Skip to content
Transcript

Transcript · 020 · Can All AI Content Be Labeled—and Who Is Responsible When a Model Changes Itself? — ToTheMoon

English machine translation of the Russian-language episode. Timecodes open the source video.

Episode overview
00:00:00–00:01:05The most IT-Sholm subcadast among all IT subcasses.
Alexander Volchek00:00:00

Hello, everyone! Uh, we're on ToTheMoon. Uh, technological news, Silicon Valley sites and the world. The subject is getting uncontrollable again. So if we say water signs can be put in an article, well, in a short notice, what water signs are in a short notice? Funny.

Tatyana Tsvetkova00:00:16

As I recall, in Europe, this has already been forced to legislate these water signs.

Ilnar Shafigullin00:00:23

If you've taken some big model, well, there's a Zuckerberg or Mistral or someone else, and you've got it trained, and you've spent less than $10 million on the pre-learning, you're not responsible, but the creator is responsible. Big model head.

Alexander Volchek00:00:40

Well, conditionally, if I'm a developer, I'm gonna have a responsibility, uh, some huge, there, I don't know, hundreds of millions of dollars, how am I gonna risk doing this at all?

Mix00:00:49

I don't understand. I mean, uh, because it's important to them or because they want to look white and pusy and compliant? That's not clear.

Alexander Volchek00:00:58

I wanted to start with the OpenAI watermarks, but we were waiting for

Mentions: OpenAI
00:01:05–00:04:37Venture investments are dead?
Alexander Volchek00:01:05

everyone, Sasha, um, to talk about VC market and to do a research on the market that, uh, it's a venture capital, um, It is a very specific situation that the venture market, venture investment, is now very specific. Oh, Sasha, you're like, uh, you're saying that people are talking about waiting for you. Uh, can you comment on that?

Alexander Mashrabov00:01:37

Yeah, sure. I mean, uh, uh, uh, uh, a few indicators on which the industry is described. The first indicator is, the intended or current refund of the funds. The second indicator is probably how many vials, how many percent of the starttaps are raised, a-a-a-a--a-a-a-a-a-a-startaps, for example, lifting A series in four or six blocks, after naturally, First-run. And the third indicator that many people look at is, uh, a percentage of start-ups that close. And all three indicators are in a bad position now. And it is clear that there were two months and a half years of crazy, 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- I mean, it would be, there was a period, uh, somewhere, maybe April 2020 to November 2021, and then November 2022 to May 2024. And I think that, uh, that during these periods there were-- there were a lot of inflated, well, the valuation was crazy and that, uh, VC, it was suffocating. Of course, someone else has to save it. Either private eq-- this is the first option that private equity will buy it, but private equity, uh-huh, while, for example, it's not marked for that, right? I mean, they usually need, uh-oh, strong business, which might not be very good at growing up. So far, frankly, um, there's not much of that business. Another option is that companies become public, but now IPO is record low, this is the last year. And the third option is Må. Må is not possible because Lina Khan, is leading this policy in the competition service. So there's not enough MVA. Of course, the Message A-possibility should be hypothetically large, because the investment in generative AI in the big seven is huge, and they need talent to support it, so that it can be sustained. To revert to their investment. But as long as none of these three ways of getting out, unfortunately, I don't. It is unclear how the VC funds will be repaid to their investments. And, uh, until there's no pre-condition, yes, what's the whole-to-the-the-the-to-the-to-the-to-the-to-the-to-the-to-the-to-the-to-the-the-to-the-to-the-to-the-to-the-to-the-to-the-to-the-the-the-the-to-the-to-to-the-the-to-the-the-the-to-to-the-the-the-the-the-to-to-the-the-the-the-the-the-the-the-the-the-gre market market. So far, I guess there's no presupposing that we'll be back in a situation where 0 per cent, uh, 100.

Ilnar Shafigullin00:04:36

Can you please explain a little?

00:04:37–00:06:44What's the problem with the start-ups?
Ilnar Shafigullin00:04:37

Is this the story the main problem for who? For start-up owners, for investors who are, uh, VC, yes, venture capitalists, or for the industry as a whole? I mean, a little bit.

