like this, at university they gave me an example of what this is'.
And it may be a piece of code that gives access to the laboratory of your own university, for example, or gives access to some service in your company that was written 20 years ago in C++, say. And this construction, it seems to me, Ilnar — this struggle, these political reasonings — as you said about this meme; it seems to me this meme is a very clear expression of everything that has been happening here for several years. And I repeated and repeat that Elon Musk himself said a few years ago: well, there is a 10 or 20 percent probability that AI will get out of control.
Guys, 10 percent is a lot, yes — that is critically a lot; what are you even talking about, these are people, these are people. And it does not mean research has to stop; it means development has to stop — but this hype disturbs me, because in such hype it is very hard to make sound decisions, economic, political and so on, especially if the president is already calling and overturning everything in the opposite direction, because someone asked him to — because, for example, money needs to be made, maybe on the financial market or for some other reason, yes — which is a very strange story: where is Trump and where is the market of the technology world; that is, to everyone, I think, it makes no difference at all what he says in terms of influence on the technology world, honestly, what difference does it make what he said, that he spoke — but it is an incredibly important alarm bell, it is simply a super-important alarm bell showing the absurdity going on in the technology industry.
Let us see, by the way, what Sam Altman says in two weeks at his main annual presentation.
As for the capabilities the companies have — you say, like, what can the companies themselves, or governments together with companies, do — they demonstrated it very well with the Navier–Stokes equations, with that Millennium Prize problem. They launched 10 thousand agents that worked for about 80 hours and solved a problem that people could not solve for an enormous amount of time. Very cool mathematicians could not do it. But they simply showed what power they have.
Who is the daddy among them.
Moreover, those tokens — tomorrow they will come and say: sorry, we are switching the tariff off here. Or: we have a limit restriction for the day now. On that day they will devour a trillion tokens — judging by 70 million requests per second, yes — and solve another 50 equations. Only, perhaps, for some other question, in another country.
Yes — that is, what other tasks they have already solved this way, and not mathematical ones — that we are not being told yet. What do our viewers think on this topic — yes, what other tasks have they solved, non-mathematical ones?
That is a politically correct statement from Ilnar. What other non-mathematical tasks have they solved?
Yes, and this power is already there at many companies. Yes, I also want to look at the last thing, probably, on this topic — what else I want to add. On the one hand, with this report Anthropic shows that we are watching all this, that yes, there are hackers who use it, and we, in fact, 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, and on the other hand, all these actions have already been carried out.
That is, yes, after the fact they say that yes, we saw that such things happen, this is how it happens — but they happened. That is, using these bots, these actions have already been carried out. Those are the ones they were able to notice. That is, 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 possibly started acting differently now, and so on.
And by the way, to finish — well, not to finish — since last time we finished with Apple, I propose to give an Apple news item.
Apple, after all, should somehow be present in artificial intelligence at all. So, Tim Cook was also present at meetings with Trump and everywhere in the government. So, Siri — Siri AI came out, and this is with the 27-generation systems, for everyone who has installed them or not. Now Siri AI can search for the needed information in messages, mail, photos, continue a conversation with clarifying questions and so on. That is, in the announcement Apple indicates 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 realise I already have the developer version 27 installed — it has been installed all these months, updating constantly. And Siri is switched off for me. And it is switched off so that I could use some other services. I no longer remember why exactly it is switched off. Because I, of course, do not use it. That is, I do not use Siri AI. And I did not even use it to test it for our episode, because I did not even understand what context this is, when I need to find some photo.
Because if I want to find a photo, I open Photos — Apple has an unreal search. By face recognition and everything else — I just type. Fish. Well, and it shows me photos with fish. Or 'Polina with a fish' — for my wife. It shows me Polina with a fish. Or Polina with a fish and with me. So, and it shows something else too. And this story that Apple has...
Sorry, here is a joke. That is, 'Polina with a fish' is needed so that it is not your photo together that simply comes up.
So, here it is... That is, if... And I realised that I do not even have usage scenarios for Siri AI. Whoever of our subscribers has them — write. Really, write. I am sincerely interested. I am such an Apple geek in general. But in terms of using artificial intelligence, how far ChatGPT, OpenAI, has pulled away from Apple in terms of artificial intelligence — I do not understand why I should switch to some, again, intermediate version now. That is, even though Apple often tried to integrate with various systems — you could choose ChatGPT through this Apple Intelligence button — I chose it, did it, but it was so inconvenient that I still went back to the ChatGPT interface.
And now, after all, it is a system that stores all my context, that knows the history of everything going on with me. And why would I, say, use artificial intelligence in Apple and ask, I do not know, where I am, about a building, or what is going on in this city, if it does not even know the context. Whereas in ChatGPT I open any chat, and in the new chat it fully knows where I am, which hotel I am staying in, when I am going where, where I am flying. A full chain of context.
Sometimes, clearly, the system gets confused, but it gets better and better every day. And here we, of course, understand — well, I very clearly realise — that the market — well, possibly Apple, of course, will no longer enter this market. That is, many say it may enter later; possibly it will not. By the way, at the time of this recording my iPhone 18 has not arrived yet, but there it is, on the 18th just as — as I say, the recording is going on and the iPhone is on its way, yes, it will arrive.
But I, clearly, do not expect from it — with the additional dual processor for artificial intelligence, or whatever it is called — I do not expect an improvement in artificial intelligence, yes. It is rather just an automatic purchase that happens every year. Although one of my partners wrote: Sasha, are you going to buy the Apple Duo? I am like: and what are you going to do? Well, but here Jensen Huang was sitting on stage with this shovel. Ilnar, take a look — if you have not seen the video, you have to watch it.
He really was sitting with this shovel; I thought, is that how he walks around with it, with an iPad, or what — a phone-iPad? It looked, of course, absurd, especially. If it was an Apple Duo, we need to find out. I think — I hope our team will confirm what it actually was, because he essentially could not find on the Apple how to turn on speakerphone, yes, or on some other phone. Here we will get to the bottom of it.
That is 100%. And see you in our episodes. They already come out practically every other day. And on our weekly podcast, of course, every Sunday — do not forget to support our channel, like it and subscribe, and write comments. All of this lets us develop, and lets as many people as possible see us. That is always especially difficult in our times. Bye everyone.