Hi everyone! This is the ToTheMoon channel — tech news and insights from Silicon Valley and around the world. Do you know what I want to start with today? I want to start with a really interesting comment that someone wrote on X: "We may be just twelve months away from an agent that can watch your work and then do it without any instructions. The combination of screen recording, agent memory and a model running right on your device makes it possible to do this right now." And then this same person brings up something interesting, some interesting comments. He had meetings there — Huang was there, you know, really serious people, here in the Valley.
And he gives a lot of interesting examples, in particular his own sense of anticipation and of what's going on. And there's also, there's a big story there too, in particular that a lot of people are talking about how they decompile the code of various existing software products that are worth — the figure is stated right there — more than $5 billion. It literally says so, right? And then they clone those products with the help of artificial intelligence. It's a really unusual time, of course, isn't it, Ilnar? A very unusual new time.
Let me remind you that I still don't know — I was supposed to by this episode, and many of you may have been waiting since Friday — about how Amazon showed that I'd spent $37,700 on servers. I set the spending at $350. It suggested that to me. I wanted to set a $150 limit, but it suggested I set a $350 limit. And I'm like: "Well, okay." I set the $350 limit, logged out, and ten minutes later I got an email saying I'd spent $37,000. What's important, and very interesting, is that the service blocked itself automatically. My access to the Amazon servers was automatically cut off, for my projects in particular.
And on top of that, Amazon also completely shut off my access to all the statistics. So I can't even see whether this money was really spent or not. It's just not there. Second story. The second story that's out there is, of course, the question: what if I really did spend thirty — $37,000? In this case it wasn't me who spent it, it was Claude Code. I mean, what if Claude Code really did make mistakes and, once it got access to the system, somehow managed to get around the limits that were in place, change those limits, get some kind of additional access, somehow tell me one thing and get something else?
Ilnar, I don't know, have you seen this story or not? You haven't seen it yet.
Not yet.
Yeah. I still can't figure out whether I spent it or not. So, for those who were waiting for an answer in this episode — well, I wasn't planning to spend it in the first place. I was planning to spend $75 or $100. For those who've been waiting since the previous episode: as of this recording, I still don't know. So, we'll keep you posted. And-
What does your bank account say, Sasha?
They haven't billed me. It just says in my Amazon account that I supposedly spent it. And, like, since I'd set a limit — like, you have a $350 limit, it says, you've spent thirty-seven, thirty-seven thousand, like, seven hundred, so set a limit of, like, $120,000 or, like, $150,000, and then you'll be fine. And they keep changing that prediction all the time. What bank account? Lots of people have already given me all kinds of advice on what to do. Well, well, what good is advice? For now I'm watching to see whether Claude Code really did this or it's an Amazon glitch.
And for now everyone is counting on it being an Amazon glitch. But we'll still see what actually happened in this story. Although, of course, when you go into Claude Code or Codex and tell Claude Code, like, here's the situation that happened, it says: "Ah, well, yes, we just had no protection here, so…"
"Yes, yes, yes, you're right," they usually add as well.
"You're right." Yeah. Well then. Yes, let's get back, let's get back to it. That was just an aside, just an aside.
Yeah. About the idea that in, say, twelve months agents will predict your wishes — before you've even given it a task, it'll start doing it. I can easily believe that. First of all, people are fairly predictable, especially when you have a snapshot of what will most likely be several billion users by then, who've asked about just about everything. Second, I think many of you have already seen this — in Codex, for example: you're working on some task. You've told it that it needs to do this and that.
After a while it comes back with an answer, and in the prompt box, the field where you enter your next instructions for it, there's already a line in this faint grey font with what you'll most likely tell it next. I mean, a guess at what you'll ask it to do next. You can press Tab, and that becomes the actual entry. You press Enter and off you go. So already now, that same Codex predicts your next action in advance. In some tasks — fairly, well, they call them straightforward, right, how do you say it properly?
In linear tasks, where the next step is generally predictable, it's probably really easy to do. But I can easily believe that this will extrapolate further to other areas of life, to other tasks and so on. When it'll know in advance, like in that silly example from the OpenAI presentation, that you need to order groceries — which groceries you'll need, depending on whether you're having guests or not — and it'll put something together, suggest you do it, and all you'll have to do is approve.
And in some cases you won't even need to approve, right? That I really can easily believe. I'd like to remind you once again — we've already talked about this "Doctor Who" episode, and in the comments there were even lots of people who recognised that episode. The sapphire planet, "Doctor Who", with David Tennant. The story went like this. A character appears who at first repeats everything after a person, gradually starts catching up, and they speak at the same time. And after a while this character starts speaking before you start speaking. And that's very much like how neural networks behave.
I mean, our experience with them over the last few years has been that they carried out our tasks. Now, little by little, they're starting to — well, the moment you want something, Claude Code already — no, not Claude Code, right, Codex, that's the one I work with most often. Codex already knows what you want to ask it for. After a while it won't even check with you that this needs to be done, and it'll say: "Yes, this is what you need right now, I'll do it for you now." So I'm quite ready to believe in this story. If there's enough compute for the number of people needed, or if it suddenly doesn't cost some astronomical amount of money, it'll most likely appear.
Whether in twelve months or not — I think it may well appear even sooner. Remember, there were already even some rumours, predictions, that in six to twelve months there'd be an AI uprising and so on. So within twelve months this can definitely be done. And you can already see movement in that direction.
Listen, what's interesting is that you kind of, sort of expanded it — you heard what I just said in your own way, because I said one phrase. Look, I said that in twelve months there'll be an agent that can watch you for a week and then do all that work without any instructions. And you expanded it and said this: that an agent watching you will not only be able to do the work you do, but will basically also be able to predict your behaviour patterns, all kinds of details, particulars and everything else, yeah.
I mean, I just want to point out that you expanded it for the viewers, and — but I support the point you're making now, in particular the watching, because when it comes to prediction — well, given the amount of data in there, a lot can be predicted. Although, of course, you know, I look at my own life — I just had a laugh at myself. Yesterday evening I — yesterday Iolanta got her driver's licence, at sixteen. And imagine, I'm adding her to the car insurance, and they tell me: "You need to pay an extra $6,000 a year."
And I'm like: "What extra $6,000 a year?" And: "Well, because you've added a young driver." And so —
Yeah, you'll have to take the hit.
