There's a big difference between five and four on the first day of work and now. I mean, it feels like I had five and four on my first day of work, and now I have five and two. We're getting a new update. The difference is, what's the proversion that's different? The ChatGPT regime should not retract your remaining facts about you and should not rely on past chat as a personal memory. Yeah, what's the problem now, I think it's a big deal in these pgreeds, all the systems. That you could get used to some quality of work, and that's not the quality of your job that keeps you. Anthropic has issued a report on the development of different, uh, professions. I mean, in what professions LLM has a great potential. Gemini has a very large aide, which is about Google-camp integration with Google Docs, Sheets, Slides and Drive, yes. Hello, everybody! We're on ToTheMoon. Technological news, Silicon Valley sites around the world.
We go out every Sunday, special releases Wednesday, Wednesday. So, look, we had a graduation last Wednesday, yeah, the last Wednesday, and, uh, the safety of artificial intelligence. Everything about the safe of artificial intelligence. Who hasn't seen a very cool graduation with Max Grigoriev. I hope we'll take more special editions with him. Oh, well, we're starting with a theme that looks like our Sunday graduation was, and we didn't sound there that ChatGPT 5.4 came out. And it was, uh, a little, and then we're gonna take it off about 5.4. But at the same time, as practice shows, it's best to wait, because there are big, big-- I have a personal difference between 5.4 on the first day of work and now. I mean, it feels like I had 5.4 on my first day of work, and now I have five and two. Well, like, like, like, like, like, five and three. Something came back. Because on the first day, I saw this imagination. So the system started working on a new one. I was still at the conference, uh, last week, and I was just saying, "Oh! It's 5.4. Look at the new type of interface." Well, we don't even expect the system to work. There was a logic in 5.4 that she gave you a certain answer in advance or said something, some information was kept in mind for a long time. And in general, this system is fun, by the way. I mean, I'm often, I'm starting to read the rizoning answers now, you know, that are temporary, which I can't see, but which, like, are hidden, can be seen. Oh, to keep it from waiting for 15 minutes, and I'm gonna see some more information, but then I'm good at getting a full answer, read it. And now they seem to have started to implement some of the similar mechanisms. And, but I've got him gone. I'm in America. Do you have one? Yeah. I have a place in America, right? I don't. Here. And I don't. You don't, do you, Tanya? Yeah. It's a very nice thing.
You still have one? I really hope it doesn't. I really like her. Yeah. We don't. It's gone. Ilnar, you mean, too, right? It's about the dance. Yeah, sure. Illar's about, I mean. And Illnar, you see what he says, "Yes, of course." He's not Tanya, who hasn't bought it for two years. And, uh, well, write the audience, like this. And I think, of course, we'll be here. This is an example of how we don't understand how models work, how chat work. It's a very funny, interesting example, when, in the idea, you might be answered immediately. That's how if we use the vocal system in 5.4 now. By the way, everyone was tested for the last voice, the last up-to-date voice system. She's been very good at work and very good quality is good. Oh, well, again. Not bad, better. I mean, she even started answering. I asked her questions, which, in fact, require resoning. Like, I'm on a trip to East now, and I asked her, like, restaurants in town in one. And accordingly, I ask restaurants in this town, there in Cairo. She answered me in real time with a voice. I didn't expect that. And she said it was good enough because I checked in a long resonance. I was surprised because, in fact, she needs to look for her. Maybe they're using it. Remember, we told one of the episodes that they were responding to some information, yeah, they're still downloading, then they're processing it. They may have started using this system and 5.4, basically, like, tried to do something, but this primitive answer is still in my hands, Ilnar, it's gone. Plus, I've got a different job of deep research. I'd say he's not working well. Write on who's got a theme. What do you want to add? I want to take some things out here. We talked a lot about the difference, and, uh, in the regime, in the whole way of the conversion. We're getting a new update. The difference is, what's the proversion in the part, uh, what's the difference?
