Hello, everyone! We're on ToTheMoon. Ah, sites, technological sites, technological news worldwide from the Silicon Valley. We're going out once a week, Sunday. Oh, sign up for the canal, don't forget to sign up. Many of you are watching again. Sign up to keep the videos off. And plus you support the channel. She's got a very strong level, isn't she? I mean, it's a PHD compared to a man's intelligence level.
What's the question of the start-ups that are being made by the assistants who want to release the assistant for $30?
I'm sorry to waste these requests.
Oh, we're talking about a lot of innovation. And last time, people wrote about Apple's innovation. We had a graduation. We'll talk a lot today. Maybe we'll touch Apple again, OpenAI, and you'll have to leave your comments. We're also turning your minds into our next editions.
Well, I'm just gonna start with, of course, the OpenAI and their pepgreade, but I'm gonna start with that. So, a couple of days ago, the news is that Google released, and Google released the wallet that allows Google Wallet to add up here, uh, his passport. And, uh, to have a man traveling, well, that's their first, at least, movement this way. Electronically presenting a passport at the border, at least it will, uh, be in the US. There's a TSA system in the U.S., it allows you inside if you're registered. I'm, like, registered with her. Tanya and I were in Chicago, and when we came back to the airport, I went in that line, and I went with my wife in a minute. And Tanya and I were waiting ten minutes, right? Although it was a little bit of a turn. Yeah, that turn was a little bit, but it doesn't matter. The system allows it.
I didn't have to break up.
Yeah, the system allows you to stay on the belt, not to take off the jacket, not to get the laptop out of the suitcase. And very quickly, I mean, I have some I.D. I.D., there's a right or a certain system coming in, or a grin card or a passport check, anything. And he went on, uh, and went on. What do they want to do? They want to inject you... There's a system now-- anyway, you're scanning. They want you to have an electronic document, so you didn't produce any card. And on the TSA TSA, they'll start testing at some airports. I think it's a very good innovation. I'm like Apple user, I hope they'll do it, too, so, next, after Guglom. I think there's a clear interface, and at least I'm pretty light on it. Here. Although it's clear that there are fingerprints, facial recognition, there's no way that the rest of it will be cleaned. But the subject is very interesting.
Look, this is really fun. We've been discussing this guy, uh, instead of ID, there's a little pad to show, yeah, some bars. And, accordingly, some kind of lad can be paid, and there's no way to get any documents even on the border. When the border is crossed, everything is electronic. That's, of course, quite interesting.
And, uh, specifically for our commentators who wrote that these apple hamsters don't know, but Android has some kind of innovation, too. See, we're following. Google, Gemini has the opportunity, uh, in Wallet to add, uh, the identifier. So we're not just following Apple.
No, we're definitely not just talking about Apple. It just happened that we all have Apple phones. Someone's got a style, someone's got security, someone's got different innovations. But we're showing everything and, and more, yes. Not just Apple. I was wondering about the situation on Google yesterday. I was at high school at my high school at my high-school daughter's, uh, choir concert. And there was a man who made a pretty serious Google app that all the people who use Androids are using. And you're standing with him, he's talking about it. And then he says, "Well, you know what that is?" Then he even says, "You have Apple or Android?" I say, "I have Apple." He got his phone and showed me, yeah. Yeah, I ordered myself, ordered my daughter. Yeah. And I forgot. So Tanya and I are going to Chicago on Friday. Now. And Tanya says, "Have you ordered an iPhone?" I say, "Ops!" There.
I didn't recognize Sasha at this moment, that he forgot to order an iPhone.
And O1. And I'm the model, uh, like many, I think I've had the opportunity to test and work with her. I don't know if anyone has it on the devices or only partially. You, Elnar, have you got it?
Yeah, yeah, I did.
Yeah, everyone's got a show, I think they're on the devices. Here. And, of course, very interesting things. We... Whoever has seen or has no access to it and who's watching our channel. It's some kind of move, I think, some innovation, at least, from the viewpoint of the visual we see. Because you can, uh, see what she says, "I think I thought six seconds," and show the context of what she thought and thought, and what it was. This model has no access to-- it's trained, limited to a date now, and it has no Internet access, at least not linked to the model. From what I saw in my tests, because some of the boxes she couldn't do me. Well, but her level, she's got a very strong level. Yeah? So he's comparable to PhD, with the level of human intelligence PhD. I just don't know what level it is, how it's described. Yeah.
Sasha, what are you going to say? What kind of feeling were you feeling?
