Sergey Brin on the Future of AI & Gemini

0:00 [MUSIC PLAYING] LOGAN KILPATRICK: Hey, everyone.

0:05 Thanks for joining us.

0:06 We've got an I/O special, Sergey Brin.

0:09 We're talking all things Google.

0:11 Thanks for taking the time to chat.

0:13 SERGEY BRIN: Thank you, Logan.

0:14 And you and I are all the time in chat spaces on all kinds of products,

0:18 but it's nice to hang out in real life.

0:20 LOGAN KILPATRICK: Yeah,

0:21 it is the— my California experience is always— it's incredibly fun

0:25 to— I spent a bunch of time with Koray yesterday and today.

0:28 And he's just— it's like you feel the warmth and the humanity

0:31 of AI progress when you spend time in person with everyone.

0:34 So it's been a ton of fun.

0:36 But we're sitting here at I/O, and I think the general sentiment from the world,

0:40 and I think also from the team internally, is incredibly great day for Google,

0:45 a ton of progress across models, across all of our products.

0:51 What's your take?

0:52 What's your reaction?

0:53 Obviously, there's a lot of stuff we still need to do, but where's your head at?

0:59 SERGEY BRIN: Yeah, I think it was definitely a phenomenal set of announcements.

1:02 Honestly, I probably didn't even know about 30% of them or so.

1:08 There's only so much time.

1:10 And I'm deep in the weeds on Gemini,

1:12 and I didn't even know about the virtual fit,

1:16 for example, for products in Google Search.

1:18 I didn't realize we were shipping that.

1:21 So there were many things that also even surprised me.

1:25 That was great.

1:27 I think the reception has been great.

1:30 There are so many things though.

1:33 I think it takes a while for people to explore them,

1:37 wrap their heads around them.

1:39 And obviously, we're very busy delivering all of them right now.

1:45 And it's a lot of energy across the way,

1:50 just making sure things are actually shipping, shipping smoothly,

1:53 that people are able to sign up for Ultra,

1:57 have all those new features, and so forth.

2:00 LOGAN KILPATRICK: Yeah,

2:00 I feel like I/O is the start of lots of work for lots of other people.

2:04 It's like at the finish line for some teams,

2:06 and then it's like the starting line for a bunch of other teams.

2:08 How do you think about— obviously,

2:11 we launched more— there's a bunch of Gemini announcements.

2:14 Gemini Diffusion, we'll talk more about later.

2:16 Deep Think, which is continuing to push the frontier on reasoning models.

2:20 I see you in the reasoning strike chat all

2:23 the time pushing on people to keep pushing the frontier, which I love to see.

2:28 How do you think about— maybe how you think about your focus

2:33 but also just generally the DeepMind team's focus around like Veo and Imagen.

2:38 And we have a whole suite of generative media models.

2:41 We just announced a Lyria, our music model,

2:44 and then also at the same time, the core Gemini main model.

2:49 Do you work on the Gen media stuff at all,

2:52 or is it mostly laser-focused on Gemini at the moment?

2:55 SERGEY BRIN: Mostly, I'm on Gemini,

2:58 the core text model primarily because I think

3:02 that that's what will lead to, for example,

3:07 self-improvement, will help US code and develop

3:11 the science behind AI even better.

3:13 So that's the number one focus that I have.

3:17 At the same time, the generative media is so amazing.

3:21 LOGAN KILPATRICK: Yeah.

3:22 SERGEY BRIN: That is just so superhuman.

3:28 With a text model, there's some math problems or whatever that I might be able

3:34 to solve that it gets wrong or something

3:36 like that, or stumbles on a piece of code.

3:38 Although, that's less and less frequent.

3:40 And I actually end up now relying on Gemini

3:44 to do some of my coding math, and so forth.

3:47 But nevertheless, it's in the range of human.

3:52 LOGAN KILPATRICK: Yeah.

3:53 SERGEY BRIN: Given my artistic talent,

3:55 there would be zero chance of me ever coming up with either an image or a video.

4:02 Just the amount of work that that would— I

4:06 imagine would take if I was an expert videographer, 3D renderer, or whatever,

4:12 special effects person, that's got to be a month of solid work to get

4:17 the thing that I can get in a few minutes.

4:20 And obviously, it's visually so compelling that just sucks you in.

