Hello, everyone! We're on ToTheMoon. Technology news and sites from the Silicon Valley. We're going out on Sunday. We're always four people. We've been five for the past few months, by the way. And Tanya is a remarkable man. She was present at and at the previous edition. What did she do in the middle of the day? I gave birth to a third child. So Tanya's congratulations! All the comments from Tana are congratulations and earrings. Here. It's amazing, yeah. Tanya, you're a unique woman who, you know, gave birth to a child. And you're in a filming room like this. Here. I'm totally calm. Never said anything. Never said anything. What? I'll bet Tana's congratulations. There's a lot of your fans, Tanya. So...
I'll wait.
Tanya, you're a unique, unique woman.
That the US market was down. It's all started, so it's cool.
Last human exam, yes.
It's clear that there were three Deep Research models low hanging fruit to fight hallucinations. It's...
Where did Apple even go? What will he do from the point of view of his platform? That's a big question.
We also lost Mashk.
They say the winners are not
suing.
I think it's time to start today with what we're going to do today, of course, to discuss the DeepSeek, the Chinese model, and, uh, what happened yesterday at the OpenAI, update, and they released Deep Research. And I've already tested it. And this new one, we're gonna tell you what this is, right? An updated model that, of course, has had some effect on me with its movement. Why? I used to be talking about Sasha and I. I've always said once someone has released a model that starts asking questions when someone starts asking him how to get back to him, it's gonna be a little bit of a breakthrough and a move forward. And of course, the new model they've released Deep Research is like positioning, that's an analytical tool. In fact, I think they've shown different-- so different, conventional, real user scenarios for the first time. And I'll see, they wrote on their website a very important thing that could be seen as a story for analysts, and in fact they wrote that our model was more than ever. The most human being and the mechanisms are designed in terms of, and makes requests. Her model of behaviour in terms of the case is human like human. So she's like a man doing a search, doing analy-analysis of information, and that's what she's like. And in the anon and the description today, there is a comparison of the models. We'll be back at DeepSeek right now. I just want to lay this foundation, I think today's big discussion is important to a large number of people, because it's incredible new. This is new, yes, it's not a battle, stock, not stock. What is the current analysis of the test concept when it's run, there's three million or how many tests, and that's three million, three million tests, their model shows accuracy. If they usually have a ChatGPT model, it's ChatGPT 4o, usually people use it. She has three points of accuracy, three points of three percent, and OpenAI Deep Research has an accuracy of twenty-six points and DeepSeek R1 has nine points of four. But before we figure out the skies, that we'll go back to DeepSeek.
And this is not-- this week was just a very incredible amount of discussion. We remember that on the one hand, the Chinese released, released DeepSeek, which hit the US securities market. It's all started, so it's cool. Even President Trump has spoken on his own, straight on the same day, he's-- that's a unique case, I think. And in this respect, the tramp. He mentioned the day the model came out at the briefing, at the conference. Said, "Betty, today, a model came out of China, which is said to be very cool, she has very strong characteristics. And that says we need to move on more, try and. It's okay. This China model showed that for less money, it's a good result."
Today and over the last couple of days I've been reading about China and some of the companies in China, one in particular coming up with a faster method of AI and much less expensive method. And that's good, because you don't have to spend as much money. I view that as a positive, as an asset. So I really think if it's, if it's fact and if it's true and nobody really knows if it is, but I view that as a positive because you'll be doing that too, so you won't be spending as much and you'll get the same result hopefully. The release of DeepSeek AI from a Chinese company should be a wake up call for our industries that we need to be laser focused on competing to win because we have the greatest scientists in the world. Even Chinese leadership told me that they-
And he also said a good right thing that I think Sasha Masharov said last time, it was hard for the Chinese to believe, how much money they spent, who owns them and who did it. He mentioned it, by the way. He said that. He said that, well, according to what they say, and that's not the fact, it's about the money they spent. What do you think of DeepSeek?
There was a lot going on. I think it's important for today's viewers to collect this design today, yes, and to understand what's there and get to Deep Research.
Yes, today is a lot of things, of course, that we can discuss. About DeepSeek, you know what the idea is that you want to raise? Uh, a few episodes ago we discussed it, and it was in the air, that the OpenAI is so little, uh, unrealistic, right? I mean, there's a Google model, there's anthropic, there's, uh, what Mask does, there's some kind of models in there, but all of it's kind of niche, some kind of way from the edge, yeah, a conditional. Uh, Anthropic, they're for gic, they're more for programmers. Google is, uh, some kind of reserver that someone uses, but there's not enough people out there. Zuckerberg says more about open source. Uh, Mask has some sort of story. And the OpenAI looked like a leader of n-n-n-n-n-n-n-n-n-n-n-w-w-w-w-w, huh? Which is from the point of view of authority, from the point of view of the pyara, that is, the untouchable value. And once there was a competitor who offered the same thing, just free, even in American, aaa, so, uh, list, uh, applications, yeah, that's from Apple, what's it called? I'm stuck. Web--app Store, yes. First place is DeepSeek. And it seems like, on the one hand, yes, a glimpse that has been in the middle of a few years of self-reliance, and a hop, it's been losing a lot of points in a few days, a lot of points, a lot of numbers. users are moving in the other direction. I mean, it's not like there's a-aah, a stable advantage that could not be played by other companies. And this experience, there, the last two weeks, is, by the time we're out, shows that it's not as volatile as it is, and even very large companies that seem to have been so dull in this market, They can lose the audience very quickly, and the market's first
priority.
