Because in all fifteen years you've sent me a GIF for the first time. But I just needed to test it for some reason. I couldn't do it just for nothing. It was very funny.
So how was testing GIFs?
Listen, it does it great, technically speaking. But creatively, I think, it's a total disaster and everything is bad. So. That's why if you write, say, a completely crazy, perfect prompt that you have to create and come up with yourself, the result is impressive. But if you just ask it to create something from some simple sentence or simple request, I don't know, like, my wife spends too much money or I'm tired at work — I mean, from some very short comment like that, then everything is very primitive. That's how I'd put it.
Listen, I can hardly imagine what kind of GIF the system comes up with for 'my wife spends a lot of money'.
It came up with one; I had exactly that request. My husband spends a lot of money.
And you sent him a GIF like that?
Yes, yes, yes, yes. He-
He thought you'd lost your mind.
No, my husband immediately wrote to me that-
Probably it's your assistants.
Yes, ChatGPT keeps endlessly posting that it has new updates for various kinds of images. So. I don't know who even follows this, but someone does. Anya here was following the news that they can now make GIFs, yes. And users of many, many platforms where this has long been possible will laugh now. But still, yes, ChatGPT is developing in this direction. And the question is when they'll, when they, well, and why they'll allocate big capacity to, say, video generation or things like that.
Yes, for generating images, GIFs and all that, it's better not to give just a prompt, like, say, my wife spends a lot of time at work, but to give some references, or make it lay those references out as a prompt that will produce something similar. Say you like the style of some GIF, image or something else. We drop it into ChatGPT and say: I want, like, or describe, like, how such an image is built. And it writes a prompt on exposure, on, like, colour grading, on some other things, on style.
Then we drop in that we want, with the same prompt, and add some meaning of our own, and get a similar image. Like that's how people make all those Hayao Miyazaki-style GIFs and so on. In general, it helps quite well to get close to the reference. First we make it describe what it sees, and then actually do the same.
Yes, well, that's definitely very useful. But I was also curious to test the speed, because they claimed it's really very fast. I've just never used it, so I don't know what fifty percent faster means, compared to what. If I were a consumer of this service, I could compare, but it really does generate very fast.
Well, hi everyone! You're on the ToTheMoon channel — tech news and insights from Silicon Valley and around the world. And still, for people the big insight is a GIF or an image, right? Because we send them to each other. So the topic is actually interesting. Following even the simple things Anthropic or OpenAI do. I think I, probably, I shot a Short on this topic. Now, at the start, we'll bring it up. The markets are falling very hard, right? This morning there was just some cosmic drop in the tech market. Anya, you probably don't even watch it, do you?
No, I didn't look.
I'll let you know. A drop, a crazy market drop. And today I've already checked like fifteen times, which isn't like me, well, because the money… how do you say it? The number is dissolving — not the money, the number is dissolving. So. And we were just sitting this morning, just now.
Oh, yes.
Drinking. Nothing good, right?
Yes.
Then you won't be able to host the episode. So.
No, well, listen, it's reality.
And the change of power in the US, a relatively small change of power. Actually nothing has changed, as they say, everything stayed as it was. But the reaction from the markets is clear right now, of course. We'll see what happens. I think nothing will happen, and everything will turn around and start moving the other way. And which topic did I want to start with? Right now overall it's like that, of course. We'll soon move on to discussing the Anthropic IPO. I think it's an important story. Too bad Max isn't here today.
Last time he told me: 'We didn't get to discuss the Anthropic IPO'. So, I also want to say. Grok released, I don't know if you saw it or not. Grok released a new feature. A very cool one. It's a kind of caller. So you can give it a task, the Grok bot, for example, call the insurance companies, these five insurance companies, and agree on what terms I'll get for, like, car insurance. Or, for example, call the internet provider and find out why I had some problem yesterday. When you need to hang on the phone for an hour, I don't know.
And I saw it on the very first day — it came out just a couple of days ago — and I see a huge number of reposts on X about it, people saying: what a cool thing! And I think, well, the idea is great, and at the same time very trivial, right? It seems simple to all of us. So, neither Anthropic nor OpenAI has this. I mean, they don't have a way to set up a scenario so that your bot follows a script, comes back to you during the script if some extra question comes up during the call, and checks it with you.
And separately, this thing has existed for a long time, but surprisingly Gemini has it. And Gemini, because I somehow missed it. I don't know, did you see it or not? I missed it.
But it says Gemini is testing it in closed mode, I mean, among some limited group, and there are very few people, and maybe that's why it hasn't launched. And Grok, given that it's on X, on X, of course, there's Elon Musk. Elon Musk keeps reposting, retweeting, yes, posts about Grok Bot all the time, and the case is very interesting. I'm waiting — our viewers, write in if you've tested this feature or done something similar, or tested it in Gemini. I'm very curious how well the system actually works.
And in general, if we discuss this topic: what's your view, how will this system develop further? Because if a bot calls a restaurant, how should the restaurant treat the bot? A good restaurant wants to know who you are, what you are, how you, so to speak, look and so on, right? And here a bot calls. How does it call? It says: 'I am Grok Bot, calling on behalf of Alexander Volchek, I'd like to book-
Sasha, have you tried it?
No, no, I haven't, I'm telling you I haven't tried it. Yes, I've been travelling a lot these days, and even the two episodes this week, on Wednesday and Friday, the ones we released, we recorded from the car, though the episodes are great. And if you haven't watched it, watch the one about splitting and merging ChatGPT and Work. I think that's a very important topic, fundamentally important, yes. And the second episode is about building your own financial accounting system, personal, and, well, or for a company, or if you're a sole proprietor — watch that finance episode from Friday.
I'm just curious; I think in Europe this should be banned. Or I don't quite understand how it's supposed to be implemented, since there's a law under which anyone who hears any conversation or voice of artificial intelligence must be notified that it's artificial intelligence.
Well, probably right at the start it'll say: 'Hello, you're getting a call from...'. That's what I'm telling you: Grok Bot on behalf of Alexander Volchek. Just look at an interesting angle. Imagine a bot calls to make a deal, and you, you're an insurance sales manager, right? And even more likely you're some call centre, some eighteenth call centre. And this eighteenth call centre picks up this Grok Bot. They'll probably have scripts: if you hear a bot, do this or don't. In principle, you don't want to talk to a bot.
You think this bot will also call forty companies. Why should I talk to it? On the other hand, if you adjust properly to the bot and understand it's a machine, you'll be able to pitch certain details into it. Because we all know how ChatGPT now and then pushes some idea or picks some restaurant, or hotel, or something else.
Uh-huh.
Based on some features that make you sit and think: who even needs these features? Well, probably only you and ChatGPT. So this topic is very interesting. How people will react and how companies will set things up. When we started filming ToTheMoon, one of our first news items was that ChatGPT got banned in Italy. Remember, Ilnar?
Yes.
I mean, Italy, how-
It was there just yesterday, yes.
I even remember that episode.
Yes. I mean, how did Italy react? It reacted: we shouldn't have this. Banned! So clearly there will be some companies whose scripts in their companies… And we have many viewers with companies too. Write in: what will you do if you get a call? What script will you set up for the case when a bot calls, right? And you've determined that it's a bot calling. On the other hand, Anya, I don't fully understand whether the bot will announce that it's a bot, because it probably depends on the country, as you rightly said, or on the person's wishes.
