This week it became clear: artificial intelligence no longer lives in a chat. OpenAI is preparing for an IPO, and that means the market will, for the first time, want to see the real economics of ChatGPT. How much do users cost? How much does an answer cost? How much does one agent cost, the one that writes code or works inside a company? Google, meanwhile, is showing not just a new model, but an attempt to rebuild search, the browser, documents, development, the whole human interface with the Internet.
Meta is cutting an incredible number of people and moving thousands of employees over to AI. China is showing its own chips instead of NVIDIA's. Anthropic is discussing the model's cyber risks not with bloggers anymore, but with financial regulators. Plus a new law in the US, about Trump and ninety days. So the main question of the week isn't which model is smarter. The main question is: who will now own the work, the money and the infrastructure of the AI world?
Hi everyone! We're on the ToTheMoon channel. Tech news, insights from Silicon Valley and around the world. And our weekend podcast. Yes, we've got a lot of episodes during the week now. Plus the weekend podcast, which we have now. Ilnar, you know which topic I want to start with? For me it's an incredible discovery. So, I was sitting there, I prepare a lot of reports. In PDF, in Excel, everywhere. And at some point, in Codex, it was doing an analysis of audio processing and recognition of my own audio for me.
Because these fifteen different systems I use, various products, software solutions — I'm sick of them by now. It turned out that going straight through Codex, through the most expensive models, costs me less than all those systems and all the people around who do this. But here's the thing. At some point I asked it to give me the answers so that I could make a quick correction and return the result to it, return the result of that answer as JSON. And it prepared HTML for me, an HTML format, a report in HTML format. I saw it in HTML and thought, damn, that's a cool view, somehow easier to look at than a PDF.
Over the last five days I've started preparing all my reports in HTML. And the world changed. The world completely changed for me. It turned out that HTML is set up completely differently in terms of report quality. We're just not used to it, we're used to sending everything around as PDF, Excel, Word. But in terms of data visuals, processing that data, structuring the data, the precision of presenting the result, you can see that the models themselves, in this case ChatGPT prepared it for me, but the model itself is far better at working with HTML than with any other environment at all.
And what's more, it can genuinely take all your requests into account, and you don't get fifty pages of some unmanageable PDF, you get a pretty interactive HTML. Which I recommend everyone do. And here it'd be very interesting to hear your opinion on this, whether you've used it or not. But what's interesting here is that this is exactly where we get to one of the most important things happening right now. We have a separate special episode about OpenAI winning in court against Elon Musk. Yes, a big special episode, go watch it, it's out.
These are our new formats. And in parallel we'll have a special episode literally a few days after this video comes out. It's a special about what Codex really is. And there's a story that, by all accounts, OpenAI is now planning to merge everything. Well, our discussion, Ilnar, the one you and I have always had. Why is ChatGPT separate from Codex, the API separate, and they want to bring it all together. And the improvements you can see in Codex now, the updates they've made — their biggest update came on May 20. So that's, what?
Four days before this episode comes out, before today, as you'll see it. For us it's, let's say, today, let's say it happened yesterday, right? And they made a very strong update there. As for what they did: first, they sped up the interfaces a lot, improved remote sign-ins, fixed various bugs and so on. That is, places where, say, the system could spin and burn tokens. That's a cool topic too, by the way: the cooler the model gets, the fewer tokens it burns, even if they've suddenly gotten pricier, say, in some new system.
And they're also improving the system so that, for example, the system doesn't spin tokens on its own. So you can see OpenAI is moving toward a unified infrastructure. If they manage to do that, I think they'll beat what Elon Musk is planning. Ilnar, did you see? Elon Musk is no longer... Initially they said a trillion — there'll be a special episode about it.
Then they said it'd be a trillion and a half, and now they've filed. The IPO is planned for June, in theory, and apparently at more than two trillion. It'll be worth two trillion. So I think when OpenAI goes public, whenever it goes, I think the price won't be lower, not lower. That's not a one-trillion valuation. So, back to HTML. It's a really important topic. I recommend everyone try it. Whoever tries it, write to us, and you can drop your own tricks and things you do in the comments too.
It's a really cool story.
Yes. About SpaceX AI, which has now merged everything into itself. They're going, as you say, for a trillion-dollar IPO. At the same time, if you look at the revenue structure planned for the coming year, SpaceX takes a huge share next year. That is, well, rocket launches, satellites and so on — that's still the lion's share of what they'll be getting in money. But at the same time, if you look at the market share they want to capture, it's AI that, over a horizon of a few years, takes a very, very large part of it, much bigger than rocket launches and the rest, at least if you look at market share. So it's really interesting where all this ends up. As for HTML instead of presentations, on the one hand it's clear why this happens.
