Astra is out, and almost immediately a scandal flared up around artificial intelligence. OpenAI announced that the system had solved a mathematical problem with a million-dollar prize attached. But mathematicians are asking: "Did the artificial intelligence really get there on its own, or did it use our ideas? What actually happened there?" And the second story, which already concerns literally everyone. Shopify is showing a sharp rise in purchases made through artificial intelligence. Essentially, you come in and the AI doesn't just help you choose a product — it really influences what you buy, where the money goes, and it affects the average order, by the way. And, of course, the new iPhone 18 and the foldable Duo.
Is there anything there at all worth upgrading for? Today we'll go through all these questions. Welcome to To The Moon! Hi everyone! We're on the To The Moon channel. Tech news, insights from Silicon Valley and around the world, and our weekly podcast. Ilnar, listen, you know where I want to start? When people are told that some mathematical question has been solved, they usually don't understand what it is, and it doesn't even get very interesting either. But a tweet caught my eye on X: someone wrote a very rude message in English to Sam Altman, and I got into this news not through OpenAI's announcement but through that insult. We won't read the insult out loud, but we'll show the tweet in English.
Well, why not? And the insult is very harsh — he basically wrote that you haven't done anything at all, nothing happened there. And it seems to me that it's still worth paying attention to this and telling our viewers what actually happened. To give some explanation of this whole story, because it's quite important. And initially, of course, this looks very peculiar to me, because when they write that tens of thousands of agents solved some incredible mathematical problem in eighty or however many hours —
Eighty. Yes, yes, yes.
In eighty hours, right? And some cosmic number of agents, right, was involved. And different news reports gave different figures, but at least there were more than ten thousand. That many agents. And I'm thinking: what kind of agents are these anyway? Who managed them? What kind of system was written, who controlled them? So in the end, was it one agent controlling all of them, or many agents? Lately this has been a topic that, I think, is very hard to discuss. But still, let's tell what happened from the point of view of mathematics, what can be considered somewhat grandiose, and what to me looks like OpenAI marketing.
Yes, look, here's what I want to note right away. Every time — well, maybe it just seems that way to me, check me in the comments if I'm wrong — when OpenAI releases a new model that seems pretty good, there's always a news item saying: actually, we have something in training that you can't even imagine, we're about to finish training it and it will blow everyone away. And now the same thing has happened. Yes, GPT-6 Astra came out, and we'll most likely mention it more than once today.
And they say: we have a great model, we're training it, and along the way we decided — why not try it on the Millennium Prize Problems in mathematics. Let me explain a little what that is. The Clay Institute, in the 2000s — in the year 2000, I think — picked several problems, seven of them originally, I believe, and said: these are super-tough problems that nobody has been able to solve for a long time. Whoever solves one gets a million dollars. My academic supervisor at university also always teased us, saying if you solve a Millennium Problem, take me to a restaurant, give me a bit of the money. But it was always a joke.
The point is that mathematicians know about these problems, especially those working in the field. They're all from different areas, but those who work in a specific area know a great deal about these problems. The Navier–Stokes equation is something I heard very, very many times during my studies, but it belongs more to mechanics. I was a mathematician, so I won't go into details right now. But the gist is this: the problem is old, ninety-plus years old, and many good mathematicians tried to solve it, tried repeatedly, and spent a great deal of time trying to understand its solutions.
Nobody succeeded. Well, the Clay Institute hasn't paid anyone a million dollars yet. Until now it was still an active problem that could be solved. Most likely our viewers have heard of Perelman, the mathematician who solved one of the Millennium Problems. And there was that story about him refusing the million dollars, or them not managing to hand it to him, but most likely everyone has more or less heard about it at some point. So this is another problem from the same list that has been solved.
Why did OpenAI, as they themselves stated in their article, take an interest in this problem? Towards the end of August rumours appeared that a group of mathematicians — actually one mathematician, Braxton, I think his surname is — and an Anthropic employee were discussing this topic. And they said they had an intermediate result on this problem. That is, they hadn't fully proved it, hadn't fully solved the problem, but had managed to make substantial progress in that direction.
As OpenAI states, this became the incentive for them to try this new super-cool model they're training on the Millennium Problems. So they set their agents loose on it. In the directions where they managed to get at least some results, they gave more compute, and took compute away from less promising ones. And so ten thousand agents, eighty-eight hours, and they managed to solve the problem, as they say. That is, roughly speaking, find a solution to the Navier–Stokes equation — that is, prove that a solution exists and that it is smooth.
And they showed it. If you look at that publication, right at the top there's a kind of whirlpool. That's the demonstration of the solution to this equation; you can take a look, it's a rather pretty picture. And so, in passing, they solved this problem. And here a small scandal arose, because the mathematicians who had said they had progress in this direction began to suspect that OpenAI had stolen their ideas. It has to be admitted that these mathematicians used ChatGPT and, apparently, Claude to check and generate ideas; they even generated small papers and sent them to each other for review. So one way or another they were putting their thoughts into ChatGPT.
And here a question arose for them: those thoughts, their ideas that helped them make progress — were they used in training? Or did someone peek — someone watching the mathematicians, lifting ideas, then setting a model on them and getting the results? Well, OpenAI says no: no dialogues after 3 July, I think, if I remember the article correctly, were used in any way to train this model. And the discussions between the mathematician and the Anthropic employee were around 15 August, if I remember the chronology right — but you can check the article.
And they say that the prompts this mathematician entered could not in any way have influenced how this model behaves, what it was trained on, or what decisions it makes. Moreover, as it turned out, the mathematician and the Anthropic employee don't have a complete solution, they only have an intermediate step. In this respect, credit to OpenAI. They say: we don't claim priority in proving this intermediate part. An excellent mathematical result. Well done! But as for the solution —
Ilnar.
