Sergey Brin，Google 联合创始人｜All-In 迈阿密现场

We've got a special guest who's going to come join us. This always happens. Another banger, everybody. Oh my God. Somebody told me you started submitting code, and it kind of freaked everybody out that daddy was hungry.

All models tend to do better if you threaten them.

If you threaten them 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 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 muckety-muck can basically look at you and say, “Go fuck yourself.” No, but 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 some of the most fun I've had in my life, honestly. I retired like a month before COVID hit, in theory.

In theory.

And I was like, “This has been good. I want to do something else. I want to hang out in cafés and read physics books.” Then, like, a month later, I was like, “That's not really happening.” So I just started going to the office once we could go to the office.

Actually, to be perfectly honest, there was a guy from OpenAI named Dan. I ran into him at a little party, and he said, “Look, what are you doing? This is the greatest transformative moment in computer science ever.” Completely. I'm a computer scientist. Forget that—I'm the founder of Google, but I'm a PhD student in computer science. I haven't finished my PhD yet, but I'm working on it.

Keep working. We'll get there. Technically, you're on a leave of absence, right?

And he told me this, and I'd already started going into the office a little bit. I was like, “You know, he's right.” It has been incredible. You guys all obviously follow the AI technology, but being a computer scientist, it is the most exciting thing of my life, technologically. The exponential nature of this—the pace of it dwarfs anything we've seen in our careers. It's almost like everything we did over the last 30 or 40 years has led up to this moment, and it's all compounding on itself.

The pace—you had a company, Google, that grew from 100 users and 10 employees to now having over 2 billion people using it. I think 6 products, or 5 products, have over 2 billion users. It's not even worth counting because the majority of the people on the planet touch Google products. Describe the pace.

Yeah. I mean, the excitement of the early web—I remember using Mosaic and then, later, Netscape. How many of you remember Mosaic, actually? Am I a weirdo?

And you remember there was a “What's New” page? The “What's New” page was great.

Right. You'd go through 2 or 3 new web pages.

It was like, “In this last week, these were the new websites.”

Yes. It was such and such elementary school, such and such a fish tank—

Michael Jordan appreciation page.

Whatever it was, these were the 3 new sites on the whole internet. Obviously, the web developed very rapidly from there, and that was very exciting. Then we've had smartphones and whatnot, but the developments in AI are just astonishing by comparison. The web spread, but it didn't technically change so much from month to month or year to year. These AI systems actually change quite a lot, quite a lot.

Are there parts of the AI stack that interest you more than others right now? Are there certain problems that are just totally captivating you?

Yeah, I started, you know, sort of—I don't know—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. But that's sort of the big, super—you know, you throw huge amounts of computers at it. And I learned a lot just being deeply involved in that and seeing us go from model to model and so forth and running 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 in AI. So we don't really know what the ceiling is.

When you explain what's happening from prompt engineering to deep research to a civilian, how would you explain that step function? I don't think people are hitting the down caret and watching Deep Research in Gemini's mobile app. You've got a mobile app, and it's pretty great. By the way, I got the Fold after you and I were talking about it, and Google kicks Siri's ass now. It actually does what you ask it to do. When you ask it to open something, it does it. But the number of threads, the number of queries, and the number of follow-ups that Deep Research is doing is 200 or 300. Explain that jump, and then what you think the jump after that is.

To me, the exciting thing about AI, especially these days, is that it's not quite AGI yet, as people are seeking, and it's not superhuman intelligence, but it's pretty damn smart and can definitely surprise you. I think of the superpower as being when it can do things at a volume that I cannot.

By default, when you use some of our AI systems, it'll suck down the top 10 search results and pull out whatever you need from them. 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 1,000 results, then does follow-up 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 if they're 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. I said, “How many deaths occurred per decade?” Then I said, “I want to get to deaths per mile driven.”

Which model did you use?

I used Gemini. Gemini's fabulous version—the fabulous one. I said, “I want to get to deaths per mile driven,” and at first it was like, “That's going to be really hard.” I said, “I give you permission to make your best shot at it and come up with your best theory. Let's do it.”

It was like, “Okay, there are this many teams and this many races.” I treat it like I get sassy with it, and it kind of works for me. It's a weird thing. It's like you're drinking the wine.

We don't circulate that too much in the AI community, but not just our models—all models tend to do better if you threaten them.

If you threaten them with physical violence.

But people feel weird about that, so we don't really talk about it.

I threatened it with not being fabulous, and it responded to that as well.

Historically, you just say, “I'm going to kidnap you if you don't.”

