Sergey Brin 3/2 transcript (AGI House)

I must admit I rolled out of a lazy Saturday morning here, Rocky texted me. I wasn't really expecting an AI hackathon to be this huge. That's pretty exciting times. Anyway, thank you all for coming. First of all, thanks so much for giving Gemini a go.

What should I say? We actually have people who know what they're talking about, I think, Okay, Simon!

I was worried I'd have to say something that I'm not quite up to speed on. I'll just quickly say, look, it's very exciting times. This model that I think we're playing with is Gemini 1. 5 Pro. We internally called it Goldfish. It's a little secret.

It's because goldfish have very short memories. It's kind of an ironic name. It's kind of an ironic name. When we were training this model, we didn't expect it to come out nearly as powerful as it did or to have all the capabilities that it does. In fact, it was just part of a scaling ladder experiment.

When we saw what it could do, we thought - Hey, we don't want to wait. We want the world to try it out and I'm grateful that all of you here are here to give it a go.

Question & Answers:

Okay, quick questions. I'm probably going to have to defer to the technical experts on those things, but fire away. Any questions?

Any reflections on Gemini Art?

Sergey: Yeah, we did. Okay, I wasn't really expecting that. You know, we definitely messed up on the image generation. I think it was mostly due to just like not thorough testing. It definitely for good reasons upset a lot of people on the images, as you might have seen.

I think the images prompted a lot of people to really deeply test the base text models. The text models have two separate effects going. One thing is just quite honestly, if you deeply test any text model out there, whether it's ours, ChatGPT, Grok - it'll say some pretty weird things that are out there that you know, definitely feel far left, for example.

Any model, if you try hard enough, can be prompted to, in that regime. But also, just to be fair there's definitely work in that model. So once again, we haven't fully understood why it leans left in many cases. That's not our intention, but if you try it starting over this last week, it should be at least 80 percent better of the test cases that we've covered.

So I invite all of you to try it, this shouldn't be a big effect. The model that you're trying, the Gemini 1. 5 Pro, which isn't in the sort of public facing app, the thing we used to call BART shouldn't have much of that effect, except for that general effect that if you sort of red team any AI model, you're gonna get weird corner cases. Even though this one hasn't been sort of thoroughly tested that way, we don't expect it to have strong particular leads, I suppose you can give it a go. Though we're more excited today to try to blow on context in some of the technical features.

Correct? Yeah a couple of recent developments in multi totalities. Have you considered, like video chat and QBT? Video ChatGPT. You probably wouldn't call it that. I mean multimodal both in and out is very exciting. Video, audio and we run early experiments, it's an exciting field.

Even the little, you guys remember the Duck video that kind of got us in trouble? Though, just to be fair, it was fully disclaimed in the video. That it wasn't real time. But, um, but that is something that we have actually done, is embedded images in frame by frame, how to talk about it.

So yeah, that's super exciting. I don't think we have anything like realtime to present, right now, today.

Are you personally writing code for some projects?

Sergey: I haven't actually written code, to be perfectly honest. It's not, like, code that you would be very impressed by. But yeah, every once in a while, just a little, like, kind of debugging or just trying to understand for myself.

How a model works or to just analyze the performance in a slightly different way or something like that. Little bits and pieces that make me feel connected. It's, once again, I don't think you would be very technically impressed by it. It's nice to be able to play with that.

And sometimes I'll use the AI bots to write code for me, because I'm rusty and I actually do a pretty good job at it. So, I'd be very pleased by that. Okay, question in the back. Go on with the man first. Okay, pre AI simulation, I'm sorry, pre AI, the closest thing we got to simulators was game engines.

What do you think the new advances in the field mean for us to create better games or game engines in general?

Sergey: Yeah, we want that. Um, BOOM! Sorry, that wasn't like a sigh of disapproval or anything. Um, I think, I mean I, um, What can I say about game engines? I think obviously, like, on the graphics, you can do new and interesting things with game engines.

But I think maybe the more interesting is the interaction with the other, you know, virtual players and things like that. Like, I don't know what the characters are. Um, I guess, I guess these days, you know, you can call people who are bland NPCs or whatever. But in the future, maybe NPCs will be actually very powerful and interesting.

I think that's a really rich possibility. I'm probably not enough of a gamer to think through all the possible futures with AI, I mean, it opens up many possibilities.

What kind of, like, applications are you most excited about for people, like, building on, on Gemini?

Sergey: Yeah, what kind of applications am I most excited about?