They don't have it, right.
I want to draw your attention here, just to make a quick insert for the viewers, so it's clear what 'stole' or 'didn't steal' means. And someone might say: "Well, the mathematician didn't do it, and they did." OpenAI spent a hundred and thirty billion tokens on this task. And if we all count how much a hundred and thirty billion dollars costs — a hundred and thirty billion dollars — first of all, they did this on GPT-6 Astra.
Tokens, not dollars — a hundred and thirty billion, yes.
Oh, that's a slip, yes. A hundred and thirty billion tokens they spent, and they did it on GPT-6 Astra. That was about two point seven million messages. The system ran around ten thousand agents working simultaneously. If you look at the price, it should be more than a million dollars. So it's unlikely, on the whole, that a mathematician could invest that kind of money into studying this question, right?
Yes.
dig, from him, right? Then, if you didn't know which way to row, you'd need not a hundred and thirty billion tokens but, I don't know, ten times more, because you'd need everything at once —
Or maybe it wouldn't have found it at all. Or maybe the problem would have been solved anyway.
Or maybe it wouldn't have found it, yes. Maybe it wouldn't. But actually —
And is that bad? Imagine, the problem has been worked on for the last two hundred years or however long. Obviously, over the past decades a huge amount has been documented and described. I think there are texts that not even all mathematicians could read, from different countries.
Obviously, yes.
In all sorts of formats, all sorts of scientific papers and everything. Clearly this system absorbed all of it, most likely, I'd assume, and started finding this solution. As far as I understand, it doesn't say that we found this solution entirely on our own. Whether the agents worked on their own here, or we took some directions for how this problem could be solved and started reviewing them, working in parallel. What did they launch, after all? It's not just ten thousand agents working in parallel, it's actually a whole force of nature.
One set of agents constantly finds hypotheses, a second set checks them, a third set in parallel looks at the end result, selects, chooses and so on. And that's how they move forward. Because on the one hand it sounds like a lot, like a million dollars, but in fact it's eighty hours and only that much in resources. And for solving a problem that, as you said, is almost a hundred years old — I even read it was nearly two hundred, right? That is, some cosmically big problem — and it's free. A free solution. The question is whether that's really so.
Yes. Look, about whether the problem was solved from scratch or on the basis of something.
I don't remember which mathematician, but there's a well-known phrase: it's not that I'm great, I'm standing on the shoulders of giants. In general, in mathematics, and in any science, that's how it is. You rely on previous results that you use, and on the basis of them you make a little bump on the body of science, as they say. There's the big body of science, and your result is most often a tiny bump sticking out of it, and everything else was already known before that. And with agents it's exactly the same story.
They process the entire volume of mathematical papers that exist, and on that basis, roughly speaking, they get — in quotes — 'thoughts' about which way to move, how it can be done. What's worth clarifying here? Not all mathematicians' work is known to everyone. Well, obviously, given how many there are. Moreover, it's less so now, but fifteen or twenty years ago there was definitely a huge number of publications that were reprinted in another language. Roughly speaking, mathematical journals published, I don't know, in the Soviet Union or in Russia in the nineties weren't translated into English and weren't published abroad. And the same the other way round. And so, let's say, less-than-honest scientists could simply pick up ideas from foreign journals.
It worked in both directions, one way and the other. And then, on the basis of those ideas, write their own paper in their own language, publish in the journals they could publish in, and get some benefits for it — say, defend a doctorate or something else. So the world of mathematics is, of course, much more globalised now. We have access to different works, but the number of works is so large. And AI, neural networks, are very good at finding interesting ideas at the intersection.
That is, you take ideas from one direction, apply them in another, and — well, not everything works out, but you get some progress. For a person that's much harder, because you have to understand both one and the other, and then try to do it at the intersection. After all, if we're talking about high science, you're most likely a very high-level expert in one area; in the others you do understand things, but not that well. You can compare it with sport: when we take elite sport, hockey players, say, play hockey well.
Footballers play football well. Yes, in other disciplines they're also above average thanks to their physical development, but you still can't play both at the same high level. It's the same in mathematics, right? You still start going deep into some narrow area, and in it you're a super expert. But being able to work in a large number of areas at once and pull things from one into another — that's already a super result. That's rare. And AI, as far as I understand, is exactly what carries these things.
But the result is insane. Again, as you remember, a few episodes ago we talked about the Riemann hypothesis, that there'd been progress on it. And now the Navier–Stokes equation. These are super-hard problems in mathematics that everyone would like to solve but can't. And AI seems to have been able to deliver a result. I haven't yet seen statements from authoritative mathematicians saying yes, we've checked —
Well yes, it hasn't been recognised as proven yet, at least at the time of recording this episode. On the site that, conditionally, tracks whether a problem is completed or not, it's still listed as unsolved, yes. We'll see what happens next. It seems to me this is partly, probably, a correctly built marketing concept. On the other hand, obviously OpenAI is a serious organisation, serious mathematicians work there too, they wouldn't just pour out some nonsense. Clearly a whole body of things has been fundamentally solved there. Right? And I think this conversation about artificial intelligence already being great as a programmer, an architect — well, not as an architect, as a junior or a great analyst.
A few days ago I recorded an episode precisely on the topic — take a look — of OpenAI releasing a study on how much time their analysts spend, money, resources on research, on the work of artificial intelligence, and how artificial intelligence helps them. And they called it an intern, an AI intern, a research intern. On the one hand, it's quite an interesting piece of material, take a look. It was on our channel on Friday. On the one hand, it's an intern — that seems to be really so, OpenAI itself calls it that — but on the other hand, for me, for example, artificial intelligence is often not an intern but a super-super-super professor, a doctor of science.