Yeah, it's really, really, I think we-- I think that a safe-- that's a really good moment. I mean, because the model's practice is, uh, well, just to say, it's three-steps. So if it's simple, it's pre-training, post-training, that's fine-tuning and like inference time. Inference time, well, most important, because it teaches the model as to think. To say, the utter-- key difference of the O1 model, from there, pre--- all previous models, that the model was fully-- with full reinforcement learning. The model has been improved by the results of some of your history. So when it's decided-- that's what she-- that's what the model was on the date of their training, there was a lot of data with specific problems. The c-- usually the solution to the problem is a few steps, right? I mean, what they're showing in the interface now. And when they were training, they'd find out at every step what the thought branches are right and what wrong. Yeah? I mean, it's like much more, it's kind of simple, you know, it's a new paradigm compared to what it was before. This is a new paradigm compared to just predicting the next token. Yeah? That's it. That's all LLM we've seen is the prediction of the next token. And there's really a hard way to differentiate. When-- now, really, it's very interesting whether OpenAI will be the leader or not. Now let me explain. I mean, if we're talking about LLM, that is, a lot of consumers, like oriented skies, for which Meta is making his open source Llama a model to make everyone feel Llamu, and Meta.
I took all the creams for my consumer skies. C, uh, with this, like full reinforcement learning and with this, and with this process in O1 a model that is called, and, like, a test time search, where it is in the process of generating decisions where she chooses many options. Such systems are obviously not suitable for a large consumer skies. The O1 model is therefore more expensive in OpenAI, so the number of requests is limited and so on. And I wonder if Meta would go this way or not because ow-- with a high probability of these models, which are likely to be high, it won't be high, even with a higher probability. open source models. Because now, in a rude way, in LLM, the open source model, at least in the west world, can be pos-- look, that's because the e-this models that, uh, can be resonant, they're demanding in there 100. I'm not sure if you're not counting anymore. I'm like a benef-- uh-this model, they're super helpful mainly in skies, and, uh, those are connected-- professional skies, like when you need a deep house.
And when you need a deep house-based examination, right? And, by the way, the current companies that produce open source the model big as Meta, they might be unsuitable. They may be advantageous to just stay, make LLM for consumer use case and not go into a deep reasoning that showed o1. Please describe what specific practical tasks o1 are better than all other systems. And the most interesting use of the case we didn't know about, we'll be covered in the
next issue.
Yeah. Well, in reinforcement, like learning, yeah, that's where it trains... Uh, anyway, uh, uh, well, as always necessary, a pro-optic model that will test the results of the first model. It could be the same-- it could be like two instances of the same model. It could be one big model, another little one or the other. I mean, there are different methods. But basically, uh, I think that-- I think it doesn't matter whether it's called Strawberry or not Strawberry. I think it's important that we see how full we can see, we see a model that's designed to reasoning with reinforcement learning. I mean, she's, like, a model that's known to learn how to do her, and she's kind of learning to do more difficult things, and she's learned to think about step by step. I mean...
It's happening.
It's a big deal in her practice. It's different from all the LLM models we've seen to date.
It turns out that's a very strong progress. And plus progress in terms of the issuance of content and the issuance of information and decisions. There's no other way. Plus, the addition of the checks, what we said, what gives a general opportunity to exclude hallucinations. And, in fact, we'll have to get a tool in the course of the year, which, in the idea, should automatically be between these models. Because it's not likely that any person will be easily chosen between o-- between ChatGPT-4o, ChatGPT-4o mini, o1, o1 mini, there, I don't know, launch a system of verification. I mean, it's not gonna be easy to switch between this.
I guess one of the following developments is one that should be when the system should determine which model is automatically delivered at a given time, right?