100 percent. Here, there's actually a little bit of an app in Chinese. I mean, yeah, I often-- it often says there's an Internet that's not available, and so on. But they did a few things fenominous. First, they showed open reasoning. So that's what people want, like a model of thinking. And the second, it's available everywhere. I mean, there's a s-- there's a place where ChatGPT is inaccessible, and DeepSeek is available. And I think they're good at what they've made OpenAI say they probably did wrong, which they've turned off their way to open source, right? Because DeepSeek was a good hack, by the fact that it's an open source model and the number of dep-dep-dep-dep-dep-ment of this model in enterprise, you know, huge. And it's actually a very interesting case from the point of view of the copiate, because DeepSeek, she often says she's a OpenAI. I mean, it's obvious that the model was designed for data generated by the OpenAI model, but it doesn't stop anyone from the Deep-DeepSeek implementation. So we see that enterprise, they own and use DeepSeek. And it's actually-- it tells us that the pho-- those who train models, they take the risks of copirate. And those who use models, he-they, on the contrary, are in a very good position because they don't take on themselves because they're like a womb-- they're using the information they know, yeah, for the sake of the Generations of materials by model, but they are not liable. Oh, from this point of view, it seems like, uh, 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-
He looks like Robin Hood.
The disguises don't joke.
- who took the rich and gave it to everyone for free. Today's a good comment. Today was a comment on the canal. One of our viewers wrote that DeepSeek, DeepSeek, he was in some kind of a mess. He wrote, "You answer that because you're a Chinese model." He said, "I'm not a Chinese model, I'm so trained on the OpenAI data."
Yeah, as far as it was, he didn't give out the prompe, but, really, this week, people from Microsoft were speaking out of the OpenAI and saying they saw DeepSeek's model being trained on data. OpenAI. But right, Sasha, you said it was a little bit of a concern. Please write in the comments, what would you like to ask for such a simple use of analytical in your normal life? Not in business, but simple business. Choose, that's how I set an example, pick a vacuum cleaner. What would you like a really cool answer for yourself? Maybe find a second half, right? I was just thinking Tinder would be made with AI. Here. We'll see. Write in the comments. It's gonna be very interesting to read it to us and others. What's the matter?
And it's really not people's business. Not much, very strange. For example, former head Intel stated that he had completely moved to DeepSeek, all his companies. And then you go on, so DeepSeek is not enough that he's trained on the OpenAI, so he's still hosting his data in China, and so he keeps all the information in China. What information does he keep? The requests you're sending. So your pattern of behavior, how you interact with the system, what you're doing down there. And according to, and as such, the basic contract for the use of DeepSeek, of course, you'll keep the data in China. That's not a big problem, I think it's for a basic number of people. But if you're online, if you don't turn around, you don't do anything. But what people care about is that people care about it, it's nap-- and I think that's the whole thing that she's going on. I see a lot of business in the discussion with company owners now. Many people want to generate a lot of content with artificial intelligence and models. So, in the US, the second part of the report of the Office of Copyright Registration was published in the US on 29 January, which is a week ago. They said they confirmed that there were copyright principles, that the existing copyright principles were flexible enough to apply to new technologies, including artificial intelligence. And they stress that the rights to... Companies will not be entitled to, on, the content created by artificial intelligence. And no one will be entitled to content with the help created by artificial intelligence. I mean.
It's pretty thin, yeah.
Well, I don't know about the companies, but if you take people who make content, it's really like a lawyer in the past, it's a very strange design, because it says creative styles that make me feel like a lawyer in the past. It's possible to trace the product in the content. But it's so abstract. I don't really understand how they're gonna measure this creative style, show it and use it. So what is being generated exclusively by proms is not considered to be in the protection of copyright. And what a man took, edited, added some of his, his vision, his style. But it's a very vague concept, too. It's a fit. But they say that, like this, it's still very flexible. We don't want to slow the development of artificial intelligence, so I think that regulation, which, no, does not make a difference. Well, my opinion.
