What about, uh, movies? How realistic is that?
The same Sorah could be used in the future. It's not like you're making a full video from scratch, it's for editing.
No, I'm not ready for that. Let someone else test him, there, a couple of years. Uh, that he's out there, you know, not going to sleep at night, not killing his masters, uh, not breaking into other cars.
They get a package of five to ten million a year, they're very cool.
Hello, everybody!
Uh, we're on ToTheMoon subcategory, where we're gathering technological news, technological news from the Silicon Valley, technological news from the world. We're always four people. A great deal of emphasis in artificial intelligence and in AI. And there were three episodes, and it makes sense to see them, because it's not about getting news for a specific week, and it's about watching the analysis, opinions, in the websites, the technology world. And today is where I want to start.
Sasha, you were telling me about this incredible internal feeling about what you saw last time, uh, how the OpenAI Sora model generated video, and you saw it in real mode. It's like a movie, yeah, on the cool screens. And they've released a skate this week, uh, maybe there's a show. The TED-related roll. Ia is so super-some video incredibly made. That's what I'd like to say. I'm in Rome now, and last week, I was in Dubai, and, uh, there's a... a rat-- dogs in Rome. And it was interesting that my buddy was alone, and I told him when I told him about it, and that's the proo of creating a video with artificial intelligence from the proms, he said... I mean, we'd-- we'd go to dinner, and he'd say, "You know, soon, movies would be, like, generating real with artificial intelligence. All of them! It's a new future, a new world." I thought you'd imagine how much resources you need, uh, and that it should be for a superb system that could think about the story, a few hours. Because now a man writes a prom and he's still being generated, and he's incredibly qualitative, super cool. And it turns out that on the one hand, there's this vau you had on the other hand, that I don't have a personal feeling that the world is even ready, uh, to create a real video of two hours. I want you to be inside, again, video production, I think it's a very strong technological movement in the world. What do you think about that?
Yeah, I saw that video, it's impressive. I also heard a stove from the studio that used Sorah. And it's clear that, really, uh, here-- that there was a huge amount of post-sale sales to get those videos, right? I mean, it's not-- what it is, um, like a lot of hand-work on the clinap and a big-- and external clouds like Topaz for the pscayling. That's the first thing to say. So we don't know how good it is, uh-a, Sora's model, and I think there's a big problem with control. Well, at least between the lines, it's been said that there's a big problem with control and that it's probably not a tomb for, like, everyday, I don't know, a content maker, to
be rude.
Sasha, what do you mean there's a problem with control?
Usually, to produce, produce, er, picture or video, a long prom must be made, and often this prom may be fifty, fifty words. Uh, and, uh, well, there's a ha-hayda out there, I don't know, almost there's a fifty-six pages for each model, how to properly blur the proms under different scenarios. It looks like Sorah's problem is, um, not solved, and you need to be able to get past the model. This in fact entails the creation of a separate human pool that is promt-engineered. And around-- and there's a very big question: how will they be integrated into the business processes? It'll be special, it'll be agencies or it'll be a mark-up, or it'll be independent like prom-engineers, who are just, the brands will be hired through, uh, I don't know, instagram, when they're in. or TikTok. Uh, as long as it doesn't work, how economically it works. The reason-- and now it's very, yes, it's hard to talk about Sorah. So, you can assume that, uh-a, the cost of creating a content is still cheaper because, uh, it's kind of a-- it's a replacement for the sales phase directly, right? So if production-day, cost tens of thousands of dollars, I assume that Sora is more economically profitable, but it's not a substitute for post-producation. Of course, at-- at my point, uh, the winners are the opt-designers who are good, who will learn to bleed the model because they will now have much more opportunities than before, right? So people who have a creative vision and who learn to smash the model, they can make movies, short films now, maybe from un-successful, non-sexual trailers. - what they couldn't do before. It's gonna be a big ankle. Uh-huh.
Oh-- well, so far, unfortunately, there are not enough success stories.