Because right now, look, right now a bot calls because you're using an external system. We'll all get such an option with our own, well, say, digital copy, right? But it will definitely appear that you'll sit and say: 'On my behalf, like, call and pass on this information', right? And the system will call. That's obvious, especially with the current capabilities of voice synthesis and voice interfaces.
In the company's place, if I knew a bot was calling, well, I'd just transfer it to a bot. Let them talk to each other as much as they like.
I wanted to say that I quickly checked. And indeed, it doesn't work in Europe, and in the US it doesn't work in Washington and Illinois.
Fine, and did you check what happens when it calls in terms of, well, you find out, Il-
The feature started working on the fourth of November and accordingly-
No, no. But how, what does it do? Does it say who's calling? I think not.
No.
Probably not.
Precisely for this reason, precisely for this reason it isn't allowed into the European Union, because a person has to be told they're talking to artificial intelligence.
Yes, but it can be configured.
Well, theoretically, yes.
Yes. But look, essentially Ilnar just said an interesting thing. An interesting thought. Let's break it down. If artificial intelligence isn't working in your company or wherever, and artificial intelligence called you, then in principle the reaction should be normal. Well, a person called to book a restaurant. What difference does it make that a bot is calling for them? By the way, Google has long been calling restaurants and, well, booking a restaurant for you if it doesn't have a direct connector, right?
I mean, in principle, it's a lead coming to you. And you go: 'No, I won't talk to bots. Beep-beep-bee'. Well, in principle, you're supposed to sell. On the other hand, Ilnar rightly says you should transfer it to a bot. And here, Ilnar, is an interesting thing. Now when I call as a human, who do they transfer me to, a human or a bot, right?
Yes, you need a classifier at the entrance. If it determines that, well, first, with high probability you'll be talking to a bot in any case, most likely, over time.
Yes.
Otherwise, if you have a classifier that understands, I don't know, like, by intonation, by, like, some other technical signals at the entrance, then you simply decide to send it to a specially trained Vasya who knows how to con ChatGPT, or, I don't know, to the next, more advanced version of ChatGPT so that it upsells it.
About the word 'con'. Right now it really is the era of the right scripts, yes? I mean, if a bot calls you, some companies will now be able to make a lot of money on these bots. Obviously, right? Whoever understands that there's some bot, and they'll manipulate through that bot. Well, accordingly, a market for manipulation will appear. And there'll be the flip side, right? That bots are calling now, adapting, saying they're people. You say it's clear from the intonation. Listen, yesterday on the road we used ChatGPT Voice now and then, and my Voice — Dima, riding along, says: 'What did you do to it? Why does it talk like that?' Mine talks like this, it says: 'Mmm, listen, so'. I mean, it's… I think I've so wrecked it in every respect that it, I don't know how it corrects itself, and this thing of its, this 'Le-e-et me check', it, it even sounds kind of, well, kind of cool.
Yes, like my friend back then, like Misha said back then in Europe: 'Who was that talking? What kind of system is that? I want one too'. And his voice talks completely differently. Well, yes, it's the settings. By the way, check your voice settings — they update them, and you can choose between different systems. And of course, well, Ilnar's voice talks pretty cool. My point is just that it's not that easy to tell from these incoming signals.
So it turns out it was Ilnar, Ilnar or Sasha, you told us in the past, I think, or maybe it was on some podcast of ours, I don't remember anymore, about how a bot talks to a bot and how they can't say goodbye.
Yes, I told that one.
So it was you who told it.
Yes, yes, yes, I told it. And they can't say goodbye.
There may be endlessness at this point, and on the other hand, you know, it's a little scary.
It's cool, because each side can squeeze the other to the maximum in terms of terms, negotiations and other arguments. I keep remembering that film 'Mercy', where the judge was artificial, or rather was artificial intelligence, and the detective had 90 minutes to prove to the artificial intelligence that he was innocent. Well, in general, I must admit, it's not a great feeling when you realise that in front of you is a mind that, without emotion, evaluates nothing but bare arguments, and, well, it's a completely different level of negotiation altogether.
So discounts from insurance companies, I think, can also be obtained quite substantially.
And they evaluate in their own way. What's the problem with evaluation now? Imagine there's a criterion, say a person's height, and it's very easy to compare, like one ninety, one ninety-five, one eighty, one seventy, you can compare. But if you're told, say, that the service is not bad, and each of them has something written on their site about what kind of service they have. And ChatGPT today compares them like this: it writes you 'service' and then takes the summary of each service.
And you look at these four companies and go: 'So, so in the end what's the difference between them?' That's if you run a comparison of Dots from OpenAI, Duet from Claude, Grok Bot — run Grok Bot in the comparison — and Muse from Meta, and write, some of you will want to compare parameters. For example, whether it keeps the context window or whether it has memory. And on this 'whether it has memory', it'll just give you some, well, some statements that you can't tell how to compare with each other.
I mean, they have no parametrisation at all. And so, of course, Anya, what you're describing, when artificial intelligence, as if you're facing such a mind — how does this mind then do the comparison? I mean, based on what parameters does it make this comparison, how much is really built into it, how much does it understand? The topic is certainly very interesting. We'll see what happens with all this.
Sasha, when you started talking about Grok's caller, you unlocked childhood memories for me again. I read a book, science fiction by Sergei Lukyanenko. The story is called 'My Dad Is an Antibiotic'. Well, it's about the future. And the point is that exactly these callers have appeared. There's the option, like, I don't know, you can't get through to a person at the moment, you set up auto-dial, and it keeps going until you get through. And there two friends agreed that they would never use auto-callers on each other and would always call themselves, because, well, it's kind of unethical, not nice to call each other that way.
And I have the feeling that with the arrival of Grok, Grok Bot and the caller inside Grok Bot, these agreements from a children's book, I think at some point they will have to be put into practice, because, well, I wouldn't want some Grok Bot calling me on someone's behalf. Well, even if it's, I don't know, to clarify some details, whether I'll come to a meeting or something like that.
Ilnar, and how do you feel if a bot writes you a letter?
Look, it all depends on the context. I mean, if it's communication with a person with whom we generally have, well, just everyday communication, I wouldn't want that. I mean, it turns friendly everyday communication into something soulless. Well, better not write then. If it's purely work, then again, it depends on whether we have such an agreement. I mean, if there isn't one, then negative. If we've agreed that yes, it's fine, and in which circumstances we use it, then yes. I mean, again, if it's regulated and agreed, then yes.
If it's the Wild West, then better not write.
That resonates a lot; it seems close to the calls thing.
Listen, you're saying something interesting right now, in particular that if they call you, it's a definite no. Me, for example, personally — interesting, we'll poll our viewers too — and support our channel, like it. I always pay attention to that, yes? And subscribe to our channel. And the comments — it's interesting to find out what everyone thinks about this. I, for example, don't mind if someone's Grok Bot calls me about questions that reasonably fit the calling system of this Grok Bot.