And a combination of them. Because what ChatGPT or Anthropic can do today — or actually, what practically all models can do today in terms of analysis — surpasses any person, their basic knowledge. Then, of course, whether these systems are capable of solving such complex problems. By the way, about Astra.
It's been a week since Astra's release — ten days. And it seems to me it will take more, as I said at the start, a week ago, probably, when we were recording our episode about Astra — I did a special episode — that it'll take several months to really say whether the system is great or not, and how far GPT-6 surpassed the results. What did I personally get? GPT-6 as a chat — I didn't see any difference at all. That is, I didn't see that super effect. I remember I had a super effect with — probably GPT-5.5, probably somewhere around there, right?
There was a real super effect there, especially in the Pro version. And even GPT-5 maybe had some super effect at some point. Here I don't see a super effect, because it's unclear how the system — although I talked about it; for those who didn't watch, see the episode we had on Wednesday on the channel this week — what improvements there are. For example, there are no improvements in web search, I understand that, I know it, and I'm conditionally not counting on it much. But there should be an effect, for example, in terms of deciding what the model can do itself rather than ask a person.
the last month. And it's much simpler than a million-dollar mathematical problem, but solving it is probably worth more than a million dollars. And I have to say that it not only didn't solve it for me, it even made things worse inside, and it spent — not just tokens, it spent money on other systems too, because it used database load, all sorts of servers, third-party systems. And it spent a great deal. And it didn't solve the task. And I'd even say that I, again, carried on.
Fable 5.1 came out, and I don't know whether it works better than Fable 5, but it works no worse than Fable 5. It works just as superbly. And I carried on solving this task with Fable 5.1. I'll say right away, Fable 5.1 hasn't solved it for me yet either, but some part of me told me I should keep working with Fable for now. And in GPT-6 — go solve some other tasks of yours. Which is what I ended up doing. Ilnar — oh, our viewers, after all these ten days, write what you think about it.
By the way, I'll note that the comments diverge between our community and our viewers. Many people write that GPT-6 is something unreal, that it's already AGI, that it's already cosmic, that it's unsafe. Today comments have already started coming in on that latest episode of mine about — we'll touch on this topic now — one of Anthropic's employees, having quit Anthropic, said that almost the whole company — well, part of the company — says there's a very high probability that humanity is done for. And what's more, this isn't just discussed at Anthropic — employees discuss it much more intensely at Anthropic than at OpenAI.
And at OpenAI it's also discussed that humanity is done for. And it's not a joke, not a prank — it's what Elon Musk — Remember, Ilnar, about a year and a half ago Elon Musk — I hope our editors find it — said: "Well, there's a ten, twenty percent probability, yes, that it gets out of control and everyone's done for." Well, he said it lightly, as it were. But here employees are saying it. And so part of the community is on that side, and part of the community is on the side that GPT-6 devoured tokens.
Crap. Doesn't work any better than the previous model. It's not even clear what they improved. Took on even more and did even less. On the whole it reminds me of GPT-5.6 Sol in ultra mode. I even stopped using that mode, because it's just unreal. So I have a question, Ilnar: how did they launch ten thousand agents working simultaneously? Because judging by Anthropic's study — Anthropic had a very interesting study, we had an episode about it too — they launched eighty agents working in parallel, and the agents wrecked each other's architecture, blocked each other, didn't coordinate their joint work — in short, partly killed each other off.
Who thinks what about this? And Ilnar will now share his opinion, because Ilnar hasn't yet had a chance to talk about ChatGPT 6. That was me recording a special episode. So let's hear from him.
Well, look, as for the ten thousand agents. OpenAI did say — what I said at the beginning — that we're training a super-cool model that's much stronger than GPT-6 and so on. That is, they always lay a little groundwork ahead. And apparently here too there's a hint that we have everything set up so well that ten thousand agents can work simultaneously in coordination for eighty-eight hours and solve a Millennium Problem. But one way or another, let's leave that on their conscience for now, until we've seen it.
Inside Codex we haven't even dreamed of ten thousand agents yet. So. Jacob, whom you started talking about — yes, he really made noise by quitting Anthropic that way, and earned himself some PR, but on the whole, of course, yes, he opened a bit of a Pandora's box, and X, I think, simply exploded after his posts. But we'll get to that. As for GPT-6. It tore the benchmarks apart, of course. The benchmarks they show — ARC-AGI and so on — in effect that benchmark no longer exists for it.
It scored ninety-nine percent or so on it. And on the whole it no longer makes sense to test it on that benchmark. It won't score more, because, as you remember, we said that on benchmarks you always have to leave some share of errors that are built in from the start. So we consider it simply maxed out there. As for my personal experience of using it — did I switch from 5.6 to 6? Yes, I switched completely. Does it solve all tasks? No, it doesn't. Even the ones that worked pretty well for me with 5.6, when I led it by the hand. GPT-6 can't do them from the start; I have to adapt to it and either feed it tasks differently or something else.
But on the whole my general feeling is that it certainly hasn't got worse, and it seems to have got even better, because for some of the tasks its answers and what it does suit me completely. So on the whole I feel we're already at a level where, as an assistant, a helper, agent systems are already good, yes. If they keep improving — fine; if they stay the same — well, just great. Give me any task.
And I think my problem — let me put it this way — I think the problem I was running there... we do understand that GPT-5.6 Astra is definitely great — oh, GPT-6 is definitely better than 5.6. No question there at all. And obviously it's not a marketing trick and so on. I'm just adding colour here. It's just that OpenAI periodically pours out peculiar releases. But I think my task itself is hard to solve, and so the question is: since Fable has problems with it and ChatGPT has problems with it, you need to approach the task from some other side, design it differently, think about it differently. Although I'm still waiting, Ilnar, look — I'm waiting for some system to come along that solves it.