So, the proversion, she's looking. Remember, Ildar, we talked about it. I saw that I sometimes had no search. She's doing a search. What doesn't make a provert? The proversion doesn't work with your chat memory. I mean, look, there are three kinds of memory at the GPT chate. Well, a lot of systems must have been like this right now. I'm always bringing an example on GPT chatting as a mass system. There are three types of chat, oh, three types of memory. First memory is the memory of the context window of one chat. So you're asking questions about it. There's one window in which you have memory in one chat, and it's always using that volume. There's a second memory. It's your past chat room. All the chat rooms you had, or parts of the chat, the queries you made, how you said you liked what you had interests. So ChatGPT is starting to draw some conclusions from this. There are strange conclusions. I mean, if I'm interested in a country, he thinks I like this country. Not a fact. If I want to go to a country, it doesn't mean I like her. Maybe I was just forced to go there. But he's making some conclusions. And there's a third memory. It's a memory that you're asking in the building that you're trying to use it as, like, it, I don't know what it's called Russian, inside, and in the personalized buildings where you are. You're telling me, you're telling me about yourself and you're keeping it. Remember, they had an old project, and they still hang, but it's a little bit of a curve. When you could say, "Save these data for me," and now they're in the third. And he said, "I kept you this data," and they're in construction there in the form of separate plates, separate like the units of such information. So, Pro theory, she doesn't use chat memory. And that was a nice thing, by the way. Illnar, did you know that? I feel like I'm using it. But it's just my experience. Yeah. They've now officially written that Pro ChatGPT should not retract your remaining facts about you and should not rely on past chat as a personal memory. OpenAI writes directly. Pro is inaccessible, first of all, apps are inaccessible. It's different applications. Which is weird because Deep Research is also app. It's finally available or not unrecognized. Web search is like app, too. I don't understand. Well, not your built-in type. Yes, memory is unavailable and inaccessible Canvas and Image Generation is not available. Yes, and Image Generation is unavailable. I mean, they're officially writing that they don't, but they're pointing again that memory, custom instructions, that is memory, is memory of your past chat. Custom instructions are what I said, these are personal designs, right? It's just not the same thing that everyone understands. I think it's a very useful explanation for you all, right? Because, in fact, in the future of the system, they must work, that there is some predicament. Well, you're in the system giving some volume and you're talking, so it's binding. And then there's another body that this system has been working with. They must be getting along, right? But still. And there's a memory of one window. And I, by the way, have occasionally wanted me to have an isolated contours. I knew that a private chat could be made in an isolated contour. Everyone knows, yeah, what's private? You know what private chat is? It's when you can open the chat that won't be kept in history, your history will not be preserved. And, accordingly, OpenAI will immediately be put to disposal. I want to point out, by the way, that this disposal will take place within 30 days, not just immediately. Yeah? But there's a concept. It's a useful thing, if you know that your system, like, is being used by other people, or you wouldn't want this request to be initiated with you, would you? And, accordingly, if you create this one, it's like a private browsing on the Internet, that's when you create independent deposits from the browser, there, and so on, yeah, that he doesn't take your cubes or anything. I mean, totally, totally ready central system. By the way, with the models coming, we see the complexity of all these things even more. Who uses it, who doesn't. There was a case now very interesting. Sent to the court on the Meta glasses that Meta kept and handed naked images to third providers in technical data. So we're getting raw data. And there were naked people. And yes, yes, yes! That's Ilnar's theme. I just wonder what the context was. Why are people in glasses when someone's naked?
Well, when the glasses are wearing, yes, Tanya.
Wait, they wear, yes, Tanya.
You were walking, you didn't record that.
What were they doing?
Shel, I didn't record that. Tanya, one thing you've been recording, another case you didn't record. They might have been recording a part. And plus there was a story, we're all over it, I don't know, the end of the trial, yeah, we'll study it. And plus there was a story that might, for example, be that if there was such personalized information, it might not be able to leave third people if it was about certain things. Although how to filter it is a big question, yes. That's why this is a subject, of course, shorter, very interesting, and it's bound to be studied and watched. With Pro's story, it's a pretty cool thing, because I, for example, often have to do some big research. I don't want to rely on past chat because I see how much trouble I see that I'm often confused, I have independent topics. I mean, I can ask in one contours about some corporate analysis and in another, ask completely, I don't know, about my own, and in the third, ask about partnerships. And I need the system to give me very independent data. Or I'm asking about my analysis in one contours, and the other about the other person's analysis. I don't want her to interrupt them. And it's weird, of course, that they still, Ilnar and Tanya, have not made a tie, yes, that could be turned off. I mean, you're obviously getting surprised, it's kind of a little construct that--
It's more of a tombler, Sas. We said we should make it all work.
