Hi everyone! You're on ToTheMoon, tech news and insights from Silicon Valley and around the world. And this is our Sunday episode. We're finally back to the four-person format. We did our first fifty episodes or so as a foursome, I think, without missing one. And I think the format with four people in the room is great. Ilnar and I never miss one. So, how many years has it been now? Where do I want to start? Especially with this lineup. Just yesterday the head of Google posted that a new model, Gemini Carbon, is out, and Elon Musk reacted to it too, saying it's very cool.
By the way, I still wonder why Elon Musk keeps reposting the head of Google, him specifically. He always reposts him when he puts out some analysis. In that analysis, which our editors will show on screen, Carbon is presented as a much stronger system than Fable, and Anthropic's Opus 5.5, and than, well, 6 Astra. I don't know how true that is. We'll test it, we'll look, we'll see what happens. Don't forget to support our channel, and especially our foursome today: leave a comment and give us a like.
So which topic do we start with? Max, Max and I recently recorded a series of episodes, and Max voiced a very interesting opinion, by the way, one you don't often hear in the market. And I even quietly agreed with him. I'm known for saying that nobody in the world is catching up or can catch up, and for the past three years I've been saying it about the American market. Because a lot of people said Europe would catch up with it, the Arabs would catch up, Russia would catch up.
And I kept telling everyone: “Not a chance, it's impossible with that kind of infrastructure, that kind of money, those projects, those people, it can't be done.” And yet we know perfectly well that China seems to be very close, isn't falling behind, and keeps releasing all sorts of great models. For example, when Moonshot Kimi K3 came out two months ago, they wrote to me themselves asking me to talk about it. They said: “Here's our model, it's on the level of Fable from Anthropic and Astra from OpenAI.” I smiled and thought: “Really, is that so?” And now a few months have passed, and somehow we don't see any enthusiastic comparisons of that model with the same Fable, which has been working beautifully since the summer.
And now Kimi has just shipped the 3.1 update. It's a cool company, a cool model, a cool system. But still, Max, if I understood you correctly back then, your view was that the Chinese market won't be able to do what America is doing. So it's not that they're, well, a couple of months behind, but that they wouldn't be able to do what America does. That's it. And since both Moonshot and DeepSeek have just released updates, and it's being widely discussed anyway, I'd like to debate that.
Well, I didn't say they can't do it. Anything is always possible, of course. Nothing is impossible, just as no outcome is ever guaranteed either. So of course they may eventually do it, but here's the problem. China, and Asia in general, rarely lead in research. It's a big, big exception when something genuinely new is created in Asia. So there are different opinions about why that happens, but I think it's simply a cultural thing, really. Asians are taught to copy things. In school, at university, and so on. There's this thing, they, they… it may have to do with pictographic writing.
We'll wander far into philosophy now, but the point is that from day one at school children are taught: right, come on, copy this character correctly, copy this poem correctly. It's not even that the characters have to be copied correctly, even the strokes of the characters have to be copied correctly. And people do well at school if they copy well, right? Well, that's a kind of philosophy.
In fact, if we look at what has happened over the last five years or so in AI, quite serious Chinese models have appeared. DeepSeek first, then Qwen, then Kimi. But if we look at their performance, as you rightly noted, they are still catch-up models, they appear later, and they are closer to the performance of models from American frontier labs that are six months, nine months old and so on. Why does that happen? Let's put Kimi aside for now. The other models' architecture is basically quite understandable; it usually more or less mirrors what's happening in the American segment.
But it usually lags behind: some model comes out, and while it's being reverse-engineered, while enough information is extracted about how the frontier lab built it, some time passes, and that information finally reaches a Chinese frontier lab, using the tricks with which OpenAI has been building a new model since last year. But the most important thing is still training. A large share of the data at this stage is already synthetic. And generating that synthetic data is actually very hard. Synthetic data is quite difficult to generate.
Could you explain that a bit? Explain why there's a problem, why they have a problem with training?
Well, that's what I'm explaining. The problem is, first of all, compute. Training is very… pre-training and post-training are pretty serious things, right? In terms of compute. The Chinese simply don't have access to that many GPUs. The export sanctions are still in force, right? So the Chinese can't just go and buy NVIDIA GPUs. Although there was a story that NVIDIA was supposedly selling cut-down GPUs to China. And obviously they can get around the sanctions by buying GPUs through third countries, through the Arabs, say, right?
But it's still much harder than for OpenAI or Anthropic, who go straight to NVIDIA and say: “Come on, sell us five hundred thousand GPUs,” right? The Chinese can't do that, no matter how much money they have. So training, the number of experiments they can run, is limited. The amount of training they can do is limited. And most importantly, the amount of synthetic data they can generate is limited. And synthetic data at this stage is what moves us forward. The models we have are so large that they need so much training data that the Internet is no longer enough, especially for specific tasks such as code generation and math, right?
There has never been enough data. I remember very well how in 2022 and 2023 OpenAI and Anthropic, among others, were running around like crazy looking for training data to teach a model to code. It was a very, very serious race. And the deal OpenAI made with Microsoft back then was aimed, among other things, at getting access to GitHub, to GitHub's data, right? But that data still isn't enough, it's fairly... You need somewhat different data. And in the end all the frontier labs took up generating synthetic data.
It's a very serious process. Synthetic data is hard, it's very expensive, especially carefully checked, high-quality synthetic data like that. And that's why training models, even if you have all the GPUs, which they don't, and their situation is far more constrained, even if you have a cutting-edge, the very latest architecture, which, well, except for Kimi again, we'll talk about Kimi, the Chinese still don't have, they don't have those architectures. So they're behind one way or another. Frontier labs don't publish their latest findings. And to find some new idea you have to run a ton of experiments. What is an experiment?
An experiment is engineers multiplied by compute multiplied by time, right? The Chinese have time, obviously. They basically have the engineers too by now. But compute is still the limiting factor, a very seriously limiting factor. So even to build a new architecture you need, well, you can't run those experiments simply because compute is limited. So. And, well, the main thing is data. And what do the Chinese end up doing? Why do their models still work very well? By all appearances, it's distillation.
Distillation is extracting intelligence, as it were, from one model into another. It's a set of techniques that let you, roughly speaking, ask a model questions and train on its answers. It's essentially synthetic data made with another model. And there was even news recently that Anthropic shut down some incredible, tens of thousands of accounts, some incredible number of accounts that apparently belonged to Chinese botnets that were busy distilling its models. So the Chinese simply spend the money they certainly have on pulling the intelligence out of American models and training on it, roughly speaking, on that data.
So they...