Well, for example, imagine my wife is driving, and suddenly she remembers she needs to turn off the iron, and she can't call herself. She says, for example, by voice: 'Dot, call Sasha and tell him to turn off the iron urgently'. And this Dot starts trying to get through everywhere, it starts calling every Amazon Alexa, it starts calling me, starts calling, I don't know, Iolanta next and so on. Or, for example, or, for example, well, I mean there are cases that are quite reasonable.
Let's take away the iron, let's even take away the emergency, well, some non-, non-, non-urgent circumstance, yes, where there's some damage somewhere, where some person, say you, Ilnar, will write, like, your Grok Bot will call and say, like: 'Sasha, I'll be ten minutes late for the episode today'. And it will call me by voice, right? Because it knows that I, for example, read messengers later. I don't get notifications, and I, for example, need to be called. An absolutely normal story, because at that moment you may be somewhere where, well, it's really hard for you to call.
And if you're a person who approaches it with respect — if your bot starts calling me fifteen times a day and asking: 'Sasha, how are you? So, tell me'. I say: 'You, Ilnar wanted to find out from you. I'll say: 'Tell Ilnar he's a fool'. Things like that I'll cancel, right? But if the story is reasonable — even I talk to my mum, like, I don't know, once a week, right? With my wife, like, I don't know, every three days, or, like, my sister, depending on other circumstances, once a week.
But, for example, if my mum's bot calls me now and says: 'Sasha, listen, please don't forget to, like, prepare my trip to America, this, this, this thing', I'll react calmly. I mean, I think it's a question of reasonable perception. And in the world one would rather not act like Italy. But maybe that's what you meant, Ilnar, no?
What's more viable, Sasha, I think, is that, say, about the iron or something else, a message was left for you, and your Dot on your side sorts it out itself and decides how to get it to you and what needs to be passed on. And in that sense the hypothetical Grok Bot calls, like, your Dot, leaves you a message or something else. And then, according to rules you've set up, Dot sorts it out and gets it to you, by a call or in any other way. Then control will be on your side.
Ilnar, the question here is this: if I put my Dot on incoming calls on my phone and don't pick up for anyone, and when someone calls at first, my Dot will filter, and then you'll be, so to speak, hanging on the line with my Dot. It says: 'Who's calling?' You say: 'Ilnar'. It tells you: 'Ah, I'll connect you now'. It says: 'Who's calling?' Ilnar's Grok Bot Vasya. Then off it goes. Right? That's the situation here. I understand that, look, I understand what you mean. It's a question of a huge number of different settings. Can you come to the door and knock, can you send a letter, can you send a message.
When I was telling this — in my family, for example, there was a very clearly defined rule. I don't read messages during the day in WhatsApp or Telegram; I may not read them. But if you need to reach me quickly, urgently, medium urgency, you can send a text. I don't know, like, my mum or my wife. If you can't, well, I don't answer, and it's urgent, you can call my assistant's phone or my driver's phone. They may pass me the phone. If not, call my phone. That's the highest urgency. But you can't explain that to the whole world, right? And you, well, there's a certain number of people who know about it.
And probably we'll arrive, Ilnar, at what you're describing now, we'll probably arrive at a system where all these will interact with each other, well, our own artificial intelligences, our own models will interact. According to Sam Altman, he wants to give everyone an incredible amount of their own compute, right? So that we'd have our own capacity there, our own servers at home. I think he's saying this ahead of the Anthropic IPO; we'll talk about that now. And if it comes to that, why not? We'll train artificial intelligence that way and learn to interact somehow better. But will we learn to interact better? That's a big question, right?
Because we often can't interact with each other by ordinary messages. Some people, to send a message, send forty messages one word each, right? Someone sends voice messages and the other person doesn't like them, for example; someone, on the contrary, likes voice messages. I mean, it's an unreal detail. Will you be able to adjust someone else's bot to yourself? Well, imagine, imagine you make a bot and say: 'Wait a minute, Volchek's bot, I'll tell you my rules for how to communicate with me'.
And it'll say: 'Volchek has prohibited other people's rules'. Or it'll say: 'Fine, go ahead'. Or Ilnar will say: 'You may call me only at the appointed time. You must call me by my patronymic, address me in such-and-such a way, come with a gift'.
But still, continuing this too, Ilnar, human contact remains. You know, I, for example, analysing even my own behaviour, always hang up if I realise it's not a human talking to me; I don't even wait for what they say next. And I receive a very large number of all kinds of offers, like, from completely different fields on LinkedIn. And I never reply to those contractors, like, providers, specialists of whatever kind, if I see it wasn't written by a human. But if there's a certain human, like, I don't know, some kind of contact, I'll always reply, especially if it concerns, for example, recruitment.
Just understanding what kind of work that is, I'll always reply, but I'll never reply — I always skip everything a bot sent. And each time I think: if I react this way, maybe all people on the market react this way. Well, let's suppose. And so the question of how effective it is to use bots and switch to them completely is a big question.
Anya, well, listen, what if a bot calls you. Imagine you initiated a return of some item, for example, and a bot called you. And a bot calls you: 'Hello, Anna, I need to ask a couple of questions'.
Very much a scenario, yes. It's as if, so-
What, what will you say? 'I won't talk to you', right?
No, well, it's as if… depending on what. That's the thing. If it's-
Or, for example, a bot calls you and says: 'Anna, we'd like to inform you that you had an appointment, like, at the dentist six months ago and you missed it. Will you book again?' You say: 'Yes, I will'. Or you'll say: 'No, let a human call me'. Well, sort of-
Well, then it turns into a kind of categorisation of scenarios. If it's technical support, help, a reminder — that seems acceptable. But if it's scenarios where you need that human contact, partnership, choosing one out of, like, twenty, selling some service, then you seem not to want to waste your time on it at all.
Listen, well, look, you, you've been in employment, like, twenty years, right?
Yes.
Let me tell you a simple thing. Probably, if someone is being hired as CEO or, I don't know, chief commercial officer, and a bot calls and says: 'Hello, I'm calling you from such-and-such a company, and now we'll do an interview'. It seems, look, it seems kind of wrong. If you're applying as front-line staff, mass hiring, I don't know, as a courier. If a robot calls you and says: 'Hello, you applied for the courier position. You'll now have an interview', you hardly need to say: 'No, I won't do an interview with you'.
But now let's go back to the CEO. What if it's a real, normal, sensible survey that lets them gather certain information from you genuinely very fast? And they called you, you said: 'Please call me back in forty-seven minutes'. It'll tell you: 'No problem, I can call you back in forty-seven minutes and four seconds'. So it calls you back exactly at the appointed time. You pick up and say: 'What did you want to ask?' It says: 'Now I'll ask you some questions. When you get tired of answering, stop answering and talking. Whenever you like, you can tell me extra things, you can tell me more, you can tell me less. Whatever you want!
I'll talk to you and do it all'. This information will then be processed and passed on internally not only to the recruiter but also to the senior people who are hiring you. Look, it seems to sound bad, but let me tell you...
It doesn't sound bad, Sasha. It sounds very customised. You've customised a certain scenario a lot, and you've customised this scenario so broadly and non-trivially. In reality-
Well, I'd like it to be!
Yes, but in reality it's simpler. You, you know, this is such an ideal picture of how it could be.
Yes, agreed.
And in this picture it sounds really cool.
And in this picture you react calmly, notice.