No, they'd better get it. Look, they'd better take out the unnecessary tumbles, like in the Instant Thinking Pro automobile inside each, inside Thinking.
Very Legacy, yes.
Ilinar, why would I--
If someone is very attached to some of the past models.
If I don't use Instant or auto, why would I have to go inside Thinking Light? So what's the normal person in Pro's version using Thinking Light? Well, that's--
So we'll get to Android when you're doing it yourself, you'll have a separate window of inspiration.
I'm not getting any more.
You see how you can't see.
Yeah. Look, I fixed it yesterday. I bought one of these, I have to take some sessions to measure, there, scientifically, in terms of heart rhythms. And there's a sophth that's got me, and there's a Cubios soft professional, and he opened me MATLAB. When I saw him installing me MATLAB, I thought, "E!" And I already paid $450 for that soft. And I thought, "U! I don't want to go there, because I'm open, so this thing that asked me to do Java, Java, means some kind of plagin. And she opened up to me, and I'm a click-- and I just opened up, "Will you make the construction?" And there I saw this Microsoft Windows eighty-six fields. I'm like, "I'm not at that age, you know. I used to be interested."
Sasha, you take a picture of your phone and you say, "What do you do? Say "Bratski."
And I was thinking,
It's my thing, it's my subject.
So, five four. Ilinar, yours, your five-four. What do you think?
I love it. They've expanded the context window. I think I read a million. I can be wrong because now many millions of people have been rolled up, I think they're as much as a million, but in very long chat, they're good at context, well focused on what happened before. Even if there's something else going through, there, I don't know, five ten messages, you go back to your original task, keeps it pretty good. At least that's what I've tried. We've been back in a long way. Of course, it's a two-way story. On the one hand, wait half an hour while he answers, on the other hand, some kind of very good quality, at least those tasks I've tried. Well, what you said at the beginning, that he, like many others, is now beginning to speak, there, so in the first stage I thought it, in the second stage I came to this conclusion, in the third stage I was thinking, I'm doing something. It's rather interesting to keep an eye on the process of reflection, even though it's not as detailed as it may be, at the beginning of DeepSeek, when it was possible to read as he thought. But there's some kind of clipping here. You see how this is going on. So in this sense, I have a very positive impression of five. I like it. I didn't even watch the baccalaureate because it was probably just a little bit of a baccalaureate. But what does it matter? Everyone's got them growing up. You're looking at the tasks you often use, and I have a feeling that quality has grown considerably. And the time of implementation is too. I hope I don't get back because you have.
I, you know, have a little smearing feeling. Because I just signed up for Pro recently, and it's no longer clear where the difference is because I have Pro or the difference because it's a new model, but it's definitely getting much better and better to work and very quickly. I missed it if I didn't have that function, that he first gave a short answer, then a more reasoned one. But in general, I really like it.
I don't regret my whole thing.
No, it's 100 percent, Tanya, actually, from the point of view, you still have this function Thinking Heavy, it's not in the usual version. I'm thinking, like, just hard work, I'm not doing anything anymore. I don't know, Ilnar, I guess, too, right? I mean, you're--
Well, I'm not always. I sometimes have some simple questions that need to be answered.
And even you know, I even, I even, I'm just a regular one, and I'm actually starting to get into heavy because I'm actually not-- the system's gonna over-check myself. Although I see we have some updates in my scout, I think we're getting some updates, right? I think that given the American American amer-American scouts, I think I've seen myself having five four Thinking Heavys sometimes started answering, uh, they've done what now? They took away some of this piece of history. I noticed even when you were on ChatGPT, you used to drive, he was having some kind of discontinuation with you.
I'm sorry.
Yeah, he stopped doing it. He's technical. That's what I wanted.
Yeah, I noticed, too.
He's technical. I had a situation yesterday where the answers were just so stupid that I was so easy, I already wrote, "I know you're a device, but it's just the dumbest answer in the world." And he ignored what I wrote and gave me just the next. And it was really cool.
Maybe Sas, maybe he put you in a pencil in a notebook?
He has a book, yes, where he's wearing it when he's got a constant AI, he'll remember everything, Sas. There's a strange story, by the way.
So many people should have remembered, so...
Sasha, I don't know how you're holding up.
How many people should I remember
then.