Alexander Mashrabov00:04:50

Yeah, sure. I mean, I think that, uh, uh, well, at LP level, that's, that's, uh, that's, uh, pension funds, endaupments, etc. But I think it's on the ear-- family offices. I think there's no big problem at LP level, because the VC is relatively small. Oh, I think it's always been perceived as a risky asset. The big problems at the VC level, because usually, well, it has been a long time since many funds have been raised new and new funds. I mean, like, like a single parent company, roughly, but they have a lot of different funds under different stages, under different verticals, and so on. This model is broken now. So, now, little VC raises new funds and a lot less than it was before. And for all the start-ups that didn't adapt to a new reality, of course, it could be very, very, very sinister.

Alexander Volchek00:05:52

What do you mean, the start-ups were adapted to a new reality?

Alexander Mashrabov00:05:55

Yeah. So every start-up has some expectations as long as he needs, how much money they have to expect.

Alexander Volchek00:06:04

So, you're adjusting to spend the money right?

Mix00:06:07

Yeah.

Ilnar Shafigullin00:06:08

In fact, uh, reassessment. Too optimistic to look at the power grids. And now we're here to see that money ends, and the companies themselves haven't been profitable yet. And, uh, the vector capitalists in this story, she's, uh, falling down, yeah, there's no such strong drink. What's the matter, um, the starttups themselves, uh, just so the ordinary viewers, uh, just calmed down a little bit if someone was worried, well, or you'd better fix me that I'm wrong. Uh, start-ups work for the funds they've been using, not their own.

00:06:44–00:08:55What happens to the start-up industry?
Ilnar Shafigullin00:06:44

Most often, I understand. Ia for start-up owners, the end of the extra funding that leads to, right? That's just this start-up, this business, it's over. But, uh, the owner of this business himself, the owner of the start-up has no way of with some huge loss, right? Are these losses on venture capitalists or not? Or correct me if I'm wrong.

Alexander Volchek00:07:11

No, mostly, of course, it's a venture capitalist. Another moment, uh, how many people, uh, enough money to move on and what's gonna be inventried. And it's not a question, I think, just, um, you noticed about investing in agenuity AI, I think there's more money in there. We're talking about a different industry, I guess, a start-up, right? And in general, what happens in comparison to a period of time, I think it's again, compared to a period of time, and it's a plus. I think the fundamental difference that has occurred has just become a space number of everything. It's like you used to be like, two thousand seven-- sixteenth year blogging in Instagram and posting a hundred-riss all the time or posting, you'd be automatically, anyway, really big. Now, competition is so big, very competitive and a big competition in the start-ups, I think it's just incredible. This is a fantastic competition. So everyone around, everyone who doesn't get lazy, they want to do it. And plus we've heard the last two years, people have gone to raise money, make money. It was a big chance to make money. And there were plenty of very strong engineers leaving. There's some requalification, it's not the end of the market, is it? I have a friend, he's private equity, and he says, "Oh, that's an A.V.C., it's just impossible, it's a dead market." And he's been telling me that for 20 years, that's how much I know him. For 20 years, he's telling me that this is a floating market and that's all. But how-- look at the pill, huh? Look how much athis company costs, like the US. Ilya's good thing has offered to start today about what was-- the story of OpenAI said that he'd start introducing watermarks into ChatGPT.

00:08:55–00:12:00ChatGPT add water signs?
Alexander Volchek00:08:55

What do they mean by that? That the text generated by ChatGPT is supposed to have, uh, inside a certain structure that can be re-identified by the same system and that it is written by means of a system. ChatGPT. I can't quite imagine what it's possible to do on some, some, maybe some volumes can be done, some kind of thing. Here. And the other side, there's a-- the regulators just, I think, at CI, they just don't, yeah. Every day, news of new laws comes. Except I was just throwing a new law out there last night, and they're talking about California. So what about water signs? This story is important to people, uh, especially to those who use it in their work. Uh, to determine-- understand that the picture is there, uh, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, on, the, the, on, on, on,

Alexander Volchek00:10:00

It's a way to determine, uh, maybe a picture, a video of the generosity. In theory. Both. What do you do? I was told by a friend that you could try the video. Here. Oh, um, what do you think of water signs?

