The Future of AI with Illia Polosukhin: The Man Who Put the T in GPT

26:55And so you don't actually know what biases it has. There's this concept of a sleeper agent. Those weights are a function of the training data. While the weights are known, if the input that determines those weights is unknown, then you're still flying blind, whether it's open or not. Correct. Yeah, you're still flying blind. Like you get the benefit of you don't need to give your data to somebody else or in case you can run it on our infrastructure. But there is this concept of sleeper agents just to, you know, leave everyone awake at night. Good. Please do. Please do. You can train the model while you're training the model, leave a specific kind of precondition that under a certain context, model behaves itself differently.

27:39And it's undetectable from like a weight inspection perspective. And so to give you an example, let's say you're using the coding model, right? You know, you're writing some code and it detects that you're writing some like, I don't know, sensitive code that does financial transactions. and it inserts some malicious piece only in that case, right? It doesn't behave like that in any way. It doesn't show up on any benchmarks. It specifically only injects it there. So that's like an idea of a sleeper agent where you could be like embedding these types of behaviors into the model. And unless you know the whole training process, you wouldn't know if there is something like that.

28:18A more normal example is like just the biases in data. For example, if it's a model by organization that is left-leaning or right-leaning, they may be reduced amount of training data from the other side or from another country or whatever. We just don't know what goes into it. And so it's really hard to evaluate. You can evaluate on some benchmarks, but you don't really have a full visibility on this. And so the next step for us is really enabling this provenance for the training process itself, which is way harder to be clear. But that's where also you have a lot of crypto economics come in. The way to really enable training and kind of like frontier model development at scale outside of a lab is really to create economic alignment where multiple parties can bring hardware, data, and new ideas and really push forward this development and then monetize it back, right, when it's created so that you have a flywheel with new models and new compute coming in.

29:18Yeah. Now, so maybe here's an opportunity to bring this kind of to a fine point for the hyper practically minded part of our audience, because you're living in the future. You have a clear eye towards the future. We also know that teams that are leveraging AI can work a lot more quickly and produce a lot more code and all that stuff. So it'd be fascinating to hear you talk about the practical decisions you make to make sure that your team is moving as fast as possible, but also as safe as possible, right? You're probably more aware of safety risks than others, and you probably feel more urgently the need to move quickly.

29:55So what does it look like? How do you think about unleashing your team with AI augmentation? And what are the kind of best practices that you make sure to enforce in your organization? I don't think there is like a one size fits all, obviously, for this, and it's evolving really quickly. Generally, there's a few things we see, right? One is, I mean, obviously, if you like any kind of product brainstorming, product development, AI is really a helpful partner. It's effectively your sidekick who can do a lot of research, who can help a lot define specifications, et cetera. But you still need to talk to customers, right?

30:34So you still need to understand exact needs from customers. And so you can stream all the meeting nodes into this kind of thing that gives you a TLDR and refines product specs from the conversations with customers. So that's kind of like that big piece of work that I think, I mean, it's accelerated just based on how much post work you need to do, right? That is done pretty quickly. I think the development side, obviously, model has been improving dramatically, right? Even this year. I think I went from, I mean, I don't code normally, but I do have AI running on the background, writing some code.

31:09So it went from, you could probably do a front end in a V0 like prototype format to now it can actually build like a proper, I would say like three, 5 ,000 lines of code app and like reason about it. And you can do really good refactoring and really good updates in the logic code basis. it still doesn't really work on like larger code base like it doesn't it's not able to maintain like a full context and the reality is that if we're looking for engineers right now we're looking for people who can actually reason at this like higher level architectural context and be able to kind of instruct the systems can i just interject a question do you think that's in general obviously you having written the paper we often on this podcast talk about what is it that humans will be uniquely good at.

32:03It sounds like one thing is to have this kind of abstraction layer that is basically bigger than the context window of the model or abstraction. I mean, it is a current limitation though. So if we want to talk about it for the future. That seems like there's an expiration date on that. Yeah, yeah. I think all of those things are going to get improved. I mean, I think we're going to a very different world if we move a little bit further. I think right now is a really interesting time because we are in this precipice where effectively people who know what they're doing gotten like 10 times leverage.

