The Era of Agents is Here: Logan Kilpatrick on Why Everyone Is Now a Builder

— Yeah. — So you you're sort of one of the key people at AI Studio. I um we will I think we agreed that we'll say it's your baby, right? Um the uh that's been a really interesting ride like that you know it's been around you know I think originally was called Maker Suite and uh you've been adding a lot to

it. Do you want to tell us a little bit about sort of what some of the big things you've been you know working on recently there? — Yeah so AI studio has been you know I think we sort of I need to come up with a better way to describe this but we have sort of the different eras of the journey. I think it was like originally it was like prompt to prototype, get your API key, sort of get off to the races, test the models a little bit. I think we sort of crossed the chasm probably 18 months ago to like really help people go to production with what they're building. Um, and I think the takeaway from that was like people, we can help so many people do more than just get an API key and sort of kick around the models and then go off and build. Like why not actually help them build the thing that they want directly in AI Studio? And I think the whole vibe coding uh wave that's happened. So we added this build tab in AI studio actually last year at IO and we're coming up on a year now of uh sort of pushing in that direction to see what the interesting thing. — So why don't you describe that? I find I'm amazed that I still find people that they don't really understand what the build tab is and you've been adding amazing stuff to it later on. So how about you tell start off telling us a little bit about it and then we can talk about what's new and stuff. — For sure. So it's a full-on vibe coding experience. So like you can go literally go from like prompt to a working app to adding a database to actually deploying it using cloud run behind the scenes um in like minutes uh which is really cool and a lot of it is actually free as well which is really exciting. So, we have a huge like millions of people building apps in AI Studio using all these services. Um, and they're not like spending a ton of money in order to make it happen, which is something that we think a lot about is trying to make sure that the technology is accessible in the hands of this like next generation of builders, which is really exciting. Um, and to your point, the team is shipping

lots of stuff. So, we just landed a bunch of really cool things like design previews. Uh, so as you sort of ask for your app, and you can kick it off right now if folks go to a. studio/build. um ask for your app and you can sort of see a bunch of different iterations of the UI and click through those while it's building and then choose which direction you want to go in. Um another

one, this is like the a classic reimagination of the I'm feeling lucky button in Google search. So like we're trying to solve this inspiration problem which is like people want to build. They're like I'm so excited. The techn here the technology can do all these cool things. — I don't know where to get started. You can come to AI Studio. You can go to the build tab. you can click one button and we'll sort of come up with your first app idea for you, have it connected to the Google ecosystem. Um, and then you can go and build that thing. — What I love is that you can also sort of say, I'm l you know, I'm going for the I'm lucky, but I want it to have nano banana, you know, I want an image model, I want a database, I want some off and it will just take all of those things and and run with it.

— And we've gone one step further, too, which is like if you start typing out your idea, we just landed this um it's called this feature called tap tap. And as you start to type out your prompt, we'll use sort of like traditional autocomplete um except it's using flash behind the scenes to like come up with the extension of your idea. So you say like I want an app that uses AI to help me organize and then AI will sort of complete that thing. You can press tab and then it'll go the next step step. And that's where the tap tab comes in because you just keep the model will help create your prompt generatively which is really exciting because I think people have a hard time actually articulating the breadth of what they can actually do as you're building apps. And so I think this like helps meet people where they are as far as like trying to get their idea into a text box um and bring it to life. It — it's changing so quickly. One of the things I find is that stuff that just wouldn't work last year, perhaps with the previous generation models or something like that, people go and tried it last year and they're like, "Oh, well that doesn't work. I'm never going to use it. " It's kind of like, "Hey, did you use it in the last 2 weeks? If not, go and check it out. " That's what I find myself telling people. And also just, you know, what seemed to take like, you know, you needed a 20page PRD before, now it the model now is much