Yes, and in this picture you really want to talk to this bot.
Well, I'd be glad. Look, a bot from Elon Musk calls you and says: 'We've decided. Elon Musk is thinking of meeting some people and for some reason decided to meet you. You need to talk it through'. What, will I say: 'I won't talk to it, let him call me personally'? Well, there. That's all there is to it, right? Whoever won't talk to Elon Musk's bot, please write in the comments, okay?
Sasha, well, take the example of hiring a CEO. Well, the funnel, how many is there? Well, ten people per position are looked at, at most. It's easier to call an HR person and talk to a human than to set up a system.
I hear you, I hear you, but not always. The funnel actually isn't ten people, Ilnar. Look, the funnel is hundreds of CVs, and from a CV you can't tell how to hire a chief commercial officer or a CEO. They'll let you know. It's practically impossible, right? Because you're hiring a person with long experience, right? Even mid-level managers are very hard to hire. And this thing lets you gather far more and make not an abstract filter based on the two-page CV people usually have, right?
Because someone decided to write it up nicely, or someone said: 'I won't write anything up for you at all, I'll write what there is'. You have the opportunity to genuinely choose from maybe a thousand people, from a thousand people, Ilnar, to really choose ten, but choose them more seriously. That's what it shows you, for example. Choose more seriously. I would apply this thing in terms of such an approach, because no recruiter is able to call… I don't know anyone who has ever called a thousand CVs for a CEO interview.
My people called hundreds of people to hire directors. Obviously, right? By voice. They called, but then they gave their personal opinion. Say Vlada is sitting and calling — you know her, right? She calls and speaks to him, this person. She didn't like how he answered her about something. She goes: 'Goodbye'. Wrote the comment 'jerk', for example, or 'weird'. Came back. I say: 'Vlada, I heard a certain person approached us'. She says: 'No, Alexander, no need to look at that one'. And that's it.
But you did, but you did trust Vlada's opinion?
Of course! Look, of course I trusted it. But this system would let her learn a bit more broadly, because we, well, since I employed her for more than ten years, right? They later saw that sometimes they could form a wrong opinion about a person, just like I could. I've said many times that this is a wrong-
Maybe just like the system could? Since it-
That's, Anya, Anya-
It also has a certain algorithm built in, yes.
A hundred percent! It's just an additional option. And here's exactly the thing. I'm not protesting. It's just that you said you don't talk to bots. I gave you a couple of cases, and right away you say: 'Well, in that case I do talk'. So.
Because, well, the cases really are idealised, and I think customising them like that, approaching it like that and explaining it like that... Well, let's be honest. In reality it works differently now. Can it work like that? Of course it can. Will it work like that? Well, super cool. All the more people will use it outside Europe.
You know what I believe in more with systems like this? I have acquaintances, fairly young, mostly guys, data scientists — what do they usually do? Well, and they're not married. In dating apps, first, they use an auto-clicker that swipes everyone in the right direction, and then they start chatting after a match. But chatting is too much effort as well. They set up a system that, with the help of, say, ChatGPT, Gemini and others, then chats with the girl, finds out everything of interest and sets up a date.
And then they go on those dates. I mean, in the end you launch the system and say that this Friday or next Friday I have a free slot, so let's meet someone. And they simply step out of this communication process completely, yes, and wait for the system to find something. I believe more that these guys will also set up Grok Bot to call girls, so as to chat with them and close the date faster, than that this will get commercialised. That, I think, will work.
I'll also tell you that over the last year, according to Greenhouse — a company that makes various hiring and HR systems — half of American candidates had an interview with artificial intelligence.
Fairly successfully, I think.
Well, I don't know about successfully. I mean, in this case we're talking about pass-through, yes.
Personally, I'd simply suggest in that case — and what you're saying, Anya, that of course the number will only grow and increase, and the number of voice assistants, people, robots will only grow.
I gave you an example of a case I'd happily use myself if I had the option of communicating within my family, including through a bot. Not to reduce human communication, but to improve the efficiency of resolving certain questions. And I think it would be very effective for companies to develop systems where they think more broadly about people and understand different processes. It's not easy. Doing such surveys at all — as you say, idealisation — isn't easy. But right from the start you have to see whether you're really able to select, check everyone and set things up, do it and so on.
And at the same time not burden the person. I mean, if a person doesn't want to — there are now people who, well, maybe psychologically can't communicate… that happens too, right? People psychologically can't communicate with a robot, for example. Say Ilnar — it calls him, he says: 'Well, I can't, kill me if you like'. It says: 'No problem, Ilnar, we'll talk by voice with a human; someone will call you back now'. If you, probably, set it up, set this thing up and approach everyone humanely, everyone, then it's a great thing.
I, for example, very often don't like that in America very many companies want to call. And I, for example, often want to text on WhatsApp or by SMS. I mean, I don't like calls. When it comes to purchases, offers, details. It's much easier to sell to me by message. It's easier to sell to me. If it's a normal message, I text normally, I see the service, it's much easier for me. As soon as the calls start, that's the end. You have to do it at a certain time, they start extracting unclear information from you, ask the same questions over and over.
You can't refer back and say: 'I've already sent you this, sent it above'. I mean, I think the question of service quality here is true multichannel communication. And not going all into one thing. Right now we see this big problem in the world. I just read a case about Airbnb, that on Airbnb all the hosts are complaining, those who rent out, who rent out apartments, flats, and complaining that their support has become entirely, say, non-English-speaking, and support that doesn't know the system, yes. That's bad.
I mean, they say that before they called, people used to know the system, answered any questions and so on. You don't understand where you even are. But I have my own Airbnb complaint. I'm in Europe now, and several times I had to stay not in hotels but at Airbnbs. Of course, I was struck with horror that at check-out you have to, you pay two hundred dollars for cleaning, and you have to take out the rubbish in Provence, not knowing, who knows where. I mean, you have to go out into the street, find the bins — you don't even know where they are — sort the rubbish, take it out and throw it away, yes. Or strip the beds.
Me, well, we rented an Airbnb in Laguna Beach with Dima the day before yesterday, and it says there: 'In our house you don't need to take out the rubbish, you don't need to strip the bed linen, you don't need to take out the towels, you don't need to do anything. Just walk out and close the door'. So. And the only thing written there was, like, the only thing, turn off the washing machine. So. These days almost everywhere you have to practically load everything into the dishwasher, unload it. That's a question of understanding service quality, yes, probably, not of bots.
And the Anthropic IPO is not just on a simple horizon; it's also right at a change in a certain system of power in the US, in the distribution of who, so to speak, can call whom on the carpet and who will allocate what money to what. Tech markets really have sagged over the last month. You can clearly see the tech market sagged more than the broad one. Meanwhile Anthropic is preparing for its IPO. The valuation is expected to be over two trillion. I have a feeling it should be over three even at the start. Well, we'll see. We'll see what happens.
I mean, I have a feeling Anthropic should take off brutally. Their last valuation was a very long time ago, under a billion dollars. And what's interesting about this construction? Here, well, from the numbers alone you can reason that they'll get a lot of money, big financing, and unlike small operators they'll be able to develop very seriously and strongly. And the small ones will start collapsing here and there and so on. That they'll supposedly have money to compete more with OpenAI or someone. Maybe, though hardly.