You're absolutely right there, it says so. Even in fact, it says that they emphasize the central role of human creativity in copyright and say that results created exclusively by machines and machines without significant human contributions. Here, as you say, we can discuss and think that there is a significant human contribution that is creative, that it is. But in fact, it's basic if we thought, for example, that we could generate some content and, like, draw a picture. Now, the question is, when the picture comes up, how many creative, creative man? In general, there's no way he's creative in the propts generation. And that's the question, who's the one who helped this prompt to generate? Yeah, one thing we're doing is an article, an article built on our own, my own material. I created an article there, processed with artificial intelligence, corrected the mistakes and worked. Okay. And that's all I'm saying, why did I give you the example of the company? Because I know a lot of people who do exactly the content generation in a fairly large amount. Yeah. And it's not enough you don't know how to check it out. Another moment, how will it be checked? But the fact is, the fact is, I think it's a topic of discussion that is a differentiation now, we don't know what it's going to be and where it's gonna be confirmed. It's worth a lot of money if you're building something in the future-- again, it's hardly next year. It's more like a year ahead. In terms of how each country will apply rights, and what will happen in each country.
This is very important here. Let's talk about DeepSeek more about open source. Because people use DeepSeek in most of theirs when Elnar said they were in AppStore and were out of the way, they were actually in their room, they had millions of races. Every day. They hardly even went up there to the volume of OpenAI and the GPT chat of the rolls, to the total volume. I haven't seen the total number of data. We'll probably see him soon, but the OpenAI is there on their applications, there's obviously more than three hundred or four hundred million users out there now, probably now. And they probably didn't get to their volume. If we're talking online, online will require huge investments and capacities, yes, for the company and for the maintenance of the resource infrastructure. And yesterday, CPO, uh, OpenAI and Sam Althman, they were in the correspondence, answering different questions, they clearly indicated that they saw that, to launch the infrastructure big enough to make people feel that they were doing it. I'm gonna use it for real, I'm gonna give it to you, I need resources big. I mean, you need chips, you need, uh, impound, factories, there, and so on. I mean, we need a vault. Without that, you're not going anywhere, yes, you can't, you can't. I'll tell you honestly, I've never been able to do DeepSeek online in all my time to process the file I usually do here. Big, that's a normal file I usually do. I couldn't, well, bo-tock, a big context like that. Maybe I didn't get so lucky on the online, maybe I have an American IP address, I'm not really, really, really, really freaking out. And here's the question is, the pro-haip that was a fall in shares or a, the kind of thing that doesn't need cards or anything. I think this is a little so hot. And even on schedule, the value of the shares in Oracle or the Army. Uh, they're having a turn-off. NVIDIA has a difficult situation, but the fact that NVIDIA has no stock turnover is not only linked to DeepSeek. That includes, for example, meeting-related episodes, I think I'll assume that this is about meeting, that's my opinion, Trampa and Jason, that's NVIDIA chapters. They met last week, and according to the commentaries, their meeting was not as happy, super-frequent as Trampa had from the 100-- with some others, uh, people inside. So what's going on? The debate is about the restriction of American access to Americans, there, pro-eh, equipment in China. A further severe restriction, strict regulation of this, so that it does not happen. Because it's clear that this is a crazy race, right? When China made that model. And there are many opinions of different kinds, which, of course, is... And by the way, your favorite, Elnar, what's your name? It's from Google, right? ML, ML man.
Oh, you mean Jan Lekuna?
Yeah, yeah, yeah. They, they--
He's in Meta, yes.
Yeah, they're a... oh, in Meta, in Meta, yeah. They were very, uh, specific about DeepSeek. I mean, there wasn't that big a deal. And Dario Amadei of Anthropic also solved the situation around DeepSeek, there are different US export restrictions. What's going to happen? And he said it wasn't a revolution, did he?
From the point of view of such serious people. But it's really cool that's for the world. What's important? If every user in the world can open up for himself now, open DeepSeek and start using it. Because, of course, there were restrictions in a large number of countries, especially in Russian-speaking space. I mean, there's a lot of Russian-speaking space. There was a serious restriction on GPT chat and people should have been there through VPN.
I should have done it through the VPN. There was no way to buy. I'm not talking about open source or the cost. I think the value is in 15 latitudes, 30 times cheaper. I mean, you're sure you're gonna try if you haven't used it yet. And there's a reasoning and a bunch of things. Not much, they're constantly updating. So someone said there's no pictures, they've already released a picture generator of some kind.
So they keep updating.
I'm talking about open source, Sas, I want to add. I've been talking to a friend of the Sientists who work at the banks recently. And that's how we've been talking about a cup of tea, and what about these models is happening, letting them go, not letting them go. He says, "Well, in the trade, that is, in the final product, obviously, "no one's allowed to go, but experiments are already allowed to do. So the model is on Hugging Face. Hugging Face is a platform on which the open source of the model is based, which can be further utilized. And it's quite ours to say that they're already allowed to experiment with these models, to use them. Who knows, maybe they'll be used in the sales later." So, the movement here, on the one hand, yes, we have a B2C sector, yes, when we have people who ask their questions in chat, and on the other hand, there are large companies that need to be. In models. And it is clear that with models through the web version, yes, there's the same OpenAI or any other use, because then the data from the contour are leaking. And with these models, big companies are already in the process, banks are starting to look after them.