more better at intuiting what it is that you want. the fact that you can sort of add in that oh yeah I want fire store I want a database or I want some off and stuff like that is that what you guys are also seeing — for sure and yeah and so the model in AI studio like has all this context about how to use Gemini how to use Firebase how to use cloud run to deploy things and so as you ask for your idea it's sort of it has that context in the background and it knows all the things that it's capable of and can sort of intelligently put them into place so we've um I like to say you know we we've we've given a more opinionated take on what it looks like to build an app. And I think it's it's there's trade-offs, but like it's helpful because it gets people to something that actually works and uses the technology faster than they'd otherwise be able to if you had to like literally start from scratch, no context, etc. So, I think we continue to like AI Studio is a story of like we're taking an opinionated look at what it means to build a vibe coding platform. Um, and we we'll keep pushing in that direction. — That seems to be really smart though. Like I find you know one of the things that I guess when a few months ago when it didn't support next. js right so now just adding nextjs where you know opinionated I can say look I want you to build something but I want it in next. js JS I wanted to have a database I want some off right okay I kind of know now that I'm getting best practices on those things and I can sort of go wild with it or more importantly not so much me but like I find other people that kind of even code can do this right that this is like a whole thing that we can touch on as well is that um but I find like you know developers that have never done front end go in there yeah — boom they they're now amazed that they can build a pretty nice looking front-end app. I think you're also integrating, like you said, the new design stuff. What What's uh — lots of design stuff in the works. Um I think we're we have like a whole slew of different things. The first one is uh design preview. So, as you're building your app, you can actually iterate and it'll sort of show you a bunch of different options actually once you are and that's like while the app is doing the initial generation. We'll be rolling out soon the ability that like once you actually already have an app generated give you multiple different like options as far as like other themes of like what could this app look like if you change a bunch of stuff. Um we have a bunch of like targeted uh sort of like edits rolling out. We have this like edit mode where you can sort of draw on the preview of your app or you can like directly select elements and then like regenerate images or change different filters around things. Um, so that will actually hopefully roll out I think this week or early next week, something like that. — It seems like every week you've got something new. — I mean, it's honestly I have a hard time credit to our team for pushing so hard and making all this stuff happen. I genuinely have a hard time keeping up with all the things that are happening because we're just we're doing so much stuff. Uh, which is it's a good place to be in. — So that that's good to hear from us outside like if you had time to give up, that's good. Um

another thing actually that you didn't even mention was you've added voice. So the ability to actually sort of talk to it and sort of I guess prompt it and get stuff going just purely with voice now. — Yeah. So we have I mean the voice models which we should also talk about the text and speech models the live models are incredible and sort of I think Gemini's really pushed the state of the art as far as audio capabilities and we have a new sort of like LIA music playground. We have a new texttospech playground. We have all the live models sort of rolling out across the ecosystem. But we also in the vibe coding experience we have this feature which u I've come up with all these goofy names for things. It's called uh it's called yap to app y to app. So you can go in and you can sort of just like say whatever random you know garble of words and ideas that you have. And then the thing that we've done is we put Gemini. It's not just like a pure text to speech model. We put Gemini in there to actually like formulate the idea that you were just putting together in a coherent way. so the model can take action and actually bring that idea to life, fill in any blanks that you forgot to mention or sort of like shape it in a way that that's going to work. Um, and so we've seen a ton of people using this, which is really exciting. I think beyond the I'm feeling lucky button, this like yap to app experience is so popular. Uh, it's like the number two most popular thing, which is awesome. — So, so what about mobile? Like one of the things that a lot of people are asking about is, okay, this is great, but I don't really need to make a website. I want to make an app. if I want to put it up on the iOS app store, the Android Google Play Store. — iOS app store is sort of contentious. You know, there's lots of conversations going on around there. We're we definitely have a lot of exciting stuff coming in the mobile space. The team has been pushing super hard um to bring AI Studio to life on

mobile. So, we'll have more to share there soon. Um and it'll definitely be the first cut. Like I think we're the future for mobile for us is like we want to let people build anywhere wherever they are for whatever platform they want. Um and so mobile's obviously super important. Also as you sort of think about like reaching this next 100 million user sort of like nextgen developer audience like a lot of those folks are on mobile. They don't they're not on a web laptop desktop whatever it is. So we need to go there as far as reaching those users and um yeah lots of cool stuff on the Android side as well. collaborating with those teams. Lots of interesting things on potential uh ondevice models to enable which is also really exciting. The Gemma 4 models are incredible. I saw a demo from our team of AI Studio Mobile running uh local models which is really exciting. So I don't know if it'll make the cut for the initial release but lots of cool things that we're poking around with which is exciting. So I'm