And you sit there thinking: how is an agent going to predict my behaviour patterns, and what I should do? For it, this is just some action, right. For it, well, she has to be added, so she has to be added. I asked ChatGPT: is it normal that my insurance just went up by some astronomical amount, by ordinary standards? It says: "Well, sometimes it's less, sometimes it's more. It's different for everyone." Meaning every company has its own, right. And on top of that it's all very unclear, really very unclear.
In the US there are tons of areas where, if you don't go in, analyse and look at the details — well, basically people have no idea at all what's going on. Some amounts just get charged to you, and people pay them automatically. So can an agent predict what my behaviour will be in situations like these, all the different ones that happen every day, right? Hmm, well, like, like with Amazon. I think that kind of thing isn't easy to predict. But basic, simple things in basic, simple systems — yes, of course, those really will appear.
A very interesting topic, yeah, it deserves a whole separate discussion. I just started with this, started with this. I really liked it. You know, I really liked that thought.
Look, OpenAI already has 1.2 billion users. Or, as you always point out too, maybe that's how many accounts there are, not unique users. Well, one way or another, if there are, I don't know, eight or nine billion people now, then they have information, in one form or another. Obviously there aren't that many active users who really pour a lot of information in, but they have information, or can get it, on, say, every seventh person on the planet. We throw out some portion. Say, yeah, in Africa, in some regions, there probably aren't that many users, but they're not very much like you anyway, right.
And on the one hand, you'd like to think that every person is unique, unpredictable and so on. But when you have that much information, if you can process it properly, then in most situations a person's behaviour can, I think, be predicted, if you have a large enough body of information.
A hundred per cent. That's, that's, that's for sure. I was just giving some particular exceptions, and — but basically, a hundred per cent. Listen, you know what, I think it's a really great topic today, and of course it deserves its own episode. By the way, in the next few days we'll have an episode with Max Grigoriev.
I'm very glad that he — well, I hope, I hope he'll be there, yeah, because we're filming with him in an hour and a half. I hope he doesn't disappear. Max loves to disappear. Well, we do see each other often. But video — just try dragging him onto a video. But what definitely deserves its own episode is the topic of today's coding systems on the level of Claude Code or Codex. Cursor deserves special attention too — it has grown quite a lot. And in general, I'll point out to everyone once again that SpaceX — in this case xAI, or, well, now they're playing around with xAI, right — is a serious global competitor. Seriously.
By the way, Ilnar, they claim they have 500 million users on X. And I looked into this. It turned out that X has actually slipped over the last five years in terms of daily audience, because the daily audience was a bit over 200 million, and now it's supposedly a bit over 100 million. Supposedly — nobody knows for sure. But why? Because they give a figure of 500 million users, and that's users together with Grok. I say: "Listen, I was looking into this information yesterday."
I say: "Listen, it just can't be that Grok has 300 million users. Grok just can't have 300 million." Well, my feeling is that Grok isn't well-known enough to have 300 million users. But I'd still keep Cursor as a strong topic. For example, Grok released Grok Voice. That's the speech-to-speech thing, the one that talks, right? And on a number of parameters it's very serious, in particular in analytics — for example, in comparisons with OpenAI's speech-to-speech. Although I think OpenAI, specifically in working with voice and with approaches to different languages in general, is top-notch.
I used to rank Google at the top too, but Google — somehow Google Gemini in my car infuriates me so much it makes me sick. I just want to — I want to see this developer and post him, you know, put him up on the internet and write: Wanted — whoever made this wretched thing, right? An absolutely, totally dumb system that understands nothing, like, of course, a huge number of services from big companies like Amazon. Right? But look, what's our topic today? Google is releasing a new update. They're releasing Gemini 4 Argon.
As for Argon as such, I think it's still too early for us to make a proper comparison. You'll give your opinion in a moment, of course. Personally, I think it's too early to compare, because judging by the experience of the Sonnet, Fable, Sol and Astra releases from Anthropic and OpenAI, it takes some time to even understand how these systems work, to really compare them rather than get this kind of marketing comparison. Or like someone going: I ran the tests and compared how it works. I mean, some time has to pass so that people can weigh in after working with it for a long stretch.
But what's interesting is that Google, Google does have Google Antigravity, and Google Antigravity isn't a new system — it was launched on May 19, 2026. But it's not just a model, and it's not just working in the terminal — it's a shell that lets you interact with the model through a proper app, the way we do in Codex or in Claude. By the way, Ilnar, I once made a Short, and it said: who still works in the terminal? And people there were like: me, me, me. Well, because it's all different.
That's very good.
It's still a different story, a different story. But still, systems like Codex or, say, Claude Code are made precisely so that people could — well, we are moving away from the world of working in the terminal after all, especially with all the connectors and plugins.
Let's put it this way, Sasha. Most users never went into the terminal in the first place, so they can't be leaving it.
Yeah, they were never there. When I say we're leaving, I mean the era is moving on anyway, but, like, well, that's already a vestige of the past, right? So, Google Antigravity is a separate, independent environment where you can get things done. It appeared at the end of May, and nobody really paid much attention to it. But now, inside Google Antigravity, on October 13 of this year — that's literally this week, right — they added, at the start of the week they added Gemini 4 Argon, and so, for me personally, Google Antigravity has, for now, joined the lineup of projects like Cursor, Codex and Claude Code.
In terms of work, in terms of what kinds of tasks you can give it there and what kinds of projects you can do there. And by the way, I want to say — do write in, by the way, those of you who've tried Google Antigravity and those who've tried Argon. Although I think Argon isn't available to everyone. It's only available for corporate use, with a whole bunch of restrictions and stuff, but-
They've only just unboxed it.
But we have different subscriptions. Yeah, yeah. And OpenAI Dots hasn't reached everyone yet either. I mean, it's still only in Pro. And again, by the way, Ilnar, OpenAI sent me a marketing email. Recommended for you. We've got a new product, Dots. We recommend turning it on. And I'm thinking — I once went to their website, and they had a vacancy for someone doing email marketing. And there, you know, the requirements were really strict, tons of stuff about email marketing, and you're, like, some kind of executive.
And I'm thinking: well, fire, fire, just fire that whole team already. How can you, how can you send out emails like that? How, how, how is that possible? Here, well, here I'm a realist, actually — whatever anyone says, I keep it real. It's nonsense, nonsense and stupidity. So, Dots hasn't reached everyone. Obviously Google Antigravity with Argon will reach people a bit later, but we have all kinds of people at ToTheMoon, and many have different kinds of access as it comes through.