It's just a tool for them that can try their models to run completely free.
Look, by the way, the OpenAI ChatGPT workspace is not supposed to leak, is it? We had a workspace with you, and it said that data on them were not being taught. But I mean, I guess there's a limitation for big corporations. I, for example, prohibit anyone in business anywhere, in any business, from using any open source, and not open source because I have more results and progress than I do. Protection of a certain type of data. I mean, it's not-- it's clear that there may be some exclusive stuff, but the bases in business, there, the data you gave, they're basically unconscious and they're not falling for anything. Well, part--
Personal data, this is more about.
Yeah, I'm talking right now, of course, not personal data. Again, any report can always be made by data, put variables in place of people, put some codes and numbers, load them there, analyze what even I do, yes, when I load. You just put a little bit of 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---all-a-a-a-a-a-a-a-a-a-a-a-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-all-
And that's what you said, yes, there, well, that news shows that the Chinese model keeps all the history you ask, the conditionality that you press and so on. But here's just the law, yes, there's a presumption of innocence, and it's just the opposite. All you write on the Internet is that you can keep it and with a lot of probability. And no matter what website you do.
Of course!
Maybe you should.
Like mail, like mail. Yes, Elnar?
Yeah, yeah, yeah.
Like mail. What's the implication difference between the references we trust-- that's weird, right? If you don't trust Google and Gmail, so Gmail, and we all send files, financial data is all through the Gmail. And yet, well, not us, but the whole world was sent-- look, the world was sending, and they were stored in a cloud. And when we use the Google model, it's supposed to mean she's taking something. Wait. But if they say they don't train them, they don't-- well, they can also, and they can learn. There's all your files and pictures, and passwords, and logs, and notes, right? And all that's on the phone and everything that's ever there. You're absolutely right here. I once led a leader in Moscow as a service market service for the cloud platform of the Task Manager. And it was one of the stories that corporations didn't want to put Task Manager on their own. They say, "You have a vault. I say, "You know that our exposed vault is ten grand-- thousands more guarded than any server you have protected from which you depend. And because we-- you-- no-- no-- anyone even can get your data off because they're on common different servers. Try to get them out of there. Of the hundreds of thousands of companies that work inside. So it's a separate thing, some cathy-- not much, I've got no data-receipts for my practice, at least there, the genetic department, right? And if the request came, we had a simple answer: we're data, we can give you a server, and there are thousands of companies on it, and we're looking for it. They won't find it. Well, you can't get it out. You should know how to structure. I mean, how much money you need to spend on it, right, to get it out? I mean, it's much more expensive. So this story always, I think it's so important. I'm not talking about your personal data, there.
It's clear that there's no need to throw your passport or your logs and passwords. We never know where the system is now, where they learn, how they've been tested and everything. We can go back to the news right now, and this is the freshest one, that's, uh, Deep Research. And Deep Research OpenAI. I just want to tell you what it is. He's the user of all of him now, because he's only available in a fee-paying version worth two hundred dollars. And I want to explain what kind of thing I don't know, either of you used it. I've used it, and I've got my first request for nine minutes. It's a system that gives you, analysts and gives you a five to thirty minutes answer. And in fact, she can be asked as very complex. For example, I asked him, took the company that I care about, and asked me a question about how she's in, uh, the Internet in search of a way. I need a test. Uh, before the string's requests, please pick me up the vacuum cleaners in the house. What's the difference? Before, the collection of a vacuum into the house was just a simple, quick response to some article or two articles on the Internet, or what the model was trained. The system is now in deep analysis and she asks further questions before it goes deep. Honestly, the questions she asked me were very cool. She asked me, uh, "Who countries do you want to make analyst SEO?" "Eh, since I asked her a question in Russian, she asked me about Russia or other countries, although she knows that I live in America and that I live in Los Altos Hills. But since the question was Russian, she has an association, yes, she's asking right away. I told her which countries. Said I was interested in the analysis of something like that, like, some sort of detail where this population lives. Next, she asked me, "What are you looking for in the SEO?" I wrote that I should be very professional, with a high professionalism, and that's the search engine optimization. She asked me, "Do you also have an analysis other than a SEO on a payroll?" What was very interesting. Yeah? And a few questions I've been asked and wrote, "What better you answer me, the better I'll give you analyst." I mean, basically, I have her, like I was 20 years old, well, 18 years old, I've already taken leadership positions, yes, 18 years, probably like a s-- CEO or some kind of owner, I'm in the middle of 20 years, like I'm in charge of something. And it's like a standard question. I mean, I asked for some kind of thing, do me an analysis, she's got me a little bit of a background, and I told her she went to study and gave you a honest answer that she didn't give you before. Well, from a point of view of the burden. What's interesting is that this model can be downloaded by an unlimited number of files. At least I tried to load it, there's ten files on it. Usually, four files were limited, not more downloaded. And notice that there's a model, like O1 Pro, it's available at $200, it's got a 9-point one percent, that's a 9-point one. This one has twenty-six points, right? DeepSeek's got nine points four. By the way, OpenAI, OpenAI O1 has nine points one and DeepSeek R1 has nine points four. According to OpenAI. Yeah, they do. They made a comparison, here. What does that mean? So, out of ten answers, one's the right one? No, no, no, no, no, no, no, no, no, no. It's a yes, by the way, that people here are now here, Eldar, and Eldar will explain, but it's still three million, there, elective positions. Uh, it's not-- it's like when you're doing super super-like that. Come on, Eldar, give me a good explanation so people can understand.