It's happened to me more than once. I remember when Fable came out, and Fable solved several of my tasks just like that — bang, and they poured out. Here's another example that seemed impossible a year ago — copying a website, for instance. I'm not even talking about what's been done now for the ToTheMoon channel, for example, on my website, or for my own channel — the semantics, the topology, it's already published on my channel, by the way. For those interested, I recorded an episode a week ago, I think, about ten days ago, about the semantics and topology of a complex data set. And you can find it in the Knowledge Base section on the site — well, I hope our editors provide the links — how uniquely the search works by itself, by its structure, the connectedness of these things. It's not some know-how or anything, but it was impossible to imagine as easy development and migration.
And how quickly I solved it with a combination of, roughly, Codex, Claude Design and Claude Code Fable, right? That is, Codex was for some simple things for me. Claude Fable was for complex work, for layout — for layout, by the way — and for complex architectural things. And Claude Design was for visualising this data. So I fully solved this unsolvable task ahead of time, right? And you, by the way, Ilnar, you yourself remember saying — I think six months ago — "Sash, right now you definitely can't just press, sit down and quickly, in a couple of chats, not even one chat, in a couple of chats copy a website, right?
Well, you can do a page, two, three, four." And that's how it was. Now we're entering a completely different era. So nitpicking Astra is the last, hopeless thing to do. But still, when GPT-6 is poured out, you yourself fundamentally expect some unreal functionality that should do something crazy. And then what gets thrown at me is: we've released four new images — ChatGPT started advertising some image thing — well, some micro gimmicks, right? But people around the world basically don't fully get it yet. Just now I was travelling around Europe with a friend, and at some point we had a long drive.
I tapped ChatGPT in CarPlay, tapped it on the screen, and it started talking to me by voice. I discussed some topics with it, and my friend says to me: "Who were you talking to?" And he's so far from artificial intelligence, and at some point he... This is, by the way, the example I told about in Madrid with Alisa. It's that example with Alisa. A person with Alisa. And he, who uses Alisa — the funny thing is he then says to me: "Listen, how do I pay for this version? I want to talk to it." And you could see that the structure of the conversation had made an impression on him.
And also how it talked. He says: "Did you train it to talk like that? It talks as if it were you." I say: "It wasn't me at all." I say: "It just uses some slang and style, and it seems to you that it really isn't a robot." Although I tell it that in a conversation I like the machine to stay a machine, not to try to adapt to me somehow — I don't know, be cheeky or something. I want clear execution of tasks, but in a normal, normal voice, a nice one. By the way, what's everyone's experience with their new Voice? Because ChatGPT Voice is still a new update.
In two months it's already improved a lot. What results are you getting, tell us. In the car, again, I get a decent effect, but I still don't use Voice, right? ChatGPT still hasn't got me hooked on it. Of course, I dictate texts or dictate tasks, but I haven't used it. By the way, write in who uses Codex by voice, or Anthropic, or Claude by voice. You, Ilnar, don't use it by voice? I just type in Codex, I don't use it by voice.
Yes, I'm not comfortable with it. If I'm sitting at the computer, it's much easier for me to type on the keyboard than to dictate. But from the phone, yes — dictating something or even asking by voice, that's yes. Well, because typing on a phone keyboard isn't very convenient, but for work at the computer, no — only by hand for now, at least.
Here's an interesting topic I want to raise today. Ilnar sent it over: Shopify is one of the main providers of online stores — well, of purchases and procurement in the US. It published some statistics on how requests to stores through artificial intelligence are growing, versus orders that came through ordinary web search. And the result was this: visits to stores from artificial-intelligence services tripled over the year. That's quite substantial — it's plus two hundred percent — and orders that came through artificial intelligence also tripled.
And what's growing there is not only click-throughs but real purchases. And visits from ordinary search engines grew thirty percent over two years. That is, essentially, about fifteen percent a year. There's a very interesting effect there: this statistic can be read in different ways. What was said there is that Shopify isn't the only one that has made a bet — it seems to me a very interesting topic — on cooperating with platforms that drive traffic, not only through artificial intelligence, because who are you working with, right? And many companies will then start abandoning Shopify, right?
And, for example, working somehow differently altogether. Though it's unclear how, because it's a kind of internal service for all purchases — procurement, wholesale, statistics and so on. But the point is: what's the interesting effect? First, why did they show that web search still remains, and it's growing too. Second, if you look at the data on artificial intelligence, it's unclear what to count. Because if, for example, last year a hundred people searched via web search and via artificial intelligence five searched, then growth of plus two hundred percent is only fifteen. And there it grew by fifteen percent over the year, and that became a hundred and fifteen, yes, and not a hundred. That's first. So artificial intelligence may be that same small thing, and it's logical that it could grow so sharply over this year.
That data isn't presented. On the other hand, there's very interesting information that people who come in through artificial intelligence buy better.
That is, they have a better conversion to purchase — visits are given there — and they even have a better average order. That is, Shopify showed data that they buy about fifteen percent more. I don't entirely believe this statistic about the average order, but engagement — engagement through search versus engagement through artificial intelligence — really, through artificial intelligence engagement will be higher. That's probably obvious, because a person who searches through search can open a huge number of different tabs, while in artificial intelligence the information is more or less aggregated, and you then narrow everything down to a choice between one and two. Often.
Sash, what's more, you — at least now — trust it more. That is, the assumption is that ChatGPT or Claude is an expert. And the expert advises you, and most likely your perception is shifted a little in favour of what it recommends. Well, by the way —
Who trusts it at all? Write in. Write in — an interesting aspect. Ilnar already trusts it. He said so — he trusts it more. Write about this aspect. I, probably, because of my versions — certain versions of artificial intelligence — and a correctly asked question, probably also trust it more than search. That's obvious. But I'll stick with the versions caveat, right? A little caveat. Write in, our viewers: do you trust it or not? Because, for example, my friend says outright: "I don't trust artificial intelligence when it comes to conclusions in analytics, in sales or in marketing, or in business." And he also says to me: "Well, you don't trust it either, do you?" I say: "I already very often trust artificial intelligence more when it comes to analytics or marketing, or business. Not to mention doctors — and there's a lot I trust artificial intelligence with."