There's a strange story with all these applications. Like, I'm a very frequent user of Nana Banana. I'm opening it almost every day and I think that today is Nana Banana's day, which is the mood I'm in today. Sometimes I just, I'm doing interiors, visualizing fast. That's the perfect thing from the semi-word, it's doing great, followed by instructions, and it doesn't have to be a difficult prom for her to write. Sometimes I just open it, do the same thing twenty times, she's just giving me some kind of trash. And there's a GPT chat in it. It's like he's starting to slow down. And I also have this kind of desire to call it a device, to say that you're so stupid, something else. And full of silence in response. I mean, it's just starting to do the right thing. And he's been there before, I'm sorry, yes, I'll never be there again, under any circumstances. That's very funny.
And that's a skirt, and they have it in their voice. I was in the voice, I had to go a while ago, and I wanted to vote, I didn't like how he used to work, and I chose what a voice I had. And in fact, among all votes, all voices like I don't know, the Marquetologists are modern or the manager's sales.
Or a teacher at some primary school.
Yeah. And the fifth grade teacher is talking to you. You're like, but I can have a normal person on the other side, with a robotic man. I found some less voice in there that I answered, like audio books, more or less neutral. What, Tanya, you say, this is a juscany, really. And yesterday, on the example, I saw that, yes, there were very unbelievable answers, but the jusy went away. And the question is, of course, we don't understand what's going on inside them and why one starts working better, the second worse, then it comes back, then you get used to it being a good worker, and then you get it. There may be a whole range of mistakes, and you already think it's working well. But that's what it is.
You're even counting on, yeah,
in the job, that you're gonna have a great visual.
What a problem now, I think it's a big one in these pgreeds, all systems. That you could get used to some quality of work, and that's not the quality of your job that keeps you. So you can get a different area better, and it's gonna make things worse. It's like a hysterical man, right? So you're either a sick person, actually, like schizophrenia. I mean, you don't know what's going on with you. Today he says A, and tomorrow says the apple, and you're rolling around, and you need to remember every time that he can always be wrong and always give you the wrong information. I mean, he seemed to be able to do it, too, and he's over--
Well, at least what I see is they're going to have deep research now, it's some weird story. Or they, uh, Codex, released an app, and the Chief of Codex says, "We have a super-opgride!" There. And puts an interface like that. He says, "We can build an interface!" And there are twenty fields of interface design. I'm sitting there and I think it's a shit? So you just put the system out as unique, moving to AGI. And for the last month, all your leadership has said that Codex is the future of AGI. And agents do a lot of things. And then you come out and enjoy the interface that's built by the interface. Well, we're two thousand twenty-sixth year. Not much of a problem with your interface is different, right? You have linen chat, and you can't figure it out. You don't have a search, work, glows all the time. I mean, and that's a serious question, of course. Well, they're probably gonna decide at some point, maybe not. What's the point you want to raise today? Although there is, of course, a very interesting subject that is happening in the world with the OpenAI installations and with all these conversations, with the teams.
A huge number of people like you leave the OpenAInthrop in Anthropic because Aic is working differently with the War Department and so on. But before we move this, uh, a very important topic for people, and not just us here. It's a subject. Anthropic has issued a report on the development, uh, development of different, uh, professions. I mean, in what professions LLM has a great potential. In fact, it shows, well, from the beginning, you can think that these are professions that will be replaced by artificial intelligence. But the report doesn't really mean that. He says how much, and now the model can reach the whole system. And for example, the most exciting story is programmers. Seventy-four and a half percent. So this is the coverage of AI in general what's going on in programming. For example, the client support service. Seventy-seven per cent, data entry operators, sixty-seven per cent, medical specialists. 66%, market analysts and markers 60%, and so on, huh? Or, for example, there's an information security specialist of 48 per cent. Look, this is not a subject to consider that this story, that the programmers have decided about 75 percent because it's basically a market coverage. I mean, what kind of data is inside there. And when we say that this volume of data is automated, it means that many simple tasks are quickly to be met. So people will be more challenging, right? And the question is, how their system will be able to capture these new challenges. I mean, how is this process going to be all the peddates, right? How these billions of sophths will be updated, which are already written. That's a separate question. The report is generally very interesting. The main message that is there is what is actually happening, with the fact that today, in a client support unit, 70 per cent coverage, while at the same time the application is interest. I mean, there's very little implementation. So the scope of the model suggests how much models in this area are generally being dealt with. And the world's introductions are very small. Well, the client support service is what the customers are, processing requests, complaints, and so on, right? And the fact that the model is really used everywhere, it's just microscopic. But I would have taken a very high look at this report. That's, uh, very, well, so big, good, mass-data. And Anthropic, I think, is a reflection of this history as part of the assessment of such areas, although it is merely a technical assessment, yes, it is all a technical assessment. Because how can you estimate the amount of all the data? They should know, in some way, what amount of data exists to assess and how they determined what amount of data exists.