curious to sort of chunk up a bit and ask you okay so we've got sort of vibe coding right and we know that sort of more traditional or older developers kind of were against that at the start like you know there was a lot and now we're hearing the term sort of agentic engineering which sounds like a nice rebranding of vibe coding um and I can get it why people are against it last year because it you know you would get to a certain point and then you would reach bugs and you couldn't fix it or whatever it does seem that's changed though, especially in the last few months, like since Gemini 3 came out. Uh what's your take on that? — Yeah, you're right. I think um folks for the right reasons and actually like we feel this inside Google as well sort of like there's a high bar for like landing code especially in like production systems and like even for AI studio specifically like what we've actually ended up doing is like there's a lot of folks me Amomar others on the team who are vibe coding things like in the actual AI studio produ like production agentic engineering not vibe coding um and we're making changes to the product and what we've had to do is like have a deeper partnership with our engineering team. So we actually have somebody now whose job is like you know our product our members of the technical staff team goes and makes changes to AI studio gets it so that all the CI is passing it's all green we run the tests things look good and then we actually hand off to the engineering team they take a bunch of these changes they get them over the line and they become the like owner and the steward of getting that code over the line and directly into the actual ASI codebase and I think this partnership ship model between folks who are like trying to vibe code and the actual like senior engineers who are want to make sure that this thing is reliable and scalable etc. I think works really well. Um and then the best part is like that person who's on the hook to like get a bunch of these aentically engineered changes into the codebase is also responsible for like how do we make sure that that cycle is better. So like what are the skills that we need to put in place? what is the infrastructure we need to build so that like we have better test coverage we have all those things so we I think you know we should probably do a u a blog post at some point because I feel like there's lots of conversation about this and I feel like our team's found a really nice sweet spot of like letting in new contributors to a production code base while also like it's Google we're holding the bar for quality really high um with millions of like paying customers using this platform so — it also sounds like the lessons that you learn from that are going to just improve for everyone, right? Not just for you guys doing it yourself. Like as you sort of learn that, oh, the models maybe guide not well here or that kind of thing. It allows you to course correct and make things even better for higher level quality code. — 100%. — That that's really fascinating to hear. One of the things so we were both at an event, private event yesterday. One of the things you said there I really struck a note with me. you talked about and I realized when you said this that I felt the same thing and I thought this is a really nice you talked about like having more ambition — like that I really like the way I