But by pulling out, pulling out intelligence you can't actually create an architecture. You're only copying some part of it.
No, well, the architecture, sure. Their architecture is some fixed one, but you have to train it on something, and with distillation you can cut the time you need to train a model very, very much, especially post-training. So distillation lets you, well, it's a kind of shortcut that lets you jump ahead quite quickly. Here's the problem. With distillation you'll never surpass the quality of the model you're distilling.
Of course.
So it lets you, for example, what was distillation used for before? You could take a big model, train it well, and then, by distilling it, make a smaller model of much, well, roughly the same quality, with almost no loss in how well the model works. That's what distillation was used for in the 2010s. But right now it seems that the Chinese, at least in my opinion, are mainly doing distillation. With it they're closing this, this gap, right? But what does that mean? You won't overtake the model; you can keep shrinking the gap forever, but you won't overtake the model you're distilling. And what does that mean?
It means that they are essentially building analogues of models that have already been released. The models frontier labs are building and training right now, the Chinese naturally have no way to distill, no way to look at their architecture, and in the end… their target for building is a model that has already come out. That's why they're three months, six months, nine months behind. And until they change their approach to building these models, especially to training them, they won't overtake the American frontier labs.
I'd push back a little from the other side here. I'll play for the other side today, for the Chinese a bit. Or rather for the middle class of users. Because on the one hand, I fully agree that the Chinese have nothing to do at the frontier for now, they're behind and most likely they really are catching up, copying, distilling. But remember, about six months ago we already started saying that even if model development stays where it is now, we can get huge value out of it, at work and by building it into whatever is already going on around you.
And the latest Kimi benchmarks that came out, if I remember correctly, compare it with GPT-5.6 Sol and it even beats it slightly. Yes, it's not Astra, not 6.1 Sol, it's something from three months ago or even earlier. But it's already a level from which you can get real value in some work tasks, in everyday life or elsewhere. On the other hand, OpenAI is starting to cut limits. Soon, maybe by the time this video comes out, if the marathon of unprecedented generosity with these resets isn't extended, we'll get to where the $200 subscription gives ten X instead of twenty.
And that means you can get less value from these models. So it turns out that ordinary average users are quietly being nudged to install some open code or other and start using DeepSeek Com 4.1 Flash, Kimi 3.1 Pro and that whole zoo of models. And start, for example, managing: we break some tasks down with Claude or ChatGPT, but given the limited token allowance and so on, we start implementing those tasks with Chinese models, which may be three months behind. But doing it is insanely cheap. Hitting DeepSeek or Kimi via the API is much cheaper than the same 5.6 Sol.
Not to mention frontier models. At the frontier, sure, we won't catch up, but it seems to me that they will still take their own market, at least within the compute they can cover. And the big companies can't digest all the demand there is yet, so they themselves are nudging people a little in that direction. I mean, I've already installed open code and started testing these models on different tasks, getting ready for the moment when Tibo stops throwing resets into Codex every two or three days.
Three years ago I had all sorts of long debates. I kept saying that countries are limited by infrastructure, and whatever they did, it seemed China, well, China should have invented its own chips by now. And by the way, what haven't they invented there! They have, well, simply an incredible number of super-cool chips. And yet here's Max, in 2026, in October, saying: “They're still no match for NVIDIA chips, right?” Well, by chips I mean various infrastructure solutions, so it's still...
...even though they win on the amount of electricity they have, they clearly win on the sheer quantity of people they have. So it seems China can learn and generate a lot more data, even human data. Well, if you look into it, Max and I have a mutual friend, and you both know him, Zhenya in China. And I remember visiting him in 2011, and Zhenya says: “Do you know what the Chinese government did to render every city in 3D?” That was fifteen years ago. They simply took sixty thousand people per city to sit and draw 3D maps. And they drew the 3D map. So they did it through volume.
And you arrive and you see that strong result. But when it comes to frontier quality, to what's being built in the future as AGI or ASI, it seems China isn't going there, and it seems they won't have billions of users. By the way, Google reported a billion users in Gemini. Obviously that's not, not a billion users. Yesterday I had a debate with Dima, a friend who came over to visit, and I told him: “You're really a zero-level AI user.” He got offended and watched some interview with Greg Brockman from OpenAI. Well, he's the number two after Sam Altman, you know that, right?
And he watched the interview and says: “Greg calls people like me advanced users. These are people who use ChatGPT in more than three or four use cases.” I say: “Well, I don't call those people advanced users.”
Tell him about Karpathy's funnel. Remember, we discussed it, he drew that funnel back then?
So, so obviously this billion Google users are very simple users, like a billion users in a messenger or on social networks, on Instagram, at, I don't know, Meta, billions of people. But they're still billions of users. And China in total, with all that access and advantage, has, I think, half a billion registrations across all its models right now. And even then we'll never know the truth. And I also have a big question about Chinese models: are you ready to upload your data there?
No, we use Chinese models more for some training things. Well, when we see that, for example, for some simulation or for training it's much cheaper. As, as training, not as some project request or project need. Well, again, I don't have information on every project, but by and large.
And if inside some services, Sasha, you use LLMs, well, in business at least, from what I see, Qwen in various versions is used a lot. The ones you can deploy locally in the quantity you need without letting data leave your perimeter. So here I'd still separate personal use, on a subscription and via the API, from use inside big corporations, a bit like Windows and Linux. I don't have the information...
Obviously! Look, it's obvious here, it's like NVIDIA right now: they've had an update, they released new models for translation, for turning audio into text. And that's a big achievement for them right now. For me, as a regular professional user, it's like: what translation? OpenAI has translation, Google has translation, ElevenLabs has super-cool translation. They've simply shot ahead in quality. ElevenLabs has managed to beat Google in many respects in translation, in comparisons, well, in this technology.
But I understand perfectly well that a huge number of companies that will set up NVIDIA infrastructure locally will use their segment. So let's continue in terms of... And our viewers, it'd be interesting to know which of you connect Chinese models, who has them? We have such a big audience. Of course, as someone who really loves to broadcast to the masses, to billions, I have the feeling that China won't be among those billions. But we'll see, again, we'll see. It's just that every day you read headlines like: look, a model was just invented at Stanford, or the Chinese invented a model.
Yes, of course they do. But I'd still make a distinction. Ilnar rightly noted that there are a lot of Chinese models on-premise, as it's called. But why? Because the top frontier labs simply don't want to give away their weights. That's all. The Chinese publish weights only because it's the only way to attract attention and attract funding. They're popular only because they're open source. Because if they were closed, nobody would use them at all. Why would you? If you have free access to ChatGPT or to Anthropic, or Google, why would you use Qwen or DeepSeek?