I mean, this is the mystery inside. Here.
As for the programmers and, yes, these models, they always talk about it. Amazon, uh, banned the junction and the muddle, uh, medjili, you know, sending their changes to the project without, uh, jealousy from senior specialists. It was probably before, but now it's very hard because many people are making code with I., especially if we're talking about juniors, they don't always understand what they're doing, which is what they're finally doing. He does. They put these pieces of code in, and then the company faces enormous problems. Amazon obviously had, and I think, in December, some services were inaccessible because they were making mistakes in design, and it turns out that AI was used in it. Here. Therefore, on the one hand, I, LLM and all these services are very well placed in programming, but this has an even large number of problems. In particular, in FAANG, yes, in large companies, Amazon, in particular, imposed a strict ban on juns and muddles sending something to the project without checking seniors.
And there's a news here that Anthropic, again, has released a new service for code review. So they specifically bury their model to make the changes you're suggesting to make, pull request the software, they're first tested with Claude, specially. I'm trained to check these things. And after that, only after his findings, after his approval, these changes have already been included in the draft. And the cost in the article, again, the anthropic had a history of between $15 and $25 for one of those checks. And here I want to, uh, be happy for anthropic, because they, uh, go to purely applied things. Because there's something, and a pool of rivers, that's what the project is, big companies have a huge number every day. And basically they're moving from the format of this, you know, first chatting, then, uh, programming systems are going to close down the next phase when you make a decision on the basis of this system, like, make these changes or not make these changes, or at least report, and what problems this change may, or, if so, bring. Yeah, and that's what happens, on the one hand, a large number of programmers start using AI places in a vain. This leads to problems that Amazon is simply a strict prohibition to make changes without verification by the senors. On the other hand, anthropics, uh, are setting up a system that will check all the pools of the rivers, that is, make a rheuma code, check the code and take the money for it. Basically, well, a flat tax. A large company can simply allocate a budget to this and then manage it, see what pools of rivers we'll check that we don't check. Development is therefore very and very intense in Ishka and programming. It's not always very happy, as people think.
Well, in particular, of course, what you're talking about checking this one. That's a very interesting example and a case. And what happens to all companies that can't check or where people don't even know that artificial intelligence has been used. But it's actually like a normal person to check. What-- what's the first thing? Did someone check the software code when the software writer wrote something that someone was checking through his mistakes? The main mass in the world of written soft and websites, and everything. The main mass is I say nine hundred ninety-nine thousand-nine ninety-nine nine hundred ninety-nine of a million. And that's the kind of thing, this kind of error is definitely not checked by the techs.
And people just do an incredible amount of bagov and mistakes and problems. So, well, we're just entering a new zone, uh, very specific things, yeah.
Just, Sasha, what's wrong? Previously, the programmers, especially the juniors, think they were on a bike, moving slowly, could not cause much damage in a short time. Now that everyone has access to the code, the code, the boat, the Ishke and the other, yes, the courseor and other code writing systems, they start to generate the code much faster and can cause harm. More. I mean, if they were checked periodically before, but they didn't get much harm between these checks, you can't hurt them very quickly.
Well, that's right. Especially when it's supposed to be quicker, you're gonna be out, right? And everyone wants to get out faster. I-- that's, of course... We've been running the canal for two years, right? Or maybe it's about two years. And yes, and two years we've had enough detail--
No, no, no. We started on Space Day. Oh! April 12th. April 12th, yes. Still in the nick.
Twenty-three months. Yeah, twenty-three months. And we're watching and telling you about trends, people's speeches, the profession, the job replacement, the students, everything. But in fact, if you look at the reality, it's like a fun tool, like they're giving more knowledge. But what--
Nooo so that something really changes or that we start living in another world or that all the lubes become smarter, and in this case, we're doing the well-being of others, we obviously don't see it. We only see an endless increase in the race or the desire of all people to use this resource for their own benefit.