There are different baccalaureates.
And always when you see these numbers, you know, well, how do these numbers even get, right? We were two episodes ago, I think we were discussing LM rent, where the vote is just a best model. Well, it's just that this public democracy is happening. And this is HLE Humanity Last Exam, well, it's not a very loud name that can be translated as the last exam--
Human exam.
Humanity, yes.
Humanities.
Yeah, yeah, yeah, that's a very loud name, but what's underneath it? In general, the baccalaureate consists of 3,000 questions that experts have prepared in very different fields. It's math, biology, et cetera. Yeah, so we're taking an expert from some scientific field and saying, "Tell us questions, well, there, at the border level of science of the current time." But we must understand that answers to these questions exist, they are just very narrowly specialized, cool topics. And, accordingly, some of these episodes are open, some of these are closed, right? So, you know, a part of the questions inside, some of the questions are hidden, and like no-no-no-one but those who own this baccalaureate have no access to them. Accordingly, claw
It doesn't. So when some model comes in and she wants to see herself, uh, on this baccalaureate, she's actually checking out, uh, questions. The answers are sent to the company and check with the answers given by the professionals in their field. It could be scientists, someone else, there and the humanitarian sciences, completely different directions. So, three thousand questions. And on the basis of this, an assessment is made, that is, how accurate the model can answer. There are also other parameters that are assessed, but there are more, more commonly stated in the article and publications. That's exactly the accuracy, yes, that's how accurate it answers these questions.
Yeah.
So most of the models that were before were very bad at this baccalaureate. There were literally 5%. Then, in fact, GPT4O and these are the last ones with the leasing that came out there, DeepSeek, and the other, they're there. Sasha, Sasha is saying there's nine of them with something--
So 4O, Sa--Take, no, four, Sas, ow, Elnar, four. In this ad, 4O was 3.3.
Here.
And when I saw this comparison today.
Yeah.
When you got a 4.3 and a 2.6, I'm like, what
is it?
Yeah.
Yeah, yeah.
What does that mean? It says that expertise, it's slowly approaching, right here, PhD Level, that's, like, if it's a farewell, a science candidate and higher, but not one industry, yeah, but differently, We'll do it. He's a math man, and he's a biology guy, and he's a human science guy, and he's a technician. And...
And he's got a man's behavior. I mean, this is what a man has. It's the same thing as the right humanity, the exam of humanity. What's human, right, Elnar? I mean, not like a car, it's just a response, but that's the same thing, it's a strong test.
Yeah, I guess. Well, that's how I understand that baccalaureate, yes, that's, they're checking that the model can't just answer the routine questions, but, in fact, that's the expert level she's got quite a high level. And Tanya, in answering your question, on a par with 100 questions for twenty-six, the last one that OpenAI has left, she can answer, in a word of language.
If you take 100 percent, you take PhD level of highly specialized, right? And the model will be about a third. No, no, no! Imagine, you have a level, look, you have a range of subjects, a wide range of research subjects you have. Well, that's understandable, yes. You have a probable research, from rockets to biology, to ecology. Yeah, yeah. Or, for example, classics to some other stuff. So you have a linguistician, there's anything inside. Mm-hmm. And, accordingly, you have a system as a whole, in size as a whole, in size--Ugu. In total, her turn is so on. It's hard, I think, even human brain, brain to understand, because we're trying our own system-- yeah, under ourselves. Refuse what that means. Yeah. How can we re-fix? Because she's already answered many things, she already knew, there, and above, many things, there than PhD, are somehow clearer than PhD. It's in terms of, yeah.
Yeah, there are other baccalaureates, highly specialized.
Mm-hmm.