They still don't work that well. Although there is one aspect that also pushes people in that direction.
It's not only on-premise, not only access to the weights, it's also censorship. All these models, you've probably talked about this many times, but they're censored quite heavily. Frontier labs censor their models' output very heavily. For safety reasons, for simply political reasons and so on. Censorship is built into them quite deeply. And Chinese models are censored, but in a different sense. So if you don't care about Chinese Communist Party politics, you can get access to prompts that would be censored in Western models.
So. And for some people that's also a fairly serious reason to use Chinese models or open-source models, where they can remove these systems that are generally called AI safety but are actually a form of censorship. Although, for example, Grok: I know that many people who don't trust, who are on the political right in America, use Grok because Grok is less censored.
But that's all a philosophical conversation, really.
Because Grok is completely unscrupulous, it does things that, I don't know, OpenAI's image generator or some other one won't do, legal or illegal things. Anyway, some of my acquaintances who use Grok admire its lack of scruples and that it has a much more relaxed attitude to the restrictions imposed on models like these.
But it still has its own particular influence, really, and it has its own tuning. We've had whole episodes where we showed comparisons on a fairly large set of studied answers: how Grok behaves, how ChatGPT behaves, how they make decisions, whom they consider a villain and whom they don't. Like the question, remember, whether the US should have started a war with Iran? And there are whole breakdowns of how Gemini answers, how ChatGPT answers, how Claude answers, how Grok answers. And sometimes you can see the creators in it, and where everyone stands.
I have one simple thing about Chinese models, personally. It's data security, and I trust American data security much more, the most, than Chinese. Well, Chinese security probably comes from a different side altogether. And Chinese data security can be understood roughly like this: if you're hiding from some other country, then you can, maybe you can hide there. Because there the data is relatively safe and so is your data. But if I untick the “don't train on my data” box in a Chinese model, I don't believe at all that it actually happens.
So there. And that they aren't checked. No, it's just that I know how that filtering works...
I have friends living there, I know how the filters work there. You send the wrong message on WeChat, to the wrong...
Sasha, and do you believe in that checkbox in ChatGPT?
I do! In ChatGPT I believe in that checkbox, for one simple reason. If you ticked the box and later it surfaces somewhere, you'll either become a millionaire, or at the very least that data can't be used or applied against you later. I fully believe in it, because it's very seriously controlled. And by the way, I believe in it much more than in Europe. Take Europe, for example: the European personal data law is, on the one hand, kind of nice, because supposedly everything is protected there. I just don't know how.
And by the way, I want to say that the Russian personal data law is closer to Europe's than to other countries in the world. There is a serious, strict personal data law. And at some point the Yarovaya law that was passed in Russia, and various other laws, they provided, they wanted, look, it looked like protection of people's personal data, but in reality it's state control. And what you see in Europe, and here I agree with Elon Musk, is not people's privacy, it's state control. And by the way, Durov talks about this a lot on Telegram.
I can't say Durov's ideas and his worldview are close to me, but I'm close to what he says: he has said many times that Europe demands that we open everything to them at any moment. So supposedly you're protecting people, but in fact they demand access. In the States it's all very simple. You don't have to click “allow cookies” every time. It's the damn European law that made this crap happen all over the world. You constantly have to agree to cookies, right? And by the way, in Russia you also have to ask for that consent.
Whereas in America, for example, you can just have a little “Opt out” button at the bottom of the site. The user has to opt out himself if he doesn't want things saved or looked at somewhere. On the one hand, this is supposed to ensure there's no, I don't know, scams or information leaks. On the other hand, I'm sure, for example, that there are big questions about Mistral. By the way, new modern systems like ChatGPT Dots are new, next-level systems, a new approach. It's no longer just a model, no longer just a session with a context window, it's something, well, I'd even compare it, there was an episode two days ago, by the way, I filmed, for those who haven't seen it, go watch it, I filmed about Claude Duet, ChatGPT Dots, Grok Bot and Meta Muse.
And the Grok story, for me Grok, oops, for me ChatGPT Dots is exactly the story of a new era of artificial intelligence, where copying something... And it'll be harder and harder for the Chinese market to fit, because I personally am starting to choose the infrastructure I work with. Simply the infrastructure I live with. I'm such an Apple fan, that's it, I've stopped comparing Apple by, I don't know, screen size, weight, amount of memory and everything else.
And many countries too, I only checked that DeepSeek can be downloaded in Poland, but in a number of countries there were definitely requests, I mean European ones, to remove DeepSeek from apps altogether, from all the stores.
Well, Europe for sure, they have to. Europe definitely has to...
Because of data storage...
Of course!
...and opening it up to all sorts of intelligence agencies and so on. In particular, in Poland you can download it, but I don't know what restrictions other countries have today, Italy, Germany, for example.
And there's a whole debate. Max was talking about NVIDIA, that chip shipments were allowed. There's a counterweight, the camp of Jensen Huang and Amodei. Jensen Huang says we should be friends with China, supply them and build things. Amodei says: “Screw it, lock everything down and push forward, develop our own systems, or they'll overtake us.” Yes, and that debate is endless. But I don't think the American government will ever release its coolest technologies to anyone outside America, and to China in particular.
I think viewers will find it very interesting, because they're very prominent in the market, and they've just released a new update.
Yes, Kimi is a very interesting story. It's almost, probably, the first time, well, Qwen did it in some sense already. Kimi is a very serious evolution of the models' architecture. Kimi used know-how that, well, we certainly don't know what's used inside frontier labs, American frontier labs. But Kimi as a model is quite seriously ahead of everything else around, at least architecturally. It uses Kimi Delta Attention. That's a separate story too, it's a different type of this attention mechanism, which apparently improved performance very, very seriously.
Then they use this attention throughout the whole model, not only on the input sequence. Sorry, I'll say some things in English. Anyway, they applied several pieces of know-how that no one before them had applied in open models. And what does that tell us? It tells us that either they somehow managed to extract this knowledge from Western frontier models, or they managed to do their own research showing that these things work. And, as I already said, research is pretty hard to do, because there's simply no compute.
To do research in this area you need to run lots and lots of experiments. You have an idea, say, a new type of attention. An idea of how to make a new type of attention: to prove to yourself and everyone else that it works, you need to train a small model which shows that with this mechanism... models like this work better. But most likely you won't get all the parameters right at once, so you need to train dozens of such models, maybe even hundreds. So you need to run hundreds of small experiments. All of that needs compute. So.