Because HTML is, after all, a programming language, it's deterministic, and there's a huge number of libraries the system can use. For it, that's code, not pictures, right? And systems work with that much better. And you've found a brilliant use for it. The first time, if we're talking about generating presentations and so on with HTML, we were doing that back on GPT-4, trying even before 4o, when we were actually working on diarization, recognizing dialogues. When, remember, we first hooked up Whisper and made it break the dialogue down, and then that dialogue had to be presented nicely, interactively.
At the time it seemed like it did that decently, but that was a long time ago, and back then it was still doing it very, very badly. Back then we even tried to build a news graph. It used front-end libraries for that, but it came out very clunky back then. What you're sending now, I don't know if you'll share some example with our viewers or not, but it already looks like a completely different level from what it was a year and a half, two years ago, when we were just starting to do this.
Or, say, you and I are having a discussion, it transcribes it, and when it transcribes it, it puts the transcription rules into the prompt, understanding what we're talking about, knowing all sorts of words and details. Then it runs diarization online a second time. I pay extra for that specifically, because that part of the recognition costs about $0.30. It runs it further, and then it recognizes by role. And the system is trained on your voice, my voice, Tanya's voice and any other speakers.
It's trained, it splits the diarization absolutely perfectly. And then it goes on and runs ChatGPT 5 Pro for me. Sometimes it does 5.5. I run 5.5 Pro. But 5.5 Pro costs about $5 extra to evaluate a video like this and look over the transcription. And, well, 5.5 costs just pennies. And it goes through. So look, it builds the HTML. And in that HTML it does a cool thing. There are some blocks where your voice and mine and Tanya's overlap. And when our voices overlap, it merges them together. So in the HTML it gives you the option… first of all, in the HTML I see that audio snippet.
That is, Codex automatically generates HTML for me with audio snippets. I just open the folder, open it, there's the audio, I can easily listen to it and I can easily split it there. Someone will say these are elementary things, such systems exist. That's exactly the point: those are external, existing systems, and they work pretty clumsily, and they often use Whisper inside. And Whisper is a separate open source thing, it's worse in quality than 4o Transcribe and 4o Diarize. Those are the latest online models OpenAI uses, right?
What's more, training a model is very hard there, you'd have to train it separately on your own servers, do it separately. And here I just get incredible quality and a cool result. And then, on top of that, because it sees the other episodes — it does see the other episodes — it can improve the whole processing pipeline itself, and it builds interesting HTML. We'll definitely show them, definitely show examples. We'll definitely show examples of the HTML we have, in particular for creating articles.
We were doing shorts, doing various things, right? Just now there were even HTMLs for analyzing large data sets. I was doing that in development environments, say, Jira, big development environments, where I'd basically throw in a database, and it presented that data to me as a report in HTML. So you see a completely different structure. But do watch the separate episode about Codex. It'll be out next week. I devote a lot of attention to this there too.
Yes, diarization, one hundred percent. Let's put it this way. We use that word, it may not be familiar to everyone. Diarization is when you break a dialogue down by role. Say the three of us are talking in an episode, and it will indicate whose line it is, who it belongs to. Actually, you can get one hundred percent diarization very cheaply just by uploading not a single audio file where all the voices are laid onto one track, but, accordingly, three separate tracks.
Yeah, each file takes fifteen minutes to upload and so on, yeah.
Uh-huh. And then the system won't even have to learn the voices. It'll just know that this track is speaker one, this track is speaker two, speaker three, and accordingly you'll get one hundred percent diarization that way. Yes, and I fully agree that this has become much, much easier now with Codex, with Claude, with Gemini, the one that just came out, which will most likely make a great move in that direction too. All that's left is to feel sorry for the front-end developers.
Well, listen, that's precisely a question of understanding. I think it's a question, a question of understanding who does what and who is doing what, right? And now this story, this week there's a new cycle, everyone's started reasoning that the question isn't… I like this headline, that the market has started saying not which model is better, but who owns the agent interface, search, the browser, the workplace, shopping and so on. I'll talk about it in the new episode. From my point of view, the agent is an absolutely old, intermediate stage, it should have been discussed two years ago, not lagging behind the market now.
And here, what else is Google doing? First, Google is about to release a new update, it's not visible yet, it'll be inside Google. The announcement, the announcement was on May 19. A test, absolutely tiny test rollout, not visible yet. It's Google Spark. It's a kind of personal artificial intelligence executor inside the Google ecosystem. But again, it's an executor inside the Google ecosystem. So I recently happened to need to do certain setup in Google Cloud with the help of Codex.