Why would I single this story out? Because when a company goes public, various regulators have far more claims on it. And we'll essentially be seeing a different kind of reporting. And even now, when filing with the SEC, they had to disclose that — it was written there — they forecast that artificial intelligence will increasingly create itself, well, in the near future, yes, more and more. And against this backdrop Sam Altman comes out and writes a post that under no circumstances can we allow artificial intelligence to govern itself rather than humanity governing artificial intelligence.
Yes, he posts it, and people write to him there, Sam, like, well, they throw poop at him. Well, what usually happens on social media, right? So they throw poop. It happens quite often on our YouTube channel too. And then he writes: 'Actually, I sincerely, from the bottom of my heart, wanted to give all people an infinite amount of compute, so that every person would have phenomenal, big, serious, real computing power in their life'. Right? What's your feeling about this movement and the Anthropic IPO?
Actually, look, the topic might seem not very important, but in reality it's a super-important story. Because companies like Anthropic and OpenAI, even SpaceX, xAI, Tesla — that's still a separate, kind of, well, kind of a world's one-off, right? Elon Musk there is altogether, it's like a separate line of business, right? Still, OpenAI, OpenAI's and Anthropic's entry to the public market with such serious real valuations. These valuations will be in the top ten, top twenty companies in the world.
That's a very serious structural change. I mean, in this case we see — there's a good saying that California is, like, top five, I think it's in fifth place, as a world economy, or in sixth, yes, place as a world economy. And you could single it out as a separate huge infrastructure altogether. OpenAI and Anthropic, of course, are becoming separate countries today, yes, in terms of the scale of their influence, resources and access to people overall. They probably already rank, will start ranking among the top five countries in the world, given the number of users and the influence they have.
That's very cool. The fact that Anthropic is going for an IPO — or OpenAI would — is definitely, in my view, a huge plus. Because for public companies, being public is exactly one of the important points. You see the structure of revenues and expenses, how things change for them, what they spend money on, how and how much money they earn. And from closed companies like these, which just shared whatever data they wanted somewhere, they turn into companies about which you can get a great deal of information from their reporting.
Right now they've published this S-1 form, and many people are already studying it in detail. About the company's profitability, about the expenses it bears, about the investments it's planning for further compute. I think they're planning half a trillion on compute up to 2030. At the same time, Anthropic is still very conservative in terms of investment and in terms of whether its own money will be enough to cover all its needs and all the obligations it takes on. I mean, remember, about a year ago we discussed that Anthropic was going very conservatively on compute, on buying compute, computing capacity.
And now the fact that OpenAI hands out limit resets to everyone while Anthropic can't do that is a consequence of exactly that policy, when they planned for a very conservative scenario, but the market went insanely optimistic and their number of users grew very strongly. They still keep developing this line. I mean, half a trillion seems like a lot, but on the other hand it's still conservative. Anthropic has enough money of its own in the sense that they've already raised more than a hundred billion, and at the IPO they'll most likely raise a hundred-plus billion more, well, in real money.
That will be enough for them. Plus even the current level of profitability will be enough to cover all their obligations. I mean, the company looks insanely reliable from the outset.
Well, what obligations? Listen, they already have almost a trillion dollars of obligations ahead, and they'll float, Ilnar, well, ten percent, they won't get more.
Well, a hundred billion, roughly, a hundred and fifty billion they'll raise.
No, well, more. No, they'll raise more. I think the valuation won't be under two trillion. Well, or maybe they'll, yes, they could raise, you're right, they could raise four percent or five percent.
Well, yes. Actually, I saw estimates that they'll raise somewhere around a hundred, a hundred and fifty billion. They also have about a hundred billion in bank accounts right now. They keep it as a cushion. So there you have two hundred and fifty billion. And per quarter the company is now making, I don't remember, thirty-six billion or so. And accordingly, with all this you have obligations that-
No, the first quarter, the first quarter of 2026 was five billion dollars, the second quarter was forecast at no less than eleven billion, and the first half — sixteen billion.
Well, and by the third quarter, over nine months, I think they made thirty-something in total. Well, again, I may be wrong.
Yes, yes.
But the estimate is that the money in the accounts plus even the company's current earnings, given that they won't grow much, or maybe won't grow at all, cover the obligations they've taken on now. Later they may take on more, but they approach it fairly conservatively. When they talked about investments, or OpenAI's fundraising and so on, the feeling was that OpenAI is far more optimistic about these expenses and obligations it takes on. Anthropic is actually continuing this story. What do I want to say here? The Chinese went for IPOs earlier, and almost all of them lost substantial market capitalisation since their IPO.
It will be interesting to see what happens with Anthropic, because all the same, given the forecasts that their revenue will grow, the number of users they have will grow, and, as Sasha Mashrabov would say, adoption of the technology on the market will grow, they'll earn more money, and it seems the capitalisation should still grow too. We'll see whether such a company can show this on the market after the IPO, because the Chinese, I repeat, couldn't do it.
Ilnar, I have a question. Besides the transparency and access to certain data, to the company's economics, which going public gives the whole world, essentially, and users — what else might change for the user? For an ordinary Claude user.
Uh-huh. The company is raising a huge amount of money, so it will most likely be able to allocate more resources. I don't know whether they're constrained right now, but it will be able to allocate more resources to developing further models, to improving infrastructure and making its models available to more users. So all the same, it looks like a useful story. I'd worry, well, with my everyday mind from my sofa, if some other company bought a large stake in Anthropic — that would most likely mean the company's direction of development would change.
Well, it seems that shouldn't happen now, at least not in the near future.
And don't these things also affect, say, how free or cheap Claude becomes in this case?
Here it's probably better to listen to Sasha on what the company wants to show before the IPO. And based on that, there-
What it, after all, it also holds up, yes, by betting on some… certain volumes in the future. I mean, could there be some dependence on that? How much could the artificial intelligence we use now cost? We often say that what we use, we actually use a bit on credit, yes, and not at the real cost.
Look, regardless of the IPO, here's what I want to say: inference is cheap. I mean, actually, a two-hundred-dollar subscription or, like, a five-hundred-dollar OpenAI subscription, purely in terms of providing the service, where somewhere a server room heats the air and you get results, most likely pays off fine even for these super-players that do all this. When does it, roughly, not pay off? When you constantly have to come up with new models and develop them, because a great deal goes into the next models, into the research that goes on.
But if everything froze in the current state and stopped developing, inference would still be profitable. Remember, we discussed last week that even at DeepSeek, with its penny-cheap API requests, GPUs generate $36 an hour. So.
Yes, yes, yes.
For other companies, at their API prices, it's really big money. But my point is that even on a subscription it most likely works out fine. Well, and if we look at Anthropic's reporting, we can most likely see it. To my shame I'll say that I haven't looked at it yet, but we can dig in, look and then answer this question with arguments.
Well, it just seems to me that very often for an ordinary person, well, roughly — I don't know how correct it is to say 'ordinary person' — but what matters to them is that nothing changes for them in terms of availability, that everything keeps developing, that it doesn't get more expensive, well, yes, within reasonable limits and so on. I mean, the first zone is that this stays stable for me. Well, and beyond that, not everyone is interested in Anthropic's economics, right?
I, for example, believe that we're loss-making, all the heavy users. I mean, I have, for example, the opposite opinion about the calculations. I'm deeply convinced that my requests are loss-making for the company, really heavily loss-making. And could you now just shut down and stop, and stop developing? In theory you could, but it's only a theory. I mean, you'd really get a certain benefit in the short term. In the short term. But in the long run, if you look at it, others will simply overtake you, and you'll disappear from the market altogether, you won't exist.
And obviously, Anthropic isn't the only company; there are other companies, and it doesn't have the option to do that, right, it can't do that. And the various tests we see and the improvements that appear cost a great deal, incredibly much, yes. And to do what we're discussing in terms of, and all people already have the desire. Anya, you say nothing will change for a person. Look at what's happening. Everything is improving for everyone. But when people are told that their tokens — Ilnar says: 'Oh, damn, there won't be as many tokens, that's bad'. Relative to Plus.
And nobody even knows how many tokens there are in Plus. Nobody understands at all what tokens are and how to manage these tokens, yes. And how many tasks were or weren't being solved. Nobody understood and nobody understands. Well, if you take it at the basic level. But people growl. Ilnar and I once bought two-hundred-dollar subscriptions when it was just a chat. And we were happy with that cool chat. Now, damn, it's not just a chat but a colossal system that does such things for you.
I mean, now you create software that costs a hundred thousand dollars in two hours, right? And this, this isn't a joke, it's a fact, it's, it's already a fact of the past, a fact from months ago, right? It's not a current fact. The current fact: OpenAI posted that it's now possible to create hardware.
They released a new engine for creating hardware, and they created something like a Tamagotchi, completely. I mean, the printed circuit board, the microcontroller software, absolutely all the code — they programmed everything and handed it to a separate provider for manufacturing. Well, we're moving towards the case I've described many times: as soon as artificial intelligence learns to program microcontrollers and then create hardware, that's where we'll see a completely different, well, kind of different approach to everything happening in the world. By the way, the topic is very interesting from the hardware point of view.
It will bring, someone will make a great deal of money on it. A huge number of companies will lose, because the market will now be restructuring insanely. Finally, I hope, the smart-home market and TV software will change, right? Because if you take TV software, for example, it's just horrible, right? It's just horrible. They learned to make the panels but forgot to make the software for them. Yes, so. The topic is insanely interesting, and I don't know how open this engine will be and how much it needs opening, because basically Codex will be doing it.
Take YouTube, take Amazon, take WhatsApp. WhatsApp — just take WhatsApp. You understand, right, that there's no money in WhatsApp. It's loss-making infrastructure. And YouTube was completely loss-making for a long time. And Amazon was colossally loss-making for a long time. It's like in Russia there's Ozon, since 1998, I think. In total, I don't know how it is now, I think they're loss-making now too. Take it from the very beginning, look, there was no money there. There was some bit, a micro-bit, some spike for a couple of years. But in terms of overall investment it's an investment in the future. Someone will ask, how does that work? Well, that's just how it works, yes.
That's how it works in terms of the world economy. We wouldn't have anything at all if there weren't, weren't such companies, such infrastructure companies. And what about Telegram? What about Telegram? Maintaining such a volume of users, look, maintaining such a volume of users for so long and earning nothing at all. They introduced, started doing monetisation through advertising, through selling various plans. Much, well, they spent a huge amount of money. And here the story is this: if Anthropic is worth, like, twenty trillion dollars, then the population, ordinary people, will have access to models far cooler than now, essentially for free.
I mean, free access to the model now is awesome. Whoever complains that their access costs twenty dollars — they can't pay twenty dollars. Well, it's just, it's just, well, don't, don't, don't eat an extra kilo of sausages and cancel your Netflix subscription or whatever you watch. That's unreasonable spending in terms of the efficiency of your money, right? And can Anthropic, by the way, give everyone subscriptions? Does Anthropic need twenty-dollar subscriptions? Not at all. It won't make their unit economics work now, at this point in time, because users on twenty dollars, they definitely more… once again, this is my opinion — in terms of memory, individual resources, they'll spend more compute. But Anthropic, OpenAI, what are they heading towards?
They're heading towards creating AGI. They're not heading towards bringing every person some super-special system. For us, for all of us it will be a commodity, guys. And paying a hundred dollars, five hundred dollars or three hundred dollars, we'll get almost nothing. But that should be obvious; it will most likely be almost nothing. It will be a very simple construction, some simple thing. To do things like we're doing now, you'll have to pay tens of thousands of dollars, well, in current terms. That's how it should be. And by the way, such statements are already coming.
Look, a huge number of people complain that on Grok, for example, tokens are used up quickly. Elon Musk writes: 'What did you expect?' And there another person says: 'Well, clearly Grok Pod costs about twenty thousand dollars a year. Like the one you have now'. That's a normal statement, an absolutely normal statement, right? It's just that right now everyone has been given a huge number of little goodies to play with, and in a year these goodies will be, I think they'll be nothing, well, in terms of where we're heading.
So. They'll be nothing. Models will be far more serious and far more expensive. By the way, what's going on with tokens? I hear that people on Pro one hundred, for a hundred dollars, have started burning through tokens very fast, and at Anthropic, those on a plan for, like, say, twenty dollars, the tokens run out fast there.
Although, by the way, Anthropic renamed its plans again, right. They're really going crazy. ChatGPT, Anthropic, they're just going crazy, I think. I, I, I… well. It's, it's unique. Maybe something will change for them, but I think after the IPO it'll be even worse. I mean, well. They shouldn't, they shouldn't… by the way, Ilnar, did you see the statement or not? This will interest you. Boris Cherny, who is one of the main developers of Anthropic Claude, said that Claude Code is outdated and, to develop it further, it has to be erased and completely rewritten.
And just recently we were talking, like, at one company where I'm on the board of directors, and we were discussing there, and it's a big infrastructure, huge, very big, yes, thousands of everything. And I expressed the opinion that the only salvation is to delete everything and rewrite everything from scratch, everything. Before, such an assumption in terms of basic development, it, people always liked to use it. Everyone always comes along, and says: 'Hey, you need to rewrite this, you've got, I don't know, a monolith here, or a pile of old code, legacy, like, you need to rewrite everything'.
Of course that was impossible to do, especially in big corporations. But now I think it's the only salvation. And Apple published information that they've cut spending per person. Anya, so don't be surprised, by the way. You tell me from time to time that people don't allocate money for development, for artificial intelligence. Do you know that Apple cut developers-
Yes, I read it.
To two hundred dollars, to two hundred dollars a day.
And for some operational staff, to twenty-five.
Yes, a day. And said it would cut it to a hundred dollars. And how can you even manage a budget per developer when now there's this AI, a machine that develops everything itself. I mean, it's an insane approach, and one that leads corporations to constructions like these. Yes, many people down to twenty-five dollars. Twenty-five dollars, mind you, is six hundred dollars a month. Six hundred dollars a month in tokens at Anthropic isn't the two-hundred-dollar plan. Well, I don't know if everyone understands this.
It's not the two-hundred-dollar plan; they get eaten up, well, like this. Or you do nothing in there, right? Well, you do some very simple things, by the way, you do very simple ones.
Maybe Apple is just talking about some kind of average, I think...
I think Apple has a problem. I think Apple has a problem overall.
Or it's average usage, just divided by headcount. Because the very allowance of an amount of money per person is very strange in general, especially per engineer. If we're talking about-
Of course.
...development, projects, teamwork. I mean, it's not even very clear how this is then implemented in terms of using artificial intelligence specifically.
Let's recall: developers at OpenAI said that they spend tokens, one session of theirs eats twenty dollars an hour. Remember, there was a price announced at Dev Day, which I attended in San Francisco; at a closed session a person said: 'Well, twenty — it's cheaper now, like, ten dollars'. So. And clearly Apple uses Claude. Well, most likely they use Claude to write code. But, by the way, we haven't mentioned Apple for a very long time. Well, more precisely, Apple, Apple, Apple — it comes up and probably appears in some storyboard of ours, but we don't mention it in terms of serious things.
And now it turned out they updated Siri, and you want to spit at this Siri, yes. I won't test it anymore, well, for now at least, I'm disappointed in them, yes. Not in the devices, but in software it's just insane, insane rubbish. So. About nothing. By the way, Dot. I liked the update Tibo posted. So, they wrote that there was an update for Dot, and there was an update. It seems a dumb thing, but very useful.
Now you'll be able to call up Dot — there's this extra button on the iPhone — and you can assign it to call up not ChatGPT but Dot. Because right now you usually have to call up ChatGPT. In ChatGPT you have to swipe, select this Dot and then go into it. Unlike, by the way, Claude Duet. Whoever has worked with Claude Duet at scale, write in too, okay. Whoever, who managed to test Claude Duet at scale. It made no impression on me at all. I stopped using it, yes.
Finally this button on the iPhone will be good for something.
Sasha, I have an iPhone 14, I don't have this button yet.
No, listen.
I didn't decide it was useful enough to change my phone.
No, look, there is one, maybe you mean a different button. There's one button on the side, above, where the volume is, and there's a button on the right that opens the camera. This right butt-
I don't have that one.
Yes, yes. And now I'm talking about the upper button on the left. You have neither one nor the other, right?
No, the left one I have, yes. For some reason I think that-
Those you have. And on the left button you set ChatGPT. I have it set, I'm very happy. It's super convenient for me. The button on the right, indeed, let me remind you, was OpenAI's idea, oops, Apple's — to create, essentially, a new interface, I mean, a certain approach to the interface, where with a button you start your phone's contact with the world around you; it can be text, it can be visual, the camera, like, audio and so on. And then you start this Siri Intelligence or Apple Intelligence, yes, Apple Intelligence, and interact inside.
It worked clumsily for them; then they started integrating it, for example, with ChatGPT and so on. Everything worked terribly clumsily. And I don't know what kind of person uses Apple Intelligence at all. And I need this button to open the camera quickly. It just opens with a single press if you don't hold it long. I don't hold it long, and it opens the camera. By the way, I should check whether you can reprogram a long hold to open Dots. That would be, that would be useful. Ilnar, have you tested Claude Duet or not?
You haven't, right?
No.
Muse from Meta, and it left a negative impression on me. I mean, what else I probably didn't like is that Anthropic, Anthropic somehow always stood apart and did its own things, and this smacked heavily of copying, half-baked copying of Dots. So. By the way, Dots is expanding in some aspects and starting to work better. I like that it has started working much better with already created chats that existed in the infrastructure.
At the same time it has, of course, a huge number of clumsy aspects, where it's unclear how to configure them, and unclear in terms of recurring tasks, in terms of what it actually wants to do and what it actually doesn't want to do. What I still don't like about all artificial intelligences is that these systems don't want to ask me anything. I mean, they don't want to improve, develop, learn even more, yes. And this, by the way-
There are some attempts, though. I mean-
Well, these, these are laughable attempts.
They ask some single question.
Yes, but this is exactly the story where they hand it over, and then you've even spent some time, and then it turns out you wasted your time because it was someone's experiment, yes. It's roughly like the OpenAI newsletter that comes by email, and we're simply not a landmark YouTube for America. YouTube — if it were, they'd hear me say that the whole email newsletter team should be fired. Well, you can't keep people like that. Well, it can't retrain, kind of, it's impossible.
It's, it's physically impossible for that to happen. So. There are so many interesting great cases, so much cool information. Even, by the way, we understand that the OpenAI newsletter is read by, I think, a hundred million people, right? More than a hundred million people. And Tibo on X is read by hundreds of thousands, maybe, or maybe not. His post, for example, right? But maybe the post spreads a bit further somewhere. So engineers there often write more interesting things or cool stuff than what goes out in the newsletter. Well, at least, at least they could repost their own engineers in these newsletters.
Tibo's target audience is still, Sasha, more of an engineering one. That's actually why they read him there on X.
I understand what you're saying, Ilnar, but he, he often posts sensible things. Or they write that a new one has appeared-
He's a very smart person, yes.
Yes. Or which ones were released. No, I mean which sensible ones. Well, yes, he writes a lot of ordinary simple things. But what engineering things does he write on X? He writes nothing engineering. For a month they were doing resets, and all he wrote was reset or no reset, or I made you a slider in an update, or the chat finally didn't disappear, or something else there. If you really look, Ilnar, then, well, for deeply engineering things it doesn't compare, I don't know, with Karpathy's posts, for example, right?
There at least there are technical, technical posts, right? Tibo has no technical… his are exactly marketing ones. It feels like he's a marketer. A marketer who sits in the company and does marketing, right? Well, generally-
And he has a reset button.
Yes. But I think he doesn't have a reset button. And by the way, I didn't like something at OpenAI. I, I just… I, I, I really didn't like it. It turns out that at OpenAI, when you press reset, it resets you — it doesn't just give you a hundred percent, it pushes your weekly corporate allowance back by seven days again.
And if you reset a day before the regular reset, your reset gets pushed back a week. What kind of setup is that with reset? I mean, well, it's, it's a real setup. At Claude, at Claude, by the way, they don't actually give out resets. They don't actually give out resets, but their reset worked differently. But they do give them sometimes, yes, very rarely, very rarely. And I, by the way, have a reset for free in Claude again; it appeared in Claude. I don't know under what circumstances they give it out and when, but it appears from time to time. But Tibo did a very cool thing.
We all devoured an unreal amount of compute on Codex, just cosmic. This month I've eaten more than ten billion tokens. I'm already at sixty-nine point six billion tokens, and I don't understand at all how that could happen.
And the last reset was already about four days ago, I think, and I'm already a bit short of it. You think: damn, Tibo, when's the next reset? And the marathon is over.
So I reset exactly a day before, and then decided no, because I've started spending too much now. Otherwise it won't be hundreds of dollars slipping away but tens of thousands of dollars in purchased tokens. And so it sometimes makes sense to wait a bit. But they gave out a lot of resets, Ilnar. On the first, the second of the month. Yes, on the first, the second of the month they didn't just reset everyone but gave extra. They pulled exactly this trick. So. They were supposed to give it to everyone from the start of the month anyway, they were supposed to give it to everyone anyway. But, by the way, they definitely increased capacity.