And this division — I think there's a very strong division in society right now. That doesn't mean I don't see — again, on the topic of artificial intelligence, I don't see — I think we'll do a separate episode on this. I don't see a threat in artificial intelligence, I don't see the problems that appear with artificial intelligence. I don't see the mistakes it makes. I don't see the things that they have, as problems that exist. Tanya often says: "It still does some crap. I mean, it still does little, it doesn't solve this task."
And my opinion remains very simple: some tasks it does — it can still do badly. Some tasks it does far better than a person, and some tasks it does certainly no worse than a person. By the way, in that latest episode of mine about OpenAI's study — and it's a cool one — they compared their analysts' work a year ago and now, it's very fresh. Take a look at how OpenAI's analysts and various researchers inside use artificial intelligence every day in their work and how much it would cost, by the way, for ordinary people. This, Ilnar, is our favourite topic — that artificial intelligence will be limited for ordinary people, right?
Because what OpenAI showed is, essentially, available today to a micro — a tiny percentage of the world economy, of the world's enterprises, so that every employee could use and spend as much money on tokens as every OpenAI employee spends, right? Because the sums are large. I won't spoil it, but that topic is interesting, and it's interesting in general how many agents each employee has working in parallel. Because when I sometimes hear statements like 'I have tens of thousands of agents running in parallel', I think: who sets their tasks?
With my huge number of projects I don't know — I can launch twenty, thirty tasks in parallel, but on the whole a person's brain is limited. How did you launch them in parallel, whom did you launch? On what tasks? On what research work?
That is, what are they collecting for you? It seems to me the whole internet could be collected with those tens of thousands of agents. There'd be nothing left to collect, right? In the overall mass and in all the general details. But this Shopify study is interesting, it's noteworthy. I still stand with the people who say web search will disappear. I've been saying since the release of the very first version of ChatGPT that search itself basically... It's not a question of search, it's a question of how a person makes a purchase decision — that will disappear. It won't disappear, it will change.
How a person even approaches what to buy, where to buy, when to buy. Of course, all of this will depend on where you live, on your age, on your income, and on a whole lot of other things. But basically, in the world, in those huge volumes that, I don't know, Amazon trades in, it will start to change fundamentally. That is, we'll start changing decisions differently. And, by the way, it's not even a question of which online store you go to, it's a question of how you decide to buy and what you redistribute your money to. I think that today, where there's a big threat to humanity and to the world economy, it's the general uncertainty about how people's decision-making chain will form and how to forecast it at all.
For example, has ChatGPT shifted me? I went on another of my three-week trips, one of them, to other countries, and so I have to eat in restaurants. I'm not a super fan of it, but I have to, because of my life, constantly eat in restaurants, right? And how ChatGPT picks these restaurants for me... I've already chosen a certain system. If a year and a half ago ChatGPT helped me a little, and a year ago it still helped somewhat, then now it's almost all-inclusive. Right? Only ChatGPT. Some aspects of managing other systems have dropped away from you, and so I've redistributed my attention, started making decisions differently. Obviously I still eat, right? That's obvious. But I make decisions completely differently.
My vector of perception has shifted, maybe even my approach will shift, and there are people with me. I have a meeting, right? And maybe my approach will shift. By the way, Ilnar, our mutual friend has arrived, right? He flew in today from Italy here to France, and he says: "We'll have dinner together tonight, with one more person." I say: "Yes, I've booked us two places, we'll decide where to go." He says: "They recommended a restaurant to me at the hotel." So what theme was he on? He was on the theme of a hotel recommendation, right?
Which, by the way, I practically never use, although it happens. In Rome recently I did use it, and it was a very good recommendation. What do you think about the search topic and this Shopify study?
Listen, well, in my view it's a very good study that shows you need to look towards ChatGPT and assistants in order to attract traffic. You've also said this many times, long ago: look at what comes up about you in ChatGPT. If before we looked at what came up about a person or a company in a search engine, how many queries there were and so on, then now, of course, you need to look at what these systems say about you, right? About a person, a company, a restaurant, a hotel, whatever else — and work on that question, right?
That is, somehow shape your image, right? And Shopify is also saying that through agents, through these chats, through chats, clients come to them. And these clients — one of the points they note, you said it too — haven't taken people away from the search engines. Yes — the search engines grew, and this grew too. Which is a message to get into this. In that sense I fully like it, I'm fully happy with this direction we're definitely moving in. As for whether search is flowing into a conditional ChatGPT — I think yes, there's movement towards it, because even when I think, OK, I won't ask in ChatGPT, I'll ask in Google, most of the time I still look at that first AI summary, and it's often a sufficient answer.
Often there's no need to dig any further. Yes, it's some technical information like, I don't remember —
Yes, but it's funny, look — that's already an AI answer, it's still an AI answer. It may be cheap and so on, but even web search has kept the AI answer, right? And this AI answer is there today in all search engines.
Yes, that's exactly what I'm saying — even when I use Google, I still don't go the old way, I don't drill down into sites. Yes, there are still such cases. Say, if I need to buy a ticket with a specific airline, I'll go to the airline's website and figure things out and search for something for it. But for reference questions this AI answer almost always satisfies me, whether it's inside, say, ChatGPT or Claude, or Google or Yandex.
Ilnar, here's the thing about the airline: look, before, to choose a restaurant, for example, I could go into aggregator apps, right?