Sasha, please tell me, in the computer capacity, right? That's if we're used to what we have, there's ChatGPT, Claude or some chat-bots that can start a prompt and get a real answer in real time, with videogeneration, right? Sora, there, there, there, there, there's no way we're gonna be tested. But your assessment is, how much it is, well, it can actually be like the experience we have with the image generation, with the text generation, with the video generation, right? In terms of how accessible this is, let's say, simple death, not large Adobe companies or anyone else.
That's a very good question. I can only speculate here. So, it's just a costly, a video, a one-minute video, a quality sora, it's probably gonna cost you a few dozen dollars. Let's round it up, maybe $100, because we don't know how many iterations. And it's not really abundant, you know, it's not really a wide-ranging circle of people that's gonna be accessible, honestly, right? Uh, I mean, uh, actually, $100 a minute video is expensive. It's more expensive than Shutterstock, for example. But as a visual picture is gonna be much better, and a special-- and you can be imploded under the skirt, but it's gonna be expensive. And so now, uh, not so-- now that's not so obvious, c-- is it, for example, useful for-- is the Sora model useful for small businesses, for example. Because if $100 is a minute of chronometering, and that's, for example, um, enough to cut some shorts, you need to spend a month on one Sora model, maybe a few thousand dollars.
And not all of this can be economically profitable, too.
I understand correctly from the point of view, that the requirements for computing capacity are that if we take a minute skate, uh, a probable 30 personnel per second, it'll be 1,800 personnel. Uh, is this comparable to the generation of this video and the generation, there, 1,800 or, uh, 1,500 images? Or is there any other ends in the video generators, not the generation of individual personnel and the video-wrapping?
I mean, you can tell that you need to generate backups, uh, maybe ten, that's, like, 10 francs a second, and then interpolate. Otherwise it's not gonna be economically viable, yes. Otherwise it's not gonna be economically viable. Ah, and I think that a lot of work will be done with this-- a lot of work that's related to it.
How is it really gonna be a big movie, Sasha? Now, with the movie industry, that's the question I asked at the beginning, and what people would be doing. Because when people see some sort of generosity video, they're just the end result of the spaceship seeing. So all the movies are made of artificial intelligence, right? It's how many people talk, so that, uh, a huge amount of robots or artificial intelligence will do, uh, just a lot of work for everyone, and people won't work. What about, uh, movies? How realistic is that not just to generode...
How realistic a two-hour prompot is, sequence, some story, this scenario, next to it, the generation is real, working, living, that supports all these lines, keeps the first. The minute-- 30 minutes corresponds to the first and so on.
Yeah, I think it's impossible because just even the best LLM doesn't have a long context length to support-- how to keep it in context. I mean, the methods, like, the vector bases, yeah, the vector databases and RAG that are,
G, which, no matter how, store all data directly in the context, and which store them in a special database. But I think, um, we're far enough from that. I think the next step is that we'll see the new content formats for a few minutes with AI generated, like-- where everyone where each content unit, uh-a, is worth a few thousand dollars in cost. The quality of this video is better than everything you're available on YouTube now, right? So it's on the picture that's probably better than even making MrBeast, for example, right? It's gonna be better than the most expensive, um, clips like Drake or Cardi B. I mean, I think that's the new type of content, and it's definitely gonna turn up and probably the musicians are the first to be the... So it looks like musicians are probably gonna get a little horny.
Musicians, you mean they have short, short lengths, songs?
Short song length, yes, because it's like an extra bone for them, right? I mean, like, do the video, yeah, on, on their song. And I think it might be, uh, so interesting, that might be such an interesting opportunity. Well, I understand that, too, if we're talking about a hierarchy-- that's for writers, for them like a hierarchy, it's probably starting with social media content,
then it goes so--- It's like a music video, and then it's a movie arthouse, and then it's like, rudely, a-a-a-a-a-a-a-a-hallywood, right? And I mean, I think that there's gonna be a adoption down the stairs, and we can see a lot of AI generated music video, in blatant words, right? So, music videos like that.