And I realized what Google did. It created an environment where a person has to die and sell his kidneys and buy ten DevOps engineers to figure out what to do there at all. That is, to set up Google Cloud, you have to… honestly, I'd like that product to die, I think, right? Because to set up Google Cloud — and it turns out you can't do anything unless you, like, set it up. And now it's a whole story. Yesterday, just to connect to the Google Ads API, then to the Meta API, you spend just cosmic amounts of time, you spend just cosmic amounts of time filling out some paperwork about who you are, what you are, what you do, why you're getting additional authorizations, why they're needed here.
A trillion accounts, just a trillion! Thankfully now there's Claude, ChatGPT, Gemini, which you can use to get the exact menu items you need to click in those systems so you never have to go in there again in your life. And Codex — I have this rule in Codex: I don't want to go anywhere at all. That is, strictly. It keeps urging me to go into Git, into Git, press some buttons there, to go into the storage, press buttons, to go into a separate database management system, press buttons. I tell it: "I'm not going to do anything at all.
I want you to do everything." And that's, that's incredible. And for me now it's not a question of, like, not wanting to use those systems. Those systems are made purely to lock you in incredibly hard, I mean, to lock you in in a brutal way. And I think many companies will lose a lot from that. But again, Google is still at the forefront, and it's rolling out, it's probably very big in the news this week. That is, they're announcing again that it becomes an agent inside, that its key model becomes, uh, the main model of Google search, that advertising is changing, the concept is changing, by the way, for the whole, whole market, the concept of SEO is changing, the concept of search engine optimization is changing.
That is, Google officially said: we're changing our approach to SEO. So SEO will work differently, right? Media, e-commerce content are changing. That is, as they write, it moves from "artificial intelligence answers" to "artificial intelligence does", right? And yet again, the question remains, we've been discussing it this week among the community, Silicon Valley and technical specialists. Will there be agents in browsers, and are browser agents needed at all, as such? So. Are they needed, and what's going to happen with that?
About Google Cloud, Amazon Web Services and the rest. First of all, all those certifications are just: "Yeah, yeah, sure, we'll just get lost then." Yes, it really is hard.
It's no accident that on LinkedIn or wherever you can put up an AWS or Google Cloud certificate, or some other platform's, and it raises your weight on the market, right — if you managed to figure that out, well done already.
You're 100%…
And as for...
I think even someone who has the certificate hasn't figured it out.
He just knows how to answer the questions to get the certificate.
Yes.
And as for the fact that, well, it really is hard: Google did say at the presentation that soon you'll be able to talk to your Gmail. I mean, you can, instead of looking for some email or whatever by hand through search, through filters or whatever, you can just tell it, like: "Hey, when is my next flight? I think I'm flying in April on such-and-such a date, find the email" and so on. So it'll be possible to give commands by voice. I think that after a few iterations, with Google Cloud, with Amazon Web Services and the rest, you'll also be able to start talking by voice.
There'll be specially tuned agents of their own that can help you with settings, and not just the basic ones, but fairly fine-grained settings too. And you'll be able to talk to them in English, or, well, in Russian.
But will you even need email, Ilnar? If you think about it honestly, it turns out that… here's my story. Right now a huge number of systems are being built for medicine, but if we really look at it, what have we arrived at? At there being dozens of platforms you have to use, pay for separately or install as independent apps. Many of them don't sync with each other, don't integrate, and in the end you live, you're left living in a world where you still have no proper storage and no understanding that you have different test results from different doctors, from different labs.
You have a lot of… I mean, there's Google Health, where you can upload things, or, oh, I don't know the exact name, or there's Apple Health too, right, where you can upload all this. And Google has now released a band, and it's like it took a hit at Whoop, at everything. I think it costs $100, but it's not expensive.
By the way, they made it together with my son's favorite basketball player from the Warriors, Curry, they have a collaboration. I think if I show my son that there's a collaboration with him, he'll say: "I want to wear that band too." They made the Whoop fitness band. Together, they have a collaboration with Curry from the Golden State Warriors. And Igor and I were at a game recently, actually. He's, of course, 15-
By the way, he's Steve's favorite basketball player too, everyone's, really.
Yes, and he lives right here nearby. He lives right here nearby, 15 minutes away. And Igor, every time he comes over, when we drive around here, going for a coffee — I drink coffee, say, he has something — into town, when we drive around here, it's 10 minutes, 10-15 minutes. He keeps saying: "I have to run into him on the street, I have to run into him on the street." So. Obviously he's a world star and, well, a pretty serious one, obviously a serious person. But still, he plays nearby, right here. So what's the point? Everyone has these Health apps. And it turns out, who can build me a proper integration today so I properly understand the data, draw conclusions?
