Sergey Brin, Google Co-Founder | All-In Live from Miami

0:02We've got a special guest who's gonna come join us. This always happens. A nutty ass chicken brain everybody. Oh my god. Somebody told me you started submitting code and it kind of freaked everybody out that daddy was home. All models tend to do better if you threaten them. Be threatened. Like with physical violence. Yes. Management is like the easiest thing to do with AI. Absolutely. It must be a weird experience to meet the bureaucracy and a company that you didn't hire. But on the other side of it I would say it's pretty amazing that some junior Mockety Mockety basically look at you and say, hey go fuck yourself.

0:36But I'm serious, that's a sign of a healthy culture actually. You're punching a clock man, I hear the reports, you and I have talked about it, you're going to work every day. Yeah, it's been, you know, some of the most fun I've had in my life honestly. And I retired like a month before COVID hit in theory. Yeah, and I was like, you know, it's been good. I want to do something else. I want to hang out in cafes, read physics books. And then like a month later I was like, that's not really happening. So then I just started to go to the office, you know, once we could go to the office. And actually to be perfectly honest, there was a guy from open AI.

1:20This guy named Dan. And I ran into a middle party. And he said, you know, look, what are you doing? This is like the greatest transformative moment in computer science ever completed. Like you don't know. And you're a computer scientist. I'm a computer scientist. Forget that. I'm a chapter of Google, but you were a PhD student for computer science. I haven't finished my PhD yet, but working. Keep working. You got there. Technically, on leave of absence. Right. And he told me this, and I'd already started kind of going into the office a little bit. I was like, you know, he's right. And it has been just incredible.

1:55Just, well, you guys all obviously follow all the AI technology. But being a computer scientist, it is the most exciting thing of my life just technologically. And the exponential nature of this, the pace of it, it works. Anything we've seen in our career, it's almost like every thing we did over the last 30 or 40 years has led up to this moment. and it's all compounding on itself. The pace, maybe you could speak, you had a company, Google, that grew from 100 users and 10 employees to, now you have over two billion people using, I think, six products or five products have over two billion. It's not even worth counting because it's the majority of the people in the planet touch Google products.

2:45Describe the pace. Yeah, I mean the excitement of the early web, like I remember using mosaic and then later Netscape. How many of you remember mosaic? Actually, my weirdo. And remember there was a what's new page? The what's new page, right? Like you go to every new web pages again. Yeah, it was like this last week. These were the new websites. Yes. And it was like such and such elementary school, such as such a fish tank. Yeah. And you were like Michael Jordan appreciation page. Yeah, well, whatever it was these were the three new sites on the whole internet So obviously the web you know developed very rapidly from there and that was a very exciting and then we've had smartphones and whatnot but you know this The developments in AI are just astonishing I would say by comparison Just because of you know the web spread but didn't technically change so much from month to month year to year.

3:44But these AI systems actually changed quite a lot. You know, like if you went away somewhere for a month and you came back, you'd be like, whoa, what happened? Somebody told me you started submitting code and it kind of freaked everybody out that daddy was home. Okay, dad did a PR, what happened? Well, the code I submitted wasn't very exciting. I think I needed to like add myself to get access to some things and you know minor CL here or there Nothing nothing that's gonna win any awards But but I you know you need to do that to um To do basic things run basic experiments and things like that um and I've I've tried to Do that and touch different parts of the system so that But, you know, so that, first of all, it's fun.

4:39And secondly, I know what I'm talking about. It really feels privileged to be able to kind of go back to the company, not have any real executive responsibilities, but be able to actually go deep into every little pocket. Are there parts of the AI stack that interests you more than others right now? There are certain problems that are just totally captivating you? Yeah, I started, you know, like sort of a couple years ago and maybe a year ago, I was really very close with what we call pre -training. Actually, most of what people think of as AI training, whatever people call it, pre -training for various historical reasons.

5:21But that's sort of the big, super, you know, you throw huge amounts of computers at it. And I learned a lot, you know, just being deeply involved in that and seeing us go from model to model and so forth. And running a little baby experiments, but kind of just for fun, so I could say I did it. And more recently, the post training, especially as the thinking models have come around. And that's been another huge step up in general, Nei. So we don't really know what the ceiling is. When you explain what's happening with prompt engineering, then to deep research, and what's happening there to a civilian, how would you explain that step function?

