Building in the Gemini Era – Kat Kampf & Ammaar Reshi, Google DeepMind

We are super excited to be here. It's been obviously a very exciting week in AI, it's been a very exciting and busy week over here at DeepMind, so super excited to chat with you about our newest models and build some demos live with you all.

I'm Kat, I work on vibe coding in AI Studio. This is Ammaar, he leads our product and design team for AI Studio. But I want to step back for a second and talk about the journey at DeepMind generally.

So what's, I think, particularly unique about Google's journeyright now is that DeepMind has been innovating here for not just this week or this past year, but for years and years with things like the Transformer, AlphaGo, et cetera. And this is obviously a graphic from five days ago because it ends with Gemini 2.5.

And we are super excited to have announced earlier this week Gemini 3 Pro. Hopefully this message has reached you all already. If not, we have a lot of work to do. But this is our latest, most intelligent, state-of-the-art model.

And ultimately what we want folks to understand with Gemini 3 is that we can really build anything. And that comes in two major capabilities. I think the first is the UI and aesthetic sensibilities of Gemini 3. It's very, very strong at design understanding and generating websites and good UIs in one shot.

And the second is with agentic tool calling. So I think this goes back to the sort of spectrum we're seeing with models. Sometimes you want a one-shot website, and sometimes you want to do really complex tasks within, you know, massive code bases.

And that's where tool calling and agentic use can be particularly powerful. So with Gemini 3, we see on theright is a SweBench experiment where it was a base agent harness across a few different models. And we can see Gemini 3 is vastly above in performance in agentic scenarios and then as well leaps above our previous models and state-of-the-art across the board.

So super excited to see what you folks build with this model. And in the meantime, we, you know, launched this on Tuesday, but there were still three days left in the week, so we had to launch something else as well.

So I hand it off to Ammaar to talk about our Pro image model.

Yeah, so at DeepMind, I think you have a few days left in the week, you choose to launch another breakthrough model. And so we're really excited about Nano Banana Pro, which came out yesterday. And it's a huge leap on our already state-of-the-art image model.

So with Nano Banana Pro, one of the things that I love about it the most is its world knowledge. So it's powered by Google Search. And so you can ask it all sorts of things, like, how do I make this tea?

And it'll actually go search Google Search, create a detailed infographic for you and diagram for you. And there are all sorts of things now with accurate information that it can do. And the other thing you're noticing here is improved text rendering.

So the text is one of those small details that if you get it wrong, you can pretty much pick it up quickly. But in Nano Banana Pro 2, it does an amazing job at text rendering. And you can see that in a bunch of examples, like here, where it wraps around the can perfectly.

And it also has a bunch of localization as well, so tons of languages, Korean on theright. So it can translate images as well and render them perfectly on the exact same reference image. On top of that, consistency is improved.

So you can now put up to 14 people in an image and then can create this group shot you can see on theright. And that, it can do more, but 14 is basically our kind of benchmark so far.

And that also enables a whole set of new use cases. And then creative controls as well. So you can see here on the left, the focus is on the woman, and on theright, on the flowers. And this was just a simple prompt.

All you had to say was, change the focus to the flowers. Maintains everything in the previous image, just changes the focus. So incredible outputs as well with Nano Banana Pro. And a range of aspect ratios. So if you want to generate wallpapers or big banners or advertising boards, you can do all of that as well.

So anyway, instead of talking, we decided we're just going to show you a bunch of demos live of what we've been building with these products over the last week. And yeah, excited to jump into it. So let's do that.

Allright, so Kat.

Yes.

Take it away.

Here we go. Kat Kampf. Cool. So for folks who aren't familiar, this is Google AI Studio. It's our home for getting started with the latest Gemini models. You can get your API key, chat with the latest models, including Gemini 3 and Nano Banana Pro.

But today we're going to be focusing on this build experience. So this is our vibe coding experience in AI Studio. You can see here we have a gallery of a bunch of example apps, a bunch of very cool, to the aesthetics point of Gemini 3, a bunch of very cool Gemini 3 examples.

But you can also go prompt to app super easily here. And this is free to use. And I think one of the unique things about AI Studio is how easy it is to integrate the Gemini API into your application.

So we can see here at the bottom, there's a bunch of these, what we call AI chips, that showcase a ton of the unique features beyond just the model you're choosing with the Gemini API. Different tools you can use, like Google Search Grounding, Google Maps Grounding.

We also let you build with our live API, so you can do one-shot examples of, I have one that lets me input a webcam of my tennis swing, and it'll give live corrections on my swing.

You also made one to improve my posture.

Yeah, yeah. Yeah, if you lean forward too much, the live API will yell at you. So it's a very flexible way to get started building AI-powered apps. And the other cool thing is you don't actually need an API key here for most of the models.