Do I have to open my email to look at all this? Or do I need Google for that? I sit there thinking: "Listen, I'll just go into Codex and tell it I've got these 10 systems where I get different test results. Let it save them all together into some folder on my computer, and then in that same Codex window I'll ask about all those details together, and it'll build me some little reports in HTML, the one we just discussed, right? And refine them, and finish them off for me."
Your email will simply no longer be for you, Sash. It'll be for the AI you connect to it.
For the AI. For the AI.
Yes. And now there-
That's not an agent, it's not an agent, it's full-fledged software, it's AI native software. That's what Macrohard is, right? Well, here I am again advertising the episode we'll have in a few days. But the point is, this is AI native software, not an agent, right? Because how, how can you call that an agent? It's not an agent, it's a full-fledged solution that delivers information to me, delivers it completely, in everything, in any form I have it, right? If, say, I'm buying myself some device now that I need for, I don't know, measuring blood pressure, say, then it… I…
have to open an app, sync it with Apple Health — so I'll say: "Listen, all of you just pour all the data straight in there." So. Or let it collect that data from fifteen, or twenty, or forty sources, and let it also periodically analyze the market. Maybe some cool new technologies have appeared on the market. Bill Gates literally a couple of days ago reported that before, well, they work a lot on helping in Africa, where children die at birth and a lot of mothers die giving birth.
And it's two hundred and fifty thousand, he gave some number of people, of mothers, and one of the issues is that you need expensive devices that do cardio scanning of the mother during labor and so on. And he showed that such a system costs, like, twenty-two thousand dollars, the device. He says: "We've made a device that will cost ten times less, for two and a half thousand dollars." That is, someone will release something, like this sensor of mine, this Polar heart sensor, right, it costs a hundred dollars.
But to actually recognize the data from it, you have to pay another thousand dollars for the software. And here you don't have to pay for software, you'll easily connect that sensor into your interface, well, into your system, if you need it. I've talked about that a lot too. So. That's why I said to Ilnar that it turns out, Gmail, what's the email for? In terms of… or if I have a plane ticket, is it because the email received the data there, is that what you meant?
No, no, it's a specific, a specific example from the presentation, where they simply explain that before, you were doing searches and other things, and here you can do it by voice, with a voice interface-
Who are they talking about? For people, the billion users, and those users fly somewhere once a year. Well yeah. And so people are, like, lost, they don't know that they're flying. They know they're flying on vacation six months before they fly.
They've got it printed out and hanging on the wall.
Of course! They printed it out, and three weeks ahead they know exactly where they're flying. What, what? What kind of use case is that?
And the ones who don't know have a personal assistant. The point is that you don't need to, well, program, roughly speaking, write special commands in filters and so on, you can do it all by voice. You know, the first time I had to use those filters was probably fifteen years ago, when I ran out of space in Gmail and had to find files, attachments larger than ten megabytes. And that's when I first got acquainted with filters in Google. Right? And now, funnily enough, you don't have to google them.
After a while you'll just be able to ask it to find all files of such-and-such size. Well, although even now you can tell ChatGPT: "Listen, which command can I do this with?"
Yes, it works very badly.
Yes.
I keep trying.
But, but still, some things work. Some things in terms of… here's a question for Ilnar: I have my email, and recently I had to buy extra space in one of the Google storages, because my files exceeded some limit again, which added another fifty dollars a month. I had to buy it, because I don't have time to offload certain documents to storage. I'd picked that they should be stored in Amazon. There's a storage where, once you've uploaded, every download, every read and write costs money, right? I mean, they store it on discs, right?
They store it on compact discs, on DVDs, on DVD discs, and then…
I don't know. Probably, yes.
Max told me, Max told me, Max told me, yes, yes, yes, yes, yes.
Are DVDs still used anywhere?
It's not that it's used, it's a solution-
It's a very good story. They don't degrade, they can't get demagnetized like hard drives.
Yes. Look, it's a super cool story. And so the only thing is you have a wait for them to load it back for you, but it's short, like two-three days. Obviously you're not a person going, "Hey, put the disc in, guys." Right? And you make a request.
And here we are talking about AGI, right?
Yes. And, yes, you make a request.
Not on videotapes, then?
Yes. But the point is, I went into AWS, I went into AWS, and Max Grigoriev, our viewers' favorite, he's coming back from his trips soon and we'll have episodes with him too. But here's the thing. Max Grigoriev says to me: "Ah, set up AWS, it's all easy there, I'll set everything up for you." First of all, he never did set it up, because he somehow didn't have time. But I can't be bothered downloading the files anymore. Because, well, what's fifty dollars a month? When I went in there, I thought I don't even want to set this up with ChatGPT's help, because, because, because it's a nightmare.