But AI generated music video, where the video video video video video AI generated, and the song is real with the real man, right?
Yeah, I think it's gonna be a lot of trouble, you know, because artists still need to sell their identity, right? It is from the point of view of the narrrativism that their voices may be. And the Vi-- the visa is very often, very often the outsourcing visa is somehow different, right?
Yeah, well, I don't know if I'm going to a virtual artist concert, paying as much money as I can pay for it, going to the living.
And that's still moving on the other side. There's more news about Adobe Premiere. What was a little earlier, remember the news that Adobe Photoshop is allowed to separate parts, uh, pictures, some photo, write some prompt and replace the background, add some details, you know, in, The image, the kind of image they look like, uh, are preparing to release in Adobe Premiere, yes, in an annex that allows, uh, editing videos that allow editing and also using, Uh, Generic models. In particular, there was a demonstration that the same Sora could be used in the future. It's not like you're making a full video from scratch, it's for editing. For example, removing a video cell, um, or, on the contrary, adding this element in a video. And here, for example, before the full-blown generation, there's two-hour blockbusters, we're probably really far away, so it's like if it's not too expensive, we're not really far away. It's getting old before that, right? When people who do edit, can not build a human, perhaps so perfect, not looking for a location that is so perfect, they're gonna have to do it now, but with that kind of thing.
Generic models edit what is the result.
Well, according to what Sasha just said, if a lot of people start using it, it's gonna be a lot of expensive for these companies. Because obviously, now the OpenAI, well, there, ChatGPT is a loss, it works a lot, and I'm already, and-- and all the other absolute tools, they all work in losses. I think I, uh, my subscription at ChatGPT, I'm always kidding, I'm eating it in about ten minutes of the month when I have it working my requests with certain ones. Ii... And then I just use it. Just, shorter than more people will use it, especially at the level of video-generation or image generation, or if you look at the Adobe Premiere, it's the editing, uh, big, large video installation. Like, I'm on YouTube or we're on YouTube editing machine doing a video installation in Adobe Premiere. It's a crazy cost, some kind of model, just crazy crutches. Not who you pay for, but who the Hungarian pays for you, which the system gave you to keep people in it and to keep no one out of the market. And from what it is, uh, what news is going on now, even in the video, what Microsoft has presented a model that, based on audio-trave photographs, makes a realistic video in real time, right? There, Adobe Research published an article on Video-GigaGEN model that allows for increased video approval. Object Drop from Google is a model that allows editing of sgenerated images, and then, respectively, to work with them. That's the endless, huge co--- everything, just going into the video, everything goes into the images. I was last saying that, uh, Meta in WhatsApp released the LLaMA, uh, instantly, images. Speed is very fast. It's crazy computing power. They're worth a lot of money. It's all done in parallel. I want to discuss this.
The silicon valley is a little bit, given that, uh, three of us are in the Silicon Valley, and Sasha is very tight in terms of engineering inside. Ah. There's a lot, uh, there's a trend, but at least there's some evidence that a large number of very strong engineers are leaving large companies, opening up their start-ups. I mean, I was dating a man last week in Vienna. No one has an outsourcing, three and a half thousand, uh, programmers from Eastern Europe to America. And he says that the problem of the start-ups means we don't need our programmers because the supposedly strong engineers are now starting to do the start-ups. Jacobs, I don't know, it's his hypothesis, right? These strong engineers can themselves find and hire individual development teams. How does this big company exit even mean strong engineers, do they have any impact? And how much is it possible that strong engineers make start-ups that we'll get a new era of super cool, little, well, new projects?