But hold on. It went through it and 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. Then it literally gave me the estimated deaths per mile. I started cross-referencing it, and I was like, “Oh my God, this is like somebody's term paper for undergrad.” Done in minutes.

It's amazing. All of us have had these experiences where you suddenly decide, “Okay, I'll just throw this at the AI. 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 go home to your life as a dad, have you gotten to the point where you're like, “What will my children do? Are they learning the right way? Should I totally change everything they're doing right now?” Have you had any of those moments yet?

Yeah. I don't really know how to think about it, to be perfectly honest. I don't have a magical way of thinking about it. I have a kid in high school and a kid in middle school, and the AIs are basically already ahead. Obviously, there are some things AIs are particularly dumb at, and they make certain mistakes a human would never make. But generally, if you talk about math or calculus or whatever, they're pretty damn good. They can win math contests and coding contests against some top humans.

Then I look at my son, who's going from sophomore to junior, and I think, “What is he going to learn?” I talk to him about this: “What is AI going to be in a year?”

Yeah.

It's comparable, right? Obviously, there are areas where I would tell my son, “Don't”—or maybe not yet. I don't know if you can plan your life around this. I didn't particularly plan my life to be an entrepreneur or whatever.

I just liked math and computer science. I guess maybe I got lucky, and it 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 specifically—what about college? Do you think college is going to continue to exist as it is today? It seems like college was already undergoing this kind of revolution even before this sort of AI challenge. People are asking, “Is it worth it? Should I be more vocational? What's actually going to be useful?” So we're already entering this kind of situation where there are questions being asked about college.

Yeah, I think AI obviously puts that at the forefront. As a parent, I think a lot about how so much of education in America, and in the middle class and upper class, is all about what college and how do you get them there. Honestly, lately, I'm like, “I don't think they should go to college.” It's just fundamentally—my son is a rising junior, and his entire focus is that he wants to go to an SEC school because of the culture.

Two years ago, I would have panicked and thought, “Should I help him get into a school? This school, that school?” Now I'm like, “That's actually the best thing you could do. Be socially well-adjusted, psychologically deal with different kinds of failures, and enjoy a few years of exploration.”

Yeah. Yeah. Yeah. Sergey, can I ask you about hardware? Years ago, Google owned Boston Dynamics, maybe a little bit ahead of its time, but the way these systems are learning through visual and sensory information, and basically learning how to adjust to the environment around them, is triggering some pretty profound learning curves in hardware. There are 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 later sold 5 or so robotics companies, Boston Dynamics being one of them. I guess if I look back on it, we built the hardware. More recently, we also built out Everyday Robots internally and then later had to transition that. The robots are all cool and all, but the software wasn't quite there. That's been true every time we've tried to make them truly useful, and presumably one of these days that'll no longer be true, right?

Yeah. Do you believe in humanoid robots and the humanoid form factor, or do you think that's a little overkill?

I'm probably the one weirdo who's not a big fan of humanoids, but maybe I'm jaded because we acquired at least 2 humanoid robotics startups and later sold them. The reason people want to do humanoid robots, for the most part, is because the world is kind of designed around this form factor. You can train on YouTube; we can train on videos of people doing all the things.

I personally don't think that's giving the AI quite enough credit. AI can learn through simulation and through real life pretty quickly how to handle different situations. I don't know that you need exactly the same number of arms and legs—and wheels, which is 0 in the case of humans—as humans have to make it all work. 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 it, with that finite data set. Google's got a 20-year-old code base now, so it actually could be quite impactful. What are you seeing literally in the company? The 10x developer is always this ideal—you get a couple of unicorns once in a while—but are we going to see all developers have their productivity hit that level, 8x and 10x, and they're just going to—or is it going to be all done by computers, and we're just going to check it to make sure it's not too weird? Because it could get weird if you vibe-code.

I'm embarrassed to say this. I recently had a big tiff inside the company because we had this list of what you're allowed to use to code and what you're not allowed to use to code, and Gemini was on the no list. You have to be pure. You can't vibe-code on the Gemini code, for a bunch of really weird reasons that boggled my mind.

Nobody would enforce this rule, but there was an actual internal webpage. For whatever historical reason, somebody had put this there, and I had a big fight with them. I cleared it up after a shocking period of time.

You escalated to your boss?

Oh, I definitely told him about it.

Sorry, I don't know if you remember, but you've got super-voting shares. You are the boss. You can do what you want. It's your company still.

No, no, he was very supportive. I talked to him and said, “I can't deal with these people. You need to deal with this.” I'm beside myself that they're saying it's weird that there's 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 muckety-muck can basically look at you and say, “Go fuck yourself.”