Well, obviously, if everyone's only used for their own benefit, then the result, what is it? It'll be, in a... alike, like, what happened without it. I want to focus on the following. Gemini has a very large pdddette that is about Google integration. Google model from Google Docs, Sheets, Slides and Drive, huh? And they, that's the whole Workspace now far deeper into the context of files, letters, web, helps collect documents, tables, slides and the Drive adds AI to the search, right? In fact, they don't separate chat-bot, but as built directly into the ecosystem. And, of course, we understand that if the same OpenAI doesn't directly start to make some very serious integration to work with the documents and everything else, if the man uses Google, well, that's me, for example. I use a lot, uh, Google layers. If Google starts to provide quality service inside, I'll just use this quality service inside and not pass, especially if I don't have to carry something out. I mean, they're very strong, well, unbelievably strong, I think they're developing in this part.
And again, write who uses Google and who has Google's got a better job, right? There was more news that Google released Gemini Embedding 2.0. It's a multimodal embedding model now that combines text, images, videos, audio and documents in one sp-- in one space and works more than 100 languages. That just happened. But it's for the developers, right, Elnar?
Mm-hmm.
It's for the developers. But it's also a very interesting story when you're all together, that's when you start working together, together with infrastructure, like a certain kind of control. Now, of course, nobody can do superpowers with this because, basically, if there's a company or a corporation, it's got some kind of automation, then you're a disconnected system, some of the data is unavailable. I mean, to integrate all this between you, you need to, or you need to get it down, teach it. How's that gonna work? So, as long as, of course, only very small-- small, small-scale projects can be seriously implemented. Well, Tana, for example, it might be interesting, yes, such systems. I mean, when, for example, you have some small-scale, uh, projects and there's audio, video, and images, and you, and you, and you get a system that automatically learns and can. I'm sure he's gonna work with him and generate it, which is very interesting.
Yeah, that would be cool.
Yeah, yeah, yeah, that's very interesting. By the way, I'm here to launch a parallel set at 5.4, uh, a request, and it's on the web, it's not a dextop app. And I got this 5.4 in response, Elnar. I've been running this message since the morning, an hour ago, or, uh, 30 minutes ago, I haven't had that intel, yes. I'd like to say the following about this embedding multimodality, which is not an old approach. It used to be an approach that was first translated into the text, then built into search, for example. Google says that the model is naively working with images, videos, audios and PDFs, and she has OCR in Vertex AI, you know, to recognize documents and get audio tracks out of the video. Audio can, uh, build without an intermediate transcribration, right? And accordingly, this reduces the number of intermediate steps where part of the meaning is usually lost. Because we have one of the stories big, we understand, yes, when we work, we imagine that a man speaks audio with a certain intonation, with a certain style, and we move on, for example, He was moved to the summit and moved on, so we lost a huge piece of the original content-- context. And if there were many different dialogues and all before, we will, in fact, at the end, receive some kind of answer that is very relatively, very relative in terms of reality. It's one of the fundamental problems that is. What's the conclusion you're getting? That's why we're seeing different answers from the models. Is the war right?
One system says yes, and the other says: no, because they have different contours of perception inside systems. Eh, Elnar, you stacked-- you wrote today just about the OpenAI buying a start-up that's in charge of security, right? And they, who use a lot of them, are PromptFu, right? It's about the safety of the launch. And plus OpenAI, it's a new layer in the code now that it's about security checks, that's the way the code's safety is, including. So they're starting to check the system code for another layer. But I'm embarrassed in all these systems, which is what it is, there's a OpenAI company, and she's already been doing a test of her systems. She's buying some company that's working on checking their systems from outside, and they're putting this layer on. I think there's a company on the market that also starts to check outside the OpenAI. It's an endless process, isn't it? I mean, he's, especially given that the system is growing and it's getting much stronger, and it's much more powerful to be checked. And that, I think, is an unfinished task. I mean, it's just some kind of pyre news.
I just want to make a joke that they used to take the money for the code generation first, and now they're taking the money to check what they've been generating. So they're taking the money just in the next phase. Just like a pool of anthropic revests.
They're taking the money for this stage, by the way, right?
Well,
anyway, you're gonna start this, there's gonna be tanches, you've got tokens, you've paid for them, and so on.
By the way, the interesting topic of OpenAI is adding. They're being tried for the next thing that means the woman who tried was sued. It's early March, right? She's loaded her lawyer's files on GPT. So she had a lawyer who was dealing with them. And she asked if she was gonna get a little stunned. Haven't seen this case, have you? She asked me if she didn't go on a gas-t--
I'm very stealing.