First, I'm gonna say that, yes, that very cool engineers, they get five to ten million a year in the bags. Strong, huh? I mean, they're not that many. AI-Gogle engineers or OpenAI--- uh, that's right top of the top, right? And they probably don't have much motivation to do a start-up because, you know, like making a start-up, getting a better outcome than that, oh-so hard. Uh, that's probably a first-serve. Second observation that it is true that now, very much loves the fands of AI engineers, but I think that technology in any business is not more than 50% of success. There are sales, there's marketing, there's a customer success of the function that, uh, someone who usually doesn't know AI engineers very well and understand. And that's why it's not the fact that this methodology will lead to new companies, is it? I mean, uh, marketing, sales, maybe marketing, sales and customer success are slightly less important in B2C. Maybe in general. Maybe. But the view-- but it doesn't show up like-- but it doesn't show any good B2C product from generative AI. Except ChatGPT, yes, which everyone uses. And besides that, I don't know any product that has such a broad adoption.
It is therefore very difficult for me to assume that it is precisely that these start-ups will show a very large number of successful businesses.
Well, look, the same Claude, for example, if I remember correctly, he, uh, uh, organized by the OpenAI and DeepMind, a lot of engineers left, opened up their companies. Of course, they're so horny, Unrecognized, achieved or not achieved. But the publication Sasha referred to, uh, was the following motivation, yes, in a big company, there's the same Google DeepMind, yes, there, or somewhere else, that's a little hard to get her innovations before they get to the bottom. market, because it's a big bureaucracy, complicated processes. Uh, the realisation, the market entry, it's a separate big bow. And now that you're making your own start-up, and you're the only one who runs it, and you're the only one who makes decisions like you can do it faster. I don't know if these people lose money, whether they lose money or, for example, whether they're able to make a company. But the publication had a note that this movement was going on, right? And I wonder how much it is, in general, mass, right? Is there any of your familiar people who, uh, have made such movements?
There's a really big subject here, actually. Ah. Big subject. Aah, what's the matter? Does it affect the big--
It's a big deal. Does this affect large companies and is there any real new ones?
Because it's a dad-- still start your start-up, it's a separate business to do. And there's still marketing and sales. That's how Sasha said about B2C's success and failure. I'll start the next topic. Uh, I want to discuss and, uh, I think she's missing, uh, in the area of people's understanding. Let's just talk about open source. So, there's an open source and, uh, Facebook, it says, "We have an open source." Mixtral-8 representatives launched another open source. Microsoft released Phi-3 at GPT-3.5 open source. Let's get a little bit clear. I don't understand personally. Open source, is it accessible to all or open source restricted? For example, an American open source can all countries use or all companies can use, or is it open source conditional?
It's clear there are different types of individuals-- uh, open source of a license like MIT and so on., yeah, but, well, often, caste licenses are used. So Lama has a little caste license to make Snapchat, for example, unusable. So there's a license she's detecting, she's putting a kind of restriction on the number of users so that Snapchat can't use these models. Uh, and, uh, Stability has a pretty unusual license, too. I think that open source exists at different levels, right? I mean, it's like a burial source is just a weight. And it's kind of fun to call open source, but it's not a real open source. A real open source is probably when you can get the way from start to end, which includes the model practice code and the dateset. And with the dates, the biggest problem, huh? I mean, I think people might not understand, that's what, on the one hand, is, like, a lot of, uh, claims. So the New York Times are judging OpenAI and so on. But the OpenAI was all in the process... trained them all our models on these dignitary data. These models are already being used actively, and a large number of people are already actively using a large number of years. And up, that's what OpenAI's got, like, a lot of intellectual property done, right? That, and that's it, this mob is no longer possible to turn around. E-I mean, on the one hand, there may be no high-quality dates with copyrights, but on the other hand, products built on these dates cannot be deployed. The adoption is very high. Even if New York Times wins the court, well, it's about some revenue share. But somehow, as copyright data are already used actively to practice models, and we're all kinds of digestive.
Sasha, please, please, when you're talking about unloading weight, is that what you can do with this routine? In the words, uh, Ilon Mac, remember, we were discussing something, saying Grok, in the open source, released the weight of our basic model. What does that mean for a regular custodian? So you can download these weights, download the model itself and just start it on your server. And that's not the way she works. Yeah, it's another thing you don't know how she was trained, how those weights were received, you can't get her to help her, there, underneath your needs. But using this model somewhere on your private server, you can. Am I right?