So you can build your application, you can share it with the world, and anyone who comes and visits your shared application will be using their AI Studio free quota. So you don't have to worry about, you know, hopefully you have an app that goes super viral.

You won't have to worry about a crazy surprise API bill or anything like that. So I'm going to actually shoot off a prompt here that is using our latest Nano Banana model. And that basically allows us to use Google Search Grounding to create an illustration of laptop stickers.

And this is one of the viral trends we've been seeing with Nano Banana Pro. So I'll kick this off. And what this will do, I have the AI chip that tells it to use the Pro model. And this will sum up my prompt and go talk to Gemini 3 to break down the task and start generating my end-to-end application.

But while that builds, I'm going to hand it off to Ammaar to show some demos in the meantime.

Cool. I think the other thing to point out here is that we're trying to think through how the vibe coding experience is also powered by AI every step of the way. So you're seeing here, even in the loading screen, it is using Gemini and thinking through this app that you're making and how you could extend it.

And so we're thinking through breaking those typical vibe coding paradigms as well and helping you iterate with the model as your partner. But anyway, let me jumpright into the text rendering demo. So when I heard of text rendering for the first time and the consistency that we were getting with Nano Banana Pro, my mind went to comic books.

And so I was thinking, why can't I now be in my own comic book adventure and also place Kat in there? And then maybe we can tell a story. And so in this app, also vibe coded, you can just upload a face of somebody.

So I've got some nervous face, of course. But I'll use Kat here and myself. And then we can choose the genre of the story and all the languages that we have so far. I'm going to do a story about us presenting at AI Engineer in New York, presenting AI Studio.

And we are vibe coding and winging our presentation. That's where we're going to be doing this comic book story. So we'll fire that off. But while we wait for that, the other cool thing about this is we'll wait for that to generate, but I want to show you the design sensibilities as well.

So you know that if you've been working with AI models and generating websites, they've been creating purple gradients and things that just, you know, they kill me as a designer. And so it's been really nice to see how this model is able to build some beautiful websites.

So this is using shader animations, flowing through all these different pages. And that's all sorts of cool transitions and effects, picked out the typography by itself. And this was the initial prompt. Just create a slick animation website. It kind of actually did say no cyberpunk shit there.

It's nice.

Just got to make sure. But yeah, you get some incredible results. And now what I love about this is so many folks who, you know, were struggling with design or might have still tried to grok their way around Figma don't have to do that anymore.

They can actually just go in, prompt their way to something pretty nice. OK, back to the comic book. OK, pretty flattering comic book here. That, you know, I'll take it. And you can see here that it's rendering the comic book.

It's got rich text rendering showing us the story. And the other thing here is that because it's powered by Gemini 3, it's actually really creative at the story it's generating. And honestly, some of these stories have genuinely made me laugh, which is the first time that's happened with one of these models.

And so you can see we're rushing to the conference. Even background details like the AI Engineer banner over here being rendered. And of course, since this is a vibe coded app, we can take this story in any direction.

So one feature I did introduce is that you can choose the direction of the story midway. So, you know, do we find a quiet corner and try to check if our API keys work, or do we just embrace it and go full improv?

I think we're going to go full improv. And so that's changed that story. And so talking about the humor here, you can see Ammaar dodged a woman carrying a suspiciously functional robot dog. So I don't know if that was announced at the conference today, but pretty cool.

And then now it's generating the rest of the story here on theright while we wait. So pretty cool to see how you can make these really dynamic, rich experiences with both the creativity of the model and Nano Banana Pro's image capabilities.

Love it.

Yeah, back to you, Kat.

Yeah, yeah. I will show, let's hope my sticker demo is finished up. Cool. So I'm going to add an API key. So Nano Banana is a new model, and it's fresh off our launch of Gemini 3. So for now, it is a paid experience in AI Studio.

But what I can do is I can see that here. I can enter different words that I want my stickers based off, or I can go use Google Search. So let's try the Google Search. I'm going to type in Ammaar's name.

And one of the other cool things about this new model is that you can select the resolution as well. So in this case, I'll just do 1K. But what this will hopefully do, but again, you saw it one shot live, is go talk to Google Search, grab their latest sources on Ammaar, build the context about what he likes, what his laptop stickers might look like.

I think it's just DeepMind, but if he were more, if he wanted to express himself more.

Oh, boy.

And so he can see here. Yeah, there he is. Weekend builder.

Right. That's true.

Yeah, and for those who don't know, Ammaar has a children's book, "Alice and Sparkle," which, yeah, is clearly he's talked about a lot because it's highly represented here. But very cool to see how it can bring in that contextual knowledge.

We've also seen this with news events getting relevant information on that day rather than having to rely on the knowledge cutoff of the model. So one other thing I'll show you folks is how we use AI Studio to build AI Studio.