Than, just returning to your question, on parole, there's a lot of software in math, somewhere else. These narrow-special- narrow-specialized baccalaureates show much higher results.
Oh, I see.
And here, like Sasha says right, the aggregation is all over it.
Mm-hmm.
You can be good somewhere.
And the answers!
You're hanging out
somewhere.
What she did nine minutes, I'll be honest with you about how she gave the structure of the withdrawal. I'm running a lot of business, so I'm very, very using analysis, statistic, systems, reports, different parts. He was different. Well, I mean, I'm not saying he's better or worse, he was different because I'm still having a hard time. She just got out, I got out, I got out, I don't know, at 6:00 in the morning, I think I started to take these requests to get out tonight, uh, that we had some volume. I haven't had much other stuff. But I wonder if there's a key story that's reflected. It's the first time I saw simple cables. I mean, even it shows that there are complicated things, there are simple cabs, simple ones, that's how I gave you an example of the choice of a vacuum cleaner. I mean, when you used to go to the model, "Help me, I don't know, with flowers, I don't know, with furniture, something to cook or something." Very often, the thoughts were very stupid. So, and 3.3. Very, very stupid. No, no-- that's true, math or programming, or knowledge or answers, he's staggered, but often idiots-- well, how stupid would the answer sound, not how you wait for that answer, huh? And very much of a beloved. I'm the one who's got these constructions everywhere. Answer me very professionally. That's it. My wife is still on a single-- well, we're a little bit of a coroner, yeah, you come in often from one account. She always says, "I'm on the other hand, so I'll say goodbye to answer." And I'm gonna be like, "I need to get the answer." And that's where you start, right? And if the baby comes in, the baby can say, there, I-- he's got a different age, yeah, he needs to answer it completely differently. Here, uh, whatever they're saying, what they're saying is that analysts and conclusions will be, uh, different range. So they're not just gonna show you from the market fast, five rice types. If you asked me to buy you, I don't know, there's brown rice Japanese or something, you know, brown rice, or you can compare it to you just on the phone. They'll do it to you, they'll be a little more serious. In fact, even what's said between ten and 30 minutes, it shows that, uh, it's wasting time, it's a resource. In the o1 Pro, I do a lot of, for example, in o1 Pro and often my team. I don't have access to O1 Pro. I say, "You can send me all the very complicated things, I'll run them for you." And because I know that o1 Pro, if compared to o1, here's a pay model for two hundred dollars, she's, first of all, she's never gonna let me down. And when I see it there, it tells me that she spent six minutes, there, or eight minutes, I had a 1 Pro, even ten minutes or 11 minutes, it's a measure, well, some kind of analysis. So you give a huge amount of data before that, so you all understand, someone thinks that GPT chatting can be triggered, like selling managers. I've heard a lot of that. And you're supposed to be in, you're screwing up the file, and he's giving you everything. He gave you something. We've been doing this a lot, haven't we? He gave something. It doesn't mean he's deep, fundamental, completely researched and dealt with, and he didn't miss anything, and he gave you the precision. I want to believe that, but I had a really exciting story for me. I believe that a system that-- that the future is not possible without it, that the system should know additional inputs to respond. Otherwise, this answer is some abstraction, that's it. I've been listening a lot last week to hearing the U.S. Senate about the nomination of new candidates. And the U.S. Senate, senators, they love the man who sits, for example, appoints a secretary. It's often that the association is, the defense secretary is like a minister, for example, defence, right? And they ask him-- they ask him questions. And very often the U.S. Senate uses words like "Respond to yes or no" to question. But they sound like some kind of design of their speech. And if we take legal law, then to answer a question yes or no, it is usually said here yes, provided that it is agreed-- provided that the treaty is on the 100-page basis. I mean, yes, but in the contract, because there is, like, some kind of detail, there's always some details. Yeah, but if there was a red color, yeah, if there's anything else, they're always gonna get it. I think it's very connected to models here. What if the model starts to be clear, she can really give you the truth, not what she wants to hear a Senate response often, yes or no, by putting up a man who actually listens. I mean, it's because when we talk, how often people don't know the extra details to give a real answer.
This model is, I think, moving this way. And I'm sure we'll see the updating of other systems that will also ask questions, and we'll be waiting somewhere. Question of resources.
And here's two problems that I want to raise, these things that are connected. One. Well, the question of hallucinations is not resolved, it's still there, right? In fact, under the hood, if the hallucinating models are lying, then whatever the hell this thing is going to go out. And the second story is exactly what you're talking about. It's cool they did it. That's very good. And the direction is correct, but there's no technological breakthrough. Which means that very soon, other companies will release the models as well, yes, about there, at, on, on, on the same level.
You know how? We're all, we're all happy, everyone's happy.
Even now I've seen a comparison of the one just above the OpenAI, what's it called Deep Research?