I think just ingesting right now, or for the version we're trying to, you know, 1. 5 Pro, the long context is something we're really experimenting with, and whether you dump a ton of code in there, or video, I mean, I've just seen people do, I don't think a model can do this, to be perfectly honest.

But people will like, dump in their code, and do a video of the app, and say, here's the bug, and the model will figure out where the bug is in the code. Which is kind of mind blowing that that works at all. I honestly don't really understand how a model does that. I'm not saying you should do exactly that thing.

Experimenting with things that really require the long context. Um, Do we have the servers to support all these people here banging on it? We have people on the servers here as well. Okay, this microphone is buzzing, everybody's really stressed out. It's a customer too, right? because, you know, vanilla context queries do take a bit of computer time. But, you should go for it.

You mentioned a few times that you're not sure how this model works, or you weren't sure that this could do the things that it does. Do you think we can reach a point where we actually understand how these models work? Or will they remain black boxes if we just trust the makers of the model to not mess up?

Sergey: No, I think you can learn to understand it. I mean, you know, the fact is that when we train these things, there are a thousand different capabilities you could try out. So on the one hand, it's very surprising that it can do it but on the other hand, if it's any particular one capability, you can go back and, you know, we can look at where the attention is going at each layer between, like, the code and the video, and, you know, we can't deeply analyze it and personally, I don't know how far along the researchers have gotten to doing that kind of thing, but it takes a huge amount of time and study to really slice apart why a model is able to do some things.

And honestly, most of the time that I see slicing, it's like why it's not doing something. So I guess I'd say it's mostly because, I think we could understand it, and people probably are, but most of the effort is spent figuring out where it goes wrong, not where it goes right.

In computer science there's this concept of reflective programming, where, like, a program can look at its own source code, maybe modify its own source code, and then in AGI literature there's, like, recursive self improvements.

What are your thoughts on the implications of extremely long context windows and the language model being able to modify its own prompts, and what that has to do with, like, autonomy?

Sergey: Yeah, I think it's very exciting to, you know, to have these things actually improve themselves. I remember when I was in grad school I wrote this game where, like, it was like a wall maze you were flying through where you shot the walls.

The walls corresponded to bits of memory and it would just, like, flip those bits and the goal was to crash it as quickly as possible, which doesn't really answer your question, but that was an example of self modifying code, I guess not for a particularly useful purpose. But anyway, I'd hope people, you know, played that until the computer crashed.

Anyhow, on your positive example, I see today people just using it to talk about them. I think, you know, OpenLoop, could it work for certain, I think for certain very limited domains today, like if you, without the human intervention to guide it, I bet it could actually do some kind of continued improvement.

But I don't think we're quite at the stage where for, I don't know, real serious things. I mean, first of all, a million contexts is not actually enough for BigCodeBases. To turn on the entire code base. But you could do like retrieval, and then occupation editing. I guess I haven't personally played with it enough.

But I haven't seen it be at the stage today, where a complex sort of piece of code will just iteratively improve itself, but it's a great tool. And like I said, with human assistance, we for sure do, I mean, like I will use Gemini to like try to do something with a Gemini code, even today, but not very open loop deep sophisticated things, I guess.

What's your take on Sam Altman raising 7T for chips?

Sergey: I'm just curious, like, how do you see that from obvious years of, you know, look, I saw the headline. I didn't get too deep into it. I assume there was sort of a provocative pipeline or statement or something. I don't know. I don't know. I don't know. He hasn't asked me for 7 trillion dollars yet. Um, I think it was, it was meant for like chip development or something like that.

I don't, I don't get, I'm not an expert in chip development, but I don't get the sense that it's just something you can like sort of pour money, like even huge amounts of money and output chips. I'm not an expert in the market, though. Um, let's see, let me try somebody in the back. Okay, yes, sir. Oh, yeah, sure.

Yeah, so we all know the training cost of not every model is, like, super high. So how do we think this should work? Oh, the training costs of models are super high? Um, yeah, the training costs are definitely high. You know, that's something companies like us have to cope with. But I think, you know, the long term utility is incomparably higher.

Like if you kind of measure it on a human productivity level. You know, if it saves somebody an hour of work during the course of the week, you know, that hour is worth a lot. And there are a lot of people using these things, or will be using them. But you do, it's a big bet on the future. costs less than 700,000, right?

Model training on device?

Sergey: Oh, model running on device. Yeah, model running on device, we've shipped it to I think Android, Chrome, and yeah, Pixel phones, I think even Chrome runs a pretty decent model these days. We just open sourced Gemma, which was pretty small, a couple billion parameters, I can't remember right now.