100 percent, I totally agree. I mean, I think that's what it means by some verticalization process. So until now, all LLM models were about the same, but even we saw it as if some models were better in one. Now, as model segregation, it's gonna increase. I mean, as many different models, it's gonna increase. I mean, we see that because this model is now, it's like GPT-3, but in the world, these R-R-R-R-L-trained models, right? So, in the world, RL-trained models are the first model, and they'll be much better over time. But we're not gonna see that GPU is-- but we still say that the number of requests is 100 times greater, hundreds more. I mean, and how fast the GPU can't be, there, the GPU won't be able to process 100 times more as a call, yes, 100 times more tokens. And so it's gonna be a little-- how much RL-trained models that are small, will be RL-trained models that are bigger, and they'll be bigger models, respectively, they'll be quite a bit more. expensive. So there's probably a few dollars worth of queries. It will be the great, huge role of the end-product developers, to help select the right model depending on the use of the case. Before that, we were talking, just before that, we were kind of-- before that, actually, it wasn't that important, because the quality of all the models there, Gemini, OpenAI, Claude, was about the same in terms of the generation. In terms of text generation, they were all about one level, and at least three times, roughly, the frontier of the model was different at the price. And now we'll see, and, a significant thing that big-- we'll see that models, those that can be resonably resonated, RL-trained models, they'll be worth 100 times more expensive, and they'll need to be there, and they'll need a more delicate one. I'm not sure how to do this. It'll take a much more delicate configuration. And it's gonna be like this, it'll make it as if it's so new vertical AI native starttaps, and maybe give a value to the market that they'll be, uh, doing both prompt engineering and some fine-tuning. Until it's not clear how fine-tuning looks for o1, but somehow, some possibility of caste-to-be will still be. And you-- you mean, yeah, that's how to choose a model, do some prompt engineering or fine-tuning, that's gonna lie on the shoulders of those companies that are developing such more vertical solutions.
Well, the idea is, they should, if they automatically do some kind of system of inter-model switching, you might have some extra necklace at some point in time. Do you want your information to be further checked? Or are you willing to wait, for example, or do you want to get a little more precision? I mean, from the point of view of the ordinary concuits. It's clear that B2B is a little different story.
There's a PI, a lot of lines and details. I'm talking about this mass-market, from the perspective of the usual consumers. When I'm told there, spend more of this money, you'll get, uh, some extra power or a chance to make further conclusions, or load more data, or you'll get some more data. More in-depth analysis. I'm gonna have money somewhere, they're gonna be, like, c-- they're gonna be writing it down, yeah, some kind of model. Although maybe as you say, there's gonna be a new tariff, yeah. It's just in consumers, B2C, it's unlikely that a tariff can be directly above the high cost, right above the big one. I mean, mass-market, yeah.
Well, I think it's critical for OpenAI. I mean, the news is, it's just been out a month, and it's like, a month ago, it was just, like, a new thing, that the monthly signature would cost $2,000,000 at OpenAI. I'd actually say that, well, it's obviously like-- they're looking clearly towards personal assistants, I think they're so advanced. And I think it's a very interesting market for them. I mean, personal assistants, that in the States, which are not in the States, they make more, you know, English-speaking markets and the CEC that work, earn more than 2,000 a month. I know the assistants have a problem that they're going to be as good as being pedgrimed, the good assistants are fast-growing and going to a bo--- a much higher-paid job. And the assistant who only gets cheaper with time, that's a very, I think, interesting, interesting proposal. I mean, I think they have a big sweat-pants here, and that's their-- and I-- why am I saying that? That it is critical for them to differentiate from the rest of the competitors. Let me just tell you what I mean here. Yet for Meta, for example, it is important for them to build a really big AI assistant in the world and how important it is for them to adopt, not to lose their existing business. And they don't have to go to very expensive, heavy models. If we're talking about Google, they never sold to the consumers something very expensive. So the only thing they sell to the consumers is that they're their last Pixel phones, but they don't sell a lot of them either. Because, after all, to sell something expensive to consumers, it's like a feeling bonus, some kind of marketing, like Apple. OpenAI may be able to achieve this, and for them it is the key way to differentiate from the others. I think Anthropic has no DNA to build such appealing as a consumer product. So they can build, disarm-- like a developer platform Anthropic can do, they can make a good decision for programmers, but I think that OpenAI can be like Apple, but like in the world, I don't know, there, with some consumer hardware. Apple is like a bonus decision, right? The segment and the OpenAI, I think, in the consumerium. OpenAI could be a consumer award in AI.
I think so, well, I think it's a very good bet on their side, considering they already have a very strong consumer brand.
You mean, you say, uh, a mass-market bonus, right? We're talking about mass-markets in this case.
Clearly, the mass-market, considering that more than half of the US is, uh, their operating system.
Yeah, I think from this point on, I think the OpenAI might be able to replicate their success, yes. Well, let's just say, the bo-more number of people who work with information, that's not really-- not everything, not the entire American population, is it? But all those who work with information and computers, that most of them in this group of people in America will be, in a rude way, OpenAI, and they'll pay them, you know, between $200 and $20,000 a month, yeah. Well, let's just say more than a thousand, more than a thousand...
Well, let's just say more than a thousand, more than a thousand dollars a year, let's just say.
Oh, between two and two-- to two thousand dollars a month. Oh, you mean a year.