forgotten exactly how you languaged it maybe you can remember but you sort of were saying that like just trying these things out you're constantly having to push yourself to have more ambition to do stuff that you wouldn't have thought was possible and then time and time again at least for me I'm like blown away that wow it works right? You know it can do it like you know I kind of thought like okay cuz the last model or cuz at some point I tried something it didn't work and this whole idea of ambition do you want to talk to that? Yeah. No, it's um I think about this all the time which is like the models to your point from before like if you haven't tried the thing in the last 6 months like — even in the last two weeks — even the last two weeks. Yeah. It's like you historically had to be like very precise about what you wanted AI to do for you and I think the models have crossed the chasm where like instead of asking for one thing you can now ask for 30 things and the model can actually do that. like you don't need to like we're artificially hampering what the technology is capable of because we're like trying to not let it fumble over itself which it used to do in certain cases. Um and so I think about this all the time. I actually think the other thread of that story is like I now feel this weight on my shoulders because like it really it's not that the model like I can't be like oh well the model can't do it so like it's fine. I'm just not going to like now the sort of like weight is on me. The onus is on me to be like I really could build this. Um, and so now there's an interesting like I'm always um at least for like you know actually for like internal work stuff I'm like oh there's some bug in AI studio I should be the one to go fix it like I can't be like oh sorry I'll just you know bank on the engineering team going and solving this problem like it feels like the responsibility falls on me. The same thing is true for like external projects where I'm now like the bar is high in my mind for what I will actually go spend as a build as a side project and it just means like I almost need more time than before. It's kind of an interesting dichotomy which is like you'd expect because the models are so capable like you know now I'm just doing something in like an hour over the weekend. Instead I'm like my idea is 20 times as ambitious. I'm like okay I'm going to need to take a week off. Yeah I got 20 ideas. I'm going to take a week off of work in order to pull this thing off because like I'm able to be so ambitious and I actually know it's another thing that's changed. It's like I know in my you know heart of hearts that like I can actually pull it off. I think before there was like kind of I always had this question of like am I going to go down trying to build this thing and then realize 30% of the way in that like I don't have the technical capability to pull this off and like I don't feel that way anymore. And so it's like uh I'm kind of yeah it's an interesting experience to like go through this and like turn that chapter. And I think we'll have we'll all have more of these types of experiences over the next you know 12 to 18 months as the models become more capable. — I definitely find that myself too that just things like you know um uh sort of multiplayer games or you know multiplayer apps and stuff like that. You think like okay that's a pretty hard thing to conone. I'm going to have to read up stuff. — That's one prompt in AI Studio. One prompt literally. Um the one of the other things I wanted to ask and this is going back I guess to something you had talked about the whole sort of next billion users right — so I have some people on my team who I'm not coders right and I but I make sure that they you know build is awesome for that right I basically say here open this build something right try something out I find that they're making software that I would have never thought to make right this is one of the things that's fundamental and to at least to me, right? And then I find myself sort of constantly going, "How did you make that? " Right? And then I find that, okay, actually wasn't that complicated. They had a conversation. They described what they want, you know, and but it just Are you seeing things like that? 100%. I think it's uh there's better quotes than I'll be able to remember, but there's so many things about like you know intelligence is like so distributed across the globe and uh like great ideas are so distributed across the thing that hasn't been distributed is opportunity and I think what is exciting to me is like I fundamentally believe the reason that it's important to build these types of products is because like we're putting the means of opportunity in the hands of people who wouldn't otherwise have been able to build this thing. Totally agree. It's going to be like a fun like the creation of software at least today is like fundamentally the most economically empowering thing that you could possibly do and that's why people learn how to write code and historically all that stuff. And now putting that in the hands of this cohort that couldn't before is going to be um we're already seeing the impact of there's like millions of these people using AI Studio to do this and other products. Um and I think we're like — chapter one of that story which is really exciting. like we're like an AI studio I think has this responsibility to now as the ability to create software actually gets solved like I'm fully convinced you know the next 12 months you'll be able to build whatever software you want what are the next set of challenges to like actually help you build that thing that you want to build like it's actually it won't be software it'll be 15 other things and so we're already starting to think about like what are the next 15 other things that we need to make sure that you can do and do them in AI Studio use the Google ecosystem and make that all super seamless. — How do you design like um you know for me AI studio is AI. dev. So if people don't know the shortcut that's what I always — AI. studio as well, — right? AI. studio too. Yes. But the funny thing is you know I'm reminded by the AI. dev is that what is a dev nowadays? — Yeah. Right. Like I — that's the tension and actually I think we feel this tension a lot in AI studio sort of as a product because AI studio is the front door to the APIs as well. Um so if you want to build the Gemini and you know you don't actually don't have to vibe code in AI studio you can just take the API go build off platform if you want use what coding models use whatever tech stack you want um and we have this sort of like dual identity now as like a builder product and as a develop like a deeply developer product like there's truly like millions

of actual businesses built on top of the APIs that we have and so we're always walking this fine line and I think it's like a it's definitely a transition Um, and so we're trying to like walk the line as much as possible during this transition, but like it's tough because like we want to build the powerful tool to your point before like developers who don't know how to do front-end stuff or don't want to do it

like you can come do that in the studio and we want to meet them where they are. We also want to enable this generation of people who like don't have any code never don't even ever want to look at a single line of code. It's really tough. It's like I it's hard work to try to find the balance. See that's definitely a challenge. Okay. So let's talk about some models right I you know I my background is in deep learning big believer in LLMs. I — you it's now about 4 months since Gemini 3 came out for — December 2025 — I um we've seen 3. 1 come along. I definitely the powerhouse of those seems to be getting better and better. The other thing that's fascinating too though is that and I think this goes back even to the original Gemini, people didn't really understand what multimodal meant, right? And they didn't understand the consequences of that. I feel now with like the new live voice mode, you know, all these other things. You want to talk about a little bit about perhaps live voice, what what's the state of that now?