So. You have to create a bunch of stuff there, upload, download, move things around. And maybe it's not super, mega long anymore and so on. But when I hear words like "create a bucket", "go into…" — I did have to figure those words out, actually. Elsewhere, you sit there like a normal person and think: what is this nonsense? Why don't you just have some button that would just tell me-
Make it good.
Migrate from your Google Drive and forget it. And I'd just press it and pay AWS the money. But look, this is a question of-
And not fifty dollars, but much more. Right, Sash?
No, no, it's less there, I was paying five-six dollars a month there, right? It's just that I'd set up two storages there, I'd set up one just like Google Drive, instant and so on. And the second storage is the one I told you about. But the point, the point is that for me it's not even a question of 50 or, like, 100 dollars or whatever. Right now, since I'm programming in Codex, I've bought up so many servers and everything that I don't even know. I think I have 800 keys. So. The question is this.
The question is that you pay for these separate subscriptions everywhere, and you don't actually understand what you fully, genuinely need. Google just released a new Google Pro plan for AI. They released a new plan in YouTube, YouTube Lite. I had YouTube Premium, which kills the ads. And I get a message: you have a double subscription to YouTube Premium, one through Verizon, one through Google. And you sit there thinking: how sick I am of all these double, triple, identical subscriptions.
It's just impossible anymore.
Try unsubscribing, yeah.
It's just impossible anymore. You stop understanding what's actually going on, yeah. So. But we'll see. So, Ilnar, tell us, please.
Ilya Sutskever — I was talking about him on Wednesday, on Wednesday actually… ah, not Wednesday. In the last episode, about the Elon Musk–OpenAI conflict, I mentioned Sutskever a lot. There's information that he's starting to be in very close contact with Anthropic. Tell us what's going on.
Ah! Well, what's to tell? It's a very cool story that Karpathy is now working, he's joined the Anthropic team, the research team. And, accordingly, he'll be working on improving, if I remember correctly, self-improvement, that is, improving the model with the help of the model itself. We can talk about that in more detail some other time. It's a technical story. Well, in general, one more way of improving LLMs. And it's now getting more and more popular. And Karpathy, if I understood correctly, joined Anthropic specifically for that task. Here, really, there's just a bunch of funny memes, that, for example, Karpathy joined the Anthropic team just to get free access to the Claude API and keep using Claude, because it's quite expensive, so that's why. So.
But there are probably several aspects to single out right away. Karpathy, first of all, besides being a good scientist, is also a very well-known media figure with a very good image, right? He's untainted by involvement in any scandals or anything like that. And for Anthropic, I think, it's also a bit of an image story. Besides the fact that Andrej himself is a good specialist of a very high level, it's been a while, I think, since he worked closely in R&D specifically. He had his own project in EdTech, about how one should learn.
He recorded brilliant videos on YouTube explaining how transformers work, built his own GPTs and put together educational courses. So still, kind of, not a frontier scientist. And, as it seems to me, for Anthropic this is in many ways an image story as well. And besides that, a bunch of jokes appeared that Andrej joined Anthropic to get access to the Claude API, essentially for free, because right now the Claude API is a pretty expensive thing, and not everyone has unlimited access to it, but Andrej most likely does now.
Well, it's important to understand that Karpathy is a billionaire, and I doubt it makes any difference to him what it costs.
That's true.
Yes, so that everyone understands what an AI engineer is, and an AI scientist.
He's a billionaire, yes. Like Greg Brockman. That's one of the problems, the details that surfaced in court. He is still protecting his 40 or however many billion dollars that he has... Probably not 40 anymore, already more than a hundred, it'll be over 100. Which topic do I want to raise? I want to raise a topic, quickly go over the China bans, chips and all the rest.
First story. There's a story going around that Trump is pushing a bill so that absolutely all models — all American models, before their release, ninety days ahead, give access regarding the update, to, well, politicians, and they'd check those models. That's the first big topic that's out there. I don't know whether the US could actually pass something like that, and whether such a story would work, because if it does work, it's a big blow overall. What does it even mean to hand a model to the government ninety days before an update? Can you imagine what that is? How can it even work, that you have to get their sign-off?
But in theory, let's imagine, in theory, thank you. Let's imagine that in theory it's, say, possible, inside. In parallel, then, the topic: US senators are preparing a bill, a bipartisan one… a bill that proposes creating a special office, a government one, a government department and a fund of about half a billion dollars, so that it's easier for allies to buy American AI models and chips instead of Chinese ones. And right now there's a very big story going around the market that Sberbank is fighting to buy Chinese chips, because they're under sanctions, and it's competing with Chinese companies right there, locally.
