Sergey Brin: Where Frontier AI Is Headed | Unscripted Q&A @ AGI House × Google DeepMind

Hey, thanks Rocky. It's a pleasure to be back here. I haven't been here in a little while, mostly just honestly hard working at the office. It's just been a nose to the grindstone kind of time in AI. Obviously, things are moving very fast and yeah, we're cranking all the time. There's basically no time for little social visits and things like that.

Um, okay, I think that it's a little bit exaggerated how much of like my presence specifically is helping with Gemini momentum, but look, I think there's a team of people there and once we got a certain critical mass and a certain momentum, it's been just incredible. I mean, it's like, who wouldn't want to be pushing the edges of intelligence? I mean, such a unique time in history, unique time in science.

Let's do it.

What's our next big move? I mean, we've done all kinds of things historically, ranging from managing fusion reactors to obviously protein folding with AlphaFold and whatnot. I think the exciting thing is that all of these things are converging to the same general models. You know, in the past we would have to have specialized models. And well, in the case of protein folding, we obviously still do. But increasingly, our main Gemini LLMs can be the state-of-the-art for math, for example, and for other kinds of scientific questions. So that convergence is, I don't know, I guess it's not something I really would have predicted at the outset, but it's been kind of incredible to see. And I guess baked into that is this concept of transfer. Just the idea that when you train for a certain class of problems, let's say you're training for coding, that that actually can help your math reasoning and vice versa. And that's been really exciting to see. The multimodal capability also is an example of that. Like can you actually get transfer from being able to process images to actually being able to think through kind of geometric text problems too.

Wow, that's a great question. What's sort of next after we hit AGI? I mean, I think everybody is like pretty focused on accelerating the growth in AI right now. What comes after? I guess you're right. We started with obviously the web and internet search, we kind of went through the mobile generation, which was another pretty big explosion. I guess now people are, you know, now AI is a huge new industry trend and what comes after that? Boy, I mean, I think if you can answer that, you'll have a fantastic company on your hands.

Um, okay. I mean, those seem like two different questions. P versus NP could be answered by humans. I mean, there's nothing that says it can't. That's a...

Oh, are you saying that super intelligence should not answer whether P equals NP or not, but you're saying that it will be able to solve NP-complete problems.

Oh boy. I mean, I think that is a very unusual point of view. I think that most people, most computer scientists probably think that P is not in fact equal to NP and therefore, sorry, I don't know how technical the audience here is, but generally speaking, that there is no algorithmic solution that will reliably solve traveling salesmen, for example, optimally. And that would include, you know, assuming those are distinct classes. Assuming that's impossible, it doesn't matter if you're a super smart AI, it's still impossible. People do try to use quantum computers for some of these more challenging problems. But a lot of people, I think most theorists would say that even though you can factor large numbers with quantum computers, it doesn't mean you can necessarily do NP-complete problems either. So I don't think I have some brilliant suggestion here, but I guess I think the community would say that super intelligence simply means sort of being able to be smarter than humans, but not in being able to do NP-complete problems.

You're welcome.

Great question. I mean, I increasingly see, you know, large companies pushing to the forefront of AI, not necessarily with like their main core workforce, but on the side, they'll be like, you know, can we design a car or a plane using an LLM or something like that. I mean, they are all running these kinds of experiments and I don't know that those have completely come to fruition as of yet, but I'm sure they'll keep trying. In addition, obviously lots of companies just take some of their day-to-day, you know, more boring administrative kind of workflows and automate those. That's very common. But for their key products and things, you know, for things that require mechanical engineering and so forth, I think that's in the realm of experimentation today, but I think they are experimenting 100%.