She's not gonna get her lawyer. ChatGPT supported the position that counsel was gas-tight.
Of course I am.
She fired, she fired her lawyers, and she started trying to reopen a closed case with a GPT chat. And she has prepared arguments accordingly. There are specific summaries of the rules, a re-launch request, and I did. What happened next? The judge refused to reopen, stating that these second thoughts of a peace agreement were not, not, the reason to reopen the case. And after that, the woman filed a new suit and then dozens of new papers. But I want to tell you right away that it wasn't in the current version of GPT. It was probably there for 20-fourth years. I mean, the suit was February of the twenty-fifth year. I mean, it's a year ago. The woman filed new lawsuits, then dozens of new papers. And she did what, basically? She made 40 or 40 complaints, described them in full, filed 14 separate legal inquiries, and prepared them with the GPT reader. And the court said that the data did not serve a legitimate procedural purpose. And then there's an example of a fiction. What happened? We're looking at hallucinations. The subject of the reference was made, there was a fake reference and a GPT chat when they asked him that you were kind of not the real deal. He said, "No, I'm pointing out the real case." And I gave a description that matched what was in the paper. I mean, imagine you said it didn't exist. He said that existed, gave a real reference to the real case, but it wasn't about what he was talking about, and it didn't mean it. And they finally said they'd put it in court on OpenAI and they'd ask for $320,000 of damage, $10,000 for another type of damage.
And plus a lot of other costs.
Yeah. But here, look at the story.
A woman was not, in fact, a lawyer. And in America, to provide legal aid, you should, for example, have a license. If GPT chats, he can express his opinion, but I'm not a lawyer, it's just some of the data I gave you, like from the Internet. And GPT chat, in fact, OpenAI in the evidence position, that's what I said. They said we weren't, we're not a lawyer. In this case, we were just providing information to a person, gathering as much as I can come in and see information in different forums or somewhere else.
In different forums or somewhere else. But it's not interested in the needles, yes. They say that, by the way, they've filed a court, not a woman on the OpenAI, but lawyers. About the gas-layt. I mean, there's a reverse situation, because there's a separate, separate, third party. I understand it was the third party that ever filed. And now, I mean, the OpenAI court, not the retrieval, but the third party in the OpenAI court, that ChatGPT has acted on this role, cannot do it. But we are going to question responsibility, yes, who is responsible. Now, Maxim and I have been talking about the security issue on the specials's release on our canal. I was here at the conference, there, Dima Greene in his canal, and right there was a discussion on who was responsible? So who pays for the banquet for this, who pays all the fines, the risks that are responsible for this situation? Or nobody's fault for this situation, and in fact, the man misused. Well, the same thing, I guess if I use the Internet, nobody's gonna judge the Internet. If I take an article independent or, there, I'll read some book and I'll go out and get out and get out for protection, and then I'll tell you what it was written in that book, no one will present the author of this book on the book. Jurisprudences, where millions of people learn. Right? I understand that nobody's ever gonna bring them.
And there was a story...
But responsibility means a person, not a tool.
Yeah. Or company. Or company.
Yeah.
And in this servant...
Physical or legal person.
Yeah, but look, it's always funny to me, and it's always funny to people laugh-- well, because, you know, guys, wait, there were books, and I was doing everything in the book. I had a mistake, for example. I kind of wrote that I did everything in the book. I don't know, I bought a book that described how to make money. I did it on the book, and I didn't make it. But there are people, well, in some countries, they are suing, but in developed countries, people in normal developed countries are not suing for it. Yeah, it's just the information that was given. Someone did it, someone didn't. Here's where this case is going to be considered, it's big for the market, because again, if some state or some judge considers it a wine, it's a precedent that starts to go on, People are starting to point out. And for world practice, yes, it is clear to world practice that people in the world are starting to use it somewhere, and some countries will use it and say, "This is how we work." Or some companies won't use the East-- I know a lot of companies that don't use artificial intelligence in any way. Actually! I'm talking to the directors or the co-owners. They say, "We can't do anything at all because any of our requests for lawyers are blocking immediately." Because the lawyer has a job? He doesn't care how the company develops. He is not getting his problem, and they all block everything, especially with regard to Europe. It's just a block is all over the place.
