What's new in Google AI

Woo!

Amazing! Welcome, everyone.

Welcome to Google I/O. Welcome

to "What's New in Google AI," and welcome to the first sessions of the day after the keynote. Is everybody really

keynote. Is everybody really excited?

[ Cheering ] Amazing!

So, I am so grateful to be here today with my colleague Ammaar.

My name is Paige. I'm the lead for developer relations team at Google DeepMind.

AMMAAR RESHI: I lead the product and design on AI studio.

PAIGE BAILEY: Very cool. And

today we're going to be attempting to condense into 45 minutes an overview of everything that we've been doing this is google in the AI space.

Going to be really hard. But

the good news is that for the rest of today and the rest of tomorrow and further on in the week we have a whole bunch of other talks that are intended to be deep dives for some of the things that we'll just be touching on lightly today.

So, I don't think it's a secret that we've been shipping at a pretty relentless.

AMMAAR RESHI: An insane amount of pace. I think if you just look at 2024 when we had the 1.5 series of models, we were just cracking the nut on multimodality, and look where we are now.

PAIGE BAILEY: Exactly. And

today we released Gemini 1.5 Flash, which we'll be talking about in a second. But it's

everything from Frontier Models to Open Models and more.

Our Gemini q m q model lineup hs expanded pretty significantly both today and over the course of the last couple of months.

We've got gemini 3.5 pro, which is still kind of our best and brightest model to use for complex problems, 3.5 Flash for kind of performance and speed

and cost. This is also the

and cost. This is also the Flash series is the default in AI studios build, right?

AMMAAR RESHI: Yeah, we use 3.5 flash as our default model for building apps And it's incredibly capable. In fact,

incredibly capable. In fact, the AI studio mobile app that'll be coming out in a few weeks, a bunch of it was built with 3.5 flash. So it's just been

flash. So it's just been amazing to see it ship across both Android and iOS.

PAIGE BAILEY: Amazing. And

then also 3.1 Flash-Lite, which is really great for kind of these use cases that require low latency but can also save you a lot of costs, which is pretty cool.

It's not just us using Gemini 3 We are blessed to have a whole bunch of customers across the industry. Databricks, GitHub,

industry. Databricks, GitHub, Harvey, Warp, and a whole bunch more that have Gemini q running in production. And Gemini is

in production. And Gemini is really remarkable in the industry because it's great at understanding all of these many different modalities, text and image and audio and video and code, but it can also output

multiple modalities.

AMMAAR RESHI: Yeah, I think one of my favorite things is because it understands so many modalities, when you're building, especially when you're building apps, I upload a video of an animation, and it can understand the frames of that

animation, and then recreate that in code. And so video understanding just unlocks a whole new set of use cases. And

I mean, so does all the other modalities. A

modalities. A PAIGE BAILEY: Yep. And the

screenshots that you can take and use to inform the front end design for some of Build's features are really, really powerful tools as well.

Our generative media models are kind of testaments to this fact, that we're able to output video We can output audio. We have

some pretty exciting text to speech models and Gemini Live models. And then also our Live

models. And then also our Live API, which we'll demo in just a second. It gives you real time

second. It gives you real time interactions with Gemini the.

And today, you know, we talked a lot about nano Banana, aka Gemini 3.1 Flash image, which

gives you high fidelity visual assets. We've got Nano Banana

assets. We've got Nano Banana 2, which you can try yourself in ai studio and today we've basically released n Nano Banana

for video Gemini on Omni Flash, which allows you to create anything from any input starting with video starting with video.

AMMAAR RESHI: And I think the other thing is editing is such a tedious task especially getting the little details right and the fact that we have Nano Banana style editing for video now is like super exciting, and in the API soon. So that'll be really

API soon. So that'll be really fun to build with as well.

PAIGE BAILEY: 100%. And if

you haven't had a chance to try it yet, there's a really nice AI demo garden in the back of this building. If you're onsite,

building. If you're onsite, that should allow you to create a commercial starring yourself

using Omni. It's called Omni

using Omni. It's called Omni Mercial. Be patient, if there's

Mercial. Be patient, if there's a line. But it's a lot of fun

a line. But it's a lot of fun to try if you're curious about generative media. Make sure to

generative media. Make sure to go to this session Build creative apps with the GenMedia suite later this week with a couple of a couple of folks from our team, and they will be able

to answer all of your questions.

All of these models are stitched together in Antigravity, as kind of like the IDE that powers these experiences. We're all

these experiences. We're all using Antigravity pretty heavily at Google internally.

AMMAAR RESHI: Every day.

PAIGE BAILEY: Every single day to check in to check in code, to help review code, to kind of design these new experiences. And we're also

experiences. And we're also using AI Studio, for everything from product ideation, to building apps, to rapid prototyping, because it is the

fastest path from prompt to apps. But I really hate slides.

apps. But I really hate slides.