32:37And so you can do a lot more, a lot faster. And the deeper understanding you have of how things work, the better you can do this. This is actually like, again, with coding, I've seen people who don't know how to code use these tools and like they get stuck, right? They just don't know what to do after a certain point versus like engineer who understands the full stack, who can actually go and navigate between it. I mean, again, you can run like multiple of these codecs. If you can manage the context yourself in your head, you can run multiple of those things and they can just write, you know, different parts of the system in parallel because it's like a very latency game now.

33:15So, you know, you type something and it's like ghost things for 10 minutes. So your flow changes. You read code most of the time, and then you sometimes say what you need to fix, and then you have multiple of those things running in parallel on different parts of the code base. I was just doing a medium-sized project. I wanted to see how far I can push it. And so I had three codexes running in parallel. And it was all my mental effectively limitations of how much context I can switch between. you know it's refactoring something i'm going deep and like improving if like one module but the other piece again this is for developers specifically but there's something about the module separation where you effectively assume that it'll just rewrite the whole module every time and so you want to have modules to be on the level of like it can fit in the context What module needs to do, its expectation test, etc., that all needs to fit in the context.

34:17And then you can just work on that level. And then the other kind of trick is cross-testing. Right now, when you're developing something, you usually test your piece, and then you assume the other things work. And so what I'm suggesting, actually, you do the opposite. You test the other things you depend on, because some AI have rewrote in it, and who knows what it does. and so you effectively like do kind of dependency testing and that's part of your specification of what you expect from your dependencies but yeah you kind of shift away from like a before you know you want multiple developers to know the code base of each module right that actually slows down a lot now because the reading now is the slowest part again it just can write so much faster.

35:05And so it's more important to just test really well what other people are doing on your side. And then assume that they potentially didn't even read their code. So that like test what they've done. It kind of shifts how you do development. Again, this is very like probably like next year or two. No, but it makes a lot of sense, right? That basically when the human ability to metabolize the information becomes the bottleneck. You need an alternative way to know whether it's good. And what you're saying is you run it. And if it doesn't break... Yeah, well, you specify, like, instead of reading all of the other people's code, you say, hey, this is what I expect from it.

35:46Like, you go more... I mean, this is... It's a little bit back to your thesis on Transformers, where basically, if I ask it a question and it gives me an intelligent answer, it's probably intelligent. Now it's if I give it a task. If I give it a task. Ask it more questions, yeah. There's also, I mean, related to, there's this met idea of intents that we've been developing. And so this actually connects with all of this very well. So intent is effectively your request, your question, your, like, what you want to achieve. And the reality is everything we do in the world is effectively some form of intent, right?

36:21You know, you go to google.com, You're effectively expressing your intent of what you want to achieve. You want some information. You want some products. You want some services. You want to get some dinner. You want a pizza. That's an intent. You want an Uber to airport. It's an intent, et cetera, et cetera. And so if we assume AI becomes, and this is kind of the vision, AI becomes the main interface, how we actually interact as computing, well, your AI still has limitations. It cannot deliver your pizza, right, and make it. it still needs to talk to somebody else to get that done. And so what is that protocol of between effectively AIs talking to each other and getting things done for the user?

37:03Well, that is what we call intent. So we have this kind of product that effectively designed to how AIs will do economy relationships between each other. This can be food. This can be like actual commerce, moving tons of steel, buying a factory, et cetera, like any level of economic activity where AI is effectively finding each other, figuring out what they need to achieve, committing to it, exchanging value, money, et cetera, settling that and then getting that done. Verifying there may be AI that's arbitrary, like a court that evaluates if something went wrong. So that fully replaces the whole economic underlying effectively contracts, billing, invoicing, payments, etc.

37:50And it uses blockchain as a core effectively instrument for settlement verification and execution of this. And it plots in into AI in a very natural way, as I said. And again, it's very much like, hey, you describe what you want the outcome to be, and then you find the counterparty, the other AI that will actually do it, that's able to do it. Can you say why this isn't happening at Google? Because what's interesting to me is you started, I mean, Google's unique structure and environment enabled you to achieve some of these breakthroughs. Why did you feel the need to leave? Why not go? Google's the best place in the world to be building this next vision of the future.

