Sundar Pichai: Gemini 3, Vibe Coding and Google's Full Stack Strategy

0:00Now it feels like Gemini is this through line across literally every single one of our products. To see it all come together, it is really special. I was just reflecting in the last couple of weeks. I think we've pretty much been shipping something every day, and so it's a great feeling. The Nano Banana Pro moment. I'm sure you've spent a bunch of time playing around with the model. People are going crazy with it, which is awesome. To me, it showed how much latent creativity people have. People are going to express themselves, and we are giving them the tools to do it the way they're thinking it in their head.

0:31And it empowers me. I feel like I've actually become more creative in the way I think about the world because of these tools. I think in about five years, we'll be having breathless excitement about quantum, hopefully, like we're having with AI today.

0:50Hey, everyone. Welcome back to Release Notes. My name is Logan Kilpatrick. I'm on the Google DeepMind team. Today, we're joined by Sundar Pachai, who is the CEO of Google and Alphabet. and we're sitting here in Mountain View. Gemini 3 has rolled out. Nano Banana Pro has rolled out and the reception's been super positive. So I feel like, do you want to sort of frame the moment of this progress to get to today where we sort of have state-of-the-art models across, not actually just Gemini and Nano Banana Pro, but VO and some of the other music models and stuff like that across the board. It's like more and more soda, the longer the time goes on.

1:28So yeah, you want to frame that moment for us. First of all, great to be here. It's been a phenomenal week, I would say. You know, when you're working on stuff inside, you kind of visualize the moment when you can actually put it all out. And when you work on products, there's nothing more exciting than that. And that's what this week is. But I think it's based on a foundation over many, many years and of all the deep investments we've built. and it was always clear to me. You could see the pace at which we were making progress. But to see it all come together, it is really special. I was just reflecting in the last couple of weeks, I think we've pretty much been shipping something every day.

2:09And so it's a great feeling. It is. Yeah, I remember back to you and I were talking probably a year and a half ago and I was complaining about something. I'm sure I was complaining about something And you were like, something to the effect of like pushing me to sort of see this long-term perspective. And I'm curious, like, I think obviously the, to me, the story of like getting to this moment that we're in now with state-of-the-art models and all this infrastructure to scale up Gemini across Google, like, is this long-term perspective? And I think especially in this hyper-competitive moment, like, what are you, like, how are you keeping that long-term perspective?

2:47And, like, because it feels so much like a rat race right now in some sense to, like, keep hill climbing 1 % on all these leaderboards. But obviously that long-term perspective is super helpful. No, I think I've always, you know, I always force myself. You know, you're obviously in the moment. We work in an industry where you move fast. You want to trade fast. and I really enjoyed that. But being able to pull back and making these long-term bets and being very focused over that time period on a long-term bet, I think it's always super important to do, I think. You know, obviously in 2016, I wanted the whole company to be AI first.

3:28A lot of what prompted that moment was 2012 was Google Brain, the famous cat paper, which had breakthrough in image classification. 2014, we brought Google DeepMind in. January 2016 was the alpha goal moment. And now people, many people didn't notice, in May 2016, we announced our first TPU. So it was clear to me in 2016, seeing all that, we are about to go through another platform shift. and that was the bet on a full stack bet on setting up Google to be an AI-first company. You know, and since then, we have been making a lot of progress. There's so many breakthroughs coming from Google, including Transformer.

4:20We were putting it in our products in BERT and MUM and making search better, launch Google Photos and so on. But obviously with the generative AI moment, and I realized the window was even bigger. Like, you know, the people were ready to use the technology at scale. Consumers, developers, and so on. So how do you respond to a moment like that? And, you know, for us, it was about, we kicked off Gemini as a project and across Google Brain and Google DeepMind. And as part of that, then deciding to bring the teams together as Google DeepMind, We really ramped up our investments in our infrastructure, data centers, TPUs, GPUs, and so on.