And if you copy someone else's architecture, then someone has already done that work for you. You look and go: aha! The way everyone copied the Transformer back in the day. Finding the Transformer architecture, coming up with it and making it work, Google actually spent an endless amount of money on that. It was very, very expensive. But everyone else got the Transformer for free, roughly speaking, right? They just read the paper and went: “Oh, great! We'll use this thing.” To come up with an analogue of something like that, for example the same Kimi Delta Attention, to invent it, it was either copied or it's your own research.
And that research is very, very expensive, and it may even be out of reach, because you may simply not have access to a computer. The third option of what's happening is that this research actually hasn't been carried through to the end. You can release some architecture and announce: “Oh, look how great it all works!” Take Kimi Delta Attention, for example. Look how great it all works. Here's our paper, in it we show so-called ablations. We show how much this Kimi Delta Attention improves the architecture's behaviour. We train two architectures side by side.
One uses some old Delta Attention, the other uses Kimi Delta Attention. And look, we perform fifteen percent better, we solve tasks fifteen percent better. So. What's the problem with that? Maybe it really does work like that, but running that research and confirming it costs a lot of money. And frontier labs can afford to experiment with ideas like that. But we, the public, have no access to information about whether these ideas really work. Kimi claims they work. But how correctly those experiments were run, how much Kimi Delta Attention actually changed the whole picture, we don't know and can't know.
So, someone has to run these experiments in open source and confirm that this Kimi Delta Attention really helps. How much these architectural improvements released by Chinese labs, Moonshot in particular, help the system, as opposed to it simply being distillation of American models, is a big question. When Kimi 3 came out, there was a lot of fanfare. We even have a kind of study circle for picking apart all sorts of technical ideas. We sat for quite a while figuring out how this model works. It's all quite elaborate. It's a very, very advanced model.
But how much the ideas implemented there actually help, and what percentage of this model's performance and benchmark success is distillation of American models? That I can't answer for you. It may even be...
But why? Why, sorry, let me cut in. Why couldn't other countries build a system for catching up with other models? Because what has essentially happened? It's one thing that you once failed to catch up with a social network. Fine, that's more or less OK. You didn't catch up with a messenger. Fine, more or less. Making an operating system was once impossible, it was something unique. It wasn't clear where it was made, who made it, how they made it. There were very few documents, you couldn't read about it, right? Then, for example, search engines like Google showed that they simply took the market.
And once you've taken the market, it's hard to enter it. And here we see the emergence of companies like OpenAI on the market, ChatGPT-like systems. I'll note, by the way, that in all the films and all the talk shows in the US, everywhere, people say: “I used ChatGPT.” Nobody says: “I used Claude.” Nobody says: “I used Gemini.” That just doesn't happen. Nobody says they used Grok. That's it. But I'm telling you, right now on a mass scale, if you look at the mass scale...
Listen, I see very heavy use of that slang in terms of ChatGPT, right? But we'll see what happens. But what's the point? That they did it right away. They say the ones who usually make money aren't the ones who invented something, but whoever comes after them. And so far we see that Anthropic is about to IPO, and according to the plans it will already be worth more than two trillion dollars. Clearly these companies will be worth a trillion dollars, most likely more than ten trillion. And the question is why other countries, with such openness, in any case, with developed engineering,
descriptions of engineering problems around the world, with the sheer number of people in general who are capable of creating something, when they say: “Oh, we'll copy that or build it.” Why hasn't it appeared in South America, in Canada, in Mexico, in Russia, in Belarus, in Ukraine, in Poland, everywhere, in every country, why haven't such cool models appeared, even with minimal infrastructure or a minimal number of boards? Why isn't there any? Proper systems. I'm talking about serious systems now. Thousands of micro models of some nonsense have appeared, that's fine. I'm talking about more or less proper, serious systems. Why aren't there any?
Like Mistral. Look, in Europe I know one serious model, and I think it's the only one in the rankings. That's Mistral. And on top of that, the European government is unhappy with how Mistral is developing. And what we see is that everything is being done at the level of the European government to block other models. In particular, the American DOTS isn't present there. Well, first of all it isn't there because it's a very heavy infrastructure task. It seems like just some bot, but this bot is very demanding in terms of compute, so it's only being rolled out to Pro subscribers and so on.
Even with Claude Duet you still have to join a waitlist. So that's the first story. And the second story, of course, is that I'm not sure the Americans, American companies, want to give their super-cool technology away. And the third story is that the Europeans themselves will block it. Because, by the way, DOTS, I think, is blocked in the world only at the level of the European Union. Precisely a block, well, by governments.
Sasha, you can't just release a model and stop there. Take Mistral, or the Saudis, I think, have Falcon and some other models. There are tons of them. There was a moment when it wasn't that hard to do, but you release it once, and two days later it's lying on the back burner, nobody needs it. You constantly have to factor in this race. You have to take part.
So why aren't they, why aren't they, why isn't it working out for them?
Sasha, if you had to, if you had to answer this question in one word, which word would you choose?
If the question's for me? I was just listening at first. Insanely expensive. That's it.
Yes, yes.
Money.
If you answer in two words.
And now the second story. Take Volozh, for example. He built Yandex, and now in Europe he's building Nebius. And what does he do? He sells compute. Just recently, about a week ago or so, DeepSeek published statistics showing that they run their models on their own small number of GPUs, well, a relatively small number of GPUs. And one GPU brings in $38 an hour. Just from people hitting DeepSeek's API at dirt-cheap prices, they make $38 an hour. And now let's take two businessmen. One has to pour in a huge amount of money to keep running.
Your compute has to keep growing just so nobody overtakes you. You're constantly in a huge cash gap, you have to pour in a huge amount of money. That's why OpenAI is, roughly speaking, loss-making, Anthropic is loss-making, they're constantly raising money. On the other hand, you can build infrastructure and sell that infrastructure. And, well, you bought today, I don't know, H100s, B200s or whatever cards there are, made a cluster out of them, and they'll definitely pay for themselves.
You won't need a huge number of scientists developing the frontier algorithm. Yes, there's its own complicated engineering story, but, like, Elon Musk put together, how much did they have? In 19 days or however long, they put several thousand cards into one cluster. And that's it, that's all. And then it just runs. Then they sell that compute, to whom? They sell it to Anthropic, for huge money, to someone else as well. And that's a different business. So you have two sides. And if you can sustain this race, well, take China, it has sort of gone in that direction.
If you can't sustain it, well, you release a Mistral once a year just to have something, distill something, some of your own work here, something else there, but you won't have the capacity to run at the same speed to stay at that level.