4:28 You can't escape it.

4:29 LOGAN KILPATRICK: The audio piece with Veo

4:32 makes it feel— I historically have— personally,

4:37 I think generating videos is awesome and stuff,

4:39 but it's always felt slightly gimmicky to me.

4:43 And I think when I saw audio in Veo 3 on stage yesterday,

4:48 I think it like— that was the moment for me where I was like, OK,

4:51 this is actually— so many people are going

4:53 to be able to— just because also practically, historically,

4:55 you could generate a video, but then you'd have to go,

4:58 and where does audio come from?

5:00 How do you sync everything up?

5:02 And now you can make humans talking and having a conversation,

5:06 and it just does it all well.

5:09 And it blew my mind.

5:11 SERGEY BRIN: Yeah, you're right.

5:12 I've been a huge fan of that.

5:14 I'm a pretty visual person, personally.

5:16 I'm not a very, I guess, audio person.

5:19 But over the years, especially with Google Glass,

5:23 the moment we added some sound,

5:26 that's just like— it adds so much richness to have sound.

5:31 LOGAN KILPATRICK: Yeah.

5:32 SERGEY BRIN: It's like you're better off

5:33 adding audio than adding 3D, for example.

5:37 Although some of the 3D stuff's cool if you played with the big wearable thing.

5:42 LOGAN KILPATRICK: Yeah.

5:43 SERGEY BRIN: But anyhow, yeah,

5:47 it's just an incredible change in perception when you get audio working.

5:53 And I knew I saw the model training the last month or two.

5:56 And I just saw it from checkpoint to checkpoint.

6:00 And I knew that, wow, this is just going to feel different.

6:05 LOGAN KILPATRICK: Yeah,

6:06 it'll be interesting to see how the fusion of those capabilities— because it

6:10 does seem like there's a lot

6:12 of similarities to mainline Gemini— mainline Gemini model.

6:15 And obviously, we landed native audio support in the mainline

6:19 Gemini model at I/O and in Veo as well.

6:22 And I was having a conversation with Tulsee

6:24 this morning about just how— are those similar breakthroughs?

6:28 Are they different?

6:29 It sounds like it's actually technologically

6:30 very different from a technology standpoint.

6:33 But it is cool that we have other

6:35 rails to do this innovation and, ideally at all,

6:39 upstreams back to Gemini in some way?

6:41 SERGEY BRIN: Yeah, honestly,

6:42 I think we've just taken a long time to ship the native audio in Gemini.

6:46 It's been in there for— LOGAN KILPATRICK: Back in the summer, right?

6:50 SERGEY BRIN: Over a year.

6:50 LOGAN KILPATRICK: Oh, really.

6:51 SERGEY BRIN: No, no,

6:52 the base model has had audio that it's trained on for at least a year.

6:57 And I don't know.

6:59 There's always— I think honestly it's just— there's just so much to do,

7:05 so much to ship that nobody has, for whatever reason, gotten it out there,

7:10 and native audio it in, native audio out.

7:14 I think native audio in has been in there even longer.

7:19 But to get through all the little hoops

7:22 to make it really work well and so forth,

7:24 I just think, for whatever reason, took a long time.

7:28 But yeah, that's finally out there.

7:30 I don't think that's done, as you say, the same way Veo does it— I believe does

7:34 the audio also through diffusion just like it does the video.

7:37 LOGAN KILPATRICK: Yeah.

7:38 SERGEY BRIN: In fact, if you watch during the training run,

7:42 you can actually see it generating videos that are— a couple percent into it,

7:48 it's these— the shapes aren't quite right,

7:50 and the words are kind of warbled and stuff like that.

7:53 But then it takes shape and it develops until at the end

7:57 of the run you have what you have, what you see today.

8:01 So yeah, I'm pretty sure that that's diffusion-based audio.

8:05 Yeah, diffusion is a really powerful technique.

8:08 As you know, we shipped text diffusion for small early test run.

8:16 I think that's one of the things that I'm grateful that we have the bench

8:23 of machine learning researchers in there that we

8:27 can pursue simultaneously different based techniques for across modalities.

8:33 LOGAN KILPATRICK: Yeah, the results of Gemini Diffusion look super,

8:36 super promising so far.