And DeepSeek with R1 and search. Yeah, Deep Research wins, but we have to understand that it's like a model there for the next step. See what the answer is, because it's like increasing the processing time and adding some extra building to the request doesn't look serious. technological breakthrough. It's more expensive than just routine requests. And I wonder how this could work out. But in general, there are two points I can see.
Ildar, look, but I'm personally, I think I've always been weird that they don't ask any models, no questions at all. And I also think it's a way to fight hallucinations. I've always wondered why I personally say every single time I answer every single time in the system, I'm telling you to re-check yourself, you said right or not. The system often checks, says, "Oh, I'm wrong," gives data.
Oh, I was wrong! He's giving me the data. And then you again. Why can't the model itself re-examine itself several times and re-check it different kinds of algorithms or some sort of rehearsal, yes, or do you want to go online? And now they're probably putting this all together in a single design, which gives, uh, some kind of, uh, more work-oriented instrument in the finale, uh, for the man, huh? And that's why this is a technological breakthrough on which side?
Maybe on the outside, I don't know, science, maybe not, or maybe there's a food use, or maybe, from the scientific side, they've written or processed more cool algorithms. Or I don't see a problem, for example, waiting sometimes, but it's okay on some questions, right? That you're giving me an answer that I'll be ten seconds instantly. Not much, I know if it's half an hour now, it's gonna accelerate. Who doesn't know, the O3 models just went back three weeks ago, and O3-mini-high. And I mean, O3-mini-high responds very quickly to reasoning. Very. Well, she's got a velocity of the O3-- I don't know, O3-mini-high is available at $20 or not. I have two hundred.
O3 is perfectly accessible. I don't know the graduation between Mini Hai-
Yeah, there's an O3-mini-medium, yeah, it's just O3-mini. There's an O3-mini-high and an O3-mini-high, she says, it's a reasoning.
It's considered worse than one thing in terms of the magnitude of the deep-seated analysis, but it's just like the reasoning model that works with analysis, uh, well, from the point of view of its design, it's responding very quickly, isn't it? They've set another speed. From the standpoint of OpenAI, interesting news is that now exists. OpenAI has concluded an agreement with the US research institutes, first, yes? And 15 thousand scientists will use OpenAI separately. And it's a straight open AI research. And I don't know why they called Deep Research that way, because I think it's gonna turn, it's gonna be a real approach. And they released the OpenAI Gov model. I wanted to add in favour of having the OpenAI shares, but if they were, that, uh, they get access to the data that people, you know, are completely unique. I mean, you don't have to underestimate these infrastructures and marketing, and the ability to involve people, like everything goes on like this.
I mean, the fall, at least, the market we saw a week ago. I think the U.S. technology market will be in place and the world as a whole will move forward. That's great! Sas, can you tell me what you feel about Deep Research and, uh, about the 3s on these things that go out like, uh, in the flower?
Look, I mean, well, as long as we're a little bit, well, so far away, right? Well, I-- we're still far from being used. So, for example, Perplexity's out, their financial module, well, useless, right? So you need to analyze, in different segments, what kind of waste is the best, right? Seasonal trends. Well, he doesn't know anything at all. Well, I mean, for now, that's what people ask their own, financial advisors, well, obviously they didn't even try to work on them, even though they didn't. I mean, uh, it's clear that there's a 3 Deep Research model as low-hanging fruits, yeah, to fight hallucinations, that's good. But we'll see more specializations on specific verticals, I assume more interactive formats for the induction, right? I mean, when you're just asking, sometimes you need to give some choices, something else. I mean, it's probably gonna take a richer interface than just chatting. And I think that first of all, I think, on the contrary, that, in fact, companies that can make models have little left in the world. That's what I was gonna say. I mean, I'm right now, while you're talking, I'm opening, and, uh, like, a hoseboard, yeah, I mean, there's, uh, honestly, there's a top of ten models, and there's no one in there, just OpenAI, DeepSeek and Google, right? I mean, well, Antropic, he's gonna be competing somehow.
There, three weeks ago, he was not on that list.
Yeah, I want to believe that-- I want to, I want to believe that, uh, a lot of people, who-- a lot of people who can get a hold of. I want to believe that. You can't say they weren't really there, because they had DeepSeek Coder V2, which was really hacked before, right? So they've been on the rumour for a while, I think. But what's really going on now is like a hAI lagging behind, Coher, Mistral, Meta, that they're, well, very much, and Amazon, of course, that they're all very much behind. And Microsoft is there. It's, of course, um, amazing in the way that we, I think, should we not be yelling at now that the model should be super small, right? There's something that Microsoft said about Phi model, I think it doesn't matter now. The question is, how much model would be, well, how well would it be to be very happy to be self-suffering with these hallucinations? Oh, so the future is, of course, models are more and more like Deep Research. I think it's realistic here until only three companies in the world have the ability to do it. It's me that's auditing. Of course, I'm very surprised by these companies, but there are only three left. It seemed like a lot of them.