5:05It is then getting the company to more at a faster cadence. You now have the technology. And once the GDM team started tripping Gemini, and you can talk about the seas of Gemini milestones we have worked through. And it's been great to have you for a lot of the journey driving it too. and now how do you make sure you manifest it in all our products? There are many, many products which touch billions of users, right? How do you get search to iterate with the power of what these models can do? And so that's been the journey. But, you know, you can step back and understand that framework. And it's so exciting because for the first time, when you have a full stack approach, each layer, when it innovates, it flows through all the way on the top.

5:55This is what I tell people about pre-training. I'm like, the fact that pre-training for DeepMind is working so well in the Gemini models, it's like post-training and RL is like this accelerant of the underlying capability. And I feel like our infrastructure is a similar story. Absolutely. You make your infrastructure better. You make the models better at pre-training, post-training, test time compute, where have you. Or how do you take those capabilities and then manifest it in products, right? How does NanoBanana show up in your products? generative UI in search with AI mode, right? So you're expressing it at all these layers and not to mention being able to take that and give it to developers who are then innovating on top, right?

6:38And that's what creates this multiplicative effect and it's always incredibly thrilling to watch. But, you know, I've always had this long-term vision of how we can do it. Some of this took time because we have a full stack approach. When we had to respond to this Gen AI moment, I remember we were short on capacity. So when we, you know, we had to invest to ramp up all these things to get it to the scale. So you had a fixed cost around it. So if you were on the outside, it looked like we were quiet or we were behind, but we were putting all the building blocks in place and then executing on top of it.

7:21We are on the other side now, right? And which is what you can see the pace at which teams are moving forward. Yeah, it's been incredible to see this. You mentioned this, like Gemini showing up in all of our products. And I think I was talking to Josh and Tulsi about this and Corey about this. I feel like actually the challenge for some of these launches now is sim shipping. And maybe not even from a product perspective, but capacity and the sort of how do you make sure that the models show up really well across all these different product experiences. I feel like it's introduced a new, like we've almost, Cori made the comment something to the effect of like, we figured out how to do the models.

7:58And obviously there's more that we need to do, but like deploying them across all of Google's product surfaces is extremely hard. And it gets me to this. And I sort of had this realization at IO this year of, and I want to gut check this with you because I don't, maybe, maybe you have a different perspective, but historically, like other than maybe your Gaia or your Google account, there wasn't this through line across this whole suite of different products that Google has, everything from cloud to Waymo to search to everything else to Gmail. And now it feels like Gemini is this through line across literally every single one of our products, which is just such a, it feels like there's something magic there that is happening.

8:40I don't know. I don't know what your reaction is to that. I think Gemini, and that's a good observation. I think Gemini to me is like a much more, it's a very clear manifestation of what is the AI first strategy. Yeah. Because now you have something tangible like Gemini, which people can understand. And Gemini, you're right, improves everything from search to YouTube, to cloud, to Waymo. What I loved about the Gemini 3 launch, you talked about SIM shipping. We shipped it across many of our products. But it was fascinating for me to see, you know, on X, it could be Copilot or Replit or Figma. You know, people all coming together.

9:23They're also SIM shipping. Yeah, yeah. Right. And to me, that's innovation at scale. Right. It's not just us. It's other companies in the world. It's been extraordinary to see. Yeah, it's awesome. One of the other threads, and obviously the Nano Banana Pro moment, I'm sure you've spent a bunch of time playing around with the model. People are going crazy with it, which is awesome. I have to keep asking, did we improve the productivity of the world? Or is it a net progress or not? The infographics seem amazing. They do. And I think as we move beyond the fun stage, I just saw on X, I think Ben Badgerin had tweeted this infographic of his core weave analysis.

10:04So it actually made me look through that thing to try and understand it. one of the things PowerPoint many years ago unleashed is people kept making more and more slides. Like, you know, I always used to collect so much information and it kept expanding. Like maybe with Nano Banana Pro, we are back to a phase by which we can kind of compress it and give it to the world in a more digestible way. Yeah. That was going to be exactly where I was going. Like there's something I've historically been personally skeptical about like how useful a lot of the gen media models would be for the world. And like, obviously, it's useful from an entertainment perspective but it felt like nano banana pro crossed the chasm specifically with infographics and actually grounding with google search in that um of like actually like i i believe i i can see very clearly how this is part of google's mission of like organizing the world's information information and actually making it universally accessible to people with those infographics like it's just it blew me away like it was so interesting and i feel like it's a great reminder of like we're gonna see those use cases and i think the nano banana team if I remember correctly, from one of the pieces of content we were doing, they were like, yeah, we weren't even like, it's not like they were trying to make infographics work really well.