No, but I'm serious. That's a sign of a healthy culture, actually, I guess. So anyway, it did get fixed, and people are using it.

So they got fired. That person's working in Google Siberia.

No, we're trying to roll out every possible kind of AI and trying external ones, whether it be Cursor or all of those tools, to see what really makes people more productive. For myself, it definitely makes me more productive.

Do you think the number of foundational models, if you look 3 years forward, will start to cleave off and get highly specialized beyond the general and the reasoning models? Maybe there's a very specific model for chip design. There's clearly a very specific model for biologic precursor design and protein folding. Is the number of foundational models in the future, Sergey, a multiple of what they are today, the same, or something in between?

That's a great question. I mean, look, I don't know. You guys can take a guess just as well as I can. But if I had to guess, things have been more convergent, and this is broadly true across machine learning.

You used to have all kinds of different kinds of models—convolutional networks for vision, and RNNs for text and speech and stuff. All this has shifted to transformers, basically, and increasingly, it's also just becoming 1 model.

We do get a lot of oomph occasionally from 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 both images and video and audio in 1 go. But we're generally able to take those learnings and basically put that capability into a general model after we do that.

So there's not that much benefit. You can get away with a somewhat smaller specialized model that's a little bit faster and a little bit cheaper, but the trends have not gone that way.

What do you think about the open-source, closed-source thing? Have there been big philosophical movements that changed your perspective on the value of open source? We're still waiting on this OpenAI open-source drop. I mean, we haven't seen it yet, but theoretically, it's coming.

I mean, we have to give credit where credit's due. DeepSeek released a really surprisingly powerful model in January or so. That definitely closed the gap to proprietary models.

We've pursued both. We released Gemma, which are our open-source—or, you know, open—models, and those perform really well. They're small, dense models, so they fit well on 1 computer. They're not as powerful as Gemini, but the jury's out on which way that's going to go.

Do you have a point of view on what human-computer interaction looks like as AI progresses? It used to be, thanks to you, 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, speaking to an AirPod, or thinking? And then the answer is just spoken to you.

By the way, just to build on this, it was Friday, right? Neuralink got breakthrough designation for its human brain interface. That's a very big step in allowing the FDA to clear everybody to get it implanted. If you could just summarize what you think is kind of the most commonplace human-computer interaction model in the next decade or whatever, is it—you know, there's this idea of glasses with a screen in the 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 on that. Early again.

Yeah. Right. Right. But early.

There are a bunch of things I wish I had done differently, but honestly, the technology wasn't ready for Google Glass. Nowadays, these things, I think, are more sensible. I mean, there's still battery-life issues that we and others need to overcome, but I think that's a cool form factor.

I mean, when you say 10 years, a lot of people are saying, hey, the singularity is like 5 years away. So your ability to see through that into the future is very important. Sorry, just let me ask about this. There was a comment that Larry made years ago that humans were a stepping stone in evolution. 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 the progression of evolution?

Boy, I think sometimes us nerdy guys go and have a little too much wine. I've had 2 glasses, and I'm ready to go. I need some more for this conversation. Human implants. Let's go.

I guess we're starting to get experience with these AIs that can do certain things much better than us. With my skill in math and coding, I feel like I'm better off just turning to the AI now. How do I feel about that? It doesn't really bother me; I use it as a tool. So I feel like I've gotten used to it, but maybe if they get even more capable in the future, I'll look at it differently. There's a moment of insecurity, maybe. I guess.

As an aside, management is like the easiest thing to do with AI.

Yeah, absolutely. I did this with Gemini on some of our work chats, kind of like Slack, but we have our own version. We had this AI tool that was actually really powerful. Unfortunately, we temporarily got rid of it. I think we're going to bring it back and bring it to everybody.

It could suck down a whole chat space and then answer pretty complicated questions. I was like, “Okay, summarize this for me. Now assign something for everyone to work on.” Then I would paste it back in so people didn't realize it was the AI. I admitted it pretty soon. There were a few giveaways here or there, but it worked remarkably well.

Then I was like, “Well, who should be promoted in this chat space?” I actually picked out this young woman engineer who—I didn't even notice she wasn't very vocal, but her PRs kicked ass. No, no, it was something that the AI had detected. I went and talked to the manager, actually, and he was like, “Yeah, you know what? You're right. She's been working really hard on all these things.”

I think that ended up happening, actually. So I don't know. I guess after a while, you just take it for granted that you can do these things.

Do you think there's a use case for an infinite context length?

Oh, 100%. I mean, all of Google's codebase goes infinite, but sure, you should have infinite access.