now? — Yeah. So we rolled out I think it was last year actually at IO our first live model and it gives the ability to like stream audio and video and text directly to the model and get responses back in real time. Um I think this enables a bunch of these like real time omniresent use cases. It's re it's actually really cool. I think it's like ai. studioive and you can sort of try out that experience. Um, but I think the use case that I think we were seeing a ton of traction with initially, which is interesting to think about, is like literally screen sharing with an agent. It sees everything that you see and then you can sort of just ask the questions. And we've seen customers like, you know, I have a really complicated product experience and like I'm a non-technical user trying to use, you know, an e-commerce platform. How do I set up a custom domain name? And like I don't know how to navigate through the product and find it. And literally the live model can like walk you through and be like click and you can do cool things like add overlays and stuff like that to be like click here and then go to this section and do that. And you can imagine I think that like omniresent you know tutor or helper is really exciting. I actually think we're going to do some interesting stuff around vibe coding too to like help you know your one-click thought partner to like see everything that you see and help you debug problems is something that we're thinking about. But the model is so capable. It — it definitely seems amazing just the personalization of where it can meet you where you're at. So it's not like I'm watching a YouTube video where I have to get through the first 30 minutes to go, "Yeah, I know all that stuff. " And now, oh, okay, this is the 3 minutes like you said, you can just power it up, say, "Hey, I don't know which button does this. " And it will tell you. I had a really interesting experience with that last year just after IO when it first came out that uh where I was staying in San Francisco, the refrigerator broke down, you know, because of a filter. — I'm like, I have no idea about this. So, I opened this thing up — and it sure showed looked at the it knew exactly what refrigerator it was. It knew uh what the warning button was. It told me I had to change the filter. tell me where the filter was and and sure enough within sort of 10 minutes I ordered one in on from Amazon. The next day I had it change and it was done — and that would have you know that was just amazing for me right at

the time. I'll share my experience actually JD who's walking past right now um who's on our team at Google we were doing a bunch of um Astra demos which was the sort of precursor to Gemini Live and this was I think at IO 2024 maybe when we first started showcasing Astra in sort of early forms and I don't know if folks know those like very fancy coffee machines where there's like handles and you like do all the things separately. I'm like I don't I'm a very simple coffee drinker. Um, and every time I see those machines, I'm like, "Okay, I'll go find a bottled coffee or something because I don't know how this works. " I opened up uh Astro, the precursor to this, and to Gemini Live, and was asking like, "What what do I actually do? " And I'm like showing the machine. It's like, "Okay, put your hand on this handle, twist it, pull it down, go here, and like literally like zero shot just worked. " And I was like, "Oh, okay. I actually know how to do this thing now. " — It really is insane. — It's very cool. — I find so many people don't know it, too. when you pull it out and just suddenly show them, they're like, "Whoa, they're like magic. "

— It's interesting you say that. I think we should talk about our text to speech model as well, which I think is more widely adopted because I think developers really understand that use case. This like generate audio synthetically I think makes sense. There's a ton of use cases. I think the challenge for live has been it's like truly. And so there isn't like an existing there wasn't something else you were using before. what you were using before was like a combination of like 50 different things. And so I think from like a product market fit perspective, it's been interesting to see like who is adopting this thing. We've had to do a lot of um a lot of education to sort of help people understand those use cases. Um but I think we're like finally it's starting to make its way into like everyday products and I think people will understand. — I find that people uh developers especially yeah that don't have no clue that it can do function calling that it can use you know search and all those things that are suddenly even though it's voice, it's still able to then do fun, you know, the model's voice is still able to do function calling, return things back. — Yeah, that's the magic is cuz like you can really use it as like a home assistant product. I think that's how a lot of folks do it. My favorite thing is like an omniresent like help button. Like you're somebody who's just like you're stuck somewhere on the internet. You click that button, all of a sudden like there's a person over your shoulder with your consent that's like it can see what you see. you can talk to it um and sort of help guide you through whatever you're trying to do. I think that that's a magical experience that I don't think is like widely diffused into the ecosystem and there's lots of alpha in building this