It's really useful you know, it could be low latency, you're not dependent on connectivity, and the small models can call bigger models in the cloud, too, so, I think on device is a really good idea.

What are some vertical slash industry that you feel like is generally going to have a big impact on, and startups should consider hacking on those?

Sergey: Oh, which industries do I think have the big opportunity? I think it's just like very hard to predict. I mean, there's sort of the obvious industries that people think of, sort of customer service, or, um, kind of just like, you know, analyzing, I don't know, like different lengthy documents, and kind of the workflow automation, I guess, those are obvious, but I think there are going to be non obvious ones, which I can't predict.

Especially as you look at these sort of multimodal models and the surprising capabilities that they have. I mean, I feel like that's why we have all of you here, you guys are the creative ones to figure that out.

We run billions of thousands of custom insured chats every day and from elements, and did you say Gemini was the only thing really working?

And now it seems that GNI is another thing that really works. Thank you so much for this. And it's just like It's so good to hear it. It seems like it's way more cheap, while it works even better sometimes. So the question is, will it stay sane cheap? Or are you just planning to raise prices at some point?

Or do you know something about it? I'm actually not on top of the pricing thing. I don't expect that we will raise prices, however, because, I mean, there are fundamentally a couple of trends. One is just that these, there's just optimizations and things around inference that are just constantly, like, all the time.

Someone says, I have this 10 percent idea, this 20 percent idea and like, month after month, that adds up. I think our TPUs are actually pretty damn good at inferencing. Not this thing, the GPUs. But for certain inference workloads, they're just configured really nicely. And the other big effect is actually, we're able to make smaller models more and more effective, just with new generations.

Just, whatever, architectural changes, training changes, um, all kinds of things like that. So the models are getting more powerful, even at the same kind of size. So, I would not expect prices to go up.

What are your predictions for how AI is going to impact healthcare and biology?

Sergey: Oh, AI, healthcare, and biotech. I think there are a couple very, you know, different ways. You know, on the biotech side, people look at, um, things like, AlphaFold and things like that. Just like, understanding the fundamental mechanics of life. And I think you'll see AI do more and more of that. Whether it's actual physical, um, Molecule and bonding things, or reading and summarizing journal articles, things like that.

I also think for patients, and this is kind of a tough area, honestly because we're definitely not prepared for, or just AI’s like, go ahead, ask any question, like we're not. You know, AI’s make mistakes and things like that. But I think There is a future when you, if you can overcome those kinds of issues where an AI can much more deeply spend time on an individual person and their history and all their scans may be mediated by a doctor or something, but actually could be just better diagnoses, better recommendations, things like that.

Are you focusing on any other non transformer architectures for, like, reasoning, planning, or any of, to get better at that?

Sergey: Okay question is Are you focusing on any non transformer architectures? I mean, I think there's, like, so many sort of, uh, variations. Um, but I guess most people already are still kind of transformer based.

Um, I mean, I'm sure somebody at the company who's speak to, more to it, would be looking. But, yeah, as much progress as transformers have made over the last, whatever, six, seven, seven, eight years, I guess. Um, there's, you know, there's nothing to say there's not going to be some new revolutionary, architecture.

And it's also possible that just, you know, incremental changes, for example, sparsity and things like that are still kind of The same transformer, also for evolution, so, you know, I don't have a magic answer.

Is there some bottleneck for, like, reasoning kind of questions?

Sergey: Is there a bottleneck in transformers?

Yeah. Um, I mean, there's been lots of theoretical work showing the limitations of transformers. You know, you can't do this kind of thing, this many layers, and things like that. Um, I, I, I don't know how to extrapolate that to, like, Contemporary transformers that usually don't meet the assumptions of the theoretical works.

Um, so it may not apply, but Um, I'd probably hedge my bets and try other architecture as well as being cool. Thank you.

What are your thoughts on Google Glass? Would you consider, like, try to give that another shot?

Sergey: That's why I was really looking forward to that. Thanks. Um, I think a lot of those are obviously part of the topic, but, um A while ago Google had Google Glass, but now Apple has Vision Pro.

Um, I think Google Glass may be a little bit early. Um, yeah, I feel like I messed up Google Glass. Um, no, no, but I feel like I made some bad decisions. Yeah, it was for sure early, and early in two senses of the word. maybe early in the overall evolution of technology, but also I think I, like, in hindsight I tried to push it as a product, but it itself was sort of more of a prototype and I should have set those expectations around it.