With Baidu, Tencent, Alibaba. And the fashionable chip now is the Huawei Ascend nine hundred fifty. And all this increasingly shows the story that, still, let's be honest. The topic we raised two years ago, we talk about it constantly — well, I'm fond of saying it — that today only the US and China have the infrastructure. And the volume the US has in terms of money — imagine, there'll now be three IPOs, all over a trillion: OpenAI, Anthropic and SpaceX. That's the top three in AI. Well, there's Google, which is already a publicly traded company, with Gemini, right, top of the world, top five in the world by valuation, and growing like crazy.
They're all raising money through IPOs, or raising money by, like, diluting shares internally, there's a huge pool of investors. And Elon Musk wants to do some incredible roadshow that nobody has ever done, and collect a mountain of money, internally. So it's big infrastructure, and the fight for compute will keep getting bigger and bigger. I hear that a lot of people have been trying to get around the restrictions in terms of sanctions, for example on using Anthropic or OpenAI.
And I see OpenAI increasingly blocking such countries and excluding them. That is, if you're in a country where using some software is prohibited and you want to use it for your company, you'll implement it. There's a huge probability that OpenAI will block everything in the next six months to a year, because at some point they'll definitely figure out you're using the system in a sanctioned place. And since this is AI, checking that is, essentially, well, I think, quite elementary, yeah.
And in parallel, here's what I want to say about China: Alibaba and Tencent. Alibaba presented a new AI chip, a very big one, as a response to the restrictions regarding NVIDIA.
And Tencent is, of course, also betting on it and, well, growing its overall AI investments internally. Just, well, to reflect the world of what's going on. China, by the way, has its own problems. So if you look at the Chinese market a bit more, study it, for those interested, you'll see that the amount of money they have and the seriousness of the projects and the seriousness of the users still lags behind. By the way, Tanya, remember last time, when you came back from China, you said that in China a lot of people use Anthropic and ChatGPT and so on.
And we said that it's partly to get around that kind of fundamental search. And I also read an article that there are resellers there, reselling Anthropic for almost zero dollars. There's this story that in China they can somehow connect it super cheaply.
Well, yes, they told me about it. I didn't quite understand the system, of course. I asked how they use it, through a VPN or something, but they said it's somehow different, but I didn't quite understand how. And all the people there who can afford American models use American models. Which really surprised me, because if you remember how Sasha, our other Sasha, kept saying that China is actually ahead and its models are cheaper, that didn't quite line up with reality, but it's quite possible it's just because the place I was in may have its own specifics and a particular field.
Well, but with all, once again, with all the aspects, it's clear that Chinese models do many cool things, but in terms of the world, honestly, well, we have to look at who the leaders are now.
If we're talking about the most widely deployed mass player in terms of AI, it's clear that Google has a certain priority. Given that, yes, indeed, a person who uses search now seemingly isn't using AI, but on the other hand, if their engine becomes AI, an AI engine in itself, then whichever way you look at it, it's AI. The question is this: will people open, I mean, what, what will the future Google.com be? That is, clearly the future Google.com should be some kind of bar, some kind of terminal, right, that you open, or some Google inside that you open, and it does everything for you everywhere, damn it, right?
Same here.
I use the chat. So with all the aspects, I understand there are people who feel good about Gemini and use it. And Google.com, what will they be able to make of it? That is, will they leave? Because search won't exist. Look, let's be honest, the very concept of, well, search engines won't exist at all.
Rambler.
Well, Rambler was a bit later, probably, even Rambler, As, As, As, I don't remember. Well, there were various ones. So you used separate search engines separately. It's like email. I mean, do you need a search engine? Because, oh, do you need an email client, in essence? Because if it works properly, email can easily be sent automatically from some other interface. Well, why would I need it? I understand that's some certain future, but still, Google today has, well, the top leadership position, just an incredible leadership position over everyone.
Will they, will Google's chief be able to hold onto that? Will they manage not to get into and not create some nonsense? And maybe even right now, when in my Pro version they've added YouTube Lite, it's like, why are they doing that? Because I think, why do I need the Pro version? I only have the Pro version to expand my cloud storage. And you go, huh. So what advantage does Google have? They can give you perks. I mean, you bought something in ChatGPT, you don't get a perk. By the way, I'm waiting until ChatGPT Pro… Well, I'll tell you this. Yesterday I had a conversation with an IT partner.