— for sure. Um another magic model banana huge moment. Uh it's only gotten better since it first sort of — indeed you know come along. uh where do you see models like that going and what — yeah I think the gen media portfolio — and we can talk about stuff as well everything else has just been incredible I think it's been like a shining light of the Gemini story it's actually back to what you mentioned before I think it's tied to the multimodal story because when you have really great multimodal understanding it lets you actually build a really great multimodal generation model um I think the direction of travel is like consolidation in the offerings as far as gen media go. So, I think like we've done a ton of these bespoke gen media models. I think hopefully we'll see sort of a lot of that come into the mainline version of Gemini in the future, which is really exciting. Um, and I think you Yeah, we've just we've got a lot of different like we have the live model, we have the TTS model, we have Nano Banana, we have Vio, we have Lia. So, there's a lot of different models. Um, and I think it introduces some complexity and — even with even though there's a lot of different models, from what I understand, and correct me if I'm wrong, it's still fundamentally the Gemini sort of, you know, key models and its knowledge is what makes things like being able to reason over images and nano bananas. So cool. — Exactly. Yeah, it is. I think there's like a slightly different training mixture and like a slight slightly different architecture, but like all the base, which is actually great. It's what makes it possible for us to take n number of research bets like this is like the conceptually what we're doing for these models is actually really similar. There are changes and then the tension becomes like what happens when you fuse these things together into a single model. There's some gains in certain places. There's definitely some losses in others and so there's been a huge amount of research as far as like how do you actually bring it all together which is exciting. — That's super interesting. any other new models that you want to announce for us on this live streaming just casually

drop that — Yeah. No, there's lots of great things. I think um people are I see lots of I see all the tweets, lots of people looking for new models which is exciting. Um I think we're pushing the rock up the hill on coding. I think there's lots of lot huge amount of investment happening from us to make a great coding model. I think we've got tons of products that will benefit from us. Our customers want this. So lots of stuff coming. Hopefully we'll have more to share. uh on coding stuff soon. — Coming soon. — Yeah, hope hopefully. Fingers crossed.

— So, uh speaking of coding, anti-gravity — also came out with 3. 0 um has really had a good, you know, a lot of good feedback around it and they're just getting I'm amazed that team is just sort of getting started right when I've seen some of the things that they're doing. Um do you want to talk a little bit about that and how that's influenced maybe AI Studio and vice versa? you're in this interesting sort of uh situation where you've got these amazing teams working on very sort of unique things, but the lessons that they learn go across multiple areas. And I think this is actually to your point, this is the beauty for Google is like I think the anti-gravity team is doing an incredible job um not only in the product side but actually like really close to research and helping push the rock up the hill from a coding model perspective from an infrastructure perspective and the benefits and the things that they're learning are diffusing across many Google products. So AI studio is a great example of this. The coding agent harness is actually powered by the same anti-gravity harness. So behind the scenes we're using like literally the same binary as what's being used. So um I think you'll see more of this sort of like cross Google diffusion of the technology and Google and AI studio and deep mind are making a huge bet on the anti-gravity team and the work that they're doing in order to like bring this to more developers which is super exciting. So I think to your point it's like it feels like chapter one of that story still and it's going to be fun to see all of the all the new stuff that they end up landing — dling out. Um just speaking quickly and one of the things that that the community is you know frustrated I think with anti anti-gravity is just like the fact of the quotota the lack of I'm sure you do. So we saw new TPUs

announced last night right and this morning um pretty amazing — architecture you know about 3x more for the inference and stuff like that. uh is that going to help to be able to serve more sort of Gemini models for anti-gravity and stuff? — Yeah, there's definitely I mean there's first of all like I acknowledge the sort of like tension that folks have and a lot of — this is the problem with being so popular, right? — It is. I mean it actually is like it's a death by success story which I think there's you know — we have way more demand across actually not even just anti-gravity but across like every Google product surface that's landing AI stuff like there's just way more demand than there is supply in order to do this. So try to be really intentional, try to be transparent with people, try to make sure that like we get the models in the hands of our paying customers. There's a bunch of trade-offs. Um, and so it's not going to be perfect in every instance, but we really are trying to make sure that it lands like the reason we do all of this is to make sure that the models end up in the hands of people and that it benefits them. And so — I get frustrated reading some of those tweets where I think you don't think if they could that they would actually work. — We're trying. It's tough. And I think there's a lot of like prioritization trade-off questions. And I think that's where a lot of the tension is. And I think there's actually like a macro like stepping back like separate from our products, but just like as an ecosystem, I think over the next two to three years we're going to go through this challenge which is like there's going to be so much demand for this technology. Um and there already is. There's so much pent up demand. How is how are how is everyone going to go through this trade-off exercise of like you know you probably will end up with a fixed amount of tokens and like where do you deploy the tokens in your own life in your business all of those like that's very much the reality of what's going to happen um and so finding those highv value use cases finding those examples of like where you can actually move the needle and get like a huge amount of leverage I think is going to be the next era of this instead of just like throwing AI at everything I think you're going to have to be more intentional about doing it in the highest value cases because there's so much demand um for the models. — Interesting. So — even with Google spending tons of money to build TPUs like we — just announced and something all the money is being spent on TPUs and it's still uh there's still a huge amount of demand. — Wow. — You mentioned like coding does seem to be where it's at the moment, right? like and I'm curious to know okay what's the second thing going to be right so like a coding I get because all the labs benefit from being you know getting better coding models means everyone can improve the models can improve what