38:34There is a pretty natural kind of situation at Google where if you're trying to build a product and you're putting a Google name on it, the benefit is you're getting millions of users right away. The negative side is if you don't have a product market fit, you effectively just murder the reputation and you're going to lose momentum and you're going to effectively get caught. And this has happened many times. I mean, Google Wave is like something that, you know, dear to my heart as an example of that. And so the reality is like, it's a really high activation cost for Google to release a product, like a consumer level product.

39:20And so it's a very standard kind of innovative dilemma, right? Again, they're making whatever, half a trillion dollars in revenue. If you're going to release a new product that's going to make like$10 million a year, it's not worse for anyone to even bother. And the problem is until you launch a product and iterate on it, it's really hard to know if it's going to have per market fit. It's the whole point why we have startups. Most of them fail. And so the reality is like at Google, it's a pretty standard thing. People leave, start a startup. And then if it's just successful and found per market fit, it gets acquired back.

39:56Yeah. Right. and then it's like upscaled and and kind of scaled up to the level and so i think what we've seen here i mean kind of with me and and with other people from the paper as well that was i mean obviously very different people different stories but for me it was like hey this is a pretty different product that it would not be likely to launch from google imagine we build a vibe coding in 2017. It works kind of shit. We launched it out of Google name, you know, market right away. Like, what is this? You know, Google doesn't know what they're doing. Some of the early image generation stuff.

40:36I mean, thankfully, Google's reputation survived, but I definitely can imagine. Yeah. I mean, even if you assume like this is hard to speculate, but Google had internally chat GPT like product. It's not hard to build that, but it wasn't great and like they could have launched it but it would be everybody was like what kind of shit it's responding like hey you know google.com is giving me good answers why is this giving me fake stuff right which was like chat gpt in 2023 right and so the reality is like you know effectively open ai validated the market for them and so now they're they're willing to bear the reputational risk that google wasn't at that time well they didn't they didn't have reputation to lose at the time.

41:18When they launched, they're like, hey, we're a research lab. This is a cool experiment. What is the Bob Dylan line? When you ain't got nothing, you ain't got nothing to lose. Yeah. I mean, like, again, ChatGPT was released as like, hey, this is an experimental product. It's cool. We use it internally. Check it out. That's like, if you go back to the marketing, it wasn't like, hey, this is our chatting thing. It just blew up as the thing. But there's one thing in a decade that blows up like this, right? And most things blow up in the other direction. here's my last question for you you started with vibe coding when you did near we talk about agent to agents now i guess it's kind of working you talk about a future obviously where we see the world through you know an agent effect how far do you think we are from actually seeing that word really materializing that my agent talks to your agent or my agent talks to the google agent and we're kind of like all live in our own little agent bubbles well hopefully it's not bubbles but actually it's your agent and it's on your side and make sure that it's your well-being that's been taken care of.

42:21I think, I mean, we are moving there. I think there's still some pieces missing. And so, I mean, that's what we've been, we've been building like past two years, we've been building these pieces kind of both from like, how do we ensure private and your AI and how do we give the AI the ability to now talk and execute with each other. The crypto also just gotten only this year effectively like not illegal in US. And so now when we say, hey, you know, your agents should pay each other in like stable coins because that's the speed of internet and the cost of internet, not, you know, in fiat, you know, again, before this year, this was impossible effectively.

43:06Now this is actually possible and acceptable. There's laws around it. So the reality is all of this stuff is just becoming possible. The hardware elements we're relying on, again, were not ever effectively launched middle of last year. And some of the stuff was released at the GTC NVIDIA this year that we're relying on. So a lot of this is barely possible. You're betting on a future before the future unfolds, which is pretty cool to see that. Yeah, I mean, sometimes way too early. So hopefully I'm getting better with my timing. But yeah, I mean, the reality is like we're just putting together the foundational pieces.