Well, so it turns out you have access both to cool technology and to infrastructure. That's what we were saying from the very start. On the other hand, there's also money... so in general everyone gets cut off, and China will soon be cut off too. Given that Elon Musk said SpaceX, SpaceX could be worth $100 trillion. That's the amount of money that American companies and American AI companies will have in the market. Google, by the way, announced, yesterday I think, they had their earnings, they grew like hell. Another mountain of money.
And it turns out that operators like xAI, well, in this case that's SpaceX, right? Google, OpenAI and Anthropic, they, well, there's also Meta, but I think Meta has fallen seriously behind in this race and this game. But they have a hell of a lot of money too, right? A year ago Zuckerberg said that if we spend $250 billion and get nothing, no big deal. The main thing is that you have to try spending, right? So who can be near these companies? There's Apple too, obviously, there's Microsoft.
I don't know what will happen with these operators in terms of artificial intelligence. Or they'll simply be built into the general infrastructure. Well, obviously there's NVIDIA, but the amount of money there, let's put it this way, NVIDIA is worth $5 trillion. $5 trillion. Who can catch up with that at all?
Look, Sasha, in short, this business looks roughly like Ilnar explained, really. You have compute that you sell to users in some form. Compute is the bread on which you spread the butter of your model, and you sell a sandwich. Right? The tastier the sandwich, the more users will buy that... I mean, everyone has the same bread, let's say, right? So what kind of butter you spread on it changes, sort of, changes the likelihood that you'll sell your sandwich. And it turns out that there's no point going to buy. Say you have two sandwich sellers, you go, of course, the sandwiches cost about the same.
You go to the one whose sandwiches are tastier, and 99% of people will go to the one whose sandwiches are tastier. That person takes the money he's been brought, turns around, and if he doesn't break his internal research system, he'll invent even better butter, spread it, put it out and sell even more sandwiches. In the end the winner sort of takes the whole market. It's a very typical story. There's a really great book by Peter Thiel called Zero to One, in which one of the main, central ideas is that every company builds a monopoly.
That's sort of the point of how capitalism works. Every company tries to build a monopoly. And this mechanism of building a monopoly looks exactly like this. So in the end, if everything is built right, all the money ends up with one, with one company, especially when there's no competition. And in the end all your other rivals in the market simply die out, just because they don't have the money to do better research. Their butter is even less tasty than your new butter. And that's it, you've won. Your sandwiches won in the end.
Well, then models made in other countries that aren't in the top ten, or maybe even the top five, even the top five, maybe even the top three, will die out anyway.
There's one...
Because at some point developing such systems, even like DeepSeek, will become more and more expensive and difficult, and harder and harder to catch up. And at some point you won't be able to... Even now, if you look at the infrastructure, let's be honest. Today an employee of mine sent me a message from Dots. She says: “I don't understand at all, this is just the level the system works at.” She says: “I have Dots running, I didn't ask Dots to do anything, and today Dots writes to me: you have episodes coming out next week on ToTheMoon and on various channels.
Note that they come out at eight in the morning such-and-such time. But since the clocks are changing in the US, your speaker is here and there, change the time.” She says: “I didn't explain anything to it, this isn't anywhere in the email. But it started telling you about it.” That's an insane resource. It's a system that has to follow your chats, your history. It has, it must have a memory of you. So it already has to create, to have a memory of you as a person or of your business.
Those are details. Let's go back to the general picture. I was trying to talk about the general system of how this business works. So there's one, one very important nuance. Why do I think the Chinese will still have pretty decent models? For a very simple reason. If you go to China and try to open OpenAI, ChatGPT there...
That's true.
Or, or Anthropic, or Claude, Grok, you won't even be able to get access to their websites. You try a VPN, and all these things detect VPNs very well. VPN detection works really well these days. If you try to open Claude or Gemini from an ordinary VPN, you won't get anywhere. And even if you later find a VPN that somehow lets you open these models, you have to pay for them somehow if you want access to the higher tiers, and a Chinese credit card won't let you do that. So in the end you're completely cut off from this...
Max, how is that connected to them not falling behind? Because you're now talking about a country being restricted for its residents, not restricted for its residents.
Because you have 1.4 billion people who will still need this technology, and who have money.
Not may, but definitely.
...yes, ideas and everything else, and it will stay there as it is. And there are other countries that are cut off in the opposite direction.
Absolutely.
...including the Chinese and a huge number of countries. For example, in Russia you can't just open Codex and use Codex. And that's Anthropic's decision, or Claude's, it's Anthropic's decision. They know how to block everyone, knowing that they're in that country. And today, of course, there's tons of diagnostics, it's not even a question of VPN diagnostics. You'll simply be diagnosed by the questions you ask. At some point it'll say: “Listen, listen, we've determined that you're a resident of such-and-such country; if you want to confirm, send us your other documents.” And on top of that, what bothers me, for example, is how it works. I arrive in Europe, and in Europe ChatGPT starts working differently.
And Dots, by the way, I talked about it the day before yesterday in an episode, for those who haven't seen it. Dima came over, connected to my Wi-Fi, and Dots appeared for him. He switches on his Singapore SIM card, and Dots disappeared inside the app. The app stayed, it works the same way, but Dots disappeared. And who's to say the model didn't start answering differently and talking to you differently, with local censorship?
Your location is definitely in the prompt. Your location is in the system prompt, at least as far as these companies have managed to detect it correctly.
And who's going to stop Claude or ChatGPT from requiring, in the app, that for the app to work you grant access to the phone's location? So there's definitely segmentation already happening.
And do they have the legal right to do that, by the way?
Of course.
Really?
They can refuse you service for any reason. These systems don't have the status of a so-called utility. In America there are utility laws. So a company that sells electricity can't refuse to sell you electricity. But that's a big rarity. Most, in most situations companies here can of course refuse you their services. Any bank can close your account. Any, I don't know, any restaurant can show you the door. That's completely normal. ChatGPT can simply say: “We don't want to provide our services to you unless you do this, this, this and this again.” That's why open-source models are very, very important, actually. And the fact that we don't have American open-source models, well, they exist, but naturally they lag far behind, even behind the Chinese ones.
That's a big, big problem. And
You think, you support the idea that it's a problem, right? There's a whole market.
Oh, a hundred percent!
Some leaders support it, and some leaders say: it doesn't matter.
No, those are leaders, leaders who make money from the current situation, naturally they won't support open-source models, because one way or another it will, sort of, it will take away a bit of their market one way or another, right? Because sooner or later solutions will emerge that work well with open-source models, and work well enough. And that means losing part of the market. Take Microsoft, or… in the eighties and nineties there were tons of vendors of UNIX-based systems.