they're building that kind of thing it does seem really interesting to see where what other areas of society are going to have these big impacts and you know I think people with sort of proposed ideas what do you see what do you think about it — yeah I think robot Otics is at this like at this crossroads that I think a lot of the coding models were probably at, you know, — 18 months ago. So I think we're probably like 18 months from maybe 12 months from like a bunch of these significant breakthroughs. And I think it's just because like the intelligence that we're packing in some of these new systems is just like so high that it just covers a lot of the edge cases that existed before. Um, so I'm excited for that. I think there's like as a customer who's pre-ordered a bunch of these robotics products, I'm very excited. I hope those teams all pull it off. Um, and there's a ton of stuff actually happening in the Gemini side from a robotics. We have a huge amount of partnerships with Boston Dynamics and others to sort of power a bunch of these nextG experiences. So, I think that field um, you know, I could be off by another six months or something like that, but it feels like it's ripe um to actually happen. I think the other one is longunning agents and I sat down with Jeff Dean earlier today and It's sort of if you look at the frontier of how long you can let an agent run without a human in the loop today, it's like depending on the use case, it looks different, but it's on the order of hours. I think as we look at the next 12 months, it's it is truly going to be on the order of like days or weeks. And there's a huge amount of work that needs to be done in order to enable that. But I think that is the direction that we're going. The models will go out and do the things they need to do a lot longer without human intervention, which is exciting.