So write in and tell us what you think of the system overall. I'm glad Google has it now, of course. And, well, first of all, it's broad availability worldwide, because Google really is widely available around the world. It's Google. Google has a lot of these bundled subscriptions. I mean, a lot of people were already paying Google for something anyway, I don't know, for Google Drive, for example. And they get access to this single infrastructure, a kind of unified system. Overall, Google has had some pretty good, interesting ideas, like NotebookLM and so on, for all sorts of things. Although I remain, alas, a devotee of OpenAI and Claude.
And we'll see whether Google can stay in this arena with a really serious system like this. Ilnar, what are your feelings and thoughts on this in general, from your professional side?
Listen, well, they announced they'd release the model a few weeks ago. We talked about it in previous episodes. Now it's finally out. I'm tempted to say: it came out — and so what. Because, well, yes, they showed better results on some benchmarks. Well, they literally added a few percent here and there on various benchmarks. It's great that they've finally, well, at least come back into this league, but they haven't turned it upside down. The previous numbered version of Gemini — that's what caused the code red, when Sam Altman said: "Drop everything, try to catch up with Google, don't let it get away, and so on."
And back then Gemini 3 really was — well, not head and shoulders above, but substantially better than what OpenAI and Anthropic had on the market. So back then they caused a stir by managing to build a much higher-quality model than everyone else had. Now they've simply, with a lag of — well, six months would probably be an overstatement — with a three-month lag, released a model that's at roughly the same level as Fable, or Opus 5.5, or, well, Astra, which ChatGPT has. So they finally have a model in the same class as their competitors have.
So you think — listen, but in your view it's the same class? Sorry for butting in.
I think it's the same class. At least from what I see in the benchmarks, give or take, it's all similar. Yes, here and there something may be a bit better, but there's not much point in dropping everything and running to Google. The result will be similar. Maybe the optimisation is better in places, maybe it'll do a bit better on some tasks. But it's not the kind of difference there is between Sol and Astra, for example, not the difference between Opus 5 and Fable when it came out.
So it's a slightly different story, after all. They've simply reached this level. Maybe they've poked their heads slightly above the current frontier, but on most tasks the overwhelming majority of users won't notice the difference. In my view, that's not the same as what happened with Gemini 3, when a lot of people I know, especially for programming tasks, really ran to Gemini, and everyone claimed — well, many claimed — that Gemini was much better than the then-current models from OpenAI or even from Anthropic.
Now, in my view, it's a different situation. But I'll back you up. It's great that they're back. A wider choice of providers who actually offer these models. Some people feel closer to Google, others to OpenAI. Either way, in such a big market — basically almost the whole planet — the more vendors offering these models, the better. It's great that Google is still in the game too and has come back into the picture.
Listen, my feeling so far — you know, I've been studying it, because it's really hard to understand or check at the very beginning, right? And because, you know, when Astra came out, you'd ask Astra some things and say: "Can you give it back to me? GPT-5.5 was answering better for me, right?" And sometimes you look and just don't get what the improvement is. Although when you look at it as a whole, everything added up, you realise the improvement is actually very serious, right?
It's just that we get so used to a certain level of quality that it seems to you the system shouldn't do anything worse at all, right? Although remember 5.5, and 5.0, and like 4.5. All these systems would periodically just glitch like crazy. And what's more, with every new version we'd say: "Give us back the old version." Every now and then, yeah. But then you look at it over the long run and realise it's a different class of system. But so far, judging by the description of Argon, Ilnar, it still seems to me that it was released, like, for the shareholders. Like, for the global trend. I mean, a lot of things are hard for you to verify.
They even have a lower price, right? Argon is roughly two, two and a half times cheaper than Fable and Astra. And Fable, Fable 5.1, and Astra — I think they're priced about the same, give or take. And I won't name these prices, because all these millions of input and output tokens are still a completely incomprehensible thing for people, right? Who understands it, I don't know. I think OpenAI itself doesn't fully understand it. Judging, by the way, even by what their developers were saying on stage, Ilnar, right?
Because there wasn't a single coherent answer about the cost of tokens. They say: "Well, roughly such-and-such a price, yeah, well, roughly such-and-such a structure." And on top of that you don't understand what their system is doing. We know a system can get stuck in a loop, we know a system can go off in the wrong direction. By the way, Ilnar, I think I broke all the records again. This week I did it the way Tibo did. By the way, guys, OpenAI's marathon with Tibo is still going on.
I think it's — unheard-of, nobody knows — unheard-of generosity, right? Someone here at ToTheMoon writes: "Listen, but I'm not taking part in the marathon. My limits get reset every day, right?" And I say: "But they get reset for everyone. Tibo is simply unprecedented generosity." Ilnar, basically, yesterday I ate up 1.5 billion tokens, and the day before, 2.1 billion. Yeah. And before that 500 million — well, basically, those are the kind of decent volumes I've got. I think the day before yesterday, or yesterday — the day before yesterday, I think — I reset it three times, did a reset.
I mean, they do give out some gifts. Yes, they give out gifts. There were also some purchased tokens, something else too. Basically, I just… And one task kept spending, kept spending everything like that. At some point I think: maybe it's just eating my tokens. I go into that task and write: "Are you sure you're doing everything right?" It says: "Oops, I went a bit the wrong way." Again, I'm thinking, you little pest. That was in 6.0 Astra. I let it loose on too broad a scope, but I want to tell you, I really did put it on a very serious piece of functionality — sales.
The ChatGPT chat itself obviously can't do tasks like that. I mean, these are long tasks with a huge amount of data, way too much data, with re-checking code, publishing text and content and everything. The system was translating a huge array of different entities, data and everything into English for me. And then it was also deploying it, obviously, within the platform. But to Tibo — just my gratitude, right? Unprecedented generosity. I think OpenAI should have been resetting tokens for people who had at least pressed the "I'm in the marathon" button, right, not for everyone.
They'd have saved just an incredible amount of everything. Or people don't know they're getting reset. Maybe people don't pay attention. This morning I come and take a look. By the way, you can put widgets on the iPhone, and I've got a Codex widget at the top of my phone. When I fell asleep, it was 0. I wake up — 99. And I'm like: "Whoa! So Tibo's on it."
Nice.