Ilnar Shafigullin00:10:13

Yeah, well, the news was, uh, the OpenAI said it would be very careful, uh, those watermarks, if they were, yeah, they'd be in trouble because it's such a very ambiguous story. And why did she start? , OpenAI, and other companies are trying to make detectors that will say, this text is generosed or written by a man. Well, no one's gonna do that, yeah. It's almost impossible to distinguish, there, text written by a man from a well-generated text. And they're looking, and what other possibilities are there for that. Well, here comes the chance to sew such watermarks. Well, you said rightly that some of the text structures, the combination of words, something else. Some kind of markers, on some such grounds, uh, you might find that this text is being generated by the model. But, uh, I'm considering that, uh, no company alone will do it because its instrument is immediately discredited to the others. Well, let's pretend that the OpenAI has wrapped these water signs, and now the Anthropic with Claude has decided not to stitch, right? And it's obvious that the flow of those who use, uh-a, Claude from those who use ChatGPT, and all other companies. I mean, if it's at the law level, you can, you know, have to do that, otherwise, there's a company that can't work, then maybe, yeah. Without that, I think it's a straight shot to the leg, uh, these companies that provide these services. And besides, we'll have to open the model source, uh, the same Mistral or, uh, what Meta is suggesting us use, uh, where it's likely that all of these restrictions can be removed or re-examined. a model that doesn't exist.

Mentions: OpenAI · ChatGPT
Tatyana Tsvetkova00:12:00

As I

00:12:00–00:15:24EI regulation in Eurosuit: boxes in Instagram for the box
Tatyana Tsvetkova00:12:00

recall, in Europe, this has already been forced to legislate these water signs.

Alexander Volchek00:12:07

Where, Tanya, is the law banned?

Tatyana Tsvetkova00:12:09

In the Eurosuit, I think. This act they adopted, there, I don't know, they've already introduced it or not, but it was written that the content should be marked with watermarks. I noticed that in Instagram, I don't know if you've noticed or not, there's already been a kind of a hunch, a splendid, that you can point out that this content was accented. Well, not necessarily. And I think that's not what I think.

Alexander Volchek00:12:40

Actually, it's a little weird, isn't it? Like, on the one hand...

Tatyana Tsvetkova00:12:44

Mm-hmm.

Alexander Volchek00:12:45

...uh, they want to determine, put watermarks to determine for something there that it's AI. On the other hand, I have a right as a person...

Tatyana Tsvetkova00:12:52

Mm-hmm.

Alexander Volchek00:12:52

... use different means to create some kind of material. I was on my way out of the book, so I cut a piece out there or a part of it somehow overphrased or did something, or asked the other people to do it. I didn't have an Instagram written that this content was generated by a 40-eight person, there, or this content is sgenerated--

Mentions: Instagram
Tatyana Tsvetkova00:13:12

Alexander Volchek00:13:12

...after I read some books to the summit, yes, or this content is being generated. That's not true. I mean, it's kind of a human restriction. And I guess someone's gonna say, wait, and why do you even limit me to using some kind of instrument? Not much of what we're implying by AI. And if humanity is going to do real-- to make real AI, yes, what we see in the movies is, yes, not what they're talking about in some words, there, AI calls anything. What happens then? I can't use the house. Everybody wants a robot. So there's a robot. I can't use robot's house to help. So this robot helped me put tomatoes. So these tomatoes can't cost money or what? Well, that's a very, very specific story in the plan, yeah.

Tatyana Tsvetkova00:13:57

Polish nerve, yes.