Yeah, you're smo-- if there's weight, you can start models at your place, yeah.
But if there's no detail about how she was trained, it might be hard to change it.
Can I, for example, figure out how a normal person might use it? Well, that's just any example.
Yeah, sure. Look, I mean, like, Databricks for-- Zaopen sourced a level model, uh, GPT-3.5 or Phi-3. Phi-3 came out of a kind of Microsoft, didn't he? And let's say, business, small business may have, uh, a story of talking to their interview-- their clients in some sort of way, and maybe some kind of goidlines, like, hanging out with them. How to act in case of disagreement, for example, is it? Well, there are a lot of ways to deal with customers' differences. And it's actually a prompot ingining level, and it's possible to decide at the level of the pinetuning. And in principle, for language models, there are a number of services that provide, as convenient, UI-definitions to make this story, right? To tell you, to explain the models that she needs a genery-- that when you ask her clients, you need to generate a kind of kind like that-- that kind of proposal, yes, in a rude way. I'm sorry you've got some bad e-spionage, so we're ready to offer you a discount. Okay, so, for example, in Yandex Taxi, for example, it was always like they used to give promocodes in any failed expirinance. There are other businesses that are acting differently. And these... And the linening, like a pos-- when, I mean, when there's a sa-sama model and there's a model practice code, that's enough-- that's the services that allow-- that show the services that make it possible to do. The pinetuning and, for example, the caste-- so that, for example, the customer can support in its way of doing business with the customers, how they work with-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-s-
Tanya, you might have a company that doesn't want or can't share its data outside, right? You have some personal information on your clients and send her to the Openai or send her to Anthropic or send her somewhere else, which might actually be a violation of the law. I see there's something to be encrypted, something else to do with it. But these models that you can turn around, they let you shut down the circuits, yes, and all the transactions, all the calculations, that's, uh, on the implication of your computer.
And then you can work with it peacefully. It could be medical records, credit data, some other information you can't release out. And here you're getting a GPT 3.5 model that's all yours, and it's not the data for its contours.
What I think is that many people underestimate that the largest amount of revenue is generated by the start-ups that provide consulting services, which are helping to plug these pips inside. Business, which, for example, makes a simple stap to start, to start any business, and a small business, and a big one, could start fast. And then they sell the very expensive consulting services to caste it. Oh, uh, and now I'm gonna try to find, and I'm gonna try to find, uh, news, but Accenture, that's very popular, I think it's a very popular consulting company, they've earned $600 million. - I'm not going to get a quarter. And in fact, my assessment, I can be mistaken here that all the other generative AI start--- we're gonna get OpenAI and, uh, Google and stuff. But if we take the start-ups, they've earned a comparable amount of money, right? So it's hundreds of millions of dollars in total revenue for all the other generative AI starttaps. All, all, all, all, all taken together. And half of that revenue was also at the consulting service, right? I mean, basically, the generative AI start-ups I'm trying to say, they're kind of competing with professional constables now. That's how I see the market. So when it comes to the fact that the start-up needs to learn to make more than 10 million revenue, it's likely that we need to go to B2B client and consult-- compete with Accenture. Because business needs, for example, to do a picture analysis, a video analysis. And they understand that, especially pictures, very often contain such confidential information. And it's scary to send data somewhere in OpenAI because suddenly Sam decides to leave the company and the company's development vector is changing, and they're gonna start re-selling the data. We don't know what's going to happen to the OpenAI. And so many businesses need to get this image analysis process back on the way, and, uh, big models inside, inside their company. And the question then goes, working with the start-up, or working as a-- with the verified fighters, like Accenture, like Deloitte, that is, those who can work very short, those who understand, I'm not sure how to work in a consulting room.
And I think corporations often have such questions.