You know, that is a phenomenal question. I mean, I am always kind of confused as to how to prompt the models. Like at what level should I be prompting them? Like you can on the one hand have a very specific, like debug this chunk of code here, or you can say, you know, can you write me a better neural net training algorithm. You know, sort of you can say the higher level thing, you can also say like, you know, what should I do today or why don't you just do something useful. And as the models get more and more sophisticated, they ought to be able to handle more general, broader tasks. But it's hard to keep up with like what are they actually capable of doing. Like it's not like even for ourselves with Gemini, we don't know exactly where the edges are. And somebody might always do something that seems, well, they're just asking the model to do the thing and that seems so straightforward but it actually works. I mean, chain of thought actually comes to mind as an example of that. That was an actual, you know, have you guys heard of this as a technical technique years ago? But you know, you just say think step by step and then you say your problem. And that just seems like the dumbest thing ever. Like why would that ever work? But it did. It actually spurred a whole significant increase in AI capability. So I guess I'm just saying that even just the plain textual interface, what you're expecting the model to do, how to ask it to do that is evolving quickly. And it is hard to keep track of. I don't have like a magical solution to that. There are plenty of people who do study and optimize prompts, by the way. But I think you're also asking like, well, as a person, how do you sort of interface to it faster? I mean, obviously there's speech and video, you can have higher bandwidth kinds of connections. You know, obviously some companies out there, the Neuralinks of the world are trying to do direct brain connections of sorts. I probably wouldn't do anything to change my biology today for the models as I see them today. I'd probably wait for things to mature a lot. But it's a very fast-moving target. I do think today's models are getting sort of smarter and smarter about being able to do general things that are helpful for you. So you don't necessarily need to sort of speak to them at much higher bandwidth and they can output higher bandwidth because, you know, they'll spin up video and pictures and things like that.

Okay. Kind of reminds me of Xanadu. Am I dating myself by saying that word?

I mean, I guess my immediate reaction is that you're kind of swimming upstream. But there's nothing wrong with swimming upstream. You know, if we were having this conversation 20 years ago, everybody was building knowledge graphs and, you know, the people who were fretting with neural nets were the weirdos, like, don't you know that doesn't work, like we tried that in the 50s and failed. So obviously the neural nets have come a long way and now everybody is doing neural nets and nobody is doing graphs except for you. So, I mean, you're playing the long odds, but, you know, those long odds are sometimes right. So, good luck.

Okay. Small questions, I guess. What is human's role in this world and what is Google going to do for the next 20 years? I think that, you know, the definition of intelligence has always shifted with what machines can do versus what people can do. And for a long time, you know, chess was the measure of intelligence and then Deep Blue beat Kasparov in the 90s maybe. And the interesting thing is people have kept on playing chess. In fact, I mean, how many people here know who the top ranked human chess player is? Anybody can yell the name. Sorry, I'm assuming it's Magnus Carlsen. I'm sure there's people bouncing up and down, but how many people here know the top ranked AI program? Okay, not too many. What, Benoit? Oh, you think it's Stockfish? I mean, that's the most popular. Is it number one? All right, Benoit claims it's number one. You don't think AlphaZero can beat Stockfish? Oh, they've adopted. Okay. Damn. All right. Well, okay. Benoit is the only one who named that top chess program. Let's point that out. I'm just saying that the fact that computers can do things well has actually not stopped humans getting better and better at them, getting more and more recognition and enjoying those things. And obviously, you know, we've sort of adjusted kind of our view. You know, it used to be like I said, well, chess is the intelligent thing. Then it was Go is the intelligent thing and, you know, poetry or painting, what have you. You know, I think we're going to find AIs can do a whole lot of pretty surprising things, but I think they also help advance people in doing it. And by the way, since AlphaGo, the game of Go has advanced a lot. Like, you know, the players that played against Lee Sedol became vastly better after, and Ke Jie after he played AlphaGo also. It has pushed the state-of-the-art. So I think people are going to be able to enjoy and do a lot of things even with AI assistance. Then the 20-year question, I don't know. I think we should let somebody else ask another question. That's a big one.