43:45It's starting to be there. And again, it's not just us. There's like a whole ecosystem. And even the big players are starting to like tap in into these pieces, right? Like, I mean, we're seeing ChatGPT also looking on like a gente commerce layer. So like it is happening in that direction. So I think, yeah, it's still probably another year, year and a half. but it's not like somewhere beyond the whale. Well, Ilya, thank you so much for joining us today. I mean, it's been hugely illuminating. A fun walk down memory lane and then also kind of a opportunity to remember the future, as they say. So thanks for making the time.

44:21I appreciate it as well. Okay, Jeremy, we just had, I think, a super interesting conversation with Ilya, who is one of the - OG. OG, one of the authors of Attention is Everything, which is, yeah, I guess, what do you call it? Is it the T in GPT? Yeah, he is the person or among the team of people who thought of the T in GPT. I mean, it is pretty fascinating. You know, you can't help thinking that it must have been fascinating to have sitting there and written that paper, and he'd written a lot of papers, and then suddenly it becomes basically the thing that the whole world is centered around. I will say it's humbling to talk to somebody, Henrik, who I think he's the first guest.

45:02we've talked to what over 50 people 50 experts world leaders he's the first person to say i was expecting chat gpt in 2017 he he was actually he was too early most of us are reacting he's one of the few people in the world for whom chat gpt was quote unquote late upon its arrival what was some of the things that that you kind of felt was kind of like sprung first of all i i think i think this is an episode worth listening to because there's such an incredible kind of walk down memory lane. I mean, he gives us such a great history lesson about how Google was organized and how the team, the kinds of problems they were trying to solve when the attention paper was realized.

45:42I think that was super cool. And just even in terms of understanding what is attention, what is a transformer, I think for a kind of a basic primer, nobody in the world better than Aaliyah probably to give that to us. So that was super fun to me. I think also, of course, learning what he's doing now at Mir and the critical importance, the reality that increasingly information and algorithms and AI change how we see reality and recognizing that leveraging blockchain technology, much in the same way that there's traceability around financial products, there can now be traceability around the provenance of information.

46:24I think he really convinced me that this is an important thing to consider and to prioritize into the future. And I would just say that that really wasn't on my radar part of this conversation. What about you? Yeah, I mean, like a little bit the same way that he got on to my radar in the Nier project, because I was trying to look at crypto in general and was trying to figure out really what is the next thing that crypto can be used for that is kind of breaking out right you know crypto has been around for a long time we talked about it as a it was a coin you know but they haven't really seen that many application where a crypto protocol was used for something you know that was that mainstream people kind of like understood and so what really struck me with this was that with money, you have an alternative.

47:15People obviously have not had the experience that he had where suddenly the bank disappeared. So most people go like, I'll put my money. By the way, just as an aside, I mean, he is uniquely qualified because of his family's experience in the Ukraine to care about this problem. So just as an aside, wow. But most people have a way of storing their value right now. They go to the bank, whatever. I think what most people do not have is a way to make sure that there is integrity in the information they consume. And over the last few years, we've obviously talked much more about it because we started to realize that getting skewed information can convince you on one thing or the other.

48:00And so this is actually the only real alternative I've heard of, of saying, hey, I am going to give all my information to these intelligent machines, these AI bots. I'm going to consume the world through these AI bots. How do I know that I can trust OpenAI or DeepSeek or Claude or whatever it is with my data? And how can I make sure that the stuff that I give to it is secure? And you really can't if you don't have a mechanism to do it. And so while crypto is super nerdy and this is obviously on a foundational level and you might never have to think about it, it actually can become quite important for how our world will kind of like evolve.

48:44Has huge implications for the future. You know, the fact that we're shifting, you know, money is power. Information is power. Information is the ultimate power. And to have people like Ilya fighting for information security and information, you know, individual sovereignty over information and assurance and verification that your source of information aren't being polluted. Or, or, um, you know, what did he call it? Secret agent. What was the phrase hidden agent? I can't remember. There was a phrase that will keep, if you listen to this episode, there is a phrase, there's a three minute segment of this interview that will keep you up at night.