Alexander Volchek00:13:57

And the idea is, someone in the States is 100 percent of some state, so they're gonna vote in California now that AI's gonna celebrate and, uh, a lot of things to do. It'll take a couple of years, and they'll start saying that a man has the right to do what he wants, and then everyone will cancel, and they'll all cancel and turn back. And they say we have complete freedom here. A man has the right to use everything he wants, where he wants, when he wants any computer technology. I mean, it's kind of paradox, I think, at some level of illiteracy, often, the illiteracy of politicians or some people, to get a little bit of a hip or a protest or something else to say.

I agree with you because, in fact, you have to forbid what violates some of the substantive human rights. There, for example, is a violation of copyright or, there, the right to life, health and so on. I don't see any of these people in here as dangerous as I am, yes, that AI will violate people's rights. And the rest is a strange restriction, which really smells like conservatism, you know.

Alexander Volchek00:15:06

Sasha, how much technology can that be? Well, hypothetically. How much in the video generators do you really know? What is it supposed to be like algorithms that will always determine it?

Discussion participant00:15:17

When you create a video or picture, you can make a noise that doesn't recognize the eye and that just puts the content like AI in a generated.

00:15:24–00:16:27Why would OpenAI do that? How to determine the content of AI, special services
Discussion participant00:15:24

Ah, it's just, uh, here for me, to be honest, the... the situation with OpenAI is completely unreal. Why-- why are they doing this? I don't understand. I mean, uh, because it's important to them, or because they want to look white and pusy and compliant. That's not clear.

Alexander Volchek00:15:48

Yeah, maybe they're like, like, trying to be, right? But here, I think it's important for ordinary users to understand that this is all very specific to do. There were such start-ups that were doing this, and a year and a half ago, someone said he was definitely there to determine a fairy from a non-Fake. But even in the article, Ildar, which was discussed, it says that-- uh, many things are very easy to get around, right? So you're taking the cha-- text, a text that's a generosized network that put watermarks, and then you're gonna start it online, uh, into another, uh, and, uh, another network that doesn't exist, like, generated.

00:16:27–00:20:16In Japan, the model of the projection itself changed its code.
Alexander Volchek00:16:27

signs, right? What happens next? To the word, about the network. So the interesting news was that, uh, in Japan, they did, so I'll just say the name. Name, because it was unusual for me to hear that. Oh, yeah, that in Japan, uh-a, there's a I-scientist, a generator model. And this genius model-- we Japan, by the way, never used, uh, or looked at it among the com-- among the countries that are included in the AI world. And yesterday, I read about South Korea, Japan alone. And what is this AI scientist who's doing, uh, research. And, uh, the creators noticed that the model started to redecorate the code itself and that it changed the code, rewrited the code for its own convenience, yes, and cycling, cycling-- added it. Python's lines were coded and thus made a cycling algorithm that was eating a lot of power or was working completely infinitely. I gave myself one, I'm sure, given myself such power to study forever. They noticed that, in fact, if these are models that are not in some isolated sand, they said that research should be conducted in an isolated sand. If there's no isolated sand in some sort of sand, it's gonna make it happen that, uh, it can create anything, even though it's not. Well, let's try to give an example that she might, uh, show us how to create, uh, some kind of tool to poison someone, like, right? Well, some of these things are so hard to do. And then there was an interesting thing about the debate, whether AI could be used in the scientific world and whether it was useful. Because if something the genetic models did some scientific research, the amount of time spent on testing this study is supposed to be human, uh, is just a little more than if a person was. I'd do that research myself. I don't know how true that is, do I? Because we assume that these models will start to do research that would never have been able to do in a lifetime. Here. And the science articles, again, say that the whole text that is generated is empty. Well, at least those scientists are talking about it. I'm gonna support something. Uh, and it's based, it has no, uh, novel or real, real, any kind of emotional stuff. Here. Well, there's a market, uh, something, uh, on the other side. I think it's very interesting.

Ilnar Shafigullin00:19:06

Well, there, actually, what's the story? I mean, like my eyes, right? That they tried to make the system, gave her a little more freedom. And this system is where it could, everywhere, uh, rigged up, so we'll talk like that. Yeah, you've given examples that she's creating separate flows and everything, the system's been there. She started to build a becup after every step and there was a terabite of data. And the system is back on. And at some point, yes, she didn't have enough time to make the process work. She just rewrite the code and increased the waiting time, didn't she? And here's all these things. I mean, reading the article would be interesting to see that something positive has been done , and there was so much in the article that there are attempts, attempts, interesting ones, some more freedom, but while they're at something. No significant results. It's really interesting that, in general, such a thing.