In America, I don't know, OpenTable; in Belarus, Relax, our close friends' service, just as an example. In Russia, I don't remember, there's something else. So there are these various aggregators around the world. That's it — for me, an aggregator of that kind no longer exists today. And I've seen that hotels, in terms of quality — look, all my life, if you remember, I always said that Booking is one of the best-made systems in terms of ratings. That is, it works if you understand what those numbers mean: that you shouldn't book a hotel below eight; that a hotel above nine is super cool if it has, I don't know, more than twenty or thirty reviews; that a hotel at eight and a half, nine is also a great hotel.
Six is all rubbish, hopeless, don't check in there at all. So it's a fairly centrally working review system, right? Good, more or less detailed, but essentially they've lost their edge, because from what I'm now seeing in terms of data, you're not just collecting reviews anymore. I complained once, for example, that Booking has a hotel with fifteen hundred reviews, and they don't give you additional information about, I don't know, what the mattresses are like — there's no star rating for the mattress; there's a star for the internet but no star for the mattress; there's a star for, I don't know, cleanliness, but no star for, I don't know, the presence of a certain kind of light. And location is also considered relatively, because what is a good location, right?
It's a certain relativity, especially in cities that are a bit more massive — I don't know, like Paris or Rome, or like New York. It's very relative where you want to stay in terms of a good location. And it all depends on what kind of person it is, what a good location is for them. It's not a small village. And here's the thing: ChatGPT today gives you the ability — it will re-read the reviews, re-read third-party platforms, re-read four services — Expedia, Booking, TripAdvisor, all together — and look through the freshest ones, and it'll also tell you: "Look, on this street, exactly when you're going, there'll be roadworks, so it's better not to go there, right?" I actually have a situation like that.
If my hotel sends me: "Be careful with traffic on Sunday in Nice — Ironman." I'm like: "Uh." I didn't know in advance there'd be an Ironman. That's a shortcoming of ChatGPT, but I understand that none of my services — what, would Booking have told me about it, or Marriott, right? Bonvoy or Expedia — would they have told me? Who could have told me at all that there'd be an Ironman? And for me that's a certain correction. I wouldn't have gone there on Sunday — I'd have gone on Monday, right?
Or I'd have gone on Friday, because I categorically dislike big crowds of people. Plus it limits opportunities of a different kind and so on. And even today my friend says: "Then we'll pop into Nice tomorrow, and then come back to it after the weekend — well, at the end of the weekend." This kind of data story is, of course, moving into the chat. And viewers, write in how it is for you. It's interesting, especially, Ilnar, with your example. Your example is just funny: you sort of go in, but you stay on the first line.
And that's the same AI agent. But for some reason you opened Google. And why did you open Google? Well, like, it's faster, right? But what's faster? The chat in extra-high, in Pro mode, already answers instantly. If it answers for a long time, that means the question is hard.
Well, there's also this about Booking and similar systems, with aggregating reviews and so on. Here, you know, I have the same feeling as with mathematics. We said, right, that it's not that I'm great, I'm standing on the shoulders of giants. The chats will still need to get up-to-date reviews from somewhere about, I don't know, the state of a hotel, how they feed you there, how they serve you, what problems there were and weren't, and so on. So anyway — yes, right now, as a layer so you don't have to deal with the reviews yourself and get some information in aggregated form and quickly, that's very good. But imagining that all these systems disappear — then this information has to somehow, in some other way —
Then, in theory, Ilnar, that's when the story begins where all these artificial-intelligence systems finally start asking questions. And even now ChatGPT could fine-tune itself and could tell me separately: "Alexander, listen, you're visiting places here. If I ask you about this place, can you tell me about it? I'll put that information out and learn from it. I don't learn from all of your information. You've blocked everything for me. I can't do anything. I can keep a little piece here for a while, or keep it only for you. But first, when you answer, give me an answer, I learn for you, I'll know what you answered. And in general, if you don't mind, I'll put this information out on the internet anonymised."
I'd tell it: "No problem, put it on the internet. Here's my real, live comment. Or save it for me." But it doesn't ask me. That's the problem: over the last two weeks I asked it about a hundred and fifty restaurants, it offers me a great deal, and not once did it actually ask me afterwards what happened, whether I went anywhere on its recommendation or not. Whether its answer was accurate or not, because I don't use that function — there's a hidden little checkbox somewhere, click to say you didn't like the answer. Or that bug thing lights up — you can write a bug report to Anthropic or report an error in Codex.
Well, write in, our viewers, maybe you use it. I don't, although I'm quite the geek, to write, report, do. But it could have prompted me. So it probably has no interest in finding that out, that information. Although look, look what a cool project there could be inside ChatGPT, right?
Yes.
Because it really knows that you, say, wanted to stay at this hotel; it can ask for information, then it has to store a lot of it somewhere, and then, accordingly — well, maybe we'll get there.
Ilnar, even the very training on the fact of answer quality. Look, it's important for them to train answer quality. They buy it from some people in Africa almost for free, for a few dollars. They ask: is the answer good or not? I'm a person who's very expensive for them. They can't buy me to ask me whether the answer is good or bad. Or you. That's very expensive. Or the huge number of people who use ChatGPT, Anthropic and everything — our viewers, right? It's very expensive to buy a live person pulled out of the network, someone who lives in some location, has a family there and a whole lot of things.
As opposed to simply buying, in some centre, a large number of people who more or less go by the standard and answer. It could find out from you: so, what did you actually visit out of all this? Or: I offer you tables, and in the end you turn off to other places — that happens. It happens that it suggests something to me. Then I still open Google, for example, and my friend says: "I found another place, and we're going there." That can happen too, right? And it could improve the quality of its answers as well.