8:38 I'm hopeful the model progress is there and it all

8:41 fully works because the demo works— we were talking off camera.

8:43 The demo looks really good.

8:45 So hopefully, the capability translates well

8:48 and everything works from that perspective.

8:50 But you mentioned this before about watching the training run.

8:53 I actually haven't seen what this looks like.

8:56 So what does it actually mean to watch the training run?

8:58 SERGEY BRIN: Oh, OK.

8:59 Well, maybe you've seen for our text models,

9:02 but we're able to test out the intermediate checkpoints in 10% of training,

9:10 20% of training, and so forth.

9:12 And the model is weak at those points in time,

9:16 but you can get a sense for the trajectory.

9:19 LOGAN KILPATRICK: Yeah.

9:20 SERGEY BRIN: And so usually, especially if you have a big training run,

9:28 that you have a lot— you're using a lot of compute,

9:31 and you have high hopes for it.

9:33 You're going to test it out in various

9:37 ways through many times throughout the run.

9:40 So you're going to have a pretty good sense of what it's on trend for.

9:43 LOGAN KILPATRICK: Yeah.

9:44 SERGEY BRIN: So that's true for the text models.

9:46 That's true for the Diffusion video model for Veo.

9:52 Yeah, all of these models have these intermediate

9:56 results that you can take a peek at.

9:58 And if you're really deep in there,

10:01 you're for sure checking them because you're nervous

10:03 and excited about what exactly it will produce.

10:06 LOGAN KILPATRICK: Yeah,

10:07 I need to come and hang out in the MK more and watch— SERGEY BRIN: 100%.

10:10 Come over later.

10:11 LOGAN KILPATRICK: And watch some of them, yeah.

10:13 One of the— I was listening to Sundar's conversation with Dave Friedberg.

10:18 And Sundar made the comment that even 15 years

10:23 ago you and Larry and him were having conversations

10:27 about— and the team at Google were having conversations

10:30 about what this future-facing AI moment would look like.

10:33 And that it's eerily close to what you

10:36 all were talking about 10 or 15 years ago.

10:40 I'm curious, what are the things that have

10:44 been most surprising to you about this moment?

10:48 And we can ground it in products if you

10:49 want to look at search or just the technology.

10:51 Or what's been surprising,

10:53 and what's been almost what you would have expected to have happened?

10:57 SERGEY BRIN: Yeah, I think from an intellectual standpoint,

11:02 you can reason your way through the singularity.

11:09 And famously, Ray Kurzweil did this, I don't know, however many decades ago.

11:13 LOGAN KILPATRICK: Yeah.

11:14 SERGEY BRIN: It kind of— I don't remember what date he said.

11:18 It was 2037— or I can't remember.

11:20 He put some date on it based on his extrapolation.

11:26 Today looks like maybe that was conservative.

11:29 I don't know.

11:30 But you can intellectually reason through it.

11:36 I think, to see it happening is altogether different.

11:41 LOGAN KILPATRICK: Yeah.

11:43 SERGEY BRIN: And I think when you're talking, whatever,

11:47 15 years ago, I wouldn't say you're joking around.

11:53 You're truly talking about it, but you're imagining the science fiction future.

11:59 But it's almost like a game.

12:04 You're just chatting with the other folks who are interested in it.

12:11 I think it's fun.

12:16 But yeah, as I said, seeing it actually start to happen feels very different.

12:22 LOGAN KILPATRICK: Yeah.

12:23 SERGEY BRIN: And of course, the way in which it happens is pretty surprising.

12:30 And I can give you an example.

12:32 The fact that language models seem to be,

12:36 right now, the way that AI is developing,

12:39 I don't think you necessarily would have known that 15 years ago.

12:43 In fact, DeepMind— in the past, and even now, to a certain extent,

12:49 has bet a lot on this physical grounding,

12:51 that it's important to have a physical world to ground on.

12:56 And we're obviously doing experiments in that vein, like Genie and whatnot.

13:02 But the fact that these language models have

13:05 come as far as they have wasn't obvious.

13:09 And an interesting side effect of that, especially with the thinking models,

13:14 is that they are also surprisingly interpretable.

13:20 You can look through the thoughts of one of these thinking

13:24 models and how it's coming to the conclusions that it's coming to.