exciting. — What are the big things? I'm guessing research is one of the sort of big ones and that's the first agent right on Gemini actually deep research. You just updated this week. — I still haven't had a chance to read the blog post. — It's good. So we uh back in December we released the interactions API and as part of the interactions API we sort of brought this concept of like models and agents both being first class citizens of the API experience. And so we launched with deep research initially. Um, last week we just rolled out with uh with two new versions of Deep Research, Deep Research and Deep Research Max, which sort of pushes even farther and goes deeper and is sort of more rigorous. Um, and I think all of this is laying the groundwork for sort of letting developers create their own agents in the Gemini API in addition to a bunch of like new uh hosted agents from Google which we'll bring to the world which is really exciting. So I think we're seeing that we've been laying the groundwork to make this happen. And the thing that I'm I really am excited about is like using the same underlying API for models and agents. I think as developers sort of make this transition from interacting with raw models to actually working with agents, we're trying to make that seamless. And I think by having it like sort of side by side in the API and literally UI now you can see models and agents um trying to sort of meet folks where they are. — Do you think that that's this generation sort of Gmail? Like Google became really famous because of Gmail. I'm not talking about just the communication aspect, but I mean as a product that was sort of like a boom, everyone had Gmail, right, with within a year or so. Yeah. And it does seem to me that like as some of these agent things start to get roll, you know, roll out, people are going to be talking about, oh, my Google agent or my Gemini agent, you know, that do this, and this, and whatever, or maybe they could give them names, whatever. It does seem like there's something a change there, right? — It indeed. Yeah, I agree with you. I think that the the correlary is maybe like every product is going to become agentic and so I think that will be like maybe the only reason what you're saying is not true is like — just Gmail becomes an agent — just becomes an agent and like just like all the other stuff becomes like Google search or like Google is an agent and you or many agents obviously um so it'll be interesting to see like the foundational pieces all become agentic which I think is the direction that we're going — all right that that's fascinating so okay let's start to wrap it up. What I wanted to ask you also was like, okay, so you've now experienced working at a couple of different companies in this field. You've been at the forefront though probably for at least the last six years, seven years. — Yeah, — maybe longer. A little bit longer. — Um, what do what are you most excited about? — Right. You mentioned the the robotic stuff and I the point I wanted to make about that is in many ways that's just another modality, right? This is where you know Gemini is being used and there are some really good fine tunes of Gemini for robots and stuff. What are the things that what do you and really you three years ago you thought we'll never do that right in my lifetime like I find myself constantly looking at things that were research problems you know six seven years ago and they're just done now right and they're just solved to the point where you know there's really not that much point of even doing sort of research in that particular you know like a specific kind of NLP or something like that kind of thing. Oh, I think I have a hypothesis and I think it's what our team is spending a bunch of time thinking about which is like similar to what happened with um YouTube in the internet era, the sort of like — early internet area where everyone was able to become a creator. Um I think that's happening with software and I think so we're thinking about this similar to like how YouTube actually thought about building a platform and an ecosystem for creators. Now everyone is a sort of now everyone's a builder. everyone can build. Um, and the technology is enabling that to happen. And there's a huge there's so much breadth and such a long tale of like actually pulling that story off. But it's what we're spending a lot of time thinking about. And it's also like the emotionally the thing that gets me most excited because I think you look at this group of people who like have not had the means to contribute in this software economy and bringing those there's like so many I actually work with lots of these people so many smart people at Google even as an example who like are right next to the software um who haven't historically created software and we have someone on my team Harrison um who is incredible and like does much more growth stuff like he's now like the number to token consumer and is like building all of these like really cool internal had never written a single line of code in his life 6 months ago and is now building all of these really cool internal tools like shipping landing pages making product experiences in manage studio and I think we're just scratching the surface of like that becoming widely distributed in the economy and I think when that happens like we actually will see this like phase change transition of like what the world looks like as more people build software and I think it's going to be for the better which is really exciting. Um and we'll also I think I'll make my last comment which is I think it will also increase the demand for traditional developers. I think is the other thing I think a lot of people look at these things as mutually exclusive. I think it's very positive sum like as this total addressable market of the number of people making software increases. It increases the demand for developers because there will be a stopping point of like how far you can go if you don't understand all the detail of how this technology works. And so there'll be a lot more cases where people want to go farther and they need somebody they need a developer to partner with and they need that technical co-founder whatever it is. So I think that direction is really positive and exciting. — How do we deal with like the doomers or like from what you're describing society is going to change right? I think that's I think it's changed pretty much all the time for the last that you know at least hundred years or so. I it's clearly going to change things like education are going to have to change the whole sort of systems around th those things you know how do we sort of as sort of professionals in this AI industry or something uh sort of convince people that it's not like everyone's going to die. It's not like you know everyone's sort of so you get these really hardcore doomers which seem to get a lot of uh you know a lot of publicity and stuff like that. uh but they're not seeing all these opportunities. things going on. — Yeah. I mean I I feel this uh this sense of responsibility that our team has like the DMI mission is build AI and make sure it benefits humanity and I think the way that we do that is like deploying the technology and I think the way that like anybody like people should rightfully be skeptical of many things that are happening in the world. The way that you sort of be people where they are if they're skeptical is like build technology, deploy it, show it and let them use it. like there's no better you know way to have the conversation than being able to put your hands on the technology and I think we have that deployment first sort of mindset as far as what our team is doing in AI studio and doing it with urgency because I think there is like if you wait too long um you know society goes through that transition and then people sort of didn't have the chance to sort of have the conversation and understand so I think this it's super important it's why we have an incredible developer advocacy relations team who's like trying and now like builder relations trying to like bring the technology to the world so that folks really understand what's happening. Um it's super important. — On that note, I think we finish up.

That's awesome. Thank you. Thank you both for the conversation. Uh it's fascinating to see where things are going. Uh fascinating to hear what you're working on. Sounds like you got a lot of cool stuff coming. Anything else you want to leave with people with? — Um I think the only last thing I'll say is like if you have feedback, please send it to us. Like we're trying to push the rock up the hill. If you need you need — paper cuts, any anything — anything paper cuts, features, whatever it is, ping us, email, Twitter, LinkedIn, whatever it is, however you best mail, uh send us a fax. Uh and we'll make A Studio better for your use case. So, please keep the feedback coming. Thank you for doing this, Sam. Thanks for coming in next and flying all the way there. — Thank you for joining us today. It's been fascinating. And thank you to everyone who's been listening. I I'm sure you'll see Logan and future things going forward. And there's a lot more content coming in next as well. — I love it. Thank you all. — Bye.