49:22Just don't let your children listen to it. Don't let your grandma listen to it. But if you listen to it, just be prepared. You're going to get your mollux three sleepless nights just out of this one episode alone. Hey, one thing I want to talk about to Henrik, which is kind of beyond, which is a point he made in a particular use case, but I think can extrapolate to any user of any AI system, whether you're building the crypto future or not, is the following. He said, and I quote from my notebook here, the deeper understanding you have, the better you can use the tool. And he was, of course, referring to engineers who can have the context and architecture of a system, you know, far, that's far surpasses a model's kind of context window.

50:05But I think that's actually a principle that we have heard is probably one of the most recurring principles on this show from all of our conversations. The deeper understanding you have, the better possibilities you can experience in collaboration with AI. And what would you call this? Because I think a lot of people that I talk to in organizations, a lot of them say, you know, do you use AI? And they go, yeah, I use AI. And, but you kind of like when you're like a super user of it, you think, yeah, but you're not really using AI. Well, so, so there's two things there, Henrik. One is how do you collaborate?

50:39Um, which I agree. I hate the word use, as you may know, I love the word work with, right? So anybody who says they use AI, I say, I know you're a problem user. I know you're a tool. If you treat AI like a tool, you're a tool. Okay. So work with it. Don't use it. But what your question was, what do we call this? And you kind of went down to use the thing that I would say, I was actually thinking about this weekend. I'm working on a new book and I was thinking about a lot of these ideas and the phrase that came to my mind just this weekend is like hot off the press, but is what I would call the human experience edge.

51:15I think there's something to mining your own expertise to say, what is the deep understanding that I can uniquely bring to my collaboration with AI? And granted, different from my use of AI. So we do need to talk about kind of collaboration, hygiene, et cetera. But there's an area of one's life that you could think, oh, no, no, this is uniquely, it is uniquely you. However, However, if you will bring that into a fulsome collaboration with AI, you are going to get differential performance. And you know what it reminds me of, Henrik? It actually reminds me of our conversation with Jenny Nicholson.

51:54I don't know if you remember this, but one of the early kind of nuggets we gleaned in conversation with an expert was, she said, your humanity is the only thing that the model doesn't have. What, you know, your unique humanity. I think there's something to that developer that Ilya was describing. That expertise is something that only that person can bring to their collaboration with a model. No one else can. And as he was saying, like people who can't code, they can make code, but they can't appreciate how an architecture hangs together in a way that sits beyond the context window of the model.

52:28I think it a little bit as a multiplier effect in the Venn diagram between what you know a lot about and how well you know how to use the models. Yes. And so what I sometimes mean are people who are very good at using the models, but they don't know that much about the problem you're trying to solve. And then obviously the other thing around. And the real magic is, of course, when you have both, which is easy when it's about yourself. So that's why it's good using AI for kind of personal problem because you're a unique call if I can talk about that. But it is interesting, I mean, we're just riffing on his point, like it is so true to me that there is this unique moment that people who understand how to use AI, they just have such a lead advantage over everybody else because they can suddenly do much more and much faster and much better.

53:14And I think to put a fine point on it, the people who know how to work with, again, not use, let's stop saying use, but the people who know how to work with AI have an advantage. And you know who is particularly advantaged among those? The people who have a depth of expertise that they bring to that collaboration. So I think that's different from, say, like a young, you know, I spent the past week with a couple of young folks who are in college studying computer science. they can bring a world-class kind of collaboration ability, perhaps, if they learn it. There is no, like, just like nine women can't have a baby in one month.

53:49You know what I mean? Like, it takes a woman to have, it takes nine months, right? And there's something about the kind of lived experience that experienced individuals possess that if, to your point about the Venn diagram, if it's brought in dynamic collaboration with AI, that's what leads to the 100x leverage. 100%. Did he have anything else? I mean, this was, it was super fun. More OGs, folks in our network. Thank you for introducing us to the people you've been introducing us to. They're amazing. People like Ilya. I mean, who are we to get to talk to Ilya? I mean, it's incredible. So keep it up.

54:28Put us in touch with your heroes. Let us know what experts you want to talk to to push this conversation beyond the prompt. and with that we have only one thing to say and that is we love you bye bye that was nice of you and bye bye