Ilnar Shafigullin00:20:00

In general, such research is going and trying to find something. But, let's just say, the second news is negative, right? What is still, at least, this article did not describe any example that would be positively decorated. Here, but...

00:20:16–00:22:32In California, the concepts of " I-Catastrof " are being introduced
Alexander Volchek00:20:16

Although it's probably there. Probably there. It's always a question of protest. There, California is advancing a bill to prevent artificial intelligence-related disasters. And the draft bill of the pre--- it involves the responsibility of the developers for incidents, where it costs more than half a billion dollars. Yeah? There are different reservations, which means that if more than $10 million were invested in this AI, then, uh, one responsibility. If you invested up to ten million dollars, another responsibility. In fact, uh, the world has been moving incredibly forward over the past year and a half in terms of, uh, what is called AI or near AI. I don't know, I feel a tremendous advantage and a lot of opportunities and more and more use, c-e-e-e-, at least, a generator-generative artificial intelligence in my life. I'm incredibly using, right? In different completely designed, models, details, any requests. And there's a huge amount of-- the market is complicated with weight-- well, it feels venturous. That's a lot of money to get. It's very difficult to develop, it's a huge job of engineers, and everyone around here is exposed to different, uh, complexities, laws, even without knowing what's coming up. Yeah? It's very... Well, conditionally, if I'm a developer, I'm gonna know that I'm gonna have a huge responsibility, there, I don't know, hundreds of millions of dollars, how am I gonna risk doing this at all?

Ilnar Shafigullin00:21:41

California is all ahead of the planet. We're a few episodes back, Tanya, you were talking about, yeah, different legislative initiatives. And now it was just the news that the emails to this legislative initiative have come in. The base was there? What if something happened-- not even, no. If something potentially could happen, but it hasn't happened, it's possible to be held accountable. Now, the emoluments are like that, only after what happens, you can be held accountable. That's the first. And the second interesting story is that if you took some big model, well, there, at the Zuckerberg or Mistral or someone else, and you got it trained, and you spent less than $10 million on the pre-learning, It's not you, but it's the building of a big model. Uh, in that sense, some of the e-mails came along, and that's good.

Mentions: California
00:22:32–00:25:26Counterfeiting -
Ilnar Shafigullin00:22:32

But in general, of course, this regulatory history will keep the market going on.

Alexander Volchek00:22:36

One of the interesting, new things that, uh, I've seen right now, this Federal Trade Commission has approved new rules to fight the U.S., uh, to fight counterfeit revotion, right? Ii, which is supposed to be impossible or will need to be very closely monitored by companies, especially by the unrealistic feedbacks that are generated by artificial intelligence. I wonder where you were when the reviews were generated, uh, lu-humans from India, for example, right? Coll centers from India.

Mentions: United States
Discussion participant00:23:09

Yes.

Alexander Volchek00:23:10

Here. Or any one from Africa, anywhere. Oh, and it's supposed to be impossible to remove. The new law will begin to be introduced now, which will not be possible to remove negative feedback. Yeah? And that's the difference between negative and when negative, it's a real negative, spoon-- it's really negative. Negative is that bulling, for example, contains inside or, or a human-related hack. I don't remove negative feedback, but if a man goes into a bouling mode like that, yeah, some kind of thing, or when he starts to do something, you know, you remove it because it's absolutely absolutely. inadequate. Yeah? So one thing is a discussion. I've been looking that way, too. Look at her, because artificial intelligence is starting to be used... well, it's clear that in the era of artificial intelligence, you're getting much better off the system. The feedback generation, yes, everywhere, and this is the subject that's getting out again. So if we say water signs can be put in an article, well, in a short notice, what water signs are in a short notice? Funny. Yeah? This is an unfinished task at all. So it's not even a definitional solution, it's an artificial intelligence that's done or not. How do you even do that, who's gonna be punished? But it says that the rules, uh... There was such a comment. Yeah, we should see that. It's-- this reality that's about to happen. It says that pro-- that this will come into force in sixty days and that the fine will be up to $50 1,000. Here. That's very interesting. For every violation. Here. And what-- and to improve the reliability of online calls. I wonder how this area will also evolve, taking into account, again, different fairies. Yeah? Now, we see, please, Trump, Trampe's team, generated the video where Taylor Swift, uh, supports him. Yeah? It's clear that it's on the grid. So Taylor Swift v. Trampa. And I'm aware that Taylor Swift and Trampa have raised the ratings. Yeah? It's all up, so it's up. The tramp was just saying how she sings songs. I won't quote.