13:29 You can't necessarily, without a huge number of tools,

13:33 inspect the weights of the model and try to infer stuff from that.

13:38 But you can understand a lot of its reasoning in very understandable terms.

13:46 So I think that's kind of— you

13:49 wouldn't necessarily have guessed that 15 years ago.

13:51 LOGAN KILPATRICK: Yeah.

13:52 SERGEY BRIN: So that's been an interesting surprise,

13:55 which I think gives a lot of comfort, not infinite comfort.

14:01 I'm not saying we should ignore it.

14:03 But from a safety standpoint, the fact that these things do,

14:08 to some extent, say what they're thinking, I think is a big plus.

14:14 And there papers out there about how they're lying and stuff,

14:17 but I think that's a relatively smaller effect.

14:21 LOGAN KILPATRICK: What's your sense just being close

14:23 to the model training process today as far as how different

14:27 or how similar it's going to look as the models

14:32 move from being text in, token in, token out,

14:35 or text out, to being actual systems?

14:38 And I think we're actually— we've already taken it— Gemini 2.0,

14:41 search is natively in there.

14:43 Code execution is natively in there.

14:45 The model is learning that in the process.

14:47 Do you think that training infrastructure or just the way we think about

14:51 models is going to be fundamentally

14:53 different as they— they're not models anymore?

14:57 They're really full systems that we're baking for people?

15:06 SERGEY BRIN: I think it's the confluence of a few things.

15:08 One thing that's remarkable how similar

15:11 architecturally all the different models are, including, for example,

15:15 Veo, which you would think video diffusion's

15:17 very different than some text language model.

15:20 But architecturally, there's a huge amount shared.

15:23 It's astonishing how much is shared.

15:26 And a lot of it, uses transformers at the core, which,

15:31 thanks to Noam's crew, we've had now for approaching a decade.

15:41 Now we are adding things like tool use and so forth.

15:44 Mostly, those things come about during what we call post training.

15:53 And post training is an increasing fraction of the overall training.

15:59 It used to be that everything was like 99% pretraining,

16:04 and it's shifting now to now maybe it's 90%, will be 80%, and so forth.

16:08 And this post training, which is some people call fine tuning,

16:13 but it includes, the RL kind of work that we do.

16:18 It used to be this just little bit d tiny bit of shaping you do in the end.

16:24 But now it's more and more material.

16:26 And the kind of things you're mentioning with tool use

16:29 and so forth come in during that now much larger phase.

16:33 And yeah, look, that makes the model vastly more powerful.

16:40 LOGAN KILPATRICK: Yeah,

16:41 I've got two more questions because I want to get you back

16:44 in the office working so that we can keep making model progress.

16:48 First one is just around the reasoning scaling.

16:52 I think we announced— we showed the results of Deep Think,

16:55 which is continuous scale up 2.5 Pro and let

16:58 it reason for longer and have parallel thought processes.

17:04 What's your general reaction to that?

17:06 It feels like we're so early in that scaling paradigm

17:09 that there's going to be a huge amount of additional unlock.

17:12 But obviously, you're in the weeds on this one.

17:15 So I'm curious what your thoughts are.

17:22 SERGEY BRIN: Yeah, interestingly,

17:24 we had about five different approaches to doing that kind of thing.

17:28 And they all converged on this Deep thing.

17:31 So it was great to see all those people and those teams come together,

17:35 and that sometimes we fragment, and it takes a long time.

17:39 But in this case, we took the best of ideas of all of them,

17:43 combined it in one go.

17:45 And yeah, it's definitely yielding stronger results, obviously.

17:52 I think that the more that continues to happen that is like a superpower.

18:03 If you can have these models— and I know

18:06 lots of the top AI labs talk about this.

18:08 But if you can have these models, instead of just thinking for a minute,

18:12 spitting out an answer, if you can leave them to go for an hour for a day,

18:16 or for a month maybe, and they actually get you a significantly

18:20 better answer to a really important question, that can be incredibly valuable.

18:26 LOGAN KILPATRICK: Yeah.

18:27 SERGEY BRIN: And that's kind of new.

18:29 And it's not nontrivial.

18:32 It's a little bit like we cracked long context for input.

18:38 And we did that a while ago.

18:40 We've had million plus contexts for a year

18:44 and a half or something like that now.