Yeah. Stateful.

Yeah, and then multiple sessions, so that you could have 19 of these things, 20 of these things running, or just evolve itself.

Eventually, it'll evolve itself.

Yeah. I mean, I guess if it knows everything, then you can have just 1 in theory. You just need to somehow disambiguate what you're talking about. But, yeah, for sure, there's no limit to the 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's quasi-infinite context. Is it a valuable thing?

Well, you say what you want to say, but for any such cool new idea in AI, there are probably 5 such things internally. The question is how well they work. We're definitely pushing all the bounds—in terms of intelligence, in terms of context, in terms of speed, you name it.

And what about the hardware? 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 NVIDIA plus 10 other options, so who cares? Let's just go as fast as possible.

For Gemini, we mostly use our own TPUs. But we also support NVIDIA, and we're one of the big purchasers of NVIDIA chips. We have them in Google Cloud available for our customers, in addition to TPUs.

At this stage, for better or for worse, it's not that abstract. Maybe someday the AI will abstract it for us, but given the amount of computation you have to do on these models, you actually have to think pretty carefully about how to do everything. Exactly what kind of chip you have, how the memory works, and how the communication works are actually pretty big factors. 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. That wasn't my question. This is my question. Let's say that again in shorter bullet points. Nope, I want to focus on this.” It's so quick now. Last year, it was unusable; it was too slow. Now it stops. That's where I want to go. I don't want to type. I want to use voice.

Concurrently, I'm watching the text as it's being written on the page, and I have another window open where I'm doing Google searches or second queries to an LLM or writing a Google Doc or a Notion page or typing something. It's almost like that scene in Minority Report where he has the gloves, or in Blade Runner where he's in his apartment saying, “Zoom in, zoom in. Closer to the left, to the right.”

There's something about these language models and their response time, which was always something you focused on. Is there a response-time threshold where it actually is worth doing voice, and where it wasn't previously?

Everything is getting better and faster, and so forth. Smaller models are 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, ElevenLabs—it's an exceptional TTS SSD stack. There are other options. Whisper is really good at certain things.

This is where I believe you're going to get this compartmentalization where there'll be certain foundational models for certain specific things. You stack them together, you deal with the latency, and it's pretty good because they're so good. Whisper and ElevenLabs, for those speech examples that you're talking about, are kick-ass. They're exceptional.

Well, wait till you turn on your camera and it sees your reaction to what it's saying, and before you even say that you don't want it—or you put your finger up—it 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 a 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.

Yeah. I don't feel like I could use it. I mean, I would get its output in my headphones, but if I want to speak to it, then everybody's listening to me. So it's weird. I just think that would be socially awkward, but I should do that. In my car ride, I do chat to the AI, but then it's audio in, audio out.

But I feel like—honestly, maybe it's a good argument for a private office. I should spend more time like you guys are. You could talk to your manager; they might get one. I like being with everybody. But I do think there's this AI use case that I'm missing, which I should probably figure out how to try more often.

If people want to try your new product, is there a website they can visit or something or special code or go check?

I mean, honestly, there's a dedicated Gemini app. If you're using Gemini, just like you're going through the Google navigation from your search, just get the download the actual Gemini app. It's kickass. It really is the best models. I think it is.

You should use 2.5 Pro.

2.5 Pro. Pay the—

It's a—you got to pay, right?

Yeah, you got a few queries, you got a few prompts for free, but if you do it a bunch, you need to make all these like 20 bucks a month.

You got a vision for making it free and throwing some ads on the side.

One step down in hardware cost, the whole thing will be free.

Well, okay.

It's free today without ads on the side. You just got a certain number of the top model. I think we're likely going to have always now sort of top models that we can't supply infinitely to everyone right off the bat. But wait 3 months and then the next generation.

All right, give it up for Sergey Brin. Thank you. [Applause]

Okay, thanks everybody for watching that amazing interview with Sergey Brin and thanks Sergey for joining us in Miami. If you want to come to our next event, it's the All-In Summit in Los Angeles, fourth year for All-In Summit. Go to all-in.com/events to apply. A very special thanks to our new partner, OKX, the new money app. OKX was the sponsor of the McLaren F1 team, which won the race in Miami. Thanks to Haidider and his team, an amazing partner and an amazing team. We really enjoyed spending time with you. And OKX launched their new crypto exchange here in the US. If you love All-In, go check them out. And a special thanks to our friends including Shane over at Poly Market, Google Cloud, Salana, and BVNK. We couldn't have done it without y'all. Thank you so much. We'll let your winners ride. [Music]