Hello, everyone! We're on ToTheMoon. A-a-a-a-a-a-a-a-a-a-a-a-a-a-a-week technological news, uh, from the Silicon Valley and from the world. There's tens of billions of dollars for, uh, being a little furious. And then the Supreme Court said that you can't do that.
LLM-kills hallucinate always. It's just that it's often enough to get into the truth.
I don't see, well, some-- some big one in it. Such, very banal advice, which I think it's hard to use.
I wonder if Trump was-- a few days ago, in an interview with Mask, he said that we need twice as much electricity for the AI in the country.
Again, in the OpenAI, uh, someone's gone, someone's here. Do you need to pay attention to ordinary people or not? Or is there anything that changes that one man, uh, left a company that one of the founders of the and, uh, one man took a long vacation? Or is it just people who decided to rest or do other business?
It doesn't look like a cho, uh, OpenAI, who's not the unqualified leader, uh, the money, like, in, uh, in--- that's what you should be holding. Yeah, so if they were to have three or four steps to AGI or some breakthroughs or something, I think the people would never leave that mass, right? It's either way. Suzkewer left first. Now, uh, there's a major salesman out there, a cofaunder gone, one took, uh, sabbatical, yes, I think it's English-language you have. It's when I left for the rest of the year, uh, vacation. Accordingly, it just says that one of the companies. There are other companies in which work can be done, both scientifically and technically, and in terms of advanced products. So for me, these news, they make the OpenAI of, uh, some unqualified leader turns into a company.
At least that's how I feel, and it's all happening.
Actually, it's hard to say, because, uh, it's hard to understand what people do to end hands-run, right? I mean, there may be no financial- that is, a few reasons. Uh, maybe there's no big financial pessayda, or maybe there's something like that that that might actually be a technology-pay for the time being and not a bit of a question. But I was most surprised that John Schulman is one of the cofounders and lead researcher in the field, uh, reinforcement learning - that he went to Anthropic. That's very surprising to me. I'll try to explain., for those who reinforce learning, in principle the most interesting question is that when there is a basic model, such as creating a, um, environment where the model can be improved, where she has a re-- where she has a certain reward for her performance. And historically, it's very hard to do. There is no one-- uh, that's the direction, it's still in itself, at the very beginning of development. And many people call it automatic data creation, automatic data generation, because Internet data that are already generated are still used three years ago, two or three years ago, And it doesn't, uh, it's already, and it's a so-called page, actually. Ah, that John Schulman left in Anthropic probably says, well, I don't think John already has a monetary motivation, exactly, and in no case does it already exist. All he was was a cofaunder and that, that he was a OpenAI cofaunder from the first days, and that he was going to Anthropic now, it was a signal to me that Antr-Anthropic has moved on to reasoning. So John Schulman has probably compared all the options to a little more and more and he realized that-- and he probably just knows that, uh-a-a, that Anthropic has moved on to reasoning. It's a very, very sy-powered signal for me.
Yeah.
But it's a little weird, though, to talk about OpenAI. It's still a big leader, yes. They obviously have, uh, inside, something we don't see, enough development. And at this time, it's still early for this company. The company is still in the very early stage, I think, right? It's a clean startup. It's all the use stories, everything. Get out of the company if you're a cofaunder and especially strong with-- , researcher, engineer. Well, that's... Maybe I just got a fight with someone, yeah, maybe something else. I just don't know what it is, it's just a transition because of the interesting--
In the clear, I understand that this is a comunity, uh, in a comunit, um, as a consensus that AGI will be coming by 2030, and they probably define AGI in two ways.
First thing, it's a system that improves itself. And the second thing that she's in most of her tasks-- most of the tasks are better than almost all of them. Almost all the tasks are better than most people. And they have only an interest in being the first to create it. I've actually heard that very, very often. And then the question arose is, if it was the first person to create it, then it would be only to determine, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, that's, security mechanisms. I mean, uh, they probably have that look. hear that.
Sasha, where's this 2030? Where did this 2030 come from?
Uh-huh.
You say, "Community 2030."
Yeah, yeah. I'm guessing that's what it's about to be 100 times more computers than GPT-4 was used because models would need to be as close as possible to themselves. It will be more difficult to select quality data for the training of ia, therefore, to require a 100-fold model. And it's about two generations, building the dates centers for that.
You're just saying,
mean, community talks about what AGI is doing.
Yeah.
Is this the community talking about it or is it your idea now?
No, it's a community, it's exactly a comm-- exactly a community.