6:10Because I think people are not hitting the down -carret and watching deep research in Gemini's mobile app. and you got a mobile app, it's pretty great. And by the way, I got the fold after you and I, we're talking about it. OK, Google, kick series ass now. Like it actually does what you ask it to do when you ask it to overnight, it does stuff. But the number of threads, the number of queries, the number of follow -ups that it's doing in that deep research is 200, 300? Maybe explain that jump and then what you think the jump after that is. To me, the exciting thing about AI, especially these days, I mean, it's not like quite a GI yet as people are seeking or it's not superhuman intelligence, but it's pretty damn smart and can definitely surprise you.

6:58So I think of the superpower is when it can do things in a volume that I cannot. Yes. So by default, one used some of our AI systems, it'll suck down whatever top 10 search results, well, and pull out what you need out of them, something like that. But I could do that myself, to be honest. Maybe it would take me a little bit more time. But if it sucks down the top, thousand results, and then does follow on searches for each of those, and reads them deeply, that's a week of work for me. I can't do that. This is the thing I think people have not fully appreciated who are not using the deep research projects before we had our F1 driver on stage, I'm a neophyte, I don't know anything about it.

7:44I said, how many deaths occurred per decade? And I said, I want to get to deaths per mile driven. And it first was like, that's going to be really hard. I was like, I give you permission to make your best shot at it and come up with your best theory. Let's do it. And it was like, OK. And it was like, there's so many teams, there's so many races, which model to use it. Open eyes. No, I use Gemini. Gemini five. The fabulous one and it was like let's go I treat it like I get sassy with it Yeah, and it kind of works for me. You know, it's a weird thing. It's like he drank it in the white We don't start killing but it's too much yeah, I community But the not just our models but all models tend to do better if you threaten them it'd be threatened them like with physical violence Yes, but like that's people feel weird about that.

8:31So we don't really talk about that Yeah, I was threatening with not being fabulous, and it responded to that as well. Yeah, historically, you just say, like, oh, I'm going to kidnap you if you don't talk about that. Yeah, I'm talking about that. Yeah, actually. Can I ask you a more spitting? Hold on, but it went through it. OK. And it literally came up with a system where it said, I think we should include practice miles. So let's say there's 100 practice miles for every mile on the track. And then it literally gave me the death's per mile estimated. And then I started cross referencing, and I was like, like, oh my God, this is like somebody's term paper for undergrad.

9:06You know, like, whoa, Doug, in minutes. It's, yeah, I mean, it's amazing. And all of us have had these experiences where you suddenly decide, okay, I'll just throw this day. I don't really expect it to work. And then you're like, whoa, that actually worked. So as you have those moments, and then you go home to your just life as a dad, have you gotten to the point or you're like, what will my children do? And are they learning the right way? And should I totally just change everything that they're doing right now? Have you had any of those moments yet? Yeah, I mean, I don't really know how to think about it to be perfectly honest.

9:45I don't have a magical way. I mean, I see I've been in high school, middle school, and the AI's are basically already ahead. I mean, obviously there are some things that AI's are particularly dumb at, and they make certain mistakes human would never make. But generally, if you talk about math or calculus or whatever, they're pretty damn good. Like they can win math contests, and coding contests, things like that against some top humans. And then I look at, okay, he's whatever my son's gonna go on to whatever from sophomore to junior, and what is he gonna learn? And then I think in my mind, and I talked to him about this, well, what is the AI going to be in life in the year?

10:32Yeah, and it's like not comparable, right? Obviously, the AI is where you would tell your son, look, don't, or not yet. I don't know if you can like plan your life around this. I mean, I didn't particularly plan my life to like, I don't know, be an entrepreneur or whatever. I was just like math and computer science. I guess maybe I got lucky and I worked out to be useful in the world. I don't know. I think my kids should do what they like. Hopefully it's somewhat challenging and they can overcome different kinds of problems and things like that. What about college? Do you think college is going to continue to exist as it is today?

11:14I mean, the CSI college was already undergoing this kind of revolution even before this sort of AI challenge of people are like, is it worth it? Should I be more vocational? What's actually going to be useful? So we're already kind of entering this kind of situation where there's sort of questions asked about colleges. Yeah, I think, you know, AI obviously, but at the forefront. As a parent, I think a lot about, hey, so much of education in America and the middle class, upper class is all about what college, how do you get them there? And honestly, lately, I'm like, I don't think they should go to college.