So Ammaar and I have a lot of ideas, only so many engineers to work on these ideas. So we love to use AI Studio to ideate and explore different concepts. So one of the concepts we've been working on is, I'm sure you folks have seen, we announced a new agentic IDE at Google earlier this week called Anti-Gravity.

And we know that sometimes these web-based vibe coding tools, they have their limits, and you may want to go into an IDE to add certain features to the application or make it specific to mobile, things like that, that might be a bit limited in AI Studioright now.

So we want it to be super easy to migrate into Anti-Gravity. So what I did here was just a one-shot prompt of a screenshot of AI Studio. I said, clone this UI as closely as possible, and then add a flow to export to our Anti-Gravity app.

So we can see it did a pretty great job of cloning light mode. The screenshot was in light mode too, of our AI Studio application and copying it and improving a little bit on Ammaar's designs. But then we see this new Anti-Gravity button that is creating my export and then exporting it to Anti-Gravity.

And I can go and open in the IDE. And I think these are the types of creative interactions that web-based vibe coding tools can be particularly useful for. Because if we had went and jammed on this feature, we probably would have constrained ourselves to existing patterns in AI Studio.

And in this case, I told the model, be creative, think outside the box. And I've played with this one a bunch. Sometimes it gives a command line interface for showing the status of the export, et cetera. So I think it's a super cool way for you to ideate on new ideas for your UI and kind of expand on your product.

But I'll hand it back to Ammaar.

Let's do it. And then the other thing that Gemini 3 has been really impressed us with is just making video games. And so this one was, again, pretty simple prompts. Make this racing game where I have a bot now at a start screen.

And so you can see I got this 3D racing game in 3JS. I drew all the things. I'm racing with a bot here. And then one thing I added for myself to cheat is I can just boost away and beat the bot.

So pretty nice. But the thing I want to tease, actually, is that all of these apps so far have been front-end React apps. And so the thing that's coming very, very soon to AI Studio is going to be backend support and full-stack runtime.

So if you want to install ShadCN, if you want to do all of those things, you'll be able to do that, again, with one prompt. And the principle with AI Studio here is that we don't want you to think about those details.

You should just be able to ask, I want to make a multiplayer app. And we know that you need to use Express and wire that all up for you and abstract all those details away. So we're going to try something a little risky here, which is we did turn this racing game into a multiplayer one.

And this was, again, a couple of prompts. So we're going to put a QR code up if you want to join us in the racing game.

We've never tried with nearly this many people, so we'll see.

Hopefully this works.

Yeah.

But QR code's up here. So if you scan that, it really should load the game. I'm really afraid of how this is going to explode.

Yeah. Here we go.

Got all these cars loading in.

Nice.

So yeah, if people have scanned that, we can switch back to the game. OK, oh my god.

So yeah, just hit ready when you're all ready.

Oh, boy.

I think this lobby is going to explode.

Yeah.

I'm going to leave.

So this is where I shouldn't have added collisions with other cars because you could clearly see that we're bouncing around.

Yeah.

19 players, 20 players. I don't know if this race will ever start, but we're all blocked on the start line. But 23 players, pretty cool. Yeah, you do all have to hit ready for us to start this race.

So we might be here all day.

Yeah.

But yeah, that is pretty incredible. I can't start this race. So do you want to wrap up?

Yeah, yeah.

Who was you all in here? That's pretty cool. The runtime didn't explode.

Yeah, and I think we're super excited not only about the multiplayer game. So next time, we'll have even more of you folks join. But also, you know, the extensibility that comes with a full-stack runtime, we want to make it super easy for you to integrate with our OneP and popular third-party APIs, et cetera.

So very exciting next few months on the AI Studio vibe coding side and super excited for you all to try it. But I think the one thing I want to step back and emphasize was what makes us so excited about this project and the work that a lot of us are doing is that we get to be the first generation of engineers who are building tools for a world where anyone can build software.

So I think what's beautiful about things like vibe coding is watching people. We were actually talking to a tech support person earlier this morning who said they started vibe coding in AI Studio after seeing a YouTube video. And we're really democratizing who can create things, and we're all getting to build those tools that enable that.

And I think it forces us to rethink the paradigms that we've become so used to. So it may not be your base IDE that people are starting from, but how can we intuit as much of the user intent as possible?

And that's what we want to do with full-stack runtime in AI Studio is make it very easy to not have to think about, I want to add a database. But if your app needs storage, it'll have storage. If you have an e-commerce app, we'll add a payment solution and make it as easy as possible to build the future of software.

So thank you, folks, for joining us. If you have any cool examples you've built or questions, feel free to ping me and Ammaar on Twitter. And yeah, enjoy the rest of the day.

Yeah, thank you.