18:46 LOGAN KILPATRICK: We need infinite context,

18:47 so you got to keep pushing this— SERGEY BRIN:

18:49 Yeah, I'm not saying a million is enough,

18:51 but that generalization is nontrivial for a model to— it's as if

19:00 you were to go through "Groundhog Day." You just get over and over,

19:05 you get to sample one day as a person.

19:07 You try this, you try that, try that.

19:08 And now all of a sudden you're meant to live your life where

19:11 stuff happens from day to day and week to week and month to month.

19:14 It's kind of a nontrivial generalization, but we figured out how to do that.

19:21 On the output side, it's also,

19:23 it's just nontrivial if all you do is, say, short little math problems.

19:31 Going from that— well, we interview people.

19:34 We ask them 10 interview questions or whatever.

19:37 And then we expect them to build these big systems over months.

19:41 And it's not clear if that's actually the right way to test a person.

19:45 But on the AI models, we do that a million times over.

19:49 We're only training them to do short,

19:53 little, clever math questions, coding things, whatever.

19:57 LOGAN KILPATRICK: Yeah.

19:57 SERGEY BRIN: And then the expectation from there that, hey,

20:00 they're actually can spend a long time to develop

20:05 something new that requires thinking over days, let's say.

20:10 That's very nontrivial, but that is a gap that we're starting to overcome.

20:16 And that's a huge leap forward.

20:19 LOGAN KILPATRICK: Yeah, this example that you gave around just how we test

20:23 and eval models is always my consistent reminder that so

20:26 much of life is a— this AI moment has taught

20:28 me that so much of life is actually an eval problem.

20:31 And the challenge of even interviewing people and trying to build a great

20:35 team and all that stuff is an eval problem at its heart.

20:37 And humans haven't solved that.

20:39 It's not surprising to me that we haven't solved the AI eval problem as well.

20:44 It is nontrivial to do that.

20:47 My final question for you, and this is just, again,

20:50 in reaction to everything that we're seeing with I/O and the pace of innovation.

20:55 Sundar had a slide on the screen, or actually, Demis did,

20:57 where it was like everything we shipped in 2024

21:00 and then everything we shipped in 2025 so far.

21:02 And it was like— I'm pretty sure

21:04 the 2025 section was bigger than the 2024 section.

21:06 So clear acceleration happening, and it feels like,

21:11 at least on a personal level for me joining Google.

21:14 I've been here more— almost a year, a little more than a year now.

21:19 Google truly feels like a startup experience to me.

21:22 And I'm curious to get your reaction

21:24 to that but also just your reaction in general,

21:28 having seen Google grow and expand and everything

21:31 that's happened in the last 20 years.

21:33 How do you think about that?

21:36 And the— SERGEY BRIN: Great question.

21:38 First of all, I think companies need to periodically reinvent themselves.

21:46 And there are different important technology shifts, I guess.

21:54 We started as a web company.

21:56 We had to make mobile work.

21:59 We're never particularly good at social, to be honest, as an example.

22:03 But now we're in the AI domain.

22:07 And I think from there,

22:09 it's exciting because in some ways Google has always been an AI company.

22:16 We've always been about large-scale data and analysis.

22:22 We are also the place that gave birth to a lot of the modern,

22:28 large-scale machine learning from Google Brain to the transformer, and so forth.

22:34 It is in the company's DNA.

22:36 So this is a shift we should be really well equipped to make.

22:43 Any shift is probably hard on any company.

22:47 But I'm feeling really good about it.

22:50 And I think going from '24 to where we were honestly catching up on many levels,

22:58 to '25, and particularly with the launch of 2.5 Pro.

23:03 LOGAN KILPATRICK: Yeah.

23:05 SERGEY BRIN: That was just a clear leap forward.

23:07 And I know, whatever, on different benchmarks,

23:09 maybe we were for a bit number one before or not.

23:13 2.5 Pro was a big step forward across the board.

23:20 And even to date it remains on most leaderboards number one,

23:25 with style control, without style control, however you measure it.

23:30 So that's been just a really exciting leap forward.

23:34 And I think it's both cause and effect

23:43 of the science engine we have going behind it.

23:49 It's going to help propel us forward,

23:51 and it's thanks to all the science that we were doing over

23:55 the past year that we were able to finally produce that model.