You open X. Yeah, you open X, and Tibo writes: reset. A full reset again, right? So, about — getting back to Argon. It seems to me, from the data presented on Argon, that Argon — supposedly there's this overall capability index, right? They describe it as exactly the same as Fable 5.1's and Astra's. It seems to me that Argon is presented… Fable, you'll agree, was originally presented as a model for very complex architectures and for digging into, well, really serious stuff. Things like big code decompilations, big architectural analyses of large systems, finding security holes, checking various infrastructures.
And I always liked how Fable does it. And I still like it to this day, yeah. Some complex — lots of very, very unique kinds of structures. And still, this thing where you look at the overall capability index for Astra, Fable and Argon and it's the same — you realise, of course, that it's a comparison, you know, along the lines of: well, you're a man, you're a man and you're a man, you're between twenty and forty. Well, something like that, right? Who are you, what are you, where do you live, what do you do? That's roughly the kind of comparison there is today. And by the way, I want to say that according to them, Argon won.
There's this test, Automation Bench. It's about running workflows through tools. And supposedly Argon does all that better than Fable and Astra. And then, and then-
And which one came out later? You'd need to straight away-
And then, you know, they also start giving examples, like how they create some contract, a document. And you sit there, Ilnar, thinking: wait a minute, so first you describe how systems like Fable break into government systems and countries, and get into environments that are impossible to get into at all. Or, remember, the Mythos system. Let me remind everyone once again. In the spring they released Fable. It was sort of an appendage to the security systems of the Mythos model.
And that this system found, like, 10,000 vulnerabilities in various open-source code and everything else. And here they start describing cases where the system drew up a legally correct corporate contract. And you're like: "Eh!"
I mean, this divide in understanding is getting insane, of course. And I like the tweet you sent today about Karpathy, right? That Karpathy said — he said it a year ago already — that this gap between the people who understand what's happening in artificial intelligence and everyone else was becoming very big, and now it's, like, yeah, just gone through the roof. And really, I want to tell you — especially those of you who've been watching ToTheMoon for the two and a half years we've been going, right.
I mean, remember, two years ago we always used to say that the gap was colossal; now the gap is simply cosmic, and it's already… I mean, if before I could still, even just for fun, discuss some AI tricks or whatever with my wife, now I realise that it's, well, pointless. Like, with her it's some micro, micro things. I can tell her where to do what. But when it comes to serious things, well, that's it, I don't discuss them with her anymore, because I understand she's not, not in the loop. Right?
Yeah. And at the same time, about Karpathy — he works at Anthropic now, let me remind you, so he has access to everything cutting-edge. According to Karpathy, what people see — even programmers with these Pro subscriptions who actively use everything there is and so on — what they see is, well, a kindergarten compared with what's in the labs. That's how big the difference is between what's already been released and what's running on the compute of those very labs. And him, Andrej, I really do believe, because if Sam Altman had said it, I'd have thought: well, okay, he's decided to score some points again, draw a bit of attention to himself somewhere.
Maybe he's preparing a story for the IPO, or covering up some other bad news with statements like that, or something else, right — scoring political points. Karpathy has definitely never been caught doing that before. At least from what I've seen, and this genuine awe of his really does make you curious. What on earth is running there, right? Well, he talks a bit about swarms of, like, thousands of agents working there, working all this out among themselves, working through complex systems and so on.
Well, that's what we saw with the millennium task, what OpenAI told us about — that's obviously going on in all the big labs. But hearing that from Karpathy was very interesting. Do read that tweet — we'll most likely leave a link to it. In it he describes how he sees this as a funnel. How many people are at which level. Pretty interesting to look at. Find your level, by the way — where you are in Karpathy's funnel. That'll make a pretty fun game too, I think.
You know, about Karpathy — I think with all his statements, you always want to broaden that understanding. Remember when he even talked about writing some software that, like, controls cameras, right? People told him back then that only he could have written that. But I think it was, remember, literally a couple of weeks before that big surge, when systems like Codex and Claude Code really made it possible to do this. I mean, for people who aren't super-developers to just go and build it. They post really cool stuff.
In particular, I think it was Karpathy who posted it, or someone else among the strong people in the Valley's AI world. The person — I think it was Karpathy who wrote it — said: I made for the dog — never mind, anyway, I made a dog door into the house, built the whole 3D model myself, then sent it off, found a company that could manufacture it for me, and produced this electronic door into the house with sensors, unique, made for him personally, and for as much as a similar system — not unique, not made to size — would cost on the market, right.
I mean, simple cases like theirs do look, of course, like those pet projects, as you and I were saying, on the one hand. On the other hand, of course, people like Andrej Karpathy see more broadly and understand what all these systems really are. And this is still what's out there on the market. You know, you hear it a lot that AI models are some kind of, well, some kind of fun little toy, right, some language models, that it's just text being generated. I mean, people try to justify this one part. And on the other hand, a lot of people think they use AI the same way everyone else does.
Like: I program too, I make mobile apps too, I make some software too, I'm developing a website, I do all the same stuff. Everyone says: "I already do all that anyway." And people are like: "I already do everything, I already do it all, I do, I do, I do." And this gap, of course, is catastrophically large — in not understanding, really, what's going on inside. A very, very interesting time, just unbelievably interesting.
Yeah. Sasha, about what you were saying — that sometimes you get the feeling the current Astra works, like, slightly worse than GPT-5.5 and so on.
It just reminds me — there's, I don't remember what it's called, a cognitive bias or, like, a paradox of thinking. Well, I'll tell you about it now, and people in the comments are sure to write what it's called. Please do write in if you're curious to recall it. People quickly get used to the level of technology that already exists, and it comes to be seen as, I don't know, something taken for granted. A vivid example of this: it used to be assumed that a computer is just a calculator. Why? Because it can't play chess.
Chess is something only humans can do, only thinking beings. A computer will never be able to play chess. Some time passes, and the computer starts playing chess pretty well. There it is playing Garry Kasparov — this system, Deep Blue — and it beats him. And what do people start saying? Nah, come on, it's just a fast calculator. It can't actually do any of that. Now, when it… And then the next step begins, and then the next step. And there's a whole chain of these milestones that the computer gradually starts catching up on.
And as soon as it reaches that stage, people start saying: nah, that's all nonsense too. Now, when it does the next thing — and then it does the next thing too. And if you scale this down, the inner feelings of each individual person are just the same. Back then, when, I don't know, GPT-3.5 would translate — well, let's put it this way, make up some stories in English, and you could have these interactive dialogues, you'd think: "Oh, cool!" Then the next thing comes along, and you already start thinking that everything that came before is just a given.