AMMAAR RESHI: Me too. Hate

slides.

PAIGE BAILEY: So we both hate slides. It's much more fun to

slides. It's much more fun to show these things in action. So

let's go ahead and go to some demos so you can actually see these tools. And we'll see how

these tools. And we'll see how well the WiFi holds up.

Excellent. Back in the playground.

AMMAAR RESHI: So for those not familiar, this is the AI Studio playground and home page Here you can experiment with all of Deepminds latest models You can tweak all the parameters to see what's what they're capable of It's an amazing way to just

get a sense of all the capabilities, and you can just take code snippets from here and go and build apps with them as well. But, Paige, you want to

well. But, Paige, you want to kick it off for the demo?

PAIGE BAILEY: Yeah, absolutely. I'm going to zoom

absolutely. I'm going to zoom in a little bit, so we can see, and then also show you how you can kind of select different media types in AI Studio, if you haven't seen this before. You

can add files from Drive. You

can upload files. Record audio,

camera, add youtube videos. And

you can also select some sample media. So if I was to select

media. So if I was to select this video, add it to the prompt, we've got gemini 3.1 Flash-Lite pulled up, which is, again, kind of our most cost

optimized model.

AMMAAR RESHI: You can say fast.

PAIGE BAILEY: It should be pretty fast. You can say create

pretty fast. You can say create a table with timestamps for all of the dinosaurs you see in this video. Make sure to include a

video. Make sure to include a fun fact about each dinosaur.

Hit "run," and we can see the number of tokens that's pulled from this five minute-ish long video. It's around 31,000.

video. It's around 31,000.

AMMAAR RESHI: Fast.

PAIGE BAILEY: Very fast.

We've got a table with the dinosaur name the timestamps the fun fact, and then if I hover over this token consumption, you can see that it took about a penny and a half in order to do all that work, which is pretty wild.

AMMAAR RESHI: Amazing.

PAIGE BAILEY: And then just like Ammaar said, if you click "get code," you can see all of the code that you would need to use to replicate whatever you just did, including search as a

one liner for a tool call this thinking config, and then the model selection that we have here up at the top. So that is Playground in a nutshell. And

it allows you to experiment not just with our Gemini family of models, but also things like Nano Banana and some of the other video generation models that we have along the way. We

also talked a little bit about Gemini Live.

AMMAAR RESHI: I love this one.

PAIGE BAILEY: Yes. And so

Gemini Live gives you the opportunity to have a real time conversation with the model.

But also share multimodal content. It's been around for a

content. It's been around for a while. But it's really, really

while. But it's really, really cool to think about building into your apps. So if I click share screen, it's going to ask me to select a Chrome tab. I'm

going to go ahead and select that dinosaur table that we just created previously, though you could select any app or any tab I'm going to go back to Google

AI Studio, and I'm going to say... "Hey there, gemini...

say... "Hey there, gemini...

could you tell me what you see on the screen?"

GEMINI: I'm looking at the Google AI studio. The main part of the screen shows a table of dinosaurs with timestamps and fun facts about them. For

instance, it says, Ornithomimus was a speedy dinosaur. Is there

anything specific you'd like me to focus on?

PAIGE BAILEY: Does anybody speak a language other than English in the room?

Arabic.

PAIGE BAILEY: Arabic, awesome. I also heard

awesome. I also heard Brazilian Portuguese or Portuguese Portuguese...

Portuguese Portuguese...

excellent. Chinese?

Awesome.

So I'm going to ask... I'm going

to ask the model to repeat what it just said but to say it in Chinese, and then we also will try another couple of languages, but you have to fact check me, if that's okay.

Cool.

Hey there, Gemini, could you repeat what you just said, but could you tell me in Chinese please.

GEMINI: [ speaking in

Chinese ] PAIGE BAILEY: Did it do okay?

Excellent.

AMMAAR RESHI: Let's go!

PAIGE BAILEY: Wonderful.

Amazing! Very cool.

And then you can also... so that

was just kind of asking it to change languages dynamically within the context of the conversation.

You can also say in the system instructions "only respond to the user in Brazilian Portuguese."

Portuguese." If I have a person who will fact check me in the audience for Brazilian Portuguese... yeah,

Brazilian Portuguese... yeah,

excellent.

And then I'm going to turn on grounding with Google Search, and we are going to ask Gemini a question as well.

[ computer sound ] PAIGE BAILEY: Hey there, Gemini. Can you tell me what

Gemini. Can you tell me what you see... Can you tell me what

you see... Can you tell me what you see?

GEMINI: [ speaking Brazilian Portuguese ] PAIGE BAILEY: It was right?