23:59 And in quick succession after that, a lot of other things have followed.

24:03 We've already gone through a few different iterations of 2.5 Pro model.

24:08 I don't know if everybody noticed.

24:10 Yesterday, we launched the new 2.5 Flash.

24:13 LOGAN KILPATRICK: Yeah.

24:14 SERGEY BRIN: Do you notice that it's actually,

24:16 on many measurements, it's number two behind 2.5 Pro.

24:19 LOGAN KILPATRICK: Yeah.

24:20 SERGEY BRIN: So we're like one and two on many different leaderboards now,

24:24 with that flash model, which I think with all the other announcements,

24:28 I just think a lot of people might have glossed over.

24:30 LOGAN KILPATRICK: It got buried.

24:32 SERGEY BRIN: But it's a super fast model.

24:34 It's really powerful.

24:35 I think it's going to appeal to a lot of use cases.

24:40 Yeah, but really, with that cornerstone of 2.5 Pro this year,

24:45 I think for us to be able to build

24:47 off of that and continue the momentum is really exciting.

24:51 LOGAN KILPATRICK: It's going to be a great year.

24:53 Sergey, I appreciate you for taking the time.

24:55 I appreciate you for pushing everyone hard.

24:58 It's a ton of fun to see.

25:00 And we have a special gift for you, which I'd love to see you unbox.

25:07 And somebody will bring it over to us in one sec.

25:10 SERGEY BRIN: Well, thank you, Logan.

25:12 And while they bring it over, I'll just say thank you, Logan.

25:15 I see you working hard all the time and making all of our customers

25:20 and partners happy and tracking the millions of issues that might arise.

25:28 It's not that easy having these models that so many people And.

25:32 Businesses want and getting them deployed and not having the TPUs melt down

25:38 and every little nuance from function calling

25:44 to caching to all the million things.

25:47 And I see you being really good about putting the customers first,

25:53 communicating the needs back to the team, being just really on top of the ball.

25:57 LOGAN KILPATRICK: The team is crushing it.

25:58 No, thank you.

25:59 The team's crushing it.

26:02 Special gift for you.

26:03 SERGEY BRIN: All right, thank you.

26:04 Shall I unbox it right now?

26:05 LOGAN KILPATRICK: Yeah, yeah, you got to unbox it right now.

26:07 We gotta capture.

26:09 The thread of this was just one of the pieces other

26:12 than all the people inside of Google that makes all this possible.

26:15 SERGEY BRIN: Nice.

26:16 LOGAN KILPATRICK: It's our TPUs.

26:18 SERGEY BRIN: This is the TPU V4,

26:20 for which internally, by the way, we call Pufferfish.

26:22 Probably not— it's probably not too secret.

26:25 LOGAN KILPATRICK: I think Pufferfish is V4, right?

26:27 SERGEY BRIN: Yeah, yeah, I never know the external translation.

26:30 We just call these— these were the hottest thing to come by a year or two ago.

26:36 We've moved on to newer generations now,

26:38 but we still do a lot of our work on these.

26:41 So this is great.

26:42 LOGAN KILPATRICK: Yeah,

26:42 hopefully we'll get a bunch of these in the MK for the team.

26:44 SERGEY BRIN: This is really cool.

26:45 LOGAN KILPATRICK: It's a real one.

26:46 They had to take it out of a data center somewhere.

26:49 It was not being used.

26:50 We were not taking compute down— SERGEY BRIN: Really?

26:51 OK.

26:52 I'm going to be careful.

26:53 We definitely need these.

26:54 LOGAN KILPATRICK: We need the TPUs.

26:55 SERGEY BRIN: Sometimes some of the early samples are a little bit faulty,

26:57 and maybe it's one of those.

26:58 LOGAN KILPATRICK: Yeah.

26:58 SERGEY BRIN: But I appreciated it.

26:59 This is super nice.

27:00 LOGAN KILPATRICK: Of course.

27:01 SERGEY BRIN: I'll give you a little zoom in on there.

27:02 Thank you.

27:03 LOGAN KILPATRICK: Thank you.

27:03 Thank you.

27:04 Thanks for tuning in for folks who are listening.

27:07 This is "Release Notes." Thanks for watching.

27:09 [MUSIC PLAYING]