And gradually, all the same, this level, the level of the norm, if you can call it that, keeps shifting. Although the current models, with these agents that launch subprocesses inside themselves, which run constantly and for a long time — that's, well, that's just something else.
Listen, that's actually a really great example. You often bring up this chess example, and it really is important to keep repeating it to people, because otherwise we'll be stuck in that mode. It's like how people piled on me just because I said that my future children won't learn to drive, they'll ride on autopilot. So the comments came pouring in: come on, it develops your brain, that'll never happen. You just live in Silicon Valley, your whole world just revolves around California, and so on.
But, well, first of all, judging by the fact that the insurance costs $500 a month, let them keep riding on autopilot. But that's, that's what it costs. It actually costs that much because of certain cars I have, right. Obviously I'll try to exclude, like, change things and so on. We'll see whether that happens or not. For now, it's my daughter — she was looking forward to having it so much, so I've already, okay, added her, let it be for now, while all this is going on. Although I think sometimes it's even cheaper to take a taxi.
But the point is, I was explaining that systems are developing, and saying that I love driving — someone, you know, people often say: "But I love driving." Well, well, okay, well, then drive once in a while. But obviously, if you drive to work for a whole hour every day, you're better off using that time to rest, sleep, read, or just look out of the window. Max Grigoriev actually came over to my place, and he says: "Listen, I hate driving. I'd be really happy if a car drove me from SF, San Francisco, to your place automatically, all by itself."
Or he says: "Next week I have to go to Las Vegas, so if only the car drove itself to Las Vegas. That's a totally, completely different thing in terms of involvement." And I told him: "Listen, I rode with a driver for about seven or eight years of my life- And when I rode in the car, a lot of people always held meetings, you know, when they sit in the back with a driver. And I always looked out of the window. I mean, you've ridden with me lots of times, right? And we'd either be talking, sitting inside. First of all, we had the chance to talk calmly.
I mean, you come out, your working day might have lasted twelve, maybe fifteen hours. Yeah, we get in, we're driving to, I don't know, dinner, we're relaxed, right? You don't think about where to park and so on. That's one thing. But I always just looked out of the window. Later, when the Zoom era began, that shifted the focus a bit, because you got this video communication, you started talking over video, and they started scheduling meetings for you in the car. Well, your secretaries just start filling up every moment you have. But this looking out, resting, right, the skill of resting, relaxing and switching to something else.
And that's why, in terms of progress, I don't want to argue with it, I want to look at the facts of what's happening, right, at what's actually happening. By the way, check out the episode we had on Friday, if you haven't seen it — there I really dwelt on the idea that everyone is afraid artificial intelligence will take over humanity. Whereas I keep saying that AI today has changed what's uncontrollable in people's lives in general. I mean, nobody understands at all how this will affect children, adults, politics, countries' economies, different businesses, decision-making and so on, right.
By the way, I tied this to the fact that $37,000 got charged to me, and I still didn't know whether the company had made a mistake or AI had racked it up. That's exactly the topic of modern errors, right, or modern problems that can come up. And there will be more and more things like that in general, right. Well, obviously, once again, obviously.
Sasha, I'd like to pay off my debts to our viewers. People even wrote about it in the comments at some point. Remember, back when I was filming the first episode in China, I said that I'd use, well, try to use ChatGPT as a simultaneous interpreter. And then people asked how come I never told you about the results. So let me quickly tell you now. The results are disgusting. Google Translate is much better than ChatGPT's simultaneous interpretation. Maybe, after tinkering a bit with the prompts or something else, you could probably make it work, but with Google Translate it was much more convenient for me to communicate in China, than with, like, several attempts through ChatGPT, because it sometimes gets mixed up and starts speaking in the wrong language. You tell it in Russian: "Say such-and-such."
It says: "Ni hao," and then continues what you said in Russian. You end up in awkward, silly situations and think: "Oh, screw it!" You open Google Translate, dictate in Russian by voice, press translate and just show it. They do the same thing on their device. And we came to an understanding much faster than when we used this simultaneous voice interpretation. Just out of interest, I was curious whether ChatGPT can speak Tatar. Somehow I'd never thought about it before either. I tried talking to it, but that was pretty much rubbish too.
Some phrases it says pretty well, every now and then it switches to Russian and also makes up some other nonsense. In short, the system is probably good, but I definitely wouldn't consider it as a translator. At least not as a voice one. As a text-
That reminds me of Google Gemini — the day before yesterday my son in the car says to Google: "Do you speak Russian?" It says: "Yes, I speak Russian." He says to it: "A-ta-ta, set our route to go home." And Google Gemini tells him: "Let me check how much these stocks cost." And I… It's a nightmare! It's just, it's just a nightmare. But what I want to say is about Voice. I do still like how Voice in OpenAI talks.
In Dots, by the way, they made it so that in Dots you can't press Voice, you have to call. So yesterday evening we went to get flowers, to the garden, with my younger son. And I pressed call Voice. Dots, call Dots. And it's an actual call going through, it rings through to them. That's, of course, of course, very, very strange. A very strange concept, right? You had a button that turns it on right away. So my little son Damian, Damian is sitting in the back and says: "Dad, who are you calling?"
I say: "Doesn't matter." So a voice picks up, and this voice, the very same one selected in the standard voice model, which is usually that little button on the right for talking by voice, starts saying to me: "What would you like?" And I start talking to it. By the way, here's the thing, look: before, it would talk to you in the chat, and whenever it's in the chat and you talk to it, it saves the conversation history with you. And here I talked to it. And then I asked it to find something and said: "Find it and send it to me in the chat later." First of all, it can't hang up by itself.
I mean, you have to hang up yourself. I tell it: "Hang up." It says: "I can't hang up." I hung up. And it turns out-
No, forget it!
And now listen to the most important part: the whole conversation isn't saved in the chat. I mean, it's not there. And, well, in theory, you could have made some kind of parallel session. Well, that's how poorly thought-out these things are. I mean, Dots is a cool thing. I filmed an episode this week, on Wednesday, check it out. I devoted it specifically to Dots, and I think I got a huge number of people fired up to use this system. I talked about how this is the future generation of how we perceive such systems in general — as an operating system, as, like, a computer, as, as everything rolled into one, right?
And, but for all that, of course, how little they've thought it through is striking. And, on the one hand, their voice is pretty good, I like it, how it works, I like how it talks now. What I don't like is that current models still can't hold a conversation with you and search at the same time — so far they do that very poorly. Although Dots, in theory, is meant for exactly that, right? Because you can actually give Dots different tasks. But what you're describing with voice — they haven't solved that case. This case you're describing, it actually,
I think, should be phenomenally important for OpenAI, because this case lets you see right away the system's ability to think at all. We, we, remember, when we came up with it back then, well, when we talked about it, there was logic to it. I even thought: damn, what a great case! That it doesn't just translate for you, but the system can also quickly prompt you and say: by the way, you know, in terms of his accent or some little quirks, I'd recommend you pay attention to this.
Or that it could have had an ambiguous meaning, or something else. And it turns out the system is being awfully dumb, right? It can't even do simple things. What you're saying right now. And that tells you, Ilnar, that ordinary people work at OpenAI. And it's these ordinary people who make OpenAI — not artificial intelligence, but ordinary people. Or the same AI that you and I have, right? And they don't have access to some secret AI. And these ordinary people think in terms of their own cases, their own constructs, their own, right?
Even now, even now you can't see on the phone which model is selected. I mean, if you're a pro user, it matters to you to see that. If you're not a pro user, okay, people there have no idea how to switch them at all, and that's that. And this shows that the people who build it just don't seem to have a feel for it. They get sent an incredible number of comments, like, bugs, details, everything. And by the way, let me ask: have you ever sent them anything, even once? And why didn't you?
Sometimes I'd mark that I-
Don't like it.
Sometimes I'd put a dislike on an answer. Yeah.
But you didn't send a response. And why not? Because my, my Google Gemini asks after a conversation in the car: did you like the conversation? I write: no. Then it suggests I leave voice feedback. I think: you just didn't understand what I was saying. And what, you want me to leave voice feedback in your computer system at Google? Well, that's just ridiculous. It's like Amazon — Amazon, in this case, their system kind of glitched, and it miscalculated my costs, and it blocked everything for me.
Ilnar, I left fifteen messages, I left messages in their various systems, I filled in all the feedback windows, I wrote to them personally on X. They replied that they'd passed the message on. I have no answer. And so, come on, what feedback? I mean, if you and I could see that — imagine, you leave feedback, your feedback gets assigned a number, and they have some cool system running — by the way, that's how I'm implementing it in my own projects.
I've got my own projects that deal with, for example, the infrastructure of YouTube channels. And on our site there's a block at the bottom called "Requests." And when a person leaves these requests, my logic is this: these requests should be analysed by artificial intelligence. For example, Codex Astra or Fable, like, Astra Max or Fable Max, for example, right? Or Opus 5.5. Well, Fable may not be needed there. Opus 5.5 Max. Well, it should check, check the attached files for viruses, blah blah, blah blah blah.
Then analyse them, see right away where there are real bugs, fix those bugs itself, notify people via the Telegram bot and on the portal that their bug has been found and fixed. Where there are bugs, where there are some of their ideas, suggestions, bring them to me in one batch somehow. At some point I should tell it: okay, this is a good one, go solve it. And then it should do that too, and notify people. I mean, in theory, the system should work in such a way that it fixes the problem itself, if there is one.
I think this is a very instructive story. It's the ability of support services to respond in such a way that support doesn't just talk to you or dig up some information for you, which today is actually an elementary, dumb thing, well, for serious companies. I'm saying, Ilnar, for serious companies this should be a dumb, basic thing. Rather, for the system to resolve the request by itself. By the way, I'd wanted to build that kind of request resolution for a long time. You know where? In one of my projects for sales managers, I also set up a certain system there, and for myself too.
As a kind of infrastructural study of everything going on. Yeah, yeah, yeah, yeah, yeah, yeah, yeah. And the system works there for the sales managers, it calculates all kinds of data. I filmed a separate episode about it, a big episode. There's a crazy amount of stuff running there. What was my logic? That if a sales manager sees a bug or something is missing, he can report it to the system, and the system then sorts it out by itself. When I came up with it, it was, I think, around July, and it was probably even a bit, a bit too early for these systems. You know how it was?
Because back then I had the feeling that some kind of visual interface was still needed for it. I mean, I wanted to build a visual interface. Now I'm building more and more systems where I don't need a visual interface. I mean, I realise I don't need to see the admin windows that AI builds for me. I mean, I try to work directly, so that, say, Dots tells me about it. So at some point Dots says to me: Alexander, when can I talk to you? Fifteen requests have come in from clients, there's something they don't like. Now that's a great thing. Listen, I was sitting and talking with our mutual friend, and my close friend and partner, Danila.
And he says to me: you've got this one registration there, and you can't enter the code in Telegram. At first I wanted to tell him — I got angry with him, I thought, why is he telling me stuff like this? Then he writes to me: am I being a nitpicker? I tell him: for sure, you're definitely a nitpicker. And it's for nitpickers that I came up with the Requests block at the bottom. And then I'm sitting there thinking that, since such nitpickers do exist around me, and they can tell me about it.
I think: I'll implement a feature so that if someone happens to tell me about some bug verbally, I'm able to say, that this person reported a bug to me, and you'll open a ticket for him, the system will tell him thank you very much for the request, and the ticket will go to him. And imagine, I implemented this feature. And now, if someone tells me about a bug like that, I tell the system. A whole message went out to him: Danila, thank you very much for your suggestion. We'll definitely look into it and resolve it. And the system really will look into it and resolve it.
And that's service, and we see this kind of service when, I don't know, I'm dealing with, I don't know, Verizon, and Verizon doesn't tell me: "Open the website and create a ticket." It says: "I'll create everything for you," right? Or at Amazon: "We'll do everything for you." They do it for you themselves. And then you get these tickets, requests. They don't tell you: "Go log into the system, call another number." You say: "But I've just been talking to you for forty minutes." That's the frontier of new scenarios.
Because look, Ilnar, scenarios like that didn't exist before. You'll agree, before it would have been impossible for a system to automatically sort through your tickets, fix them and push out the code. Well, like, it, it, well, it was impossible. Was it possible to do this five years ago? How? It, it was impossible to implement, right? I mean, you could write a solution for some tickets and things, but you couldn't foresee a huge number of them, especially when it's a bug. I mean, if it's a bug, how would this system solve it, especially a layout bug, or not being able to fill in fields, or something else.
It's of course very interesting and instructive, very interesting, instructive. I really like cases like this, and I think they develop this new, new perception of the whole infrastructure in general. I mean, a new perception of AI. And, by the way, one of my latest disappointments — we've talked a lot here about my robot vacuum. So some people turned up who said that a person vacuums better than a robot. My wife laughed, and we went in and looked at the stats together. My robot, specifically my previous vacuum model, Ilnar, has covered forty square kilometres, and I lied when I said it vacuums three hours a day.
It vacuums only an hour a day. No, probably an hour twenty, if you divide by the time since I bought this vacuum, over several years — look, over several years! That's very important, right? And if you take 700 days, then over 700 days it's an hour and a half or two hours, right? And who was it, who was it saying that a person vacuums better? And forty square metres every day.
Each individual time, maybe, but at scale-
Ilnar, not even each individual time. Because if your system is set up to go over everything again and do it perfectly, well, you'd come with the very same vacuum anyway. Obviously you'll wipe the dust off a shelf better or give a painting a wipe. No question. But I was talking about specific things, I was talking specifically about a robot vacuum. And so, after those episodes, I decided — the next day I went in and upgraded, I bought the very latest version of the robot vacuum, because the previous version annoyed me, it was bad at building 3D maps.
Roborock, by the way, is, I think, the world leader in robot vacuums and specifically in map building, interesting systems and everything else, Ilnar. But still, I understand that it's a corporation. I started building maps and thought, what idiots they are. I mean, I've got so used to the fact that with AI I can do a whole bunch of cool things. And when you start seeing these limited solutions for twenty-square-metre flats, I get it — they probably make the robot vacuum for twenty-square-metre flats, and a huge number of things just aren't provided for, simply aren't provided for.
For example, that you might have five spaces with no doors. Well, it's America. Or you might have different transitions in these spaces, or you might have lots of windows, right? And you can't mark them. And they've got furniture, in the furniture they've got a wall for the cat, but for some reason there's no, like, piano in the furniture. Well, and so on, well, in short, a bunch of dumb things like that. It recognised me very poorly, roughly like a model from three years ago. And I thought, even though this is a vacuum with new lidars.
Well, it's just insane, right? It shines its light around. Of course it does things way better, it works way, way better. It washes those mops itself at 100 °C, changes the water itself, adds the cleaning liquids itself and a bunch of other stuff, but there's just a total mess of glitches. And right away I thought to myself: could I upload a piece of code into it? Can you imagine, the thoughts I'm already having. I'm already, I'm already just, just a total geek, right? My friends and partners, when this kind of thing happens with me, they listen and say: "Sasha, you're crazy, just, just crazy." And, well, the ones who aren't like that, not that into IT, right?
And it's just that my life is already pretty broad, and I'm talking about uploading code there so that this robot could do some other, more serious things and run my scenarios, my use cases. I mean, they add smart voice conversation, they add this thing: find the pet in the house using the camera, there's a camera in it. It can, like, drive off and find someone somewhere for you. And, but still, the software interface itself is very bad. And this is the world — what this blogger posted, sorry, I don't remember his name, I really liked his conclusions.
What the American blogger posted. It's Greg Isenberg. I don't know how I stumbled on him. And I want to say that this story about decompiling code worth $5 billion. And, by the way, did you see Nikita Bier — he's the former xAI leader, an acquaintance of a friend of mine, who actually met Elon Musk through him. Here, right here, just minutes away. And he posted this story, that we live in a fun time, when it's now been announced that $2 billion was spent developing the new GTA.
Well, something like that. And he says: "And right now we're exactly at the point when AI can write a game of that level." I really like it when people of that class, developers, and it's still a high class of developers, right? Well, the ones who created xAI, guys, that's, that's, that's a serious class of people. They make statements like that. I mean, it's a whole other world. The fact that we can now make such systems, develop such systems, and I hope that with this everything will get better and better, and better, and better.
So. But there's this security problem, of course, right? Because someone just posted about exactly this — I talked about it on Friday's episode, check it out, the one with my $37,000 — that, these models, imagine, if I can upload some small piece of code to the Roborock, that code could plant a virus in there. And basically, this is the problem of all these systems getting infected in general. And I actually read an article on this recently. I forgot, damn, listen, I forgot the name of this type again.
There's this type of virus that will spread like that among home appliances. I mean, when it has left a micro system there and, basically, someone can later get into your place through that micro loophole, right? Although I think that now-
Fridge, TV.
It's, it's insane! Yes, it's insane. How you can scan everything. I'm constantly unhappy with Starlink, you know that, right? I have two Starlinks installed, and the way it works just drives me nuts. And I keep scanning my home network with Codex. And the way it scans systems, finds all the devices, goes through everything and so on, I realise that breaking into my place… Well, I mean, today it's not even clear which systems you can connect to at all, where you can connect, how you can do it, right?
I mean, what did you download, what did you install, which app did you download? This app is supposedly from a trusted developer, but whether it's been checked, whether it's trusted or not. Take Apple: yesterday a friend of mine, who's very much in IT businesses, in mobile apps, and we talked about this yesterday, that Apple now has a huge problem, that there's now an incredible number of developers who make mobile apps and release them, and these apps are just trashy. And there's a huge number of problems, a huge number of bugs, a lot of scams, and it's unclear how to check them, how to track them.
Apple is shutting down both normal and not-so-normal apps, just by the ton. So, imagine, you'll download an app, and it'll seem kind of cool. The developer didn't even mean to cause you any trouble, but this app may have a hole, a loophole, and you, it turns out, and none of these systems bear any responsibility. You know, right, that Amazon suspended my servers. Their contracts say that they bear no responsibility if something happens, even through their own fault. And you know what else I loved, what Claude Code told me? I said: "And could this, the server, not stop and keep running?"
It says: "Well, no, it's supposed to — I launched it, it ran, ten minutes, and it's supposed to stop and shut down by itself. That's how the script works. But! If the server hangs, it may keep running." And you're like: "And who's responsible for it hanging?" And so
Recently I was working on a task that takes, computationally, on a server, well, like five or more days.
I mean, you need to run computations on GPUs, pull data from the database, then export whatever comes out. And it's a very large dataset, and it takes on the order of five days of computation on a fairly big server. And this process, well, it can't go smoothly, and you still have to go in periodically and check whether it's crashed. Because it can crash for a thousand different reasons. Maybe, I don't know, your colleague on that server grabbed the GPU while your process had freed it up, and now you have nowhere to run it.
Maybe the database glitched and returned an error to you. And you can't foresee all of this in advance. The process is long. If, for example, you left it in the evening, I don't know, checked it before bed and you think: okay, now in eight hours there'll be good progress. You wake up in the morning, and it crashed at, I don't know, 11:50 pm and wasn't running all night. You think: damn, my deadlines are pressing. What's going on? For some reason it crashed. You start restarting it, looking at what the errors were and so on. And so here's an interesting thing.
It has most likely been around for a long time, but this was the first time I actually used it. It's when you ask ChatGPT, well, Codex, to become a kind of nagging poller, like, every, I don't know, fifteen minutes, every thirty minutes, to check the status of what's going on. And, accordingly, if such-and-such events happened, then to react to them, right, restart some things, do something else. Well, I mean, there's a set of actions you allow it to take. And it does that throughout whatever period you need.
That's a bit like what you're telling us about Dots, right? But this has already been built into Codex too. It's a really cool thing, at least for these very long-running things, when you need to make sure the process is working. Well, like the example you just gave with the server, right, say, if it hung and kept running, and, say, Codex found out about it, it would clearly be competent enough to shut that server down, if you asked it to. And it checks that, makes sure, and terminates it. But what's another problem? This thing runs inside Codex.
If your Codex is running, like, on a computer, on a laptop, and you, I don't know, closed the laptop lid, it went into sleep mode, the process stopped, then this process won't run either. But it turned out this can be worked around very easily. You spin up, I don't know, Claude or Codex on a server and connect it to your own system. I mean, inside your Codex you get, sort of, a version that runs on the server, and you only get, like, the results, some communication and so on, but you close the laptop and it all keeps running. So it turns out you can have this kind of Codex that's running all the time.
Only, like, yeah, you can access it from your laptop, from another computer or from something else. It's sort of a cloud version, but not a cloud version inside OpenAI's infrastructure, rather on your work server, for example, or on your personal server, or something like that. And it turned out to be insanely convenient. I think this is a bit of a parallel with Dots. Yeah, in Dots, what you're describing here, they apparently do similar things, but this can be done without leaving Codex. And I really, really liked it.
So give it a try. It's not very hard to set up, to configure. You can do it with Codex holding your hand. And if you program a bit, it's really easy. There's documentation on the OpenAI website about how to set up these connections to a server over SSH in Codex. And if you have long-running tasks like that which you'd like to keep an eye on, I recommend trying it. A very interesting thing, in my view. What else can I add here?
A friend of mine uses Claude more than Codex. Claude has five-hour windows. Well, I mean, five hours when you have a limit, then it resets. If you take a normal working day, you wake up in the morning, start working, your window starts, and then it's like five hours and then another five hours, and you only get two windows in a working day. What did he do? He built a system that periodically sends a message to Claude, well, an absolutely minimal one, like: "Hi, are you there?" Or something like that.
And that way, at night, this five-hour context window resets. And when he wakes up, he already has a window that has started. And then he has, I don't know, two or three hours in which he uses up the entire five-hour limit, in that stretch. Then he has another five-hour window, which he uses, and then the tail end of the next one. That way he, sort of, uses the limit of three windows within one working day. And that's also pretty easy to do, when you restart the system with these pollers so that your limits start over.
Like, like someone from OpenAI suggested and asked about the Pro version. He was asking: "Would you like us to double your tokens, or-
Yes, but keep the five-hour window. I'd have picked double the tokens, of course. Well, I don't know. That's what I'd pick. And people voted somehow. He says: "I don't understand your answer. I thought double the tokens would be better. Why did you, like, vote evenly?" And I never did figure out how many people voted there, right. And maybe people on X, you know, click the button for fun. I think, you know, someone just clicks the button for fun and presses it without any point, right, does it for no reason.
And because, of course, I generally think these different kinds of rules aren't a problem. Because if you really use the system all the time, then you'll, well, you'll adapt to one-hour windows, and to three-hour windows, and to five-hour ones. Well, a five-hour window is a normal window. Right? You manage fine with it. It's probably even better than, like, an eight- or ten-hour window, because otherwise your working day could, like, well, you could accidentally lose your whole day.
Plus they have these extra resets. Plus you can buy a few more tokens, do something somewhere. In theory. And people who don't use the system often, well, their tokens won't run out. What difference does it make to them what limits they have, whether there are, there are limits, basically, or there aren't any. Well, by the way, me and the Ultra Fast mode. Remember, my thing was that I was all for the Ultra Fast mode, and I really liked it, and because of it I switched to Pro 500.
By the way, I'll point out, Ilnar, that it was Pro 500 that let me spend so many tokens, billions over these days, because I clearly have some volume, well, some decent volume. They state one amount, but we don't know what it really is, right? And I want to say that I used Ultra Fast only once back then, and then once more when I needed to speed things up. And this thing is cool, and you have to use it very selectively. And of course not forget about it. I mean, when you want, when you've got something super urgent, right?
And you're, I don't know, stuck on the line with support, or your system has died and you need to bring it back up. I mean, then you can, like, say to yourself: "Okay, then I'll buy some more, buy some more tokens from you and use some extra mode." By the way, this Ultra Fast mode, you'll be able to use it. I think it makes development eight times faster, if I'm not mistaken.
Eight times faster for six times the price, the cost.
Yes, it's, it's out. It used to be only in 6 Astra, now it's appeared in 6.1 Sol too. So. Overall, overall it's an interesting thing, but again, once more, it's interesting either when you have a $1 million budget for tokens, or when you work at OpenAI, or when you get some endless token resets from them. And on that day, for some reason, you decided to do it. Well, like how I sometimes have fun here, knowing that these days I'm just working at it intensively, and that's why I'm burning these tokens, because I know I might drop out, right, and be busy with some other tasks.
And you can't always come up with something. I, you, I don't know, does that happen to you, when you know the limit will reset, and you see that your limit will reset in the morning and you still have tokens left. You launch some silly task like that and say: "Double-check whether it's there or not." Doesn't that happen to you? For me, since I have a whole lot of-
You set Ultra, you set Fast.
Projects, tasks, so I launch them. I launch them then. Let's go there.
A parallel reserve.
Yes, yes, yes. It'll be interesting to hear what our viewers say about this, who has what, what experience too, experience with all the cases we've told you about today. Well, and see you on our Sunday podcast with Ilnar in about a week. And watch all the other episodes. We now have an episode coming out every day. So. Great to see everyone. Bye, everyone.