00:25:26–00:27:58Counterfeiting -
Alexander Volchek00:25:26

So it's better, it's better not to quote.

Discussion participant00:25:26

Alexander Volchek00:25:26

Again, it's a use of artificial intelligence, right? I mean, using artificial intelligence, including at this very serious level. And that's not gonna be avoided, it's-- it's not gonna be avoided. Yeah? I mean, if we see what they do in these things, it's not gonna be avoided. Plus you always say I didn't. Well, you're generating support, you're saying you didn't even do it, someone did it. If people can make wars in the world of disaster and say they didn't, yeah, it was someone else who did it, that we had nothing to do with it, right? But then it's uncrossed somewhere, in, fifty or seventy, or 100 years. What happens when people start mass-generating stars. That's, um, I think that a story that can't even be traced.

Ilnar Shafigullin00:26:12

I guess people just have less confidence in what they see.

Ilnar Shafigullin00:26:16

Before, years, I don't know, there, 15 years ago, probably for the lawn and the IP, well, in Russia, there was a need to have a press, right? And the seal was, well, there, as proof that this document is correct, it is signed, and it is still worth the seal. At some point, they just said that the seal was not mandatory. And partly why did that happen? Because the seal is generally not so hard to fake, yes, and the credibility of it is getting lower. And the same thing happens to some people there, I don't know, video evidence, voice evidence, and, by the way, they're not gonna be considered just as necessary evidence of something, just not. They'll trust that, uh-oh, instrument, so let's say.

Tatyana Tsvetkova00:26:57

Yeah, didn't we go through with the photoshop when he showed up and when the faces were professionally changed, put a different face on the other side, and so on. That's what happened. It's just that the spiral on the new turn is much more progressive.

Ilnar Shafigullin00:27:14

Yeah, you're just having a hard time, so, making these things is getting really low, and the instrument is getting more accessible.

Alexander Volchek00:27:22

It's a very interesting thing, but it's very complicated, yes, it's complicated to understand. And, uh, here, because we all have different contacts with these technologies and at the level, I don't know, business or professional, and at the standard of normal life, or at the level of things that are, It's a good thing to see and understand at least a little, and where there's something useful, real, and where there's a real threat.

00:27:58–00:29:14How do you feel about that?
Alexander Volchek00:27:58

Well, they-- they take it easy, right?

Ilnar Shafigullin00:27:59

Now you're saying that we've all got different contact with these points. I'm very interested, uh, in our viewers, uh, what is the relevance of all these episodes: watermarks, uh, I don't know, to the prohibition of use, yeah, there, to fairies, to all that. We've been or haven't encountered it. I read the comments periodically, I'm kind of curious. We have a very good audience, so please write in the comments. It's gonna be very interesting to read.

Alexander Volchek00:28:25

Thanks for being on ToTheMoon. We're going out once a week, Sunday. There's already been a huge number of episodes. We're always four men, me, Tanya, Elnar and Sasha., technological news, various sites, opinions, in the technological world, mainly the U.S. Ring. In general, of course, different countries. Sign on the channel. Thank you for your review. Ah, we're happy to support the channel. It's when you leave comments and liquor. It's all moving. And ask your questions and tell me the different topics. We'll talk about it. See you again!

Mentions: United States