I'm listening, and I understand that this is the idea that there is a GPT-4 probable, and it's trained, and inside there's just a huge number of wired dates through it and data, and there's always a breakthrough and development, and Education, training, education and development. And on the other hand, someone is, like, in my own company, and, uh, you start training on your data. In fact, you're learning such an an ant, and they're starting to build a spaceship and a conditional break-in or future appearance, I don't know, there, GPT-5 or, there, AGI, or something, it's actually, Re-transfers the market because what can teach small businesses or even medium-sized businesses is totally uncomparable with-
Absolutely not compared to ur-- with the dates of the nets that teach OpenAI their models, it is simply not comparable at all. It's just a different world. And a very cool point is, when I'm using, there, GPT-4 or whatever, any completely lu-- from a major corporation to, uh, sys-- LLM-ku, I, uh, still get a conclusion based on some kind of corporation. space quantity of data that is still reverified by the space number of people. I mean, it's just some other story. And according to what the global change in the I-- in the I-, I don't know, two years, there's a huge amount of exercise that people are doing now, uh, to teach their own models, it's, if, Honestly, it's going nowhere. If the business is to see it, it could be nowhere, I'll say it could be right in nowhere. And, of course, in the view of the idea, something should happen and come up with some new mechanism, or it would change the whole thing, would change the market very much. Or a large number of companies will suffer very much and die very much. Because what I hear from those people, companies that pay, uh, there, companies like Accenture, Deloitte and any big consulting company, it's very often a political contract. When the company says we need to introduce something, we don't know what, we need to report to the board of directors, and the board of directors is just introducing something. I mean, it's a lot, at least, from the outside. I guess a lot of people are putting things in, but it looks like it.
About modeling. Uh, someone who didn't see the first edition, reconsider. Sasha, here, Sasha Masharov, you've just given a very good example of this, uh, that you can learn your model not on your dateset, but you can do hundreds of thousands of GPT requests the same, get answers from it. And you have a dateset you can do your model, yes, teach. And when we use some kind of open source, it seems possible, open source model, we never know what's under her hood.
She may have recycled data that were collected some way before, which we know nothing about.
That's interesting. There was now article one, which wrote, which means, OpenAI was observed using cheap force from Africa. And then there was a comment that it was good that some of them, there in Russia, are teaching some of the models real, real physics, real chemicals, real teachers. So, about, uh, the story for closing the video today. Tanya, you have a husband, two kids, a nanny, two dogs, a cat. And Boston Dynamics updated, uh, and he said he had updated robots, updated robots. Tell me, are you gonna let a robot in your house or not?
I'm looking at what.
Well, it's a real one, Tanya, real. For the clinic.
Well, some kind of robots, I understand there's some kind of entertainment, like, there's a dog robot, and there's some kind of function they're cleaning up or, uh, windows, like you, Sas.
But it's not a robot, it's all a toy, it's all like a toy. I'm talking about a robot humanoid. This is a new era, right? This new era of robots is getting bigger and more.
Mm-hmm.
I'm talking about them, which is where Ilona Mask has one, getting in the car, which means he's going, and he's going in. Here's Boston Dynamics. Here's Boston Dynamics, huh? You're gonna let a real man like that?
No, I'm not ready for that. Let someone else test him, there, a couple of years, that he's been there, you know, not going to sleep at night, not killing his masters, not breaking into other cars. Then yes. So-- doesn't collect any more information and, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, well, that's not the way it's possible to get any information. Oh, well, I'm not like that, you know, an innovator to run for every kind of thing like that in the first place, totally new. I wouldn't have agreed to test.
Given that a man has the greatest fear is death fear. Do robots that people don't fear that this robot will kill you.
I have a picture in my head that he's killing me at night. I don't know why. He's got a program, and he's killing me.
Maybe he had a program like that. Many films show that.
Well, I probably looked at them a lot, so no, I'm not ready. But it's probably our future. I think it's already inevitable.
Yes, which is inevitable, we will discuss our next episodes. Oh, a great meeting, I think it's very professional, deep. You'll have to leave your questions, your comments here on this YouTube video. Plus we started running Instagram a channel within Instagram. And I'll meet you in exactly a week.