Cool! Amazing!

And since it's got grounding with Google Search, I could also say... "Hey there, Gemini, could

say... "Hey there, Gemini, could you tell me what the weather is like today in Mountain View,

California?"

California?"

GEM Gemini: [ speaking ] PAIGE BAILEY: Was that also correct? And you can also see

correct? And you can also see the sources that it pulled in to help ground its outputs. And

just like before you can click get code and see everything that you would need in order to replicate the experiment that we just tried.

AMMAAR RESHI: That's awesome.

PAIGE BAILEY: So this is and python and TypeScript and dot net, whatever your favorite language might be both programmatic or computery, and

AI Studio is really just a magical place to get started with all of these things. But

it's not just the playground.

AMMAAR RESHI: Not just the playground.

PAIGE BAILEY: You want to show us builds?

AMMAAR RESHI: So if you don't want to click get code and actually take code and build in an IDE, we offer a completely batteries included building experience, and that's build mode here. So I'm just going to

mode here. So I'm just going to click that. And build mode

click that. And build mode effectively allows you to prompt any idea into life, and you'll be able to pick between all the capabilities that we have to offer. So you can do text to

offer. So you can do text to speech, music generation, databases. And a new thing

databases. And a new thing we've added that we're super excited about is building Android apps directly from AI studio. I

studio. I PAIGE BAILEY: Woo!

AMMAAR RESHI: So what I'm really excited about is you won't have to deal with any of the setup of android apps or getting a simulator or anything like that working. You'll just

be able to prompt an app to existence. So an idea I had was

existence. So an idea I had was around this piano, but I wanted to do it for this foldable phone I have over here.

PAIGE BAILEY: Which is very fancy. How do I get one of

fancy. How do I get one of those?

AMMAAR RESHI: You know, you just expense one. So all you have to do is say, "I want to make a foldable piano app for

fold phones where the top half of the screen is a tutorial and the bottom half are piano keys

that I can play."

PAIGE BAILEY: Awesome.

AMMAAR RESHI: And the other thing I'll just show you is we just left it for a second, and it'll even give you suggestions that you can tab into your prompt.

And so this just helps you build out a more complete prompt if you don't know where to go, or if you're just curious about where the model could take you, which is always really fun.

PAIGE BAILEY: Yeah. So you

can tab your complete your way into a startup idea. Is that

what you're telling me?

AMMAAR RESHI: Exactly.

PAIGE BAILEY: Amazing.

AMMAAR RESHI: But you know, I don't like that idea so I'm going to keep that away for now And so I can go ahead and hit build and it'll go and start to build this idea out.

And so the other thing you'll see is AI Studio says right here it's generating a theme. And so

what's happening is, while your app is building on the left, AI Studio is also able to design a bunch of design themes for your app. And super fun to just see

app. And super fun to just see the inspiration kind of come through from the model. So you

can see it's kind of giving us some interface ideas on how it's designing the app...

PAIGE BAILEY: Oh, that's pretty.

AMMAAR RESHI: You know, we can choose between a few different ones. Even a video

different ones. Even a video tutorial idea, which is pretty cool. Or this elegant dark one,

cool. Or this elegant dark one, which I really do like.

PAIGE BAILEY: That is very pretty. I like the pastel

pretty. I like the pastel colors.

AMMAAR RESHI: Yeah. Super

nice. And I like the little glow ab it added. So these this help you kind of you know I can hit select this design and it'll just add that prompt in or I can just skip that and let it build that first version of the app.

But while that builds, I actually did start this before this demo. So I do have it

this demo. So I do have it here. I'm just going to refresh

here. I'm just going to refresh this, and then we can go and build that one as well.

PAIGE BAILEY: And this app is generated natively in Kotlin, is that right?

AMMAAR RESHI: It's natively in Kotlin. And so again similar

in Kotlin. And so again similar prompt You can see my first prompt again build a piano app, foldable thing, And if I hit code you can see all the Kotlin code. And the other thing you

code. And the other thing you saw earlier was we said that Antigravity export is also in the app. So you can click

the app. So you can click export and then take that over and go from there.

PAIGE BAILEY: And it will also remember all of the context, all of the history, all of the kind of back-and-forth that you had in AI Studio as it brings it into Antigravity.

AMMAAR RESHI: Exactly. So

the whole conversation kind of flows through, which is really nice.

Okay, so we have this install button. I'm just going to zoom

button. I'm just going to zoom out for a second. And it allows you to basically connect any phone and then install it via USB. The USB can sometimes be

USB. The USB can sometimes be fiddly on stage. So we're going to try something.

I'm just going to kill the USB connection and unlock my phone.

Let's see if this works.

PAIGE BAILEY: Amazing!

AMMAAR RESHