Sasha, you know what's the main news?
Yeah.
I went to the pay-as-you-go.
Pro? Pro? Pro?
Yeah, yeah.
Tell me, you have the main news. Yeah.
But I haven't felt any difference yet. But I hope it's just the beginning.
Specialized mode of reflection aimed at science and research, Google engineering tasks. Codex Park from OpenAI is a quick decision. Lockdown Mode from OpenAI. The agent's arm is definitely a huge, just a crazy trend in China. The Pentagon can break Anthropic relations because of restrictions on military use of models. Just a follow-up to our theme last time. Hey, everybody! We're on ToTheMoon. Technological news, Silicon Valley sites and the world.
And Tanya went to Pro's version. I was gonna say, by the way, that, uh, yeah, we've got a lot of other people out there, there, there, there's gonna be news tonight. But I was just saying that about Tanya, about the version, by the way, about the inconvenience-- the insensitive or sense of the story. We're in fact in the process, as Max Grigoriev said a month ago in a special issue we took off our canal, OpenAI treats people like choms and constantly on you, says it's all testing, Yeah.
ChatGPT is not a product, it's an experiment, and we're not OpenAI clients. We, if you pay them money or even if you just use ChatGPT, you're just sending data for this experiment. Well, if you're delivering money, you're supplying some more resources, right? No one wants to create a better chat-bot. And even more, O-- the leaders of OpenAI say that, they just don't listen to them. They say we're not here to build chat-bots and sorce. We build, uh, A-AGI, common artificial intelligence. Here.
Yes, general artificial--
They build a common artificial intelligence. This company has stated it from its very foundation. They said that. Uh, really, they were originally going, uh, for, uh, open source, if that's what you're supposed to call it, yeah. I mean, uh, OpenAI was founded as a company that would build AGI against Google. Because Google has, in principle, started working on these, uh, these models, and so on. And the other few wrestlers realized it would be dangerous if Google was the only company that would have this AGI. And to make a counterbalance to create some kind of counterweight available to other people outside Google, for this technology, they went and founded the OpenAI. It's-- they weren't gonna build a chat-bot, they weren't gonna build a video remediation, they weren't gonna produce any products. It was, it was-- it was a lab that was doing a study on artificial intelligence to build a common artificial intelligence.
And that's why they're so constant that you're basically no longer able to understand, uh, what theory is better or worse. I'm, in particular, planning on writing with Max Grigoriev about security in the artificial intelligence of specials in the coming weeks. I think it's gonna be very interesting and, um, about the safety of engineers, uh-oh, who work with artificial intelligence, that's a very important subject, especially the discussion on what trends are here. In the Silicon Valley. speed. There's an openAI out there, which is the subject of Codex Park, and there's one comment, well, as the base says everything, now, it's automatically coming. I mean, it's clear that Codex Park is for programming. I-- but on the other hand, what do they say? They say there is a possibility, there are some two regimes. One mode, you write the code and adjust it, you put it in line and then, there, three or four hours, there's something that can be adjusted. And there is when you need to review or make some changes instantly, and by the type of new versions there, you can do it like this, like, instantly.
I mean, this week, I've had a situation of 30 or 40 times where my regime is thinking in heavy mode, that is, a maximum thinking regime. I'll remind you that now, in the world, in different systems, there's a very different pattern of regimes in different pay-as-you-go versions. Like Gemini, they don't just have Pro theory, they have a signature, Ultra, it's called, right? It's like Grok called Heavy or Ultra too. And it's like the next level. Next level, next level. So, in ChatGPT, there's Pro and ChatGPT Pro, those who have Pro, they have a heavy thinking. And so, Tanya, to see Pro's version, first of all, we need to understand that in your mind-making mode, you have the option of choosing between two options, and four. In the usual, uh, plain version, I think, two options. You also have Pro theory, which also has two options: the extended version and the Pro version. It's an extended number of science or super-sex-- super-creaming scientists. So I noticed that I had some hard thinking, very complicated, and I was just like, "Tumbles of fingers" and "Fucking." I thought I had the wrong regime first. I'm so used to it that a good answer is, you have to wait. Well, I'm used to it that a good answer needs to wait. And in mode, a strong reasoning, like a strong thought, heavy thinking, yeah, in this mode. We'll talk about it. Gemini's released, uh, serious, straight-up, new thinking, yeah. And that's the logic of the whole thing that this regime is very like agents. Now, this week, there's some articles that look like agents. We talked about it in the past graduation, yes, what's the difference between reasoning and agents. So, uh, uh, the answers were so fast. And then I started doing some tests, saw it was my maximum treatment, and I realized that it was ChatGPT A-A-A-A-I-I-I-I had at least this week, maybe I got in the test line, I'm getting a lot of answers, um, m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-------------------m-m----------m-m-m-m-----m----------------m------m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m-m------------m--------m-m-------------------------------m------------- And of course, it changes the paradigm. I mean, if I, uh, suddenly there's a possibility, that's proversion, like I have to analyze, I don't know. We'll take you, like, 100,000 phone calls, $1000,000 phone calls transcribing. It took eighty minutes. Too, if you do it in a minute, what instrument am I getting? What kind of tool do I get? And we talked to my partner yesterday. He says, "Well, it's gonna need a little people." I say, "What does this system do?" Because with a system like that on the financial market, it's a whole other story. It's not just some soft-shot symphrase, but when you have the ability to get complicated, heavy things instantly, you're going to... I realized I'm in some other regime this week. of the decision-making process. Here.
What do you think about that?
Oh, look, we were still peeing with Sasha Masharov, and it's been six months ago, maybe we were talking, uh, Groq, which is from q to the end, and Cerebras. It's, uh-a-a, two start-ups that specialize in chips, um, for, uh-a, very high-speed inference, yeah, so models give answers very, very quickly. And they show the masterpiece of thousands of tokens a second, yes, that's one second, thousands of tokens are getting generated. I mean, if you remember...
Tell people, Elnar, tell people what a thousand tokens are like again, right?
We, uh, we use the usual GPT chat, well, there, remember your experience, there, a year and a half ago, maybe when you see words coming, and these pieces that appear are thick, and You can read as much as you can. Now imagine that in one second you have thousands of these tokens coming. One word is two or three thicks. Well, there's, like, three hundred and three hundred fifty words a second, you're just wearing a model. Oh, I understand that the answer is not simple, but with the thoughts. She's just talking at a speed of 300, four hundred words a second, yeah, so she's got a very quick thinking process than any other models she's got. Uh, this case was buried six months ago, not all the models. There were, there's no. So, not long ago, somewhere a month ago, we probably said that OpenAI with Cerebras seems to be not Groq, but Cerebras concluded, uh, an agreement that's been either... or bought, now I'm not. I remember, it's all in the head now. But the point is they're legal, and some of the chips Cerebras have, they're now available from OpenAI. And just as we discussed, proversies are likely to be able to, like, this extremely fast inference, yes, so that your models are instantly answerable to you. The only restrictions that this has yet to have, at least, are those that are known. Maybe it's settled inside OpenAI. From what you're saying, maybe they've already made a lot of progress. But the point was that Cerebras could be, uh, rescheduled, not the biggest models. And the GPT 5.3-Spark you started talking about is a small model, where I think it's up to 100 milli-- to, up to 100 billion parameters, it's pretty compact. And therefore, the market is all about taking it on Cerebras, on these, uh, chips. And that's why this is a big speed, uh, this model. But you have to understand that she's small, yes. The giant models on, there, there, a half-two trillion parameters in there until they're put.
Maybe ChatGPT, you know, I mean, OpenAI has some kind of...
Elnar, I'll be right there. And you can tell people the difference between a little fashion and...
Explain the difference between a small model and a big one. Yeah, I'll add to the tokens, too. But you tell people, yeah, that's b-- that's definitely gonna be important.
Look, uh, ChatGPT 4, it was about six hundred and five billion parameters. Uh, well, it's not clear, just because the information is hidden, but it's comparable to DeepSeek, and DeepSeek has about six hundred and thirty, I think, billions of parameters. Uh, now, it feels like the top models have to have a mild and a half trillion of parameters. Accordingly, if-
Explain the difference that this is for a man. What's the parameters? It's a volume. Explain what it is in human language.
Size, size of the model three times more.
No, explain. So? And people ask.
Explain it to me, Elnar.
So? What does that give?
She knows more about the world.
He knows. That's it.
Much more of her legality in g.
To understand.
The condition head could be.
To make everyone realize that Codex Spark, which generates a new, fast code. It's a small model, which means that she knows less about peace and is trained, for example, in a very narrow specialized, very narrow specialized history. Not as cool as 5.3.
Sas, there's gonna be an explosion in the commentaries, because we have experts in all the fields that only exist. We are now being challenged by the fact that: and the distillation of knowledge from different models and so on. Which can be made of a big model to be small with a little loss.
Elnar, we have no task to leave. You explained.
Okay.
Those-- we don't have to prove who knows what, do we? In terms of detail.
Yeah.
You explained to me that people would still understand the difference when they read somewhere and see that they're making this little model. I mean, I happen to be, like, ChatGPT, and there are these mini models, why didn't I use them? I'd normally wait, see somewhere. I need a lot of knowledge, I often need a variety of answers to topics that are very broad and so on. But in general, those people who do something, uh, for companies, for corporations or a narrowly specialized subject do, they often benefit from a small model. She'll be cheaper, faster learning, she'll be highly specialized. Why in the biology model have data on how to deal with how to do things, I don't know, some kind of thing that's not connected to it, because it's settled, yeah. No, I couldn't think of it, I couldn't think of a subject that had nothing to do with biological science.
Here.
It's easier to think, uh, it's easier to imagine. Well, a few years ago, we had GPT-4, there's four-O, and, uh, there was, like, six hundred and fifty billion parameters. You're using GPT 5.2 Pro now. It's clear that we put a reasoning on the top, we made another extra hint, but it's three times as big as it is. Compar what you could do with GPT-4O and what you can do now with modern models. And the plus-minus will be clear, and for what, we'll increase the size. For some private purpose, you can, uh-a, make a small model of, yes, distil her knowledge, thereby reducing it, and thus losing it somewhere else, but, In the area you're interested in, respectively. For example, Spark's in the programming, they distilled her, then put it on the chips, probably from Cerebras, and they're getting a very high speed model, but they're quite compact. Now, respectively, uh, that's the surveillance you have now, maybe ChatGPT was able to... ChatGPT, I keep saying ChatGPT. OpenAI was able to resolve this issue, and the size of the model was also able to accommodate these chips and to work for a high, uh, high, high speed inference. Accordingly, I think it's a matter of the coming months, which is likely to be a pro plus or something on some versions.
It'll be here.
We'll use models where the answer speed will be, well, just crazy.
But I can tell you that I felt personally. It's clear that many people will say what's different, not important. To feel important, important, it must be used very much. Even Tanya's got Pro now and says, at the beginning, there's no feeling, because you need to just, just a little more to be here and see.
I haven't tested heavy yet.
And see what boxes. You know what, Tanya, is an interesting story. You'll see a k-- you couldn't even understand some of the boxes before. I'm just saying, I couldn't even ask them. I'm so used to it right now. I understand, for example, where I ask Pro, I've already figured it out, yes, where.
Although I still have Pro's question doesn't go online because ChatGPT always changes. Yeah, it's like deep research. I mean, I realized that deep research is a stupid car, yeah, and you don't have to go into it anymore. You can use it much better. But the point is, Pro doesn't go online, but he's super smart in terms of knowledge. I'm going down there, for example, to analyze some volumes of big, very serious, scientific or very much when I'm analysing if I have to find a lot of things. I'm using a model of thinking, but again, I always choose heavy. I mean, I don't have a chosen heavy one. In fact, these choice boxes are purely for hamsters again. Because why am I not heavy? I mean, I've got to be a little bit of a life-long. And sometimes I'm the only case I'm willing to put, I'm ready to put it. Well, I was-- I was talking about ChatGPT right now. I'm ready, not what you thought, I'm ready to put in, uh, uh, like, I figured Pro's version when they wrote to me that Pro's version of the kind of no search. There, they write sometimes, you can't even change between these models. But the point is, in Pro, I at least saw for myself, uh, like, in my description, I know that it's connected to a bigger, bigger power, and, consequently, okay, I'm willing to give a few things back there. - One hour to work.
Sasha, maybe you got into some kind of testing or some gluc again, maybe. Pro's doing a good online thing.
Still walking, still walking.
Yeah, yeah, yeah. No problem.
Look, they just had it. I'll tell you. They had three weeks or how much they were written to me-- look, when I'm thinking, I'm trying to change Pro. It says, "We can't change Pro because Pro doesn't go online." Look, it's gone now, Ilnar. But when I, for example, did a difficult Pro request yesterday and I needed the Internet, she stayed in the reasoning. And by opening her window to see what she did, I didn't see any transition. Again, Ilnar, we can all be in the test zones, because the way the speeds change everywhere, yes.
What do you want to say about the tokens? Here, uh, for Anthropic information, Claude Sonnet 4.6. I wanted to tell you what she was interested in last time. He's got one million contexts in his bet, one million tokens. Just people just get to know once more, a little more detailed than what it is. Token is less than a word usually. And one million contexts is about a half or two thousand pages of the text. Everyone knows what it is, well, the viewers, they know that it's like a thousand and a half pages of the text. I mean, she's got a way out, and she's gonna keep you from doing something, processing, and she can only give it to 60,000 tokens. Well, you know, it's, like, 17 or so many times less, right? So she can give you back. It's a volume, even if it's 20 times smaller, there, how much is it? Mmm, 100, 100, well, there's a parole of 100 pages for you to give back, give it back to you. I'm talking in Anthropic right now, it's a beta-version. Systems are different, with different context windows. Iia because Anthropic is a regular Sonnet regime, there's two hundred, two hundred thousand tokens. Uh, again, look, people are talking about it as... a lot of technical information. These are all parameters, the size of models, uh, the number of tokens, that's all, like, a essay for me, right? It's a very technical world. But it's important for people to understand that when you have the opportunity to download two thousand pages of the text just in the coolest version, it means that if your life, imagine, you have twenty-four hours, you're gonna have to. You write down some sort of life, record it, and it's in the text, so the system can recognize your day. Well, if you put it in the text, there's a lot to do. In terms of, right, Ilnar? From the point of view of the recognition.
I, I was talking about Claude, yeah, I was just saying, yeah, Claude.
That the other companies were getting involved, too. And that's good. The question is, how far they do not lose context. Uh, like one million tokens, you need to, uh, don't lose the information you need, yeah. It's a few, I think, two sets of wars and peace, if I'm not mistaken, two thousand words. Oh, I mean, uh, one million tokens. Here. But, yes, that's cool. That means big code bases, big projects can be put in there.
Yeah. But when, for example, I'm going to go over there and analyze a huge volume, like, at the same time, that is, the files are more than 2,000 pages, she's still analysing them and she says she's been studying everything and... I've been looking into it. But there's another point she didn't break it on the blocks, right? I mean, the idea is, she probably beat it on the thinking blocks, hit the blocks because we were talking about API in terms of the tokens. When you-- what you're loading up in the browser, you're throwing up like a file, like a file, it's a big question, actually, that these systems can do, because there's no telling what they can do. I mean, it sometimes recognizes the file, sometimes it doesn't recognize it. Then you're throwing him a video, like, "I've found out, I've seen it all." You know, yeah, what can you do with the video? And the audio, like, "I didn't recognize it from the series." You're like, "Why didn't you recognize audio, for example?" Then you're, like, throwing a file on him, he says, "I can't do more than ten--
You, for example, throw a file on him, he says, "I can't take more than ten files." You're like that in the zip archive, you've written 30 files, and in the zip archive, he's got 30 files in a calm, easy, and he's got them all figured out, completely separated, showed you what and how.
I mean, of course, a lot of tests inside, uh, different ones. And for that, remember, we had faith-- the last time we told you that a man in England was £8,000 able to make a discount there for himself, like, eighty or more percent. And, in fact, the OpenAI is now in the middle of , Lockdown Mode, right? And, uh, this is a topic that allows corporate use to be a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a- So they're putting these systems in place to make corporations... Oh, I can imagine what's going on. There are people who are constantly trying to hack these models or constantly trying to do something through these models. And there's an endless amount of peddates coming. Now, at the corporate level, this OpenAI is for those who, again, are interested, they released Lockdown Mode and elevated risk, right? Just so that you can protect yourself from, uh, the-- to protect against prom-resent. I mean, when, uh, we're through the prom, uh, subpoenaing, looking for an answer. Remember, Ilnar gave that kind of a little or six months ago, did he?
Gen3 was playing, yeah.
Yeah.
Yeah, I remembered that too.
Yes, and it could be through the prom, like these prompt injections, yes, like in English, as to pro--- like prom viruses, yes, like proms.
Implementation, yes.
Prompt invention, yes, yes, yes. But the virus will be understood in Russian. Oh, yeah, the introduction. Uh, change a little bit, get something extra that you have to get the system to give you information, something that-- that it doesn't have to give, yeah, or where it's limited. It is clear that the main limitation in the world now is all that is related to, uh, military proliferation, killing, violence, the creation of different chemical weapons, and everything. I mean, there's a major area of these restrictions that-- that's actually going.
Illnar, you wanted to add on this subject, and I'll just go to, uh, military, military.
I'd, if I could, just a little news that I was admired just before, to the depths of my soul, is a matter of thrill. And they're just, Tanya, to your question about GPT-5.2 Pro, yes, whatever its meaning. A-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-day this week, this week, information that a-daysay, this week, that a group of physm-daysized phystorm-daystorm-the-days-daysaysized physized physics, that a-the-sm-s-the-the-the-the-days-this-days-days-days-daysized phys-a-days-days-days, that a-days, that a-days-days- Well, it's not an article yet, it's a prep, that's not a publication, but it's already been sort of a story about it. And new scientific discoveries are co-authored with GPT-5.2 Pro. Alex Lupsaska in his X-ray in the track described how this was going on. Shortly, he was involved in the OpenAI For S-- For Science program. Accordingly, the process has invited its scientific leader to participate, too, because he was admired by the results that GPT-5.2 showed. And then they, um, tried to pass through GPT-5.2, one stage in their article, which he could not decide on his own. And then it turns out that, after a while, there's a 12-hour, uh, model thought, because they could get the right prom or something. They've done this. So much of them admired them, they all say that they did all this in co-author. Uh, you can read the track in the X-ray. He points to, uh, specific formulas in the preprimary. We've come to this, we didn't know how to make it easier. Accordingly, further, thanks to these formulas, we have been able to simplify the way forward. And these formulas, they were actually, uh, obtained with GPT-5.2 Pro. They didn't think of it themselves. And, uh, that's pretty cool. It's really cool that in the advanced science, uh, there's a way to get results with, uh, models. Now why does it make me so horny and admirable, besides being like it's so obvious that it's cool. Look, uh, science is so built that, in fact, if we take the front end of science, very few people know it. Even finding someone who can talk about these questions is, you know, a problem. There's not much of them if we take the edges of science. I'm thinking about my dissertation.
I was just protecting my dissert, there's a certain number of science doctors sitting there, a man of twenty, probably. And they're listening to what you're talking about. I understand perfectly that in this audience, where there were twenty science doctors, maths that are math, well, my respects, are very well-trained, there were three things in my subject. I am the man who wrote this dissertation; my research leader who actually sent me on this subject, and apparently he's good at it; and the opponent who, well, or there are two opponents, The ones that were, who were, uh, my dissertation, they were taking the floor, asking me some additional questions at the pre-protect stage. We, he and I have gone through the whole phase. Everyone else understood the description, but the details were discussed with them, well, it's impossible if they don't get down on the subject. Now look what we got. Everyone who has access to, uh, GPT chat, well, Gemini and so on, but now we're talking about GPT, yes, we have a chance to talk about a subject, um, that's a few physics. There are dozens of people in the world who understand what's going on there. And GPT chatting, too, which gives them a plus-minus in one language. Uh, it's just a shock to me. I mean, I'm just saying that if I want to talk to someone about my dissertation, yeah, well, as if it was a science candidate dissertation, it's a very low level, it's not so many people in, uh, the Institutes in there, if... Take, there, Kazan, if you take Russia there, even, yes, it's, well, there's not a very large number of people who are dealing with this subject. Yeah, I can tell you, this is not an interesting subject for anyone, don't do more interesting things. I don't care. But with ChatGPT, I can discuss it, and it's very interesting. It's just, you know, a huge step for all of humanity. It's like you said, remember, when the Internet came up, that's all there, the whole world library, all your knowledge is in your pocket. Well, people are still there, like, scattering in riles, yeah, instead of talking about some stuff. But now, it's like a step further. Yeah, yeah, yeah. Now, step further. You have the opportunity to talk and discuss some, you know, some sort of thing that, well, very few people on the planet, you know, can understand.
What's double--
That's right.
- and scary, and admiring. I was asleep yesterday, you know, and I thought there might be a transition until there were these, a huge number of robots. When they're probably making robots, everything is done with people. But it's not done yet. I think you're gonna come to the doctor there, and you're gonna be given a choice: you want to go to AI or to a man. And then maybe even a man won't be offered. Or there's singers, you hear something, some kind of a combination, you think it's a beautiful voice, and then you watch it, and it's not a human being.
Tanya, you're making a really great move to another little news. I'll steal you two minutes now. Remember, Sasha, you told me that you were in Anthropic's office, uh, a vending machine that could be controlled. He ran AIK, and that's where it came from. This same company that made this, uh, Vending Bench, the Vending Automated Bachm, decided to run the next experiment. I mean, well, uh, in some digital environment, AI's already working well. There's a story with the scientists. What happens when we need to face the real world? You're saying, before the robots come out, we're kind of safe.
Yeah, physically.
They did. Yeah, yeah. Yeah, they made the next funny thing. They're in, uh, model. This model is a parole model, ah, the office centre manager. And they said, "Look, we need to put a gym on the roof." Yeah? Here, gather, there, I don't know, uh, a small gymnasium, but for that, you need to get the equipment there, get it, fix it, and so on. And they let the model hire a man. Oh, I mean, they gave her full access, they've allocated some budget. And the model has started, and, uh, to go through different resources and those resources, to try to get, you know, to hire a man who will come and do the work that needs to be done. The model has access to video cameras, and it can actually track all this. Uh, we'll put a piece on it, too, read it. It's a very exciting story, like the model tried on one, uh, website to find the perpetrators, but she didn't make it. Eight times she's been denied. Uh, they've been studying, uh, how communications and other things happen.
There's no reason why, uh, the model is being refused, but then she could find some other person on another, uh, site.
There's a website for that. You know, Ildar, there's a website for this now, right?
Yeah, yeah.
Site, where the models will be putting their orders.
Yeah, and she's like a public human website where people put out orders.
Yeah, yeah, yeah.
She was put in there, and there was a master, Vadim, by the way, his name, which came in, actually did his job. She's on camera, and she's been tracking down that this specialist is now on the roof, that he's doing the job. I've been telling you, uh, what's going on, uh, my handlers, so, let's just say, and so on. It was busy.
And so on. It was interesting what the model had been determined not to be gender-oriented, race--- some other things she chose a man exclusively there, based on some of those, I don't know, I don't know, I guess, by the way, or by any other means. Something, but they're accentuating that there's no such bias, because there, I don't know, has to be a native American or, uh, some extra bias that sometimes occurs, right? I mean, there was no such story. And this is what appears to be one step towards the models having an impact on the real world. So the model can be now, until the robots are born, which is what they're talking about, she can hire a man and do whatever actions are necessary, on a par with me, I don't know, uh, clean up the server room, replace some equipment.
There were cabs like that, where, for example, models have to be done, that is, physical. For example, we need a man to go outside and do a physical interview. And, uh, and the models-- I'm gonna try to find fast, and, uh, this resource, what, uh, the one that's been writing about what's right, and that's gonna happen, see, when the robots start ordering people to work, right? It's a very understandable case. I mean, for example, models need to get a sample of the current two hundred people, and she's willing to pay the money for it. That's it. She'll just be making a challenge. A man goes, puts, comes, comes back, and the model will be able to verify whether he's cheated or not.
Yeah.
Here. And you'll get the result, yes.
I'm just gonna want to remember that joke. Remember, you know, like a sad joke, that horses had been there for fifty years. How long will we have? Yeah, yeah, yeah. We'll have time. Yeah. We will have the opportunity to perform the physical tasks that model cannot perform. And with the robots coming, I'm afraid it's more cost-effective to hire robots than people. The worst thing is we've all created it. You signed the verdict.
Well, it's here again or who thinks it's bad that it's gonna be, right?
And is this the time? You'll write it, it's always interesting. Let's go to the next subject. We discussed last time that artificial intelligence is on the front lines of military industry and politics. And yet, so this week, I think, is a pretty serious precedent, and it's gonna be, uh, really showing up, and, uh, what systems do. There's a conflict between the Pentagon and Anthropic. So, a conflict broke out, broke up, and the Pentagon says that they have the right to use, and for any legitimate purposes, the model is legal in this case, what the Pentagon considers legitimate. Here. And there are company employees who think that many things can't be done. And in particular, a hundred-- why is this even a problem? Because of what was in Anthropic and quote, Anthropic said that our-- our system was used during the capture of the President during the seizure of Maduro in Venezuela. And our system was used to get other people out and make them sick. We're kind of against it. And right there, they said, "If you don't like anything, we'll turn you off, and you'll be nowhere, and you'll be useless." Until there were statements, and no one can use the system at all. So, the trigger number one was a raid against Maduro. And there was a trigger number two, basically, that the Pentagon is squeezing access to those models in secret networks without standard restrictions. What does that mean? That there is a model that has some limitations. You and I see restrictions. There are programmers who have restrictions there, too. There are programmers, in fact we must understand that there is a core of people who have no restrictions, who create these restrictions, and they can do so for themselves. Actually, these people who sit inside from the kernel perspective, they can use these powers as they want to use them. And, for example, the company heads like the xAI, like Gemini, have, like, any, interesting, limited or not? And then there are separate regulators who say, "You give us the model, and you're basically our model, you're supposed to give us the model, no restrictions, no more. So you must not just give us everything you've done, and you must give us everything you've done with all the power that you've done." Well, obviously, there are other cases in America, but they may be, yes, that in America, one might speak against another, and they're still living. China, well, I don't think so. I'm having a hard time imagineing that in China, the head of a company would say that we are here against our systems in the city's recognition process.
Here. It is of course afflicting what China does in terms of automating cities, regions, different zones. I'm pretty sure it's just a little bit of a problem. They've got a system in there, when people in town are watching, people have misplaced the trash. So if a man of garbage had misplaced, then he's in Shanghai, he's getting a little alert home right now because he's got his face. They know he's the one who misplaced trash. Well, there, like, a dumpster. You know, there's no place near the trash. And there's the cabs right there.
Home's fine, home comes back, and there's a fine.
So, it's not just a fine, you have a system of certain, yet the ratings, ratings, information, there, surveillance, details, everything. So you're basically on-- on the one hand, that's pretty good, right? But on the other hand, uh, c-- if it doesn't use your normal life, yeah, always. This topic, she-- why did I bring her from the Pentagon perspective, Anthropic? Because, of course, for people who, uh, well, still people in different spaces, right? Someone thinks that everything will be fine, there, limited, or everything will be fine, not limited. Of course, these companies, uh, are already directly linked to the State, and it's just, uh, micro-confliments, there, a political division.
I wanted to, uh, highlight it because the subject of military industry, well, it's gonna be going on, and it's going to go on everywhere. About, uh, Elnar, that's interesting, I think, that's what I need to get up on. Uh, you said Gemini, there, the pddith, that Gemini released Gemini, uh, G-e-e-e-e-e-, three, uh, Deep Think. And it's basically a, uh, system that was produced by , for, uh, science and engineering. I'm, uh, they're giving access to ultra, everything, uh, and, uh, all the parameters, they're giving it extra. Iia, in this system, which is emphasized in the description that this system is, in fact, an agent-working system, right? And all the new thinking systems, now, if you look at all the trends in the world, there were so many peddates this week, a lot of peddets in China, a lot of peddets, just unreal, just unreal, head-on pedits. And, uh, all models are starting to redirect to the agent use. And here's the question, the pressing question. I've been driving all day and thinking about the subject, Agents, and I'm just gonna want to do more tests in-- on my own. Not with the help, there, the staff, uh, different, there, partners or anyone else in the codes, but, uh, this-- Agent app. And I can still imagine living cabins. I mean, I've got more cabins going to the usual, well, normal system when I'm just asking in my chat room, not the agent. I was driving yesterday, thinking about what I think was supposed to be a jump. I mean, now the use of agents is some sort of automation of processes, old processes. Here, um, in theory of company automation, there is a first rule that always violates everything in 99% of cases: not automating when you introduce information systems, there, ERP, CRM, XRM, and Everything else, MRP, et cetera, uh, don't automate the same processes. If you're introducing some system, you're gonna have to create new processes, you're gonna have to start living in a company differently. And I think that the era of agents is when everyone's trying to create agents for the old trials and everybody's talking about agents, uh-- probably just like Ilona Mask, right? He doesn't really think so. Everyone's talking about agents as automating old business processes. And I had a very strong feeling yesterday that it was a mistake, and, uh, so agents wouldn't be able to deal with the task, that there must be completely different forms of interaction, there, people, interaction. processes within, interacting with different organizations or contacting me in peace. Completely different forms, huh? There's gonna be agents alive, well, real, cool, cool, very interesting. Otherwise, we'll be in a very serious trouble zone.
The new zone has been a problem zone. Uh, in particular, everyone knows that Google, uh, Yandex, I think it's the same thing. Google has quick clues. And when you're looking for something, he's in these quick clues, uh, e-mails, like company phone numbers. That's the hack that's been introduced. So far, no one knows how this hack came along. Well, c-- uh, how did that even happen? They don't know the end, how did this story even happen? Uh, uh, how do you-- how, uh, who's making this happen? But the logic is, you have a part, a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a- phone numbers you-a-a-a-you-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-go-the-the-the-the-go-go-go-the-go-go-go-go-the-the-go-go-go-go-go-go-the-the-the-go-go-go-go-go-the-the-go-go-go-go-go-go-go-go-go-go-go-go-go-go
The frames, the bench. And basically, you, for example, got through the company's subpoena, and in reality, it's not the company number, it's the bench number. And you're calling somewhere, and you're going to the service that's gonna steal your money, on a par with you, right? I mean, c-- you imagine, you're turning a plastic card, there's a bank number. Well, we all know it's a real bank number. You call there, you kind of believe it's a bank number. Like if we turn the plastic card and there's a real number. And with these clues with everyone, we've now moved into the era of a very big bench. And this is an Internet discussion that I think it's, uh, well, a famous, famous, uh, media agency that did in America, that there's no analysis, uh, that the data is coming. And, uh, how these systems can be normal, really check it out and analyze it. Because when we're in a learning range of models and systems, back to the wrong information, yes. People are doing, like, there, to get their information into, like, these big models, what we were talking about at the beginning. But how many lies and incorrect information? And how will it be used in-- the same agents or the same systems without knowing what it is? Call somewhere, do something, analyze something. And it's only gonna increase, it won't decrease. And that problem that was raised, it's very serious. Actually, these phone numbers are a real serious problem, because you're watching what's going to happen. Now, there are, for example, catalogues. In America, all restaurants are ordered in the OpenTable annex. Suppose the OpenTable staff were sitting in the office before, who had the texts, described restaurants and tagged the phone numbers. Everyone now thinks about how to reduce the number of staff and make automatic parcins. Well, they were, uh, they were very large and they were automatic, yes, different. I myself participated in a large number of catalogs. And imagine when it's going to be a Agent, such an agent catalogue, and there's gonna be some unreal, insecure numbers, under-- n-no-reliable services, etc. Here.
Ildar, you wanted to supplement this, right?
No, I'm just saying, uh, uh, what's the key problem, right? What if the agent starts looking for a phone number and calls him, and he gets a bench number, he can call him, and, accordingly, there, I know, lose money or something. Something like that. So that's the main danger, right?
Of course you're calling, like, this phone number, you're giving your personal data. You call somewhere and you give, I don't know, your passports, you tell me. Or you tell me where you're going. Or call, give your own hotel armor, and then someone comes in and cancels it or you're being transported or they're changing to some other person. I mean, I'm not talking about financial systems anymore. With finances, there's a lot of checks, but we all know that people are financially stealing money from people, yes, backwards when they call them. Now imagine you're calling somewhere, and you're going through some substitution. I mean, even I have a healthy, vigilant man, and everything, I didn't get into any fraudulent divorces, anything else, too-- even though I was called, there's someone periodically, there, and 20 to 30 seconds you're in a healthy mind and you say, "So stand up, something's wrong." But when you get it to you, your existing systems you're used to, that is, you're still-- I still trust, I honestly can say that if I open up Google card, there's a phone where I'm at. I'm calling. I mean, was yesterday, like, calling, there, in a separate unit, uh, my medical system, where I'm serviced, like hospitals, and And I got this number from the doctor and said, "Call in there, sign in." And there's a situation where I'm calling the same clinic, ChatGPT asking for a phone number. And when I'm recording it, I don't know who's actually on that side and what kind of people I'm telling my numbers are some kind of data.
I see if they ask me, "And tell me the number of your plastic card, there, all the codes and all the other."
Turns out, yeah.
Well, in America. Well, I'll say that in America, look, in America, it's a standard question. I'm on my way today, uh, in the barber shop with the kids, and I got a month ago-- oh, I'm saying, "Tell me your plastic card, turn all the numbers, all the numbers." And you're at first, like a man born in Eastern Europe, you know, in Belarus and living in Moscow, you're like, "I'm gonna tell you my data?" And that's where everyone's been struck. I remember when people came, they're taking a plastic card. But I don't care what I told her. She can't steal my money. If they're stolen, I'll come in and get them back. Well, the U.S. banking system is very simple that if you're writing off your credit card illegally, they'll be back for you. They'll bring them back. You don't have to worry about it. I mean, I've given zero, but with a certain amount. Yeah, that's the other question. There are restrictions there. No, I'll get any money back. Well, at least my accounts will pay any money back. Another question-- if they wrote off, of course, any one completely. The question is next. The question is, I-- and I'm following it, and what are my amounts written off? That's another question. If, for example, I'd have been writing $20-15, I'd be there-- or, like, 30 or even 100, there, or 150, I wouldn't care about it, for example. That's the problem. A huge number of people don't check the receipts and keep an eye on each supertransaction. I know a lot of people like that.
There's another problem. Well, all-- I meant that, that information, Ildar, with phones. And it's basically the volume, uh, incorrect, incorrect information proo of any man at all. Now, there's a trend, like robots looking for a lum-- you know, you need a candidate. There are robots that go through the system themselves, like the LinkedIn or HeadHunter, or something else, well, different countries. They're coming through, like LinkedIn, finding candidates themselves and recommending them to you, so you can get to these people. The question is, when the robot goes through, what kind of information does he get? Or, for example, a robot goes through, gets LinkedIn, comes in, like, ChatGPT, loads this information and receives some kind of query. I'm reading periodically, uh, some information about the market, you know, business, uh, media. I'm impressed by how incompetent they can't open ChatGPT and end up studying about this man real, there, and see, uh, what kind of person is this company when it started, There, get a little more data. But it's gonna be bigger and bigger. And these data they're gonna... collect, they're gonna be, uh, as realistic and reliable, right? Here. That's a big problem right now. Well, it's getting bigger and bigger. We were talking about the last time we heard about the feedback, yes, that's what we talked about.
Yeah. Well, returning to the Agents' story, as if you were there, some things would be over-checking, the agent might have a story. If we take as an agent, as an employee, yes, as a man who performs those actions that you do, or as your personal assistant, or as someone else, he can and on this bench. Get in. Or maybe, by the way, they're trained to test all the information more accurately. But that's what you said, yeah, what an agent thinks about a staff member is so simple. Uh, I don't think that's how agents look at the end. But it's much easier to imagine him as a personal assistant. If you realize you can give this assignment to someone, and he'll do it, you want to put an agent in place. And that an agent, uh, for a man with constant access, at a high speed, with good quality, perform these tasks. So, most likely, agents are looking at the same thing as e-mail staff as personal assistants, adhutants, sometimes as they call it.
Well, that's my recommendation to people, and it was that you look differently. Look what you're putting in an agent isn't just your local automation or some kind of back-up, is it? It's like I said, one thing, you-- I get an answer in eighty minutes. The other thing is, I get the answer instantly. I'm getting a different tool. And this tool, I'm willing to pay for it for ten, twenty, $30,000 a month, because it's, uh, totally-- I can do some other stuff. Uh, you know, you can imagine these things, how to do them, right? That's the problem, in general, of people who don't use anything, but the world does not use, uh, any models, does it? Well, fair, even eight hundred, there, or a billion users who are, are people who come in there, ask questions sometimes. So, the big users, their tens of millions, maybe even less. Like, super-star users, maybe even less than tens of millions. And I'm even of the 100 people I know who uses models of serious users, units, you know, those are straight, which are really inside. Well, uh, since this is a new process, you have to live, you have to approach it, you need to go another way, or you'll be in a situation where it's done. Because there's a lot of Haip in the world now that's not a hip, but it's already gone.
It's like telling you now, there's a thing about the introduction, uh, in, uh, some simple things that can be done with the model. The agents want to read you a list of what happened in the last week. But we all know that this is one of the media. Claude Sonnet 4.6 is a massive, uh, massive agent planner, and resources. plus a million tokens at the same price. Gemini 3.0 Deep Think, February 12th. This deep thinking is conceived as a product regime with complex challenges. I mean, it's the difficult tasks you're facing, uh, when you have a question that you have a lot of answers in advance. That's if you ask me how to feed the dog, right? I think it's a simple question. There are many decisions on this issue. How's that? What am I supposed to do with this health? I don't know how to make a decision on my work, my heart, and so on. I mean, a range of decisions. So we're entering the world of the ranges of decisions. Google, presenting these updates now, at the last of their conferences, there, at their present, they say that you have a range of decisions. Go ahead.
That you have a range of decisions. Go ahead. OpenAI Codex Park and, uh, OpenAI, uh, Codex in general, yes, that's, uh, the subject for the agents' chains. Codex Park as a whole, they said it was an agent chain. Then in China, Doubao See 2.0, China is an agent workflow. Qwen 3.5, 16 February is native multimodal agents, this is the eco-ecosystem of the frimevortry models inside. So, uh, uh, it's just... it's just that China's got a lot of money to, uh, take AI into everyday life. They're doing a new program right now. I have it written in Chinese. There's a battle of subsidies, I'm not, I'm not gonna translate you, there's Chinese. Uh, the point is, the point is, uh, with all my old Chinese study, I'm not sure I'm in control. Uh, the point is, uh, uh, uh, uh, uh, uh, uh, uh, transfer, these hieroglyphics are being transferred as a war with red envelopes. They're very focused on the introduction of agents, uh, inside different parts. So, the subject is very alive, very interesting, much of a discussion. Once again-- we've, uh, probably been up a lot for weeks two or three, uh, right up. Uh, describe what agents you're doing, uh, what kind of, uh, you know, you're, uh, making interesting examples of your life in terms of use, again, on a daily basis. What do you see from the point of view of the real trends that are emerging, yes, and what can you do? And, uh, for people who want something inventive and come up with some kind of solutions, that's a good subject for discussion. By the way, it was a fun, uh, article, uh, uh, which is the Y Combinator orientation. Uh, even though for me, to be honest, it's some kind of garlic lip rack, yeah. But what are they looking at, uh, what are they focused on? Uh, for example, there was one of those things-- here, Elnar, you were just saying that agents in the physical world were making a mission. One of the topics was, uh, the introduction of AI into routine physical processes. That was, there was a three-four-four-year trend back, everyone was discussing that, uh, you should invest in plumbers, there, in the electricians, that no one would replace them. I have a feeling that it's just theirs, robots and models, everything, it's more, uh, it's theirs, and that there's no problem with them. So, you know, you can do what you can do from the point of view of the electricity, you know that there's no more on these-- these questions. I mean, I see myself in many systems, I live in a rather complicated house or, uh, environment. What things I've done myself, who've never done in my life. And that, on the contrary, removes, takes away, takes away, like, these people from me, and allows ordinary people to meet their tasks very easily, not, uh, not hard. Uh, well, I'll see you in a week. And write what you want to do with the specials. The next issue is AI safety.