Or community or c- or community? Oh, the community has been talking about a new deadline, 2030.
By 2030, that's right. Yeah. I mean, the idea is, if GPT-4 was 25,000 H100, it's now built a few clusters, and that's about, uh, maybe six or eight times, that's where it's got more computers., and that's six to eight times more cluster than GPT-4 practice. If they use, uh-a-a, roughly not 100 days of GPT-4 training, but 300 days, that is, almost a year, toho, you can get about 15 to 20 times more, well, like capacity, right? Here's the current clusters. And on this new cluster, like the one Mr. Mask builds in T-Tennessy, he needs another generation to get this AGI. I mean, 100,000 H100 won't be enough, probably there, 200,000 of these B-B200s, which are NVIDIA. And that's gonna be a victory. And actually, it's not surprising-- and actually, I don't think it's my-- not that it's even my opinion, a narrow circle of people in San Francisco, I think the big CEOs agree, too. They all said, "Tsukerberg and Satya Nadellah and Sundar Pitchai, "Well, they'll continue to invest, "Eh, ah, ah, we're probably talking about a lot of billions of dollars." I'm gonna go to the D.C. Ah, and they see-- and they see it as the biggest opportunity in decades. And they all-- they're all kinds of giving a sense that they don't expect instant results in a year or two years. They said it was not done for a quick return. So, reversing the GPU investment, it's gonna be, uh, probably not fast.
What people do you think, uh-- what do you think people are gonna expect, uh, next progress?
Uh, well, that's so strong, but before AGI, that's because it's 2030 and AGI, it's a mythical story. I mean, she can show up early, I don't know, I have a feeling that she might not even show up in our lives. Uh, or people on-- with a new-- new word, will they? Because what is AGI? It's a little abstract, uh, anyway. Well, at least it sounds like some abstraction. What can be expected in the nearest year, for example? You see that in a year that's enough...
It's seen that they've improved enough, um, to improve systems, uh, chat-bot search, right? They've really improved. Or we'll be in them, uh, uh, uh, see, uh, direct improvement and change. I wonder if Trump was-- a few days ago, in an interview with Mask, he said that we need twice as much electricity for the AI in the country as, uh, The whole country of electricity. Oh, he says we're clear for the AI twice as much. He told Mask about it. The mask didn't tell him any comment on the subject.
Well, he said, like, "Well yes." Here. He said, " So we have to fight with China, electricity needs a lot. " So, together with the chips, what you're saying, you have to do these chips already, so that the electricity is so much.
Yes, electricity is needed, and you need to, uh-oh, it uses a special type of cab-- well, like a special s-s-- type of architecture of the data centers to get these videotapes to connect effectively. And this requires a large number of physical workers. Ah, yes, from all of these, uh, if you look from this perspective, I think America's getting desperately behind China. Why? Woo-- because China probably has a lot of cheaper., I think China is building a lot more nuclear power plants. If I am not mistaken, China will have more nuclear power stations before 2030 than any country in the world. But we need to check it out.
I'm gonna ask ChatGPT right now, or I'm not.
If you're on your first question you've said, what's it gonna be? I'm personally, if it's about my expectations, I'm waiting first to figure out the question of hallucinations. Because, uh, growing model size, yeah, it doesn't solve the question in principle with the answer, does it? He's going through anyway. Well, there's a thing like that, you know, the idea that, uh, LLM-ki hallucinates always. It's just that they get enough to get into the truth now, but how do they hallucinate or when they write the truth, as we think they work on the same algorithm, huh? And that is a matter of principle, if not resolved, there will still be a limit to the use of these, uh, technologies, right? And just a larger scale would result in, uh, a more expensive, uh, resources, electricity, more, but, accordingly, more, in order to answer the question. The result is still not 100%, is it? And I want to make a decision on this. If this matter is solved by some company, I think it's certainly gonna be a huge breakthrough even without an askeiling, yes, without some super-natural increase in the size of models.
Well, we're wondering, in the notes to this video, people wrote that the LLM series is hallucinating and the search is giving out the real data. And there's one man who wrote, "Who told you that the real data are being produced in search of the data? The search will simply be made of some resources that can also be given an absolute fairy. " Well, that's what happens often.
Yeah, but I have a point here, uh, that search is still taking data from resources. It's clear that there are now a lot of sites where content is generated with LLM, and some kind of hallucination just takes one step more. But if we say that we're looking for information from the website, it's a probable, producer, with some more information, then the information is, in fact, the information, from the first mouth that came. And when we say that LLM is answering, it's something that's missed through the model, through hallucinations and so on. And in that sense, I'm a little more confident in finding out if he's still there from the first-hand source.