11:14It just happened as the model got really good and text rendering capabilities improved so much, which is fascinating. Fascinating. The other thing that shows me is how much latent creativity is there in the world. Yeah. So one of the other beautiful things we're witnessing is I think people are going to express themselves and we are giving them the tools to do it the way they're thinking it in their head, right? And so I think that, I think otherwise we've been constrained by the tools in front of people. You may not have realized it, but we are creating more and more expressive tools and they're more and more accessible to more and more people.

11:52And so watching that, it's incredible to see as well. Yeah. I've got another question about that in a second, but one of, I'll give Tulsi credit for this. Tulsi had suggested when we were talking yesterday to ask you about this, because she was curious, what sort of like, as you see these launches happen in these like large tentpole moments for Google, what is your success barometer for these moments? Is it like the reception online? Is it like, you know, what does the adoption look like on day one? Or like, how are you measuring, like, is this thing actually moving the needle for Google? You know, look, I'm pretty active on launch day, trying to understand what's working.

12:31I'm looking for feedback. Both, you know, an example, I'm on X trying to understand, like C is people, average users, how they're experiencing the product. I'll probably ping people back saying, look, you know, this is a valid point. We should address it. So in some sense, I'm looking at that, trying to assess it. You know, it's clear to me, teams internally are using Gemini itself to kind of collate. We have great dashboards. And so I try to take in across a variety of sources. I'm one of those people, I need to feel it firsthand. So I get reports, but I'm out there trying to understand how people are using it, what they're posting, right?

13:14And I think that is important. But I walk over to some people who all have these big screens with multiple dashboards, looking at QPS and understanding how the usage is. I worry about capacity. But all that gives you a real sense of what people are doing, saying. But that's how I do it. It's a combination of monitoring stuff online, talking to people, walking around, sitting down with people. I want to get particularly the first day. It really helps me get a sense of what's working, what's not working well. Yeah. I feel like you can feel the excitement in the office right now too. I can't walk out anywhere without some version of some banana.

13:57I don't know who did that, but kudos to the events or facilities team for somehow bringing 100 ,000 bananas into this building and making it happen. The exciting part is this is just sort of the first chapter or the first page of the Gemini 3 chapter. We don't have Flash yet. We don't have any of our other sort of models in the 3.0 category. we shipped Gemini 2.5 Pro. And actually, as I was looking at a bunch of the benchmarks, like even 2.5 Pro is not sold on everything. Obviously, competitors have caught up, but even 2.5 Pro right now is still like best in class at a bunch of capabilities and sort of taking it a step farther with 3 Pro now.

14:362.5 Pro was Google I.O. Yeah. And, you know, and you could sense that it was a big step up. I think one way I feel great is, and, you know, them as CoreIon team, the GDM team is on a good cadence, right? So we are kind of every six months or so pushing that frontier. And it gets harder, right? Because you are, yeah, you're right, 2.5 Pro is a very good model. So to kind of clearly take a meaningful leap from that, I think it's hard, but that's what, you know, that's what makes it exciting progress. I know you're always excited about Flash, which is working on and is coming. You're excited about Flash because it allows you to serve more people.

15:23And in that parade of frontier, it really makes a difference. So I'm excited for 3.0 Flash. I think it's going to be a very, very good model. Might be our best one yet, right? And we'll see. What's great is teams are internally. You know, our pre-training teams are thinking about the next version. And so this culture of relentlessly innovating and shipping, I think, makes this moment special. Special and definitely feels like there's going to be a lot of exciting progress going back to the full stack at all layers of the stack as we go into 2026. Yeah. I have a goofy question because Josh sort of inspired me with his goofy answer about launch day rituals.