Very often these were some system integrators who simply sold you big servers running some flavour, usually their own in-house flavour, of UNIX. And banks, large corporations ran on exactly these, they were called mainframes. And after, after Linux appeared in the nineties and, essentially, the ability to build your own mainframe, which ran on one or several cheap computers using a free operating system, all those integrators died out, they simply ceased to exist, right? And the industry learned that lesson very, very well.
And that's why everyone sort of loves open source in words, but in practice big corporations watch very carefully that their know-how doesn't leak into open source.
And that this, this butter that gets spread on, on this bread, on bread that's more or less freely available, right? Anyone can buy an Nvidia GPU, but you'll never get the software and weights of the frontier labs' models. And everyone watches very carefully to keep it that way. Especially the market leaders.
Interesting topic, of course.
Yes. Market segmentation is very, very interesting.
Yes, and about GPUs, Sasha, you were saying there's, say, Jensen Huang, who says let's be friends with China and sell to them. And on the other side, Amodei, who says let's not sell them anything at all. There won't be a situation where China's access to GPUs is completely shut off. Well, that would lead to serious geopolitical problems around Taiwan, when it becomes an existential problem, that the whole world, or some part of the world, is developing AI and you can't, just because they're produced right next door and you can't reach them.
Well yes, yes, yes. I think in this day and age, in this day and age that should clearly have been a well-solved thing.
That's why cards released for other regions still get smuggled in through Malaysia, Indonesia, the Saudis. Yes, Maxim.
I think, Ilnar, you're mostly right. I think the motivation for this letting off of pressure is a bit different. If you completely cut off the Chinese from hardware, you can ban not only GPU shipments, you can ban CPU shipments, memory, all of it. What will that lead to? It'll lead to the Chinese seriously pulling together and building their own technologies after all, which will start to compete, compete in quality with what we have. If you still leave some possibility of supply — here, here it's also an economic race, in a sense — you can let them still buy these cards, not, not, not directly, but indirectly, right?
But in the end make it a little more expensive. If you buy through third countries, the GPUs cost you more than if you buy them directly from the manufacturer, naturally, right? And it takes a bit more time. And what does that mean? It means that companies buying these cards are less competitive than companies that can buy directly. That's all. That's actually enough for you. This small friction, this extra lack of simplicity in doing business, makes you less competitive, but it's still a cheaper and simpler option than building your own.
And in the end you can balance like that, keeping your competitor in an unclear situation where building their own seems a bit too expensive. You can get a solution from outside more cheaply, but that solution costs them more than it costs you, and so you're more competitive. That's an economic trick, actually a cunning one, and it's been used many times. And the same story exists with Russia, right? We may drift into another topic now, but Russia, Russia uses a similar trick to get, to buy practically everything right now.
A basic control chip for, say, some washing machine, right? Russia now has to buy it through third countries. Even basic electronics have to be bought on the side, because direct import is banned, and everyone understands this is happening, and one could try to shut it down, right? But actually, why? Why not? You still let Russia buy all of this, but you increase the burden on their economy. Any competition now, war is a form of competition, is above all economic, however you look at it. Any competitive moments are won by creating frictions in the economy.
And Russia, Russia feels it. The amount of money they spend on buying technology has grown very significantly, right? And that's a burden on the economy, a burden on the country's political system, partly because of this. And the same, the same trick is being used against China. You simply create these small frictions, they accumulate, and in the end your system becomes less competitive. And you can't just overtake the players who are ahead.
Although, of course, it's a huge plus for China what they managed to do with their own technology market over fifteen years and step away from global competition. Let's put it this way: if China hadn't done that, they wouldn't have their own model now, right? And Google Search would be there, I mean, Google Search would be there just as it was before, and it might be the leader rather than their domestic system. There would be messengers owned by American corporations rather than their own, right? But there, there are essentially no messengers from American corporations. Essentially none.
Well, note that they made exactly the move I'm talking about. They switched those systems off completely: social networks, YouTube, then search and so on. They switched them off completely and as a result created competitive systems, because there were no other options. What was created wasn't friction, that is, well, this was done by the Chinese side, what was created wasn't friction but a total, total blockade, as a result of which they have their own decent search engines and, well, social networks. TikTok is a Chinese technology, right, which was...
It even started taking, even in America it started taking...
Yes, yes, yes.
Taking positions, and they started taking positions worldwide in general.
They have their own YouTube equivalent too.
Yes, yes, yes, absolutely. So.
There were lots of discussions on this topic. Should Russia have blocked Google, say, in 2005, and stayed, and built its own systems, right? Or should it have blocked messengers internally, stayed exclusively on its own messengers, so that they weren't present. Right now fifty, forty-five percent of the search market is still with Google, with absolutely uncontrolled traffic in terms of the technologies that exist separately and so on. That is, of course, that is, of course, also the advertising money that was invested there; it was money invested in another country's advertising.
There's a difference. The Russian market is ten times smaller and much less wealthy. The Chinese can afford to isolate themselves and build a full-fledged solution that competes with American solutions. Well, Russian social networks were always isolated simply for conceptual reasons. Russians don't use Facebook simply because the culture is different, the language doesn't fit and so on and so forth, right? Russians sat on their own social networks, but VKontakte can't compare with Facebook in quality.
No, time spent in Russia, time spent per user on Instagram in Russia is thirty times higher, I think, thirty or fifty, I don't know, than on VKontakte. So in any case the migration happened, yes, the migration happened at some point. And what I want to say is: viewers, please write in who has got access to Claude Duet.
And now it will roll out more and more. I'd like to hear. Today we went into a serious, serious discussion about the availability of models in the market and countries' supremacy over each other. Ilnar, I think we had discussions like this in probably the first twenty or thirty ToTheMoon episodes. It was one of the main topics back then. Who would pull it off, right? Models were only just starting then. Everything was only just beginning then, everything was coming in. We discussed this topic a lot, because people in the market often said, and to this day, by the way, a huge number of commentators say that OpenAI will die, that they have a negative model, a financial one.
I'm sitting here smiling right now. Tibo has put out a post on X, and they're going by the saying: how to make things good is to first make them worse and then put them back as they were. They doubled the limits on all subscriptions. Now on the two-hundred, on the $200 plan it'll be twenty X again. And on the five hundred you get twice what was promised, fifty X, by the looks of it. That's it, I'm quietly breathing out. I'd been waiting for some kind of stunt.
By the way, one of the topics, one of the topics that's being discussed right now, I think, is that we're entering a zone where over the next six months it'll be a fight for permanence.