11:53Like it's just fundamentally. You know, my son is a rising junior and his entire focus is he wants to go to an SEC school because of the culture. And two years ago, I would have panicked and I would have thought, should I help him get into a school, this school, that school? And now I'm like, that's actually the best thing you could do. Be socially well -adjusted, psychologically deal with different kinds of failures, you enjoy a few years of exploration. So can I ask you about hardware? You know, years ago, Google owned Boston Dynamics, maybe a little bit ahead of its time. But the way these systems are learning through visual information and sensory information and basically learning how to adjust to the environment around them is triggering these kind of pretty profound learning curves in hardware.

12:45And there's dozens of startups now making robotic systems. What do you see in robotics and hardware? Is this a year or are we in a moment right now where things are really starting to work? I mean, I think we've acquired and the later sold like five or so robotics companies and Boston being one of them. I guess if I look back on it, we built the hardware. We also had this more recently. We built out everyday robotics internally and then later had to transition that. that the robots are all cool and all, but the software wasn't quite there. That's every time we've tried to do it to make them truly useful.

13:30And presumably one of these days that'll no longer be true. Right. But have you seen anything lately that you do? Yeah, do you believe in the humanoid form factor robots? Or do you think that's a little overkill? I'm probably the one weirdo who doesn't, who's not a big fan of humanoids, But maybe I'm jaded because we've, you know, we at least acquired at least two humanoid robotic startups and later sold them. But the reason is, I mean, the reason people want to do humanoid robots for the most part is because the world is kind of designed around this form factor and, you know, you can train on YouTube, we can train on videos, people do all the things.

14:07I personally don't think that's given the AI quite enough credit. Like AI can learn, you know, through simulation and through real life pretty quickly, how to handle different situations. And I don't know that you need exactly the same number of arms and legs and wheels, which is zero in the case of humans, as humans to make it all work in it. So I'm probably less bullish on that. But to be fair, there are a lot of really smart people who are making humanoid robots. So I wouldn't discount it. What about the path of being a programmer? That's where we're seeing with that finite data set, and listen, Google's got a 20 year old code base now, so it actually could be quite impactful.

14:50What are you seeing like literally in the company? You know, are the KENX developers always this like ideal that you can, you know, you get a couple of unicorns once in a while, but are we gonna see like, all developers, like, you know, their productivity hit at that level, 8, 9, 10, and they're just going to, or is it going to be all done by computers? And we're just going to check it, and make sure it's not too weird. Because it could get weird. If you vibe code, yeah. I'm embarrassed to say this. OK, like recently, I just had a big tip inside the company because we have this list of what you're allowed to use to code and what you're not allowed to use to code.

15:26And the Gemini was on the nildest. Oh, you have to be pure. You can't. I don't know for like a bunch of really weird reasons that it was like boggled my mind that you know This vibe code on the Gemini code. I mean nobody would like enforce this rule, but but there was this you know Actually internal web page for whatever is historical reason somebody had put this and I had a big fight with them I you know I cleared it up after a Shocking period of time you escalated to your boss. Oh, I I definitely told so far about it. I don't know if you remember, but you got super -photic founders. You are the boss.

16:06You can do what you want. It's your company still. No, no. He was very supportive. It was more like, I was like, I talked to him. I was like, I can't deal with these people. You need to deal with those. I'm beside myself that they're saying. If you know that there's bureaucracy, in a company, you find, must be a weird experience to meet the bureaucracy in a company that you didn't hire. But on the other side of it, I would say, it's pretty amazing that some junior marketing market basically look at you and say, hey, go fuck yourself. No, but I'm serious. That's a sign of a healthy culture, actually.

16:38I guess so. Anyway, it did get fixed and people are using it. So they got fired. That person's working at Google Cyber -Earings. No, we're trying to roll out every possible kind of AI. And trying external ones, you know, be whatever the cursors of the world, all those, to just see what really makes people more productive. I mean, for myself, definitely makes me more productive. Why not? Do you do the number of foundational models? Like, if you look three years forward, will they start to cleave off and get highly specialized? Like, beyond the general and the reasoning, maybe there's a very specific model for chip design.