At least on Pro five hundred I started to feel a bit more capacity. But I really did recently read a comment — I think someone wrote it on our channel — that it feels like only on Pro for five hundred dollars do you still get a normal amount of tokens. And probably that's noticeable, because I eat even the five-hundred-dollar tokens. In general, I told Dima Mugenov that for normal work I need the amount of tokens that's in the Claude subscription, I think it's now called Ultra, for two hundred dollars, and in ChatGPT Pro for five hundred dollars — I need that volume of tokens every day.
I mean, in the two systems I need that volume of tokens every day, and I'd come out roughly even. I'd come out roughly at zero-zero. And every day.
Well-
In money, in money it'll be very expensive. In money it's hundreds of thousands of dollars, yes, a month.
Yes.
In money, in money it's hundreds of thousands of dollars a month. And of course-
Well, actually-
It's a little scary here.
Well, they'll most likely block you there. By the way, for those who don't know, OpenAI officially allows registering several accounts. They've even just released a new feature. Tibo wrote about it. Well, an awesome engineering feature: you can easily log in and log out of several accounts. They officially allow it, right?
Anthropic, as far as I understand, doesn't officially allow it. And Anthropic keeps up its mass bans around the world, right? And there's this story that many… let me remind you that there's a big difference between using a corporate plan and a personal plan at Anthropic. Anthropic may send you a letter asking: are you a company or a private person? Just so you know, don't rush to answer them that you're a private person, because Anthropic officially allows using an account in your own business if you're, say, on your own, and you have the right to buy an ordinary simple account, not a business one, and pay for it at Anthropic as a business. That's an important, by the way, important allowance.
Many think that I, well, start thinking that you're not allowed to do this. And, for example, Claude detects what software you're developing, that you're making it, for example, for a business, or some business processes — I don't know, you're analysing financial statements, like, marketing, sales and so on. Just so you know, this isn't prohibited at Anthropic. If they suddenly contact you — and lately Anthropic does a lot of KYC, yes, and, well, a lot of internal checks, that's what's called know your customer — they request details from you, who you are, and may even ask how, where you live, in which country, what passport you have and so on.
Your task: don't rush at first, breathe out and study a little how this system works, what rules it has, and answer it according to those rules. Well, as for the violators there are — there are plenty of violators in the world who misuse Anthropic. Let me remind you that Anthropic, in the big picture, of course, fights the use of its system, kind of, against humanity, right? Or against the United States of America, or against Anthropic itself, in terms of, against distillation, right? We talk a lot about distillation.
For them, in fact, it's important to understand all the time why they contacted you. And for them you, as an ordinary person who is a sole proprietor or has, like, five employees, and you do some reporting, calculations — there's no difference at all. So study these rules, okay. They may, at least in the Russian-speaking segment, I see a lot of this now, a lot of account bans. And they ban accounts not only in, I don't know, in Russia, for example, or, like, in Belarus. Accounts get banned all over the world, and in America too they send account bans.
And you need to know how to correspond with these contacts. I had a very serious discussion with YouTube about where our channel, our current channel, is located. And a really huge amount of correspondence, a three-month discussion, yes. And an interesting thing worked: they weren't interested in documents, or American companies, or anything at all. They found a trace, some old one, they found a trace and, for example, found an old trace, I don't know, a Russian one, right. And until I sent them very precise dates of when and where I lived, in which country — as soon as I sent them, everything was resolved at once. Before that I'd been explaining to them that I'm in America.
All my documents are in America, and all the registrations, like, well, absolutely everything, yes, of all the companies, businesses, my personal addresses, whatever you want. They weren't interested in that until I told them the real, kind of, chain of my life. As soon as I told them, everything immediately, immediately became, well, calm in the systems. By the way, that's a big problem in the world now in terms of… many systems. The system no longer understands where you… sometimes, sometimes it doesn't understand where you're coming from, where you are.
People use proxies and VPNs not, often, to deceive someone, but, for example, to protect their information or because it's corporate use. It's a big topic overall, of course. A big, difficult topic. But what's clear is that these systems more and more easily determine, in fact, who you are. And, of course, from the texts you enter into Claude. Do you know how you can tell that OpenAI and Anthropic don't train on our texts? If they did, they'd have blocked, I think, millions more accounts. Well, in principle, in principle, right? Because, I think, it's very easy for people who are...
That's being blocked now, Sasha, one after another.
Yes, but it's very easy to determine from the content inside the system, right. And-
But that's exactly what they do.
Yes. But I, you know, Ilnar, there are probably some triggers that fire. Here's what I see: from time to time Anthropic, I have a lot of Russian, right, because I have a channel in Russian, I have an English business, there's stuff in Russian, a lot of Russian. And it's astonishing that Anthropic already knows a trillion times over that I'm American, that everything around me is American everywhere. It keeps adapting any Russian-language product of mine to Russian legislation right at the start. It's really astonishing.
I mean, it isn't interested in American legislation, or, or European, I don't know, legislation in terms of personal data. It starts out by telling me about Russian laws, and you're just stunned by it. Like, you, well, why do I even need this law? Why? This is a product that has nothing to do with, for example, location, right. And I wonder how the model works. I mean, essentially, the model itself has some very strong language binding, for example, right.
In the training dataset, yes, Sasha, that's how it was built: when, roughly, you communicate in Russian, yes.
I mean, if I launch a Belarusian landing page in Belarusian now, for example, it will be absolutely sure that I'm trading from Belarus and need to comply with Belarusian law. Although, obviously, I can launch a landing page in Belarusian, like, I don't know, for selling Belarusian goods in Silicon Valley, Valley.
By the way, I want to say that for any query, any comment, any research, everyone first of all gives me the angle of Poland and Polish law. Even if it's completely irrelevant and doesn't fit the topic under discussion at all, it seems as if my artificial intelligence definitely knows.
Well, by the way, Anya, look, I think, I think here it will be very strongly adapted to the location you're in. Not for nothing did I tell you that when Dima connects to my Wi-Fi, he has Dota available, and when he connects via a SIM card from other countries, it immediately disappears for him, right. I mean, Anya, it still primarily sees, in fact, where you are. And here let's imagine you're in Europe and then start making a product in Chinese, take an interest in China and so on.
I mean, how will it react at that, at that moment, right? That's, that's interesting, yes. Or how will it react, I don't know, to Vietnam, or how will it react to Brazil. I mean, will it understand that… will it start telling you local laws? Do you know what I told it recently? You worry so much about protecting the personal data of Europe and Russia — maybe you should worry about the other hundred and fifty, like, or however many countries of the world, right? And you know what it did?
It didn't answer me. It disappeared. It kind of, its system, because, well-
It broke.
No, well, because I tell it: 'Well, this is nonsense, this is absolute, kind of absolute rubbish'. I mean, piling on, piling on, if you... And then, you know what the story is? In the States, for example. In the States you sell, but in the States everything depends on which state you're selling to. And then it starts. If you still don't sell too many goods in another state, you don't need to adapt to it. But if it's too many — off it goes, right. So, well, it's like that everywhere, right? Like that everywhere.
You have to know, you have to know the rules, but not slide into this paranoia. We're waiting for your comments and likes, and see you on our podcast in exactly one week. Don't forget to watch our episodes, which come out almost every other day. Bye, everyone.