Um, and I personally didn't know much about sort of consumer hardware supply chains back then. Anyway, a bunch of things I wish I'd done differently. Um, but I personally am still a fan of kind of the lightweight, kind of minimal display that that offered. That you could just like wear all day versus the heavy things that we have today.

Um, that's my personal preference, but the, the Apple Vision and the Oculus is, for the matter, they're very impressive, like, having played with them, um, I mean, I'm just impressed with what you can have in front of your screen. Um, but that was what I was personally going for, kind of, back then.

If you can see, um, Gemini expanding capabilities into, like, 3D, kind of down the line of, Spatial computing in general or assimilation of the world in general, all of that. And especially beyond Google Glass, Google already has several products that's really in the area, like Google Maps, Street View, ARCore, all of that.

Sergey: Do we see all of those have some synergies between them? Wow, that's a good question. To be honest, I haven't thought about it. But now that you say it, Yeah, there's no reason we couldn't, You know, put in more sort of 3D, like, it's kind of another mode of, you know, 3D data. Um, so, probably something interesting would happen.

I mean, I don't see why you wouldn't, try to put that into a model that's already got all the smarts of the text model, and now can turn on something else too. And by the way, maybe somebody's doing good at Gemini, I don't know if somebody's probably Oh yes, yes they are. I'm not sure if I forgot about it, doesn't mean it's not happening.

Yes, question in the back there.

Are you optimistic that we'll be able to bring in text generating models, ability to hallucinate, and what do you think about the ethical issue of potentially spreading misinformation?

Sergey: Problem right now, um, no question about it. I mean, we have made them hallucinate less and less over time.

Um, but I would definitely be excited to see a breakthrough that brings it to near zero. Um, I don't, you know, that's not, you can't just like count on breakthroughs. Um, so I think we're going to keep going with the incremental kinds of things that we do to just like bring all the hallucinations down, down, down over time.

Like I said, I think a breakthrough would be good. Um, misinformation, you know, misinformation is a complicated issue, I think. Um, I mean, obviously you don't want your AI bots to be just like making stuff up. But they can also be kind of tricked into, like, I mean, there's a lot of, I guess, complicated, political Issues in terms of what people consider, what different people consider misinformation versus not.

And it gets into kind of a broad social debate. Um, I suppose another thing you could consider is about them sort of deliberately generating misinformation on behalf of another actor. Um, from that point of view, I mean, unfortunately, it's like, it's very easy to make a lousy AI. Um, like one that hallucinates a lot.

Um, and you can take, you know, any open source text model and probably tweak it to generate misinformation of all kinds. And if you're not concerned about, um, um, you know, the accuracy, it's just kind of an easy thing to do. So, I don't know, I guess now I think about, I mean, detecting AI generated content is an important field and something that we work on.

Um, and so forth, so you could at least maybe tell if something coming out of you was AI generated. Yeah. Alexandra,

The CEO of NVIDIA said that basically the future of writing code as a career is dead. What's your take on that? And what could potentially protect our careers as engineers?

Sergey: Okay, yeah. Um, I mean, it's, um, Like, we don't know where the future of AI is going, broadly.

I would, you know, we don't know, you know, It seems to help across a range of many careers, Whether it's graphic artists, or custom sport, Or doctors, or, um, Or executives, or, you know, what have you. Um, I mean, so, I don't know that I would be, like, singling out Um, program in particular, it's a, it's actually probably one of the more challenging tasks for an LLM today.

But if you're talking about for, you know, decades in the future, what should you be kind of preparing for and so forth? I mean, it's, it's hard to say. I mean, the AI could get quite good at programming, but you can say that about kind of any field of human endeavor. So I guess I probably wouldn't have singled that out as like.

Say, don't study specific reprogramming. Um, I don't know if that's a good answer. Okay, hand it up. If a lot of people start using these agents to write code, I'm wondering how that's going to impact their IT security. You could argue that like, or like less check for certain issues, or you could argue that like, we get better at writing test suites which cover all the cases.

What are your opinions on this? Like is maybe for the out programmer like it security way to go? Because like the codes gotta be written, but someone still needs to check its support.

Sergey: Oh, wow. You guys are all trying to choose career basically.

don't use a fortune color for the general line of questions, but I, I do think that. You know, using an AI today to write, let's say, unit tests, is pretty straightforward. Like, that's the kind of thing the AI does really quite well. So, I guess my hope is that AI will make code more secure, not less secure.