I told him: "Lyosha, I'm waiting for ChatGPT Pro to cost more than a thousand dollars a month, and then a huge number of people who have an advantage, well, certain advantages, will drop off." Maybe I'm, maybe I'm reasoning… Look, my reasoning may seem to some: oh, what kind of reasoning is that?
Right now people just, people just don't understand what they can buy for two hundred dollars. They just don't understand. As Ilya says: you can buy five accounts in parallel and get the volume. How much? I mean, in an hour I eat up some cosmic amount of their data. That hour of my requests on the two-hundred-dollar Pro plan clearly costs more than an hour of requests I'd make to the API. Not "possibly", guys, it costs thousands of dollars. What I do, thousands of dollars, really. Over the course of a day.
And I'm just waiting for them to make the plan expensive, so that as few people as possible use that plan, less than one percent of those who use it now, so fewer people program and build cool systems. I'd probably be ready to pay even five thousand a month for the Pro version right now.
Sash, it's more like this, you know. Want to know what it's doing?
Do I look like it?
No, no, want to know what OpenAI has been doing for the last six months? Look at what Anthropic is doing. So, what did they do? They allocated you a certain quota of tokens on the subscription, and everything above that you pay for from your account with money, via the API. That is, roughly, on the hundred-dollar version you have some number of, well, not requests, let's say tokens, that you burn through once you've picked it. And it's much cheaper there, if you calculate tokens to dollars. If you take the subscription as a whole, then everything above that threshold you just start paying from your own balance.
And I think OpenAI will more likely come to something similar if it runs up against compute. Maybe they aren't running up against it yet, or for them, before the IPO, it's an image thing, so you don't hit a ceiling in the Pro version. But going forward, if that happens, I think this outcome is more likely: there'll be limits on the Pro version, and everything above- Yes.
I like them all.
Actually, I like them all too. And I'm going to check the architecture of various solutions using Opus. Right? I've already said this and I'll demonstrate it soon. I'll demonstrate it when I throw a couple thousand dollars in tokens in there so it analyzes a huge volume of code for me, because that's exactly where the problem is. What's the problem now? Right now you can't give it something very serious as input and get it out as output. Everyone, everyone has this problem. Like, looping, hallucinations, well, everything. Everyone has that problem. And I'd like the system to work through it on its own. Well, Sam Altman doesn't have that problem.
But here's what I want to say once more. Once more. Look, there's this news, it kind of looks funny from the outside, but, but it's important context. Greg Brockman is officially taking control of OpenAI's products: ChatGPT, Codex and the API. What are they going to do? They're making a Super App. Well, you can laugh now, because ChatGPT isn't managing to make a Super App. Usually, well, more precisely, they usually can't build a lot of functionality at the same time, develop it simultaneously, they develop everything bit by bit.
But if they do it, then still, Ilnar, most likely the monetization and the pricing policy will have to become completely different, different. Because the point of this two-hundred-dollar plan originally — they made it to separate out, like, premium users and show people that there are tiers. I don't understand why they still haven't put out a thousand-dollar plan. Ilnar! Why? To separate further, to show people some exclusivity. I said a year ago already that that's what they should, that's what they should do.
They should show people that there's a certain exclusivity, that it's not us, the astronauts, hogging all the coolest systems, which only researchers like Ilya Sutskever or the US government have access to, but so that ordinary people too, like me, an ordinary person, could easily buy themselves slightly more expensive access to, well, a cooler computer, right? And we'll see, of course, what they do here, what the plans will be. What you're saying about tokens is very interesting. In principle, a token, a token policy is the right thing for the world.
When the token becomes the world's currency. I'm talking about the token in AI now, right? What Jensen Huang talks about, for example, right? It really is the world's currency. On the other hand, to have a billion, two billion regular users, you need a plan. And people are used to paying for a subscription, like with Netflix. And people want to have it like utilities. By the way, it should become a utility. I was talking about how there's this American word, "utility". It's still — my wife and I were driving along — it differs from the phrase "communal services". It's a broader story in terms of the whole infrastructure.
And she tells me that in Minsk her mom's Internet already comes in with the utility bill, right? So, whereas in America it's still a big question. Is AI a utility now or not? And that's one, one of the huge topics of discussion. And all the special episodes of ours, well, that are going on now, apart from, well, apart from this main podcast of ours, are about that. About the fact that if artificial intelligence becomes a utility, then you really will have a plan, you can use it at different times, like electricity, cars.
My cars charge at night, right? Not during the day, because the rate is more expensive, more expensive during the day.
In winter it'll be cheaper.