Note, right, that Astra is better in that it needs to ask the person less. Though my question is: GPT-5.6 didn't ask me anyway, and your Astra carried on not asking me, because it thinks it doesn't need to ask me. And that's a fundamental problem that exists today. And someone, by the way, wrote in our chat, in the comments, that Astra has started asking questions. Personally, Astra opened that window for me and asked, I think, once or twice. I don't know, Ilnar, does it ask you?
Well, periodically, but literally two or three times a week, I suppose, I've come across it.
And Anthropic's Claude, by the way, has stopped asking me altogether. Before, Anthropic asked me often — I remember, endless multiple-choice questions, answer options, everything. Now it's a real big question whether the system will additionally ask you again, additionally clarify. Yes, ask you about it in some interface rather than hand you an answer like this. An answer like this. It doesn't fit on my screen. I deliberately showed my hands so nobody can see it. I can't see it either. And somewhere in there, maybe, a question is written. Somewhere up there, maybe, a question is written, and then you — It tells you: "But I asked you a question, Alexander, you just didn't answer it."
So you asked me a question. So how was I supposed to find a question in that illegible text of yours? For now I see, of course, a big problem here for working with a person, but still I had a precedent — I mentioned it in the episode on Wednesday, I was talking about Astra. For those interested, watch my opinion on Astra there. The episode was on Wednesday. I said I'd like Astra to become a model that does more for the person, right, and understands the person more, hears the person, knows the person, rather than solving these tasks like 'we launched ten thousand agents and solved a mathematical problem'.
Launch at least something for me too, at least some little piece of my life. Get to know me more, get to know me wider.
Sash, if we fantasise a bit and rely on the rumours going around, the next stage of development that awaits us is this: from chatbots that work after your
request, we move, on the contrary, to a system that's always running on a server 24/7 and works with you proactively. That is, right now GPT-6, Claude or whatever else work when you give them a task. If you don't give a task, they sort of don't work. A different concept: suppose GPT-6 works always, it has a set of instructions — I don't know, let's say GPT-8 to accompany your life. And it constantly makes requests to different systems, right? Where you are, what you're doing, what's on your calendar and so on.
And it processes them in the background and periodically comes to you and says: Alexander, an Ironman is planned in Nice on such-and-such dates. I see you're planning to go there. I don't recommend it, right? Let's change your schedule a little, because this news came up. That is, it's no longer you who starts asking, but during those dates, along the way —
Of course.
— of the trip, what will happen; instead this thing runs in the background somewhere on a server, that is, dedicated compute separately for you and so on. From the start this definitely won't be a mass story, simply because, say, OpenAI can't apply this to a billion users.
No, they don't have that much compute. It's like Anthropic's safety — they started spending twenty percent of compute on safety, right? That is, it's...
So. But my feeling is that if some plan appears at a thousand, two thousand dollars, then it seems like an excellent candidate for such a project. It's a kind of permanently running process that works on — I don't know — on your life, on improving it, on its comfort.
By the way, I want to say — I'd pay two thousand dollars a month right now if there were a thing that constantly processes the chats I have, and I'd probably get even more involved in everything — one that comes back to me about them, comes back, asks, clarifies, redoes, says: "Can I help you here too? And I found some more information here," and so on. That would be an interesting thing for me. It's not the story where I write an agent so that it checks something for me, or give it a task: "Please check my mail every day, right? Or please check the weather in such-and-such city every day."
Well, requests of that kind, right? OK, not the weather, let's take something serious. Check a set of quotes on this stock exchange for me and report certain data to me. But when this system itself, inside my requests, starts with its intelligence — because its intelligence is much greater than mine for setting a task — that's very important, I understand that. I try to explain this to many people, by the way. I mean employees, various analysts — I often tell them that this system understands a person or a manager far better on the whole, in volume, and it can generally throw in various ideas, various additional details, find out certain things.
And then, of course, there has to be some interface, right, that's made conveniently or inconveniently for you. Maybe they wanted to solve this somehow. Remember when that daily feed appeared — what was it called, I don't remember anymore.
I can't recall either.
Yes, and we can't recall, right? And everyone advertised it, Sam Altman included. But I think that feed is on the whole a similar construction, where you were supposed to tell it what you're interested in learning and so on, but it's just dumb, linear. This thing for two thousand that we're talking about could have been cooler. But, Ilnar, for OpenAI it will cost more than two thousand. That's the whole problem. That is, what you and I are discussing doesn't cost two thousand a month, it costs almost two thousand a day, right?
And, well, if you convert it into real tokens and the volume of chats there is, then the people willing to pay two thousand are most likely people who don't ask two or three chats a day, right? They have volume. And essentially that's again a minus for them. But maybe on such people the system would speed up in terms of creating itself, right? But again, here that risk arises — perhaps, again, the risk that if the system gets unlimited access to questions to a person, it gets the ability to fine-tune itself without OpenAI's control. Because here the question arises: if OpenAI doesn't train on my data but gives me the ability for the artificial intelligence to learn from me. By the way, maybe that's their own blocker. Because imagine you've opened artificial intelligence up to learning from people, from the people themselves, in terms of asking them questions. Isn't there a chance that at some point the artificial intelligence went into what both Anthropic and OpenAI are trying to block, seriously block? We'll definitely record a separate episode on this topic. I think it'll be critically interesting.
We, of course, didn't tell today — didn't have time to tell — about... this is, after all, not artificial-intelligence news but tech news.
Although there's artificial intelligence in it. The new iPhone 18 that came out, and the iPhone — that clamshell, Duo from Apple.
No, tell me your impressions. I'm insanely curious what you think about the clamshell.
I want to say about the eighteen first — that I asked it.
Go ahead.