16:14With his dry Cheerios? With his dry Cheerios, which is weird. And I need to keep putting Josh on blast because I think it's funny and it's... I'm like, next time I see Josh, I'm going to be walking around with a carton of milk to... To help him so that he doesn't have dry Cheerios anymore. Do you have any weird, interesting launch day rituals or is it just trying to make it through the day? I, it's, you know, my, normally, I mean, the morning ritual I almost always have is I kind of pick up and catch up on what's happening in the world. Yeah. I consume. In fact, I don't even check, like, Google email because the way I think about it is if something interesting has happened about Google, it'll be there in the news anyway.

16:55So I kind of, like, try to step back and take in the news. That's what I do. So launch day ritual becomes our products when we are in the news. So trying to understand your earlier question about how it's worked. So that's my main routine. And I kind of try on launch days to have a bit of a less structured day so that I can go spend it time. I love walking to the teams which worked on the products, maybe seeing them, seeing how they feel about what they shipped. So that engagement to me matters a lot. Yeah. Yeah. I have another sort of question along this line, which is I have an interesting observation.

17:36I think Demis and others have talked about this, maybe internally, but there's this micro kitchen inside of the Grating Canopy office where a lot of the sort of action is happening from a deep mind perspective. And every time I'm there, it makes, obviously Google is big and it's huge and global and there's all this stuff happening. the the mk that blue mk makes google feel small uh to me and i'm curious if there's like something interesting there which is like how you it feels small and intimate um and i'm curious how you've oh it it so reminds me of early google um you know obviously you know i i i i go there quite often and you know it could be you know you have sergey there and you have people like i mean jeff and sanjay pair program still and they are making their uh espressos how you people nothing encapsulates our culture better than what than watching the precision with which people make their espressos in that micro kitchen right i would never dare make an espresso there right I know a lot about how to make a good espresso, but I feel a bit intimidated amongst that group.

18:47But, you know, Nome was there, Cori, Oriol, just last week alone walking around. You know, it's very dense in talent. People are exchanged. Constantly people are visiting. There's a very great exchange of ideas. And so I love that. You know, kind of reminds me of how the company used to be in our early days. some of our serving team, Emma, and people are there. And that's where when I mention, I'll probably go look at the QPS. I'm hovering over these people's screens trying to understand what's going on. So it's definitely a favorite part of, I think, how the company works. Yeah. My Google feature request is we need to somehow remake these MKs across all the PAs or something like that.

19:34I don't know how that happens. I mean, there are. Or like, you know, there are other teams which have versions of these. And I think, and it really helps pull people back into the office because you realize the value for when you're there, that exchange of ideas. You can still go back and have focused time wherever you're working. But, you know, that moment really helps a lot, I think. Yeah. A lot of this AI story that you've talked about so far has been like us making these really, really long term investments and sort of setting the company up for success like 10 years ago. And I'm curious how you think about like right now or obviously and the bets are right.

20:14Like cloud and cloud has worked out really well. Waymo has worked out really well. Quantum hopefully will work out really well. Just announce a bunch of other like the quantum stuff goes over my head, but I keep vibe coding experiences to try to understand what they're talking about. It was one of the ways in which I was like testing, asking Gemini 3 to help understand a layer deeper on these topics was just fascinating. Yeah, it brings anything to life, which I'm happy we went with that tagline. But how do you think about like what are, is it just like infrastructure that's like this next 10-year bet?

20:48Or like do we like, we know AI is the thing, so that's what all the eggs are in that basket now? or like I'm curious how you think about like what are the next, you know, future facing 10 years look like and where we're making the bets now to set ourselves up for that next layer of success? Oh, look, I think it's always important, right? You know, going 10 years back, it was the bet on AI, you know, in the deep, in a full stack way. It was a bet on saying we're going to build other big new businesses, diversifying the company, bet on YouTube, betting on cloud. Google was as cloud native as a company you can ever imagine, but we weren't fully providing it outside.