Whatever you do, you're suggesting something to be introduced, and you're blocked. But in small companies, it's clear that this story will move completely differently.
Look, we talked about Jan Lecuna in the releases from time to time. He was until, mid-decade, or twenty-fifth year, the head of the artificial intelligence unit in Meta. In general, very well-known in the field of scholars. Yeah, he's a huge development, and he's made a huge contribution to the development of artificial intelligence, and, here, machine training, everything that's connected. And he left, actually, from Meta, opened his start. AMI is called. AMI, yes, there, Advanced Machine Intellect, I think it's something like that. And accordingly, attracted about a billion dollars in the three and a half-degree, I think, a billion. That's literally the last time he got the money. And the difference between the other start-ups and the other ones is that he's not making LLM.K. He's all over the way saying LLM is a dead end, that you don't have to move this way. We will not be able to deal with hallucinations, and in general, the increase in LLM-ki is no longer profiling, but it goes to conceptually another history, to create world models, that is, to create models that can be rudely, I'm gonna run experiments on the inside and make some kind of decision on these experiments, give me some kind of answer. Well, that's a lot of good practice. If we ask ChatGPT what happens if the glass falls, he'll probably say he'll crash. Where did this information come from inside ChatGPT? She shows up on the basis of what, well, it says somewhere on the Internet. I mean, based on the full range of data on which ChatGPT was trained, mostly textual, if we take, early models, yes, if we take no multimodal approach, then that's on the basis of that. Some text information model makes a conclusion and, as a prediction of the next token, it can also be hallucinated and, accordingly, it can come up with the answer itself that happens. If we're talking about world models, the point is, if we ask what happens to the glass, this model has to make a simulation that should, in fact, drop the glass, see what's wrong with it. And then you'll be answered. The idea looks interesting. There are some companies that are developing this way. The only thing that's gonna understand is that all this lab research is. I mean, in a rude way, there are some scientists who are developing this concept accordingly. The concept is interesting, but the models on open access that could be felt are not. Lekun himself in this sense says that the future is behind these models, and it is clear that he needs to attract money and he has a start-up in this direction. But, he says, there's almost all the big companies in there in six months or so that we're not LLM doing, but we're building world models. It's a question of hallucination, but accordingly, it requires that within the model, in blatant words, the laws of physics be there, yeah, there's all maths in there, some other stuff that's gonna be inside, I want the model to simulation and give some response on the basis of this simulation. Oh, I wonder if it's brighter and brighter than these new directions, not LLM, that they start to sound like money is being invested. So, a billion is practically there, in the first round, with three and a half of them, that's a very good result, too.
I mean, Ian Lekun is a very prominent person in this direction.
Look, I wonder what, uh, there's a whole series of people that are known to be very famous. And that's how Mira Murati, for example, is it? They, uh, started for a few days, there's a new integration, well, a new project with NVIDIA. I mean, they do a lot of different technical projects, but of course, influences in terms of big, big systems, like they had before, when you're doing for a billion people, they still have one of those. There's no direct influence right now. So they're doing very strong, uh, scientific work, very hard research, and they're doing some serious development, technically complicated, but they're gonna have a less impact on, uh, The end-user. Because influence on the end-user is still influenced through a massed market. Or do you have another opinion here?
You're absolutely right. There are people who want to do the product and that's how to see the result of their work, which I don't know is that eight hundred million people use some kind of thing you added every month, yeah, some kind of thing. The function that came. And other people want to do science, yes, that's, develop and find some new things that haven't come before. And to do the product because the company said we'd make more money in this direction, not in this direction. Let's just put it aside for the company, more in principle, to get some product out of here and all the efforts. And it's like a person who likes what you want. I don't know about Mira Murati, how they broke up, uh, with Sam. That's the whole story, right? Uh, she's left, she's made her own start-up or not, anyway. But as sure as there are people who are interested in research, and they go to this field so that, in a rude way, these forms are new on the website. Are their influence on the market diminished? It's definitely getting smaller. If they can do something revolutionary, then they'll be in history. If not, there's just some scientist who's been doing some research that maybe no one's gonna remember.
Write it if you're interested in some projects, by the way, getting more information from the perspective of these people. And or you know some interesting people we should mention here at The Moon, we'll do it. And I'll see you in exactly a week. ToTheMoon is technological news, Silicon Valley sites.