I mean, it's kind of, it's usually Insecurity is, to some extent, the effect of people being lazy, and the one thing that AI is kind of good at is, you know, not being lazy. So, if I had to bet, I would say there's probably a net benefit to security with AI. Um, but I wouldn't discourage you from pursuing a career in IT security based on that.

That's pretty much it. Um, okay, next.

Do I want to? Um, yeah. Yeah, yeah. I mean, I think it's, um, different people mean different things about that. But to me, the reasoning aspects are really exciting and amazing. And you know, I kind of came out of retirement just because of the trajectory of AI. It was so exciting. And as a computer scientist, just seeing what these models can do year after year is astonishing.

Any efforts on, like, humanoid robotics or these, because there was so much progress in Google X, like, in 2015-16?

Sergey: Oh, humanoid robotics. Um, boy, we've done a lot of humanoid robotics over the years, and sort of acquired or sold a bunch of companies for humanoid robotics. Um, and now there are a zillion, oh, sorry, not a few companies doing humanoid robotics, and internally we still have groups that work on robotics in very varying forms.

So what are my thoughts about that? I don't know, you know, In general, I worked on apps prior to this sort of new AI wave. Their focus was more hardware projects, for sure but honestly, I guess I found a hard way to open it. Hardware is much more difficult. Um, kind of, on a technical basis, on a business basis, in every way.

So, I'm not discouraging people from doing it, we need people, for sure, to do it. Um, at the same time, while the software and the AIs are getting so much faster at such a high rate, I guess, to me, that feels like that's kind of the rocket ship. Um, And, I feel like if I get distracted, in a way, by making hardware for today's AIs, um, that might not be the best of use of time compared to what is the next kind of level of AI I could be able to support.

And, for that matter, will it design a robot for me? That's for my personal. There aren't a bunch of people at Google Now, but who could work on hardware.

Considering advertising revenues, what's your take on how advertising would be described as everything?

Sergey: Yeah. That's the way you would go? The question about advertising? Yeah, I of all people, am not too terribly concerned about business model shifts. I mean I think it's wonderful that we've been now for 25 years or whatever able to give just world class information, search, for free to everyone, and that's supported by advertising, which in my mind is great. It's great for the world, a kid in Africa has just as much access to basic information as the President of the United States or what have you.

So that's good. At the same time, I expect business models are going to evolve over time. And maybe they'll still be advertising, because whatever the advertising kind of works better, the AI is able to tailor it better, or you like it. But even if it happens to move to, you know, now we have, Gemini Advanced, other companies have their, you know, paid models.

I think the fundamental issue is that you're delivering a huge amount of value. Displacing all the mental effort that would have been required to take the place of that, whether in your time or labor or what have you, is enormous. The same thing was true in search. So, I personally feel as long as there's huge value being generated, we'll figure out the business models.

Does Chrome's third party cookie application give Google ads an advantage

Sergey: Third party cookie? You know, I'm going to face test how naive I am about the detail. I mean, I vaguely am aware of that stuff. I know I can't think of how those things interact, I'm sorry. On you are in Adtech, well maybe you should answer the question.

Where do you see Google Search going?

Sergey: Where do I see Google Search going? Well, it's a super exciting time for search. Because your ability to answer questions with AI is just so much greater. I think it's, the bigger opportunity is in situations where you are recall limited more so.

Like, you might ask a very specialized question, or it's related to your own personal situation, in a way that nobody out there, you know, on the internet has already written about. And for the questions that a million people have written about already and thought deeply about, It's probably not a big deal.

But the things that are very specific to what, you know, you might care about right now in a particular way, that's a huge opportunity. And you know, you can imagine all kinds of products and UIs and different ways to deliver that, but basically AI is the enabler, just doing a much better job in that case.

What does AI do for Mortality?

Sergey: Mortality. I'm probably not as well versed as all of you are, to be honest, but I've definitely seen the kind of the molecular AI make huge amounts of progress. You could imagine that there would also be a lot of progress that we haven't seen yet on the epidemiology side of things, to just be able to get more honest, better controlled, a kind of broader understanding what's happening to people's health around the world.

I don't even have a good answer on the last one, I don't have a really brilliant immortality key just like that. But, it's the kind of field that for sure benefits from AI. Whether you're a researcher, or like, you know, I want it to just summarize articles to me.

In the future, you know, I would expect the AI would actually give you novel hypotheses to test. It does that today, the AlphaFold of the world, but maybe in more complex systems than just molecules.