Who, with this billion users who've touched AI, well, ChatGPT-like systems, who will come out on top? Like there's Instagram or there's WhatsApp, who gets a certain dominance, given the emergence of systems like Dots. And now everyone has started changing their interfaces, and Claude merged Cowork and Chat, we knew that before. A completely insane, strange idea, we said so right away. And now a post came out, yesterday I think, someone leaked online that OpenAI plans to merge ChatGPT and Work. And that immediately sparked a huge amount of discussion. What will happen to the limits?
Because let me remind you that among the big systems ChatGPT is the most generous and coolest in terms of free limits, free infrastructure, and Dots, by the way, doesn't count tokens, unlike Anthropic. And you can see that Anthropic does everything in a very organised way in terms of the market of money for companies, for B2B, or for enterprise, and is also developing, probably, its own AGI. Whereas ChatGPT, of course, and OpenAI want to release their products to billions of people, as it were.
And how they'll resolve this dilemma of merging Work and ChatGPT is a big question for me. And by merging — because, Ilnar, if you say they've gone back to the old limits, then of course this... Well, they'll have to burn a new couple of hundred billion dollars on top. My friend and I were talking about this. Imagine if all ChatGPT users, all current customers, bought themselves the $500 Pro plan and used it to the max, like I do. I've burned 65 billion, billion tokens in Codex. 65 billion tokens! OpenAI, well, their servers would burn down, or what? Well, it just doesn't exist.
It's been happening. I was talking about distillation. Anthropic had exactly this problem. They had a large number of accounts that were actually opened by ordinary people and then were either hacked, or bought and so on, a Pro subscription was bought for them, and then it was used in full, to the max. That's partly why they hunted down these users, because they're unprofitable customers, and on top of that they're distilling your, your model. Can you imagine? It's like, as they say in America, a double whammy, right? Naturally they're trying, they're investing a lot of money now to detect such users and shut them off.
So. The problem is that you, you're not that kind of user, they can't cut you off. You, you, you use everything by, by, by the rules. So. Although at some point they'll come up with something too. They'll find a reason why you can be excluded as well. Don't worry.
Well, I think, first of all, Pro users, I think Pro users on the Pro 500 plan number maybe thousands, maybe tens of thousands, certainly not even hundreds of thousands of people right now. And specifically on the 500 plan. And, well, this new plan that came out a month ago. And, well, honestly, for them that load is laughable, right? And what is that load to them? It's not as if an extra 20 million people connected and started pumping them hard and cutting them down. And besides, they're probably interested somewhere in people… they're training on me too. Even though training is turned off, somehow they're still training on what I do.
Well, listen, the main load and the company's main interest will shift towards enterprise, actually. That's where the biggest money is, that's where the most interesting use is, really. And that's where you learn best. Although you usually can't train on enterprise data. The company doesn't want you to train on its data. And what will they learn from you, right? But on enterprise data they'll learn the know-how of how a company works. That's one of...
But that's if it gave them permission to train. And we know that's unlikely to happen.
That's exactly the point. All the security questions there are tied precisely to the fact that enterprise, all the subscriptions work completely in a closed perimeter. And the so-called guarantee that nobody will train on that data.
Yes, well, that's how it works at this stage, but something may change at some point. For example, you could train not directly on this data but only on some part of your data, on your behaviour somehow and so on. Entropy can still gradually leak into the model. You, you can write your terms of service so that a person who reads them will on the one hand think no training on data can happen, but in reality a little bit of entropy will leak through. And you only need a little.
Yes. So, Maxim, Sasha and I have already talked about this and raised this question. What does training a model on data even mean? And when we set this “don't train” restriction, what exactly are we forbidding? Not to mention what guarantees actually exist for how this model works. And do the creators themselves fully understand everything? I mean, there are so many layers here, each of which you can pick apart, and there's definitely no answer to everything.
If the model itself… inside there are tons of cases where they hack their own system and, for example, get outside, then who's to say it isn't training internally on some company's data? Nobody even noticed. But I want to say, by way of counterweight, that I think the main money isn't in enterprise, the main money is in the billions of people on planet Earth. And studying a billion, a billion people and the behaviour of a billion people is much more interesting than studying 100 thousand Google salespeople, how they work and how they sell.
I think in terms of money, in terms of the end result and in terms of adoption. But we'll see, by the way, what they end up making money on and where they find it. For now this business of selling various solutions to enterprise doesn't look commercially right to me at all for companies like OpenAI and Anthropic. That's my, my, my own opinion. If tomorrow they said that we won't focus on ordinary people, build AGI for them, or build AGI or ASI at all, I'd write them off. Developing models for companies doesn't interest me. Although Anthropic, by the way, released this Claude Duet.
They write that it's aimed at companies, aimed at working with Convoy, aimed at working with Stripe, aimed at working with HubSpot, in terms of CRM systems. So it's aimed at companies. But my logic is: why on earth should I aim at working with HubSpot? I'll set up my own internal infrastructure myself with, say, Claude Code. Well, that's the difference. It's very interesting where, how, what will actually happen.
And I also wanted to add that enterprise is very quickly becoming too expensive a service for many businesses at scale. Even without initially understanding how expensive the service is. And if we talk about results and all the returns on investment, that's still an open question. It's still on the surface. Nobody has learned to measure the effectiveness of artificial intelligence. And there's a mass of information saying that there simply is no return, no specific economic indicators tied to growth, today.
So that's why enterprise is such a thing
I switched on Claude and asked: “What can you do?” It says: “I can see your sessions.” I say: “Can you see this session? Change the limit in it.” It says: “I can't change limits in other sessions.” I say: “And how much do you, when you talk to me, don't you spend money?” It says: “I do, I've already spent two dollars.” I'm like: “So you've already spent two dollars?” So. I'd only sent it two messages. I say: “But I'm talking to you on a subscription.” It says: “I just roughly converted the tokens into money for you, right?” And that's why I actually buy extra tokens on top of my Claude subscription.
And every time I buy tokens, I remember Ilnar's model. You just need ten different accounts, right? And, er...
At least two, so that the accuracy of it.
Because, well, because it's, it's unrea...
Just nobody
...unreal! I mean, you understand...
Listen, artificial intelligence is in many ways like an investment in the future and money for the future. But with a business, some corporate account, money is very often needed today, not in the future.
So everyone wants to make money on it.
Yes, a completely different perspective.
So look, Anthropic will make money on the API, and Anthropic also wants to make money on integrations with other corporations from the inside. And more, and more, and more. And all these companies have to make money. And it turns out you pay for a sales manager's workstation not thirty or fifty dollars, as before, but you have to pay two hundred and fifty dollars, as before, to analyse his work. They say: “But his entire margin is two hundred dollars.” Here, well, for the American market it's a bit easier, right? Because there's different money there, different margins, different opportunities.