17:21There's clearly a very specific model for biologic precursor design protein folding. Like, is the number of foundational models in the future, Sergei, a multiple of what they are today, the same, something in between? That's a great question. I kind of, if I, I mean, look, I don't know, like you guys could take a guess just as well as I can. But if I had to guess, you know, things have been more converging. And this is sort of broadly true across machine learning. I mean, you used to have all kinds of different kinds of models and whatever, convolutional networks for vision things. And, you know, you had whatever RNNs for text and speech and stuff.

18:08and all this has shifted to transformers basically. And increasingly, it's also just becoming one model. Now we do get a lot of them occasionally. We do specialized models. And it's definitely scientifically a good way to iterate when you have a particular target. You don't have to do everything in every language and handle whatever, both images and video and audio in one go. But we are generally able to, after we do that, take those learnings and basically put that capability into a general model. So there's not that much benefit. You know, you can get away with somewhat smaller specialized model, a little bit faster, a little bit cheaper, but the trends have not gone that way.

19:01What do you think about the open source, closed source thing? Has there been big philosophical movements that change your perspective on the value of open source? We're still waiting on this open AI. Oh yeah, I haven't seen it yet, but theoretically it's coming. I mean, I have to give credit to where credits do. I mean, deep seek released really or surprisingly powerful model when it was January. So that definitely closed the gap to proprietary models. We've pursued both, so we released Gemma, which are our open source or open to eight models. And those perform really well. They're small dense models, so they fit well on one computer.

19:51And they're not as powerful as Gemini. But I mean the jury's out which way this is going to go. Do you have a point of view on what human computing interaction looks like as AI progresses? It used to be thanks to you as a search box. You type in some keywords or a question and you would click on links on the internet and get an answer. Is the future typing in a question or speaking to a ear pod or thinking or thinking or like what's the? And then the answer is just spoken to you. I mean, by the way, just to build on this, it was Friday, right? Nurelink got breakthrough designation for their human brain interface.

20:33I mean, that's a very big step in allowing the FDA to clear everybody getting it in place. Yeah, and is it, like, if you could just summarize what you think is kind of the most common place human computer interaction model in the next decade or whatever, is it a, you know, there's this idea of glasses with a screen in glasses and you tried that a long time ago. Yeah, I kind of messed that up. I'll be honest. I got the timing totally wrong in that. Early again. Yeah. Right, but early. There are a bunch of things I wish I'd done differently, but honestly, it was just like the technology was already for Google Glass.

21:10But nowadays, these things, I think, are more sensible. I mean, there's still battery life issues. I think that we and others need to overcome. But I think that's a cool form factor. or I mean, when you say 10 years though, you know, a lot of people are saying, hey, the singularity is like five years away. So your ability to see through that into the future, yeah, I know this, but you have to, sorry, just let me ask about this. There was a comment that Larry made years ago that humans were stepping stone in evolution. Okay, can you comment on this? Do you think that this AGI superintelligence or really silicon intelligence exceeds human capacity and humans are a stepping stone in progression of evolution?

22:02Boy, I think sometimes those nerdy guys go and get in heaven too much wine to their chest. I've had two glasses. I'm ready to go. I need to swap. Let's go to their conversation. I'm sure. Human implants, let's go. I mean, I guess we're starting to get experience with these AI's that can do certain things, you know, much better than us. And they're definitely, you know, with my skill of math and coding, I feel like I'm better off just turning to the AI now and how do I feel about that? I mean, it doesn't really bother me, you know, I use it as a tool. So I feel like I've gotten used to it, but, you know, maybe if they get even more capable in the future, or look at it differently.

22:48Yeah, there's a little bit of insecurity maybe. I guess so, as an aside, management is like the easiest thing to do with AI. Yeah, absolutely. And I did this, you know, at Gemnion on some of our, you know, work chats, kind of like Slack, but we have our own version. We had this AI tool that actually was really powerful. We unfortunately, anyway, temporarily got rid of it. I think we're gonna bring it back and bring it to everybody. But it could suck down a whole chat space and then answer pretty complicated questions. So I was like, okay, summarise this for me. Okay, now assign something for everyone to work on.