That's a great question. You know, I've been asked that a bunch of times. Transformers have been weirdly flexible, like, you know, we use them for image and video in addition to text, sort of they've exceeded their original capability. Now to be fair, along the way they've also changed. I mean, we have whatever, sparse transformers. I mean, there are a lot of little details that have shifted along the way. So it's not like the exact same thing as a transformer paper. If I had to guess, could something close to that be AGI? I would say yes. That's just my guess, just because they've been able to evolve so much. But like I said, they are changing. It's not like the exact same thing as the original transformer paper.

All right, bringing down the goalpost. Okay. I mean, look, you know, we today at Gemini are pretty internally focused, honestly. Like, you know, we focus on the utility of the tools to develop the tools. And I think it is surprisingly similar to how many other people who are sort of at the forefront who are not even necessarily super technical, just thinking of a friend of mine who started using OpenClaw six months ago, for example, and like automated his own life in a variety of really impressive ways to me. So I think, you know, inside the organization, same kind of thing, you know, what kind of things can you have the AI do, whether it's monitoring training runs, for example, or generating its own training data, or all those things. You start to use the tool to build the tool, and that's where we've shifted a lot of our energy and that's most of what we spend our time on these days.

Great question. I mean, yeah, I'll be honest, I'm a little bit of a rabble rouser. I mean, you know, Demis and Demis and Koray on Gemini, and I, because I'm spending so much energy on Gemini and spend a lot of time with Koray, but it's like his job and responsibility is to actually deliver the goods. And I would say I poke and prod him and the team about, hey, are you really doing that? Are you really doing that? Sometimes a little bit disruptive, not going to lie. But no, Koray actually organizes, you know, the groups and has them deliver stuff and I guess I view my role as more of reminding them like what priorities they might be missing, what ideas they're not paying enough attention to, and so forth.

Wait, your name Boris and you have kind of a French accent.

Oh, okay. Is that a common French name?

Okay. All right.

Okay. I'm Sergey. Wait, what was your question again?

Oh, how world models can... Oh, world models. Yeah, I mean, world models are like video, basically, models. And I guess there's like a couple, you know, people talk about AGI kind of pretty broadly. I think of AGI as kind of the idea of the AI can actually improve itself. But other people, and I think probably those people are more correct, sort of think AGI means, well, the AI needs to be able to do anything a person can do. And those are two different things. So to do anything a person can do, you absolutely need to be able to understand and interact with the physical world. So for that, you know, being able to dream, imagine what's going to happen in the world if you do something and comprehend it is obviously important. So I think the world models, yes, if you're going to do everything, and that extends to robotics and things like that, world models are key. And yeah, you guys have probably had more time to play with our Gemini Omni model, honestly, than I have, because I'm deep into the self-improvement game. But yeah, we've been working on that for a long time. Omni is the latest version of that. Omni is also pretty cool because it's just the same Gemini, like we train it also with all the text and all the other things, trains exactly the same way. The fact that these converge is kind of amazing. But yes, you need that capability for this ability to interact physically.

Um, okay, great question. How do I feel confident or not? I mean, I think, yeah, I feel very good overall. If I simply took the temperature, you know, every month and then said, oh no, you know, now whatever, such and such model has shipped, that's it. Then, yeah, then I would probably lose my confidence very quickly. But, you know, we've seen things shift around. I actually feel very good about where Gemini is. I know that everybody now is focused on coding and I think we were a little bit late on really focusing deeply on that. I think, you know, the Gemini sort of 3.0 and 3.1 launches, you know, six-ish months ago, they were on top of kind of across the board. But I know the other labs definitely have made strides then, particularly in coding. But, you know, you watch it day-to-day, like everybody was super excited about Opus for the last few months, but I would give GPT 5.5 the edge now on kind of deep coding. But I'll just pitch you guys on Gemini 3.5 Flash is way faster. I mean, at some point that does also matter. But I want to give kudos to our competition in terms of, you know, if you want to leave an overnight task, it seems like 5.5 is able to do incredibly well. But for interactive, quick, rapid iteration, we just deployed 3.5 Flash and you'll see some more great models. But yeah, in hindsight, I think we probably should have focused on code a little bit earlier, but we are now very much focused on code, just to be clear.

Yeah.