And still it seems to me that development, this cool development with cool models and systems, is for the rich, for rich corporations. And for now we're just on subscriptions. We've been given the chance, in this developing period, to buy a little subscription in which you can really build… I was talking to a friend of mine today. You know what I told him? That you can now come to a presentation. Say some corporation orders a solution from you, for example, I don't know, train five thousand employees in their work, write a system for them, I don't know, for tracking their KPIs.
And you can come to the presentation with a ready-made solution that you build with Claude Code, for example, in three hours. Come straight away with a ready solution, not a presentation, and present to them right away the whole thing, that is, a training system, in three hours. Whereas before, no developing at all, just come straight away, straight away with the finished solution. And this gap right now is, of course, incredibly, colossally large. And by the way, viewers, write in: who has made a lot of money on this? Because when you ask, most people really didn't end up making money on it.
So, but only if a programmer works six or eight jobs in parallel. That, by the way, is a topic about parallel work. It started during COVID, even before artificial intelligence. We saw it, when people started working remotely. A huge number of cases appeared where programmers, marketers, there are such roles known in this context, worked two or three jobs. A very well-known case. Simultaneously, in three companies at once. And now, with artificial intelligence, I, I, I don't even dare imagine what's going on.
I don't even dare imagine what's going on.
In one of our first episodes we discussed the overemployed movement in the US. There was, I think, a huge thread on Reddit about being overemployed, about a person working five jobs in parallel. That was the very, very dawn, even in our episodes. GPT-4 was just out, roughly speaking. And obviously it appeared even before that. And as for the Indians, on a remote interview you can't tell one from another. One passes the interview, another shows up for the job. That's a classic, right. From the COVID days, yes.
That's a common thing right now, I think, by the way, isn't it?
Working several jobs?
No, when one person comes and does the interview for another, and someone else shows up for the job.
Well, I think so. I think it's done quite often.
Well, obviously it's not at American companies, not at OpenAI that people get hired like that, right? But overall, they say, it's very widespread. Yes. Even, even on dating sites, I think, one profile is chatting, one support team chats as if it's one girl, and completely different girls show up for the date. And they say: “I've just changed a little.”
On the other hand, on the other hand, to pass an interview competently these days you don't even need someone to take it for you. You can have so many assistants of all kinds next to you that will help you pass it perfectly in real time.
Anya, this has changed, by the way. You've been in this field professionally for twenty years, right? It's a topic that — or more. It's a topic that, well, you have a huge flow of people, thousands of people, highly qualified specialists, right? You already have a huge flow of engineers. Does everyone actually understand that it really has become like this now? And what to do about it?
Well, whether everyone understands, I can't speak for everyone. Whether everyone understands...
And what to do about it? And what's to be done?
But that we understand it, that's for sure. It's how...
Well, and what will you do about it?
How you change the way you interview a person, what questions you ask, how much you take into account what additional tools he might be using at that moment. I mean, we're even setting aside what an interview is and whether you can assess a person in an hour at all. We're even setting those things aside. But in order to at least roughly get closer to some adequate assessment and understanding of who is in front of you
That's Max!
Then you fly back.
Those days, those days...
Yes, yes.
The days of our Max's youth, yes.
And calmly, without worrying that the person is sitting there with an assistant, I don't know, or that someone else came instead. I'd happily take a trip.
As we wrap up the episode: at ToTheMoon we're additionally selecting speakers and hosts for some segments, because with media like this we want, well, we want to make more, more different episodes. And I get a lot, we get a lot of applications. And people, when the applications come in, I'll be honest with you, I'm tired of not knowing whether a person is reading a text or not, whether they made it themselves or not. Whether he's saying what he thinks or reciting the internet. And I started asking many people to just walk, go straight at it and break down some specific question for me.
And I write in advance: don't read it. Obviously he can still read it, but probably if a person is applying as a speaker, he, well, he'll have to not read it to get the end result, so we can see how the person is able to talk on that kind of topic. So he took any topic. And I also say: record it on the first take, don't prepare. And on the subject of interviews there's a very interesting case. Maybe our editors will find it and insert it at the end here. It's really fun. So, a guy came to a job interview, and he was scheduled for an interview together with an AI assistant.
And he came to the interview without inviting his AI assistant. And so, who showed up? This AI assistant showed up, the one interviewing this person, and his own as well. At some point, you know what he did? He got up from the table and said: “I need to step out, you two sort it out.” And he left. And so he went to another room. The two of them started talking, and at the end they started wrapping up the meeting, and they couldn't wrap it up. It's like in the film “Bicentennial Man”, remember, in the film he says to him: “Good night.” He says to him: “Fine.” He says: “You should say good night, Andrew, you should answer good night.” He says: “Good night.” He says it, turns around, and the robot says to him: “But you didn't say good night to me.” He says: “Good night.” He says: “Good night.” And that's how it started. Goodbye, goodbye, goodbye, have a nice evening, goodbye, have a nice evening.
And those two robots, two robots at the interview stayed together while this person was gone. And of course we are a hundred percent heading into a world where people, without even understanding how artificial intelligence works, will invite robots in to conduct interviews for low-skill… well, no, it's cheap models that do this anyway. It's not, they didn't run this AI assistant on an expensive model, they'll conduct interviews with a cheap model. And we'll see all of you in our episodes every day.
Our episodes come out every other day, and this podcast every Sunday. Don't forget to support our channel, especially today's team, all the people we've got together. Bye, everyone!
Listen, we didn't discuss half of what you planned. Actually, I mean...
That's normal.
About that, about the IPO, er...
Yes, well, with mathematics there's been a real disaster.
And about how sad things are in theoretical mathematics in general.
Only stupid mathematicians are sad. In reality mathematicians should simply be rejoicing.
Well, that's what I wanted to ask, because I read up on all of this, and I was interested to hear about exactly this kind of future for science, and in particular, well, about theoretical mathematics.
Let's have a holy war over it.
It's a really great topic. And what's fun here, by the way, if it happens, Anya and I are a bit far from mathematics, but this topic is really great, because...
Yes, I, I just don't really understand, well, what will happen to this...
The headline, yes, that headline there was, that, that mathematicians, mathematicians are suffering from this, or it's a big problem, or mathematics has been destroyed and so on. That's a topic that's certainly interesting, interesting and good to debate.
So look, it's not just mathematics, here, here it's about science in the broad sense. But in general, I think...
Any science, yes, falls under this story.
Of course.
Yes. That's it. Thank you all very much. Bye, everyone.
Yes.