But a man must be clear in 2,000. It doesn't mean it's unreasonable, but it has to get a very strong value. Why don't I know it's unreasonable? Two thousand people might seem to be numbers, uh, very large. I remember when Apple was giving out glasses, someone said the glasses were very expensive. Three thousand, there, dollars. But the glasses could cost, uh, $30,000. They could be worth a car. If this system, which replaces a man, uh, fifty percent of his life, creates an interface in his apartment, a chance to communicate with other people and with all others. Well, what we see in some movies. Again, I'm talking about some abstraction of the future. They could cost $100 thousand, yes. And, and a man buys them, there, on credit for his life, and just buys them sometime. So, $2000 could be either a thousand or three thousand dollars reasonable. Question is, what are the start-ups that are being created by the assistants who want to release the assistant for $30? So, in Sas, the reason you give me is that it's impossible to release a personal assistant or replace a staff member who will be worth some small money. And a good moment. You said that, in fact, personal assistants cost more. But I think if we go into the phase that AI the assistant will, uh, generally win the quality of the assistant of the present, the question will not be who costs or costs, but the question as, as a matter of course, That's it.
I mean, if I have a AI assistant doing quality, uh, no worse than a personal person, I'm willing to pay for it and more, because with AI as assistant at least--
Yeah, but it's on its own, you know, yeah, I mean...
I have twenty-four seven, twenty-four to seven opportunities to interact. I'm not limited to his dismissal, illness and a huge number of other questions. I mean, I'm willing to pay more money for it, not less.
Yeah, but that's interesting, too. I mean, Apple, he's kind of a modern-day camera-diggist, as I understand, and they're worth about as much as an iPhone, maybe a little more expensive, but, like, the popularity of the cameras is pretty much... I fell, didn't I? B-- same thing with the media. There were mediocratics who gave quality better. I mean, it's important, besides giving a good enough operation, it's important to build this consumer brand. And I wanted to praise or celebrate the OpenAI. I'm not sure they need this, but, anyway, out of all the AI companies, I think they're the only ones who built a strong consumer brand. The others obviously didn't make it right now. None of the other companies have any kind of consumer companies-- AI companies have been able to build consumer brands.
Who's the closest to them right now, because you're so comparing?
Yeah, well, I think it's Tammum, I think it's, frankly, even as it is, and it's a huge difference that, yeah, even that one doesn't have to talk about it, if it's honest. I'm here now, um... I mean, it's on the OpenAI that he's like, I think that by the number, they have a lot of traffic we see, they have this organic. And, well, Gemini might be trying to catch them, but I think the difference is still a few times.
It's an interesting thing. Uh, here we are, looking at a lot of people. And you must sign up for the channel to keep your skates and lick, write your comments. You are very supportive of us and of the development of the channel.
This little action, we'll be grateful for it. I mean, there are people who write that there are other systems besides ChatGPT, yes. And our job is not to make sure that everyone is absolutely to use ChatGPT. We're definitely not OpenAI marketing agents. Our task is to show the many different ties and opportunities that this system offers and will give in the future. Give this new vision of this daily application. And I guess we've always been doing the right bet, which we've been talking about OpenAI more than all the other models. Although we've been talking about other models, because why tell us about who's behind, right? Or why tell me about someone who doesn't have that big power. And if we're talking about Apple showing as the best history of integration-- integration with a company that makes models, she says, "I'm integrating ChatGPT now as a basic story." She also says, "I'm looking for Google integration." She doesn't say I'm int-- they're probably intact with AI and Google, too. The question is that ChatGPT is number one, yes, in the discussion of details, yes.
Well, here I would add that, uh, actually, the last month-and-a-half, I probably went from ChatGPT to Claude because with the advent of 3.5 Sonnet, I can say that in the many tasks he was in me. I need, uh, OpenAI's gone down. I was so curious to test what a model might be. Because, for example, there's a simple crypt, you know, in terms of some automation, there, basic programming, and, like, Anthropic, I think, even more and more. The quality was shown. What happens next? I wonder if I can see.
I wonder what a comparison you'll make. So that's what Claude did, yes, you and what ChatGPT didn't do, and what it is now that ChatGPT o1 can do what Claude did or what Claude couldn't do. I mean, considering that if you're talking about things that concern, there, math, programming, science, they're
still oriented on one model, right?