I told it: "Can you compare the important things for me — but really important, or something I don't see but is also important, right? Say, plus thirty minutes of battery makes no difference to me. But on the camera, for example, things interest me, or on the processor, right?" And it started comparing. Well, processor number nineteen or twenty. Like, no difference. But, for example, the focal length in the camera — for me, as someone who shoots and photographs a lot, OK, for me that's a tick, a certain plus. Although it seems to me that for most people in the world it's a relative tick, a relative plus, because you still need to know how to use it.
And people rarely know what focal length is, and hardware-wise too. That's a separate question.
Listen, I thought it would all work out of the box, that it would set the aperture itself.
Well, OK. Whether it will work, how it will work, well or badly, the details and all the rest, right? But what was interesting there is that they have this new processor system for artificial intelligence, and it's higher quality. That is, before there was one of something, now there are two. I'm not a super expert anymore, but this isn't 1998 or 2003 or 2004, when I knew all the processors and all the names of all the memory manufacturers and everything else. So they increased the volume — that's the artificial-intelligence story. Yesterday, walking with my friend, I said to him: "What's increasing in artificial intelligence on Apple, if it has no artificial intelligence?" Like, will it do text checking faster, or will it remove a person from a picture faster?
It does that very slowly. Well, that's some rubbish that doesn't interest me. So that story, it seems to me, is a losing one for them for now. That's it, Ilnar, that part. As for the clamshell, it's not interesting to me, it's not relevant to me. It's not relevant because for me a phone is still a phone. Yes, I need a screen, it has a certain size, I'm not against a big one or a medium one — this Max — and I'm ready to get used to any, probably. I like the Max more, but anything bigger is that story of moving into an iPad.
I've had all kinds of iPads; I have, obviously, the big iPad Pro, and I watch films on it on planes. So that's my iPad for films on planes, roughly like that. Or for some small trip. It's like a MacBook for me for trips, because I use screens — Mac Studio, or earlier iMac — they're everywhere at home, in offices, everywhere there are those. So this thing looks irrelevant to me, Ilnar, and I think Apple has gone to a market that isn't Apple's. There's a theme here. Apple has gone into a story that they seemingly don't need.
Maybe I'm wrong, because it seems like some super-phenomenal product, and people will buy it. But I can't even imagine anyone around me who'd buy one for two thousand dollars. First, these have to be people with money, who aren't just ready to update their iPhone every year. I, for example, easily update my iPhone every year. These have to be people who don't update iPhones, who are either fanatics of this kind of thing, or have spare money for it. And that's a very small part of the world's population.
Well, Sash, remember Apple Vision Pro. It's not the first time Apple has moved in interesting directions.
Do you like the Duo, by the way, Ilnar? I'm curious about your opinion. Would you buy a Duo?
Yes, I'll tell you now. Here's what's interesting: if with Apple Vision Pro Apple went in a direction where there weren't yet any big products and made something nobody had made before, here it's entering territory where foldable phones have already been road-tested for several years. Samsung has the Fold; the Chinese have them; everyone has them. And so there's already some audience of people who use them. Among my acquaintances there are people who really like foldable phones. I glance at them from time to time and see that crease in the middle and think: "My God, how do you use this? You can't unsee it, right? It's always in front of your eyes."
And at the presentation the first thing I looked at was the screen. Whether there's a crease or not.
What did you see?
Obviously, yes — in the presentation everything may be touched up and unclear. I badly want to turn it over in my hands. I'm unlikely to get one, but I think the expansion of the line-up is quite interesting. They kept the old iPhones too — I mean the form factor — both the regular one and the Pro Max.
Maybe, Ilnar, it's that it's the most-bought brand in America, for example, right? And you need to have this simply for the comparison, right? Samsung has one, because I have people who use Android, they have this clamshell, right? But it's their second phone, and it doesn't cost that much money, as far as I know, if I'm not mistaken. Although maybe Samsung's clamshell does cost that, but this story is very geeky, terribly geeky. And even in terms of the convenience of this phone, it's a big question. But we'll see.
I asked about the cameras. I'll put it this way: if there were great cameras in it, super-great, I'd even buy one for fun, just to use it. But I'm just not ready to carry two phones, right? My life has already moved on from the days when I had two or three phones in my hand; I'm not ready to carry several phones. And if it at least had a camera of the level — and there are nuances there, right, something's smaller and that's it. It immediately becomes uninteresting to me because of that. But maybe I'd walk around with it, use it and even, possibly, come up with a scenario for myself or load it up. Well, if I didn't have an iPhone like this one.
Obviously I'd get used to it. I'd get used to this screen, I'd get used to all these things. It's very interesting what our viewers will write.
Yes, yes, yes.
In particular about the Duo — do you want one? Who's waiting for pre-orders? It'll be at the end of October. Write in who's waiting to order the iPhone 18. It seems to me that story is purely an update for fun, or for those whose iPhone hasn't been updated. There seem to be no differences here at all, on the whole. It's no comparison with the iPhone when they released the first Apple Intelligence button. The sixteen, I think, right? And we discussed it two years ago. And I said it might become a new interface for people to interact with the world through artificial intelligence. Nothing of the sort happened, it didn't work out.
It didn't, yes.
I use that button to launch the camera. So. And some people maybe don't use an iPhone. Although, once again, I'm such an iPhone adept — an Apple adept, not an iPhone adept, an Apple adept. And Ilnar, by the way, is also actually an Apple person, well, partly, quite strongly. Although Ilnar, Ilnar has more than one phone, right. He has both this one and that one. He can constantly test them and constantly ask. You have an Android, right?
Yes, of course. Almost my main one.
He keeps it. Oh, he keeps it. But he keeps the iPhone too, as one of his phones. Well then, see you on our podcast in exactly a week; our special episodes come out all the time. Be with us on the channel even more. Recommend our channel to your friends. Bye everyone! Until next time.