With regard to neminga, Sash, I think there's a game of name in there, because Google literally recently released his own Deep Researcher. There's DeepSeek. Yeah, yeah, yeah. And then the OpenAI is just trying to move the others and to say that now that's our name and as loud as possible on all sides of it.
Maybe that's the word that's deep, you mean, right?
Gugla has a Deep Research or Deep Researcher right now.
I'm just saying, at the beginning, when it came out, I had this approach. I think the Consymers, though they don't care what they care, they'll let GPT5 out there and put it in, they'll make it all, and that's it. And it's just a consuser base from the beginning, it looks like a huge, superanalytic crowd. And in fact, I saw right away, the other way, as the simplest basic things that you're doing, a little more broadly, you can analyze now and much better. I mean, it's a big-- wide-ranging perception. What you've actually tried and done while I'm waiting, for example, this is how I used to lead the vacuum cleaner, I don't know, about the rice choices, it's simple, very simple things that people do every day. Some choice, and now you're getting the chance to start it. Yeah, in limited mode, it's still available for $200, there's 100 requests a month of all of these, you know, 100 of these analytical requests. I wonder if each request for clarification would be considered a request, I don't know where the counter is.
Yes.
It's gonna be like a series in the Simpsons, yeah, which we've been thinking about.
But questions are very interesting. Or two hundred vacuums. For me always, it's always important for me to look at the reasons and the consequences of some things. The Yandex once around, I think eight, seven years ago, he was released under the search line, they're starting to get a sub-standard in the test, like an addition. So you ask the question, you're told, "What do you mean, or is that?" So you get extra and, like, some sort of clarification, it's called clarification. Then it disappeared. And, well, it's not at least a mass trend, it's not super detailed. There's some sort of clarification, there's a man who can choose to take pictures, videos, and the deposits, like all the browsers. And it was always amazing for me that the search engine didn't go into the right cool details. At the same time, I'm kind of a business man, and I know they don't do it because they make money on a fucking search. So all the money is Yandex and Google, and we're not gonna help everyone find everything, and we're gonna help you get confused, and we're gonna spend more time with us, so people spend more money on the clicanium. various commercials inside that are located. It's changing the advertising market, right? From the point of view of finding. The emergence of systems that are now like Deep Research, that is, it takes a fundamentally different way, it's totally different to take a search and now, by the way, I'm getting an article of the great she's got. I did, it says exactly what resources it was based on. All these resources are on the inside, you're always on the line. So you can see why a conclusion was made, the probability where that information was taken, yes, or how it was obtained. I wonder if it would not only be an outside site that would provide as a source of information, but it would say that information was received, I don't know, from a book like this or in Africa, from some kind of stuff, or at that time. analysis of a large number of data. It's just a conclusion like a combination. Maybe we'll come to this, but it looks, uh, really good and useful. Here's the complex. And of course, such systems will not only be modified by the market AI.
I believe in changing a very strong search for a man and how people are, what kind of devices people will use in their terms.
The devices will be used in terms of their daily life, daily life, and this is their daily search. That's very interesting, that, on the basis of those applications, what application would be, what would it be? And it's gonna be like, "OpenAI or Google," or "Mexoft inside their infrastructure," or is there anyone who's gonna be using their models? Because we remember, you and I were talking a year ago about the starttaps that come, and now we're all talking about the starttaps that come. Because, in fact, it's getting all the start-ups, and just me, who was watching the start-ups two years ago, ninety-nine percent of them, they're, like, eating the same system themselves, you know, the same ones. The owners are huge, right? That's how companies like OpenAI, there or like Meta, they're inside the books-- themselves, they're doing this whole system. And this is fun, like OpenAI put the OpenAI Gov out for the State, yes, a special system for the State trained in public documents. I mean, if someone's doing this system to draw conclusions on... I can see that the start-ups are all over the world. In the Russian-speaking space, if you know, for example, the 1C has its own system, there's a consultant Plus, where you can ask legal questions, for example, and get some answers, in fact, before. It was a separate product of the B2B, and by launching a system like OpenAI Gov, they destroyed the U.S. Competitors' market because they said, "We're in the public law, not just in the public domain. American, world. We'll fully complement it, and you can use all these systems inside at the state level." I mean, it's a way in, uh, very, very differently, a little different, a little bit of a move. We'll watch what happens. I want to remember that today we didn't mention where Apple was, and where Apple, guys, would be there, where does Apple even know what he's gonna do with his platform? That's a big question. That's a big question.
We also lost Mashk. I think he'll be back in the info field again in the nearest releases, and we'll talk about him, too.
Yeah, yeah, yeah. Well, fifty minutes, we're having a good graduation. That's it! Thank you very much, everybody.