Yeah, there's two guns. Aaah, a little tour of childhood. You may have had the same thing when we went to sea as a kid, there was a camera and there were, like, thirty-six skilled gia out of the 30s, thirty-two, probably the last day, 'Cause all the previous days you're just saving and you think, "No, you have to leave, don't waste." And you're only gonna start taking pictures on the last day. And when O1 showed up, with a 30-man-week-limited model, I said I was a feeling. I'm sorry to waste these requests. I think, "Well, I'll do it on the usual. "I'm still here, "well, no, well, it's not that request to spend on O1." Here. Aah, and that's the same thing. I have a tour for childhood today. There was a series of Simpsons, um, somewhere deep in my childhood, somewhere in school, where Homer and his neighbour went to Nepal, to the mountains to ask questions to the wise man. And they're halfway up the mountains. A lot of deprivations, a lot of pain. And there are only three questions that can be asked. Well, Homer comes in first and says, "You're the wisest person?" - "Yes." - "And you can answer any questions?" - "Yes. - "And you're not lying anywhere?" - "No." All three questions asked.
Let's go home. And now that I see this O1 with restrictions, there's, like, 30 questions, and then this limit can be either smaller or more, I'm always waking up these associations. I'm always a little sorry to waste these requests. It's one story. And the second story is about how I see it technically. Oh, on those boxes I've tried, which I haven't had a bad time to ask O1 for, the feeling that I've just given control, and, uh, a GPT chat dialogue to another algorithm. So we could have asked a question before, then say that here, it seems you did something inaccurate. Or, yeah, think about some strategy first. Like, we're in there, we're gonna do some kind of little project. We say, make a plan for this project, then, in a phased manner, drop out requests, process each and collect some sort of common document. Well, spending it, uh, not one request, but, uh, there's ten requests and spending some time with it, some attention. Now it looks like this job that could be done on its own is just a little bit of an algorithm that is still being further trained. Uh, on the one hand, it's a step forward, right?
On the other hand, remember, you and I were talking about the concept of the black box, yes, when we don't understand how the answers are coming. It just happens here that the black box is put in another black box, yes, there, and it's where you're not aware of how it works, and you're going to move on, and you're still on the line of, uh, the process itself. Not good, not bad, because, well, like in the car, when we're driving, it's obviously not the same car Henry Ford made up, yeah, when we were all in sight and we were all in the hands. Here. But somehow, it's a sense of losing control, it's here, too. But I'll repeat myself again until I feel like it's rocket science and everything, there, the industry has changed. I wonder what the Antropic will have to do. Uh, I'm gonna get a little bit of a drink for them today because, remember, we were just talking about the key staff in Antropic leaving. Claude, in general, I think it was a few months ago that showed better results. How the empire strikes back. Let's see what Claude and I have.
Sasha, it's still, uh, just adding algorithm. That's how Elnar wrote that a combination of certain requests is summarised or
Certain inquiries are summarised or something different from the structure, uh, not just the generation, as you said, uh, next token, some parallel flows, and so on. Because it's hard to deal with this here, more like, "simple" with simple words.
What is it, if you believe OpenAI and their statements, is the first model, that is designed to be able to meet the challenges of a few steps and take the right action on each of the steps. to meet the challenges, yes, consistently. And we don't know, of course, from the point-- and we don't really know why O-- why they didn't put O1 in a big model. I mean, if they have small models, they should probably have a big
model.
I-- well, I have a feeling, too. I don't think the OpenAI did that. A stupid read code that'll be very close. And I think that at least what they're describing, and even though they called her, it's something else, yeah. So they and, uh, other symbols were placed in the name of the model. They didn't make ChatGPT, uh, 4B, C, D, there, E, there, anything, yeah. I mean, y-- or there's ChatGPT 4+. Uh, they've created some new answer. And it's probably just a response to make a man choose what tasks different. Although again, we'll see, yes, because Elnar, er, doubt or skepsy, or conclusion, well, absolutely adequate. We never know. Maybe that's three years, yes, it's gonna happen. Like I think there's a limited number of requests. Many of the things that the company has said, for example, are not done, are not done or are not released. We still don't see Sora on video-generation, we still don't see the search on the Internet that was said, uh, and they just don't smell. Maybe he'll be released on October 1st. Sasha said today that on 1 October, uh, OpenAI should tell you something. And that's what he's gonna say? Yeah, about B2B or is he gonna tell you about mass market B2C? That he'll-- let you out, we'll see, yeah.
Uh, thank you all very much. Ah. ToTheMoon is technological sites and news from the Silicon Valley and the world. We're always four people, we're going out on Sunday. Uh, wait for you at the next edition. Bye.