21:25So that was a deep, big scale bet on cloud. And Waymo, these things take time. Waymo is a long bet. And I think we are now seeing that inflection point around Waymo. There's always these future bets, right? Quantum computing is an amazing bet. I think in about five years, we'll be having breathless excitement about quantum, hopefully, like we're having with AI today. But I'm constantly thinking in that time frame. One example of this is Project Suncatcher two weeks ago where we announced we're going to build data centers in space. Obviously, it's a moonshot. Some of it looks crazy today. But when you truly step back and envision the amount of compute we are going to need, it starts making sense.

22:12And it's a matter of time. And so how do you make progress on that? You work back and have 27 milestones and get underway. So in 2027, hopefully we'll have some TPU somewhere in space. Maybe we'll meet a Tesla Roadster, which is going around there too. So it'll be fun. But that's an example of the kind of long-term projects you want to undertake and do. Isomorphic with AlphaFold and Wing with drone delivery. we are cooking up exciting work on robotics. So you take that long-term view and keep making progress. When I saw TPUs going to space, I pinged Demis, and I said we should fund a moon rover and put Gemini on device and then have it sort of explore the moon.

22:59It'd be a great marketing campaign, even if it's not scientifically super useful. Who knows? Maybe the product is underway somewhere. Yeah, I'm sure it is. Um, this, you mentioned this thread before, which is the capabilities continuing to go up sort of like, uh, raising the floor for everyone to go and be creative. And I actually personally feel this way. Like I don't feel super like artistically creative by default. And yet I can like tackle these tasks that historically I would have had to have been. Um, and it empowers me. Like, I feel like I've actually become more creative and like the way I think about the world because of these tools and like not being worried about not being able to do something.

23:35I think vibe coding is this like massive one. And there's this like moment where this force that is like one of the most economically powerful things in history, which is creating software and being able to code is now in reach of so many, so many more people. And I'm curious how you obviously you do vibe code sometimes. I'm curious how you've been thinking about that moment of like AI builders being able to more than just like traditional software engineers be able to build stuff. What an exciting moment, it's almost like how the internet, suddenly blogs appeared, many more people became writers, if you will.

24:14And what YouTube did, many more people became creators. And you can feel that happening on the coding side. Even within Google, there's a sharp increase in a set of people who have submitted their first CLs. And it's because of these tools that are making it more accessible. Maybe you're a product marketing person, you have an idea, in the past you would have described it. Now maybe you're kind of wipe coding it a little bit and showing it to people. So you can tangibly see that come to work. I was just speaking with someone on my team who doesn't code, but was trying to teach his son Spanish conjugation and just like one-shorted in Gemini 3 animated HTML page to describe this to his son.

25:13See, when you hear stories like that, and that person is in our comms team, right? So you can kind of see how people are beginning to do things. And so it's very, very promising. In the limited time I have and I played around with it, it's almost like not just wipe coding, but just these IDEs now, it's making coding so much more enjoyable. Of course, I'm not working on a large code basis where you have to really get it right. The security has to be there. So those people should weigh in. but I definitely think, I feel things are getting more approachable. It's getting exciting again. And the amazing thing is it's only going to get better.

25:57There's something I always used to tell about Waymo whenever people would talk to me about it. Remember, this is the worst Waymo will ever drive. Right? It'll only get better. Yeah, yeah. It's a version of all these tools which we are doing. You know, wipe coding with Gemini 3 in AI Studio. you know, it's both amazing to see and it's the worst it'll ever be. Right, both are simultaneously true. And so I mean it in a sense that you're going to see a lot of progress ahead. So I think it's definitely an exciting time and can't wait to see what people out there in the world come up with it. Yeah, it is awesome.

26:32I think my last question is like, what's next? What can we be excited for? Lots of cool things in the pipeline, but anything top of mind? I think some folks need some sleep. I probably, you know, hopefully the team, the teams, all of us, you know, get a bit of rest. But look, I think I'm excited for the Gemini roadmap. I'm excited at how it's coming through all our products. We're also shipping new things, right? I love Flow. I've been playing around with Flow. Notebook LM, you know, it has a passionate, growing community. and it's amazing seeing journalists work on it, people doing their PhDs, like really doing all the research in it.

27:19So there's a lot more to come. I'm excited. Thank you, Shindar, for sitting down. And thank you, everyone, for watching Release Notes. We'll see you in the next episode.