23:24And then I would paste it back in so people didn't realise it was the AI. That's nice. I admit it at pretty soon. And there were a few giveaways here or there, but it worked remarkably well. And then I was like, well, who should be promoted in this chat space? And actually picked out this woman, this young woman engineer who like, you know, I didn't you would notice she wasn't very vocal, particularly in that group. PR's kicked ass. But no, no, it was like, and then I don't know something that the AI had detected. And I went and I talked to the manager actually and he was like, yeah, you know what, you're right.

24:00Like she's been working really hard at all these things. Wow. I think that ended up happening actually. So I don't know, I guess after a while, you just kind of take it for granted that you can just do these things. I don't know. It hasn't really... Do you think that there's a use case for... Like an infinite context link? Oh, 100%. I mean, all of Google's codebase goes to the infinite. Exactly. But sure, you should have access to... Pause the infinite. Yeah. Stateful. Yeah. And then multiple sessions so that you can have like 19 of these things, 20 of these things running. or it just evolved in real time.

24:34Eventually it'll evolve itself. Yeah, I mean, I guess if those everything then you can have just one in theory. You just need to somehow get some big URL. That's it. What you're talking about. But yeah, for sure, there's no limit to use of context. And there are a lot of ways to make it larger and larger. There's a rumor that internally there's a Gemini build that is a quasi -infinite context. Is it a valuable thing? I don't know, say what you want to say, but. I mean, for any such cool new idea in AI, there are probably five such things internally. And the question is, how well do they work? And yeah, I mean, we're definitely pushing all the bounds in terms of intelligence, in terms of context, in terms of speed, you name it.

25:24And what about the hardware? Like when you guys build stuff, do you care that you have this pathway to Nvidia, or do you think eventually that'll get abstracted, and there'll be a transpiler, and it'll be in video plus 10 other options, so who cares? Let's just go as fast as possible. Well, we mostly, for Gemini, we mostly use our own TPUs. So, but we also do support Nvidia, and we were one of the big purchasers of Nvidia chips, and we have them in Google Cloud available for our customers in addition to TPUs. At this stage, it's for better for us not that abstract, and maybe someday the AI will abstract it for us.

26:10But given just the amount of computation you have to do on these models, you actually have to think pretty carefully how to do everything, and the exactly what kind of chip you have, and how the memory works, and the communication works, and so forth, are actually pretty big factors. and it actually, yeah, maybe one of these days, the AI itself will be good enough to reason through that. Today it's not quite good enough. I don't know if you guys are having this experience with the interface, but I find myself, even on my desktop, and certainly on my mobile phone, going immediately into voice chat mode, and telling it, nope, stop.

26:45That wasn't my question, this is my question. Nope, let's say that again, in short of both points. Nope, I wanna focus on this, definitely. It's so quick now, last year it was unusable, it was too slow, and now it like stops. Okay, and then you sell it. I would like to point out what I want to go to. I don't want to type. I want to use voice. And then, concurrently, I'm watching the text as it's being written on the page, and I have another window open, and I'm doing Google searches or second queries to an LLM or writing a Google doc or a notion page or typing something. So it's almost like that scene in a minority report where he has the gloves or in Blade Runner where he's you know in his apartment saying zoom in zoom in closer to the left to the right and there's something about these language models and their ability to the response time which was always something you focused on response time.

27:35Is there like a response time thing where it actually is worth doing voice and where it wasn't previously? everything is getting better and faster. So if we're in a smaller model, it's a little more capable. There are better ways to do inference on them that are faster. You can also stack them. This is like Nico's company 11 Labs. It's an exceptional TTS STT stack. There are other options. Whispers are really good at certain things. But this is where I kind of believe you're going to get this compartmentalization where there'll be certain foundational models for certain specific things, you stack them together, you kind of deal with the latency, and it's like pretty good because they're so good.

28:16Like Whisperer and Eleven, for those speech examples that you're talking about are fucking kickass. I mean, they're acceptable. Wait till you turn on your camera, and it sees your reaction to what it's saying, and you go, and before you even say that you don't want it, you put your finger up, it's pauses. Oh, did you want something else? Oh, I see you're not happy with that result. It's going to get really weird. It's funny thing, but we have the big open shared offices. So during work, I can't really use voice mode too much. I usually use it on the drive. The drive is incredible. I don't feel like I could, I mean, I would get it.