Automatic pricing. Well, we don't have the concept, you know, right? There's no concept of winter, but you have it. Or, say, in Russia-
Cooling servers is cheaper in any case. I mean, whether it's forty-degree heat outside or zero.
Well, it all depends on where the servers are, right? And whose servers they are. Yes, Elon Musk actually said, with Tesla, that you'll be able to power up your Tesla and use it as a server. Yes.
I'd add something about pricing policy too. On the one hand, you're right that it should go up, but on the other hand, the market has so far moved in the opposite direction. Remember when Grok 3 Heavy appeared, I think it was you who took the subscription, it cost three hundred dollars. So Musk was already pushing above two hundred, but then Anthropic rolled out its hundred-dollar one. It's getting more and more popular. And some episodes ago we discussed that OpenAI is also bringing out a hundred-dollar subscription, that they cut it from two hundred to one hundred.
No, it doesn't want to! Ilnar, you won't believe it, it doesn't want to. Its site already says "new plan, Google AI Ultra". I like that it's called a new plan. From ninety-nine ninety-nine. So it already, yes, yes, it already has.
The market's going to a hundred anyway.
Look what it gives. It gives up to 20x limits in Gemini and the Pro plan. It gives this Google Spark I was talking about, the new one that's coming out. Well, early access to innovations and so on. And it gives all sorts of Google benefits. Actually, for whoever uses Google's infrastructure for development, there are cool benefits, because it gives a lot of everything for all its systems inside Google. I mean, there are things that are the basics. For me, say, YouTube Premium matters, right? So that ads — I never see them in my life, I don't need them.
So I have YouTube Premium, and it gives NotebookLM at the maximum tier, Notebook Flow with full maximum access, Google Antigravity, Google AI Studio at maximum, Google Search — ah, there's a cool thing there, that in Ultra and, and in Pro too, by the way, you can use Google search via the API, with the most super-powerful capacity there is. Though the agents there only work in America, but still, it's there. It gives the very maximum access to Google Finance, Google Chrome, to Jules, to development, to Android Studio, the biggest Google Photos and so on.
So there it is, well, they pile on this, what's it called, this junk, which I don't know whether it's needed or not, well, someone probably needs it, again, right? The Android person who's fully absorbed into all this Google infrastructure and Google Cloud, right? The one who's also DevOps-
Managed to set it up.
I just, I just spent a little time there, and I'm simply amazed at it, well, it's ninety-seven, probably ninety-eight for me. In terms of development interfaces and the ease of visuals, and generally of getting anything done. You'd just have to be sick, well, you'd have to be, you'd have to be, uh... And I, well, whoever knows me — Ulyana knows me very well in terms of the technical work I do, Tanya knows me very well. For me all this is always very easy. I mean, like, pressing forty-eight buttons, memorizing actions. But Google Cloud stunned me, it just...
I mean, such, such trash I've never seen, such trash I've simply never seen. It's an anti-system, purely, of UX/UI, and an anti-system to confuse you and get you hooked. And I think that if you've gotten hooked on Google, there's, like, no way back out of it, because nobody will be able to get anything out of there at all. I, I don't even understand how to do all of that. So. And when you… and you're also granting access everywhere. We live in such a funny world now. Everyone is restricting access now.
Just now there was a story that AWS explained how to restrict access in browsers. And I sit and think: does anyone even understand, when you grant all these accesses everywhere, all these JSONs, API keys, tokens, publishable token, refresh token, access token, you allowed the agent to do this in the browser, and forbade that, did this. Does anyone at all understand this, among ordinary people? Not Ilya Sutskever, right, or not Greg Brockman, who... And even they, I think, don't understand it either. Well, Sam Altman, we know for sure, doesn't understand this.
And you go: "Then how am I supposed to develop anything at all? Your programmers already leaked all the tokens to everyone long ago, completely, and can do whatever they want with them. And your security team thinks it controls everything. And the programmers have a bunch of tokens they do whatever they want with."
Sash, the amount of source code of various systems that's been uploaded in there over the last year — it's, it's just insane.
It's just insane. And note that even companies like Anthropic or OpenAI have leaks. Although Anthropic is a top-secret organization, right, where it's secreter than secret, secret secret. Yes, I mean, they've already classified the oranges, right, or the food in their cafeteria. I actually ate once in their cafeteria, where the programmers eat. Well, I'd have been better off not eating there. I had to meet one of the directors there, and it was a very strange meeting. In cafeterias like that.
But that's the tech world, and I ate some salad there, yeah.
A secret one.
A secret, secret salad. Well then, see you on our channel in exactly a week. Watch the special episodes on the channel, subscribe. And we look forward to your comments and discussion, because we read all the comments. Until next time!
Bye.