So it depends on, uh, what c-- what exactly is the primary source. I mean, if someone on their website ratings, uh, movies or ratings, I don't know, clothes or houses, it's the same, uh,
Some kind of human model, yes, yes.
...unwanted than realistic-- than just taking naum. We were sitting here yesterday, watching the U.S. universities, and I was asking about ChatGPT, and he's giving me, and my nephew says, "Look, these are different ratings." And I understand that ChatGPT is giving something. I say, "Queen, take five different ratings and make them, there, some list." Because you come in when you're looking, but you've got different ratings at different times by different agencies, yes, very different parts.
So, Sas, did you look at us talking about, um, how's China gonna take AI?
Yeah. So they're much ahead of the number and power of the power stations. I've been throwing into our chat, uh, information to be able to, uh, add an info.
Well, that's an interesting thing. I mean, basically, we see how much we discussed in the episodes, uh, ToTheMoon about China's having enough serious technology in terms of, uh, visual-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g-g Well, there, or work with pictures. And perhaps it will be now, this year or next, not possible, but it will be guaranteed to introduce new models. Question: Whether they will be available for-- as ChatGPT, it is not available to all countries, but in some countries it limits itself. China will clearly use something exclusively inside its country. Well, there was a clear news and description that the government is putting the Chinese, uh, LLM in, right? And there's gonna be, uh, inside a system that, uh, censor everything there, at least, carry some agenda. Yeah, I mean, there's a LLM in the mood. Although, in my sense, ChatGPT is censored, it still censored. They censor too. They say, "We're limiting data, not giving, like, some kind of, I don't know... How to make a bomb. Well, that's right. I think it's right, too. Well, uh, the same thing seems to someone. So someone creates some rules, some rules that a person can give, that there is no way to give, what the volume is available, no one. And, in fact, every manufacturer will do it. Or the country will issue its laws and, uh, say, "We can do this." Well, that's right, what's going on between Euros and, I think, I think, just a little bit of a lot of people, right? And, uh, uh, uh, uh, XAI, and OpenAI, and Google and Apple, everybody with the Euros.
Especially with regard to, uh, use, uh, well, especially with regard to LLM, LLM. By the way, by the way, Google. So Google lost the court, um, quite a big deal about-- the Supreme Court about, uh, monopoly search, right? And there was an interesting figure that they, uh, had 90% of the market, yes, if I'm not mistaken. I don't know, I don't know which one of you looked, I didn't see. 90% of the market. Well, some big number was given. And, uh, significant. And that, in fact, nobody can enter this market at all because they pay huge money to everyone around, even Apple. There's tens of billions of dollars for, uh, being a little furious. And the Supreme Court said, " You can't do that. " And it almost says Google will have to sell some of the a-- part of the assets. So how are they gonna do that at all? Uh, considering they've made very much progress in AI. When? Yesterday. Yesterday, yes, there was a Google conference. At least we're in the middle of a-- um, at the time of the filming, there was a Google conference yesterday, and Google, uh, has moved again. He's made a new voice. Uh, he's released a new voice, uh, and it's pretty cool, as far as I can tell by the description, even though I'm not a super-hogle system. He's released a new, uh, new opportunity to work with the photos.
They showed it on the VR's usual cell. I don't know what it is, something AR, VR, and a change of image, right? I mean, they're moving very much. Uh, what's gonna happen, uh, and, uh, we were just doing a little probing about, uh, ChatGPT being really looking for it right now. Well, they're at least promising it. We don't know if it's gonna happen or not. But Google's doing a good update, right? Chrome was the last time we talked about the last time we were in the final edition. What they did in Chrome did, at least they did an update or it will, here. And now they're letting them inside their phones, and everyone says Pixel's a hot phone at Google.
Yes, Pixel is very-- Pixel is a very good phone. I'm impressed with their gay-- plan and editing of pictures and videos. It's really interesting. I wanted to talk about it a little bit. I mean, uh, that's why-- that's why I'm explaining to myself that TikTok, uh, became popular because they had three factors. The first is active spleon with creator, the second is coinciding for creators, the third is creative tools for creators. They've done all this and creative tools for the creators. The third paragraph is exactly what they did to CapCut. CapCut is getting more and more paid, and she's moving under-- subscription. And for Google, it's really an interesting opportunity, uh, to release this kind of thing, isn't it, like, a phony, right? I mean, the thing is, again, that generative AI and all these AI-Mechanisms of editing are expensive, uh, but there are companies that have a huge com--
There are companies that have a huge amount of cash and are ready to invest. And I can see that Google understands and moves actively. As for-- the tandem and training, which is as relevant as inference, i.e., the introduction of specific solutions for final, final user, for end-user. And that's very, uh-oh, great. Ah, I think they're being irrigated faster than Apple. I mean, I-- it's hard to compare them to OpenAI because the OpenAI has another release cycle. When you wanted, you showed it, and so on. So Google must still have to make his own phones.
And of course, it's clear that we don't even know what the operators are gonna be, who's gonna be the leader who's gonna do anything else. I think this year has proved. If you remember a year ago, as we discussed, uh-uh, there was no one there about Apple, so there's something else. We don't understand it. Mm-hmm, it's clear that, first of all, all the leaders are there, everyone is present. And generally, it's pretty good, at least what you can do, well, what's clear now is that you can do. It's clear that we're not discussing AGI, here. I'm gonna see if there's any new cameras coming up, I don't know. Only if some super-cup is, and I don't think there's anything that Amazon can do or not. So they got something in there. I'm not using it, by the way. Tanya and I were just discussing that Amazon in the US showed up, uh, at least they're AI. I don't use it. Tanya, you're using it?
I don't see, well, some big one was working in it. It's a very big tip, which I think it's hard not to use.
Well, anyway, you're gonna use your system long enough, huh? It doesn't matter who uses it separately. I don't know, I'm using ChatGPT, yeah. I keep using it to find, uh-oh, Amazon, you know, to deal with some kind of product or position, or some kind of search I can do unique things when you don't know what to find. I had to go a while ago, I had some basketball rings, and I had to find some kind of a basketball ring for one of those hooks, yeah. I couldn't just find this producer, and then I'm there, and I've sent all the photos to him, there, board, pole, all over it. And he says, "You have a built-in structure, there, one like that." And he found me. I got it. At least I was able to figure out, find and understand that I need to change, uh, all the ring, because it's not happening anymore. Here. Well, part, yeah, not all over it, not all the design, ch-part. But not just with that, that I'm having a grid. Here. That, um, while the tools are inside, Amazon, of course, that's not.
Yeah, well, for me at Amazon, so far, what's the most reliable is the feedback and the pictures, because I'm more of a visual. I mean, I don't know, I'm looking at the feedback from real people. The only thing that helps is when at the beginning of the description he makes a kind of summary of what people say, that's good, that's bad. But it's so much of aI. I can't say it's great.
Well, maybe they're using, yeah, maybe they can make a better rating with that.
I think one of the v-- of the most important stories in the labels is the ability to make better ratings. Iia because the rating, uh, is incredibly important and necessary. And in Amazon, the re-reviewing of the poa is very special... actually, that number and stars are a very hard story in the ratings. I've ever done enough ratings and visualizing these ratings. I was pushed by the zia in the first place, I remember I had Andrei Rogachev's partner, and he pushed me to see the Black mirror, I think season three, first series. And there's a future rating.
Possible danger--
And this future rating, he--
Prosecution--
Good gives a structure and understanding of what's happening, yes. I remember, we've been studying a lot of ways to create a real super-crunch rating. Ilnar, you and I seem to be doing a lot of work on this, too, yeah, about the ratings. I think Amazon, LLM, LLM system, and that functionality, it can certainly improve very much. Well, Tanya, you said it well, yes, the rating. Not only can a description be given, but it's more clearly a person to give a meaning, number or, perhaps, a set of parameters, which can be decided. It's still a product. Aah, it's just that you're not buying it personally, you're not buying it properly. Because what-- yeah, because it's working out that it's really strong, uh, depends. It's like you go to a restaurant, he's got a cool rating, super cool, and I don't like it. I understand it's just that this restaurant has some sort of crowded people that I'm not just a stranger to-- well, which I don't find. And what I like is that these people, too, well, I guess that I usually like food, a lot of people don't like it. Thank you very much, everybody. Ah, we're having a great graduation., by the way, in our past audience, we, uh-a, have had such a very interesting discussion, especially on the link that will be taking place with the technology market, taking into account changes in the structure The government, especially, well, is very influenced by the presidential elections in the United States. A-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a-a- So if you didn't see it, you'd have to look at this graduation, you know, uh, uh, and I'll see you on our channel on Sunday. Guys, thank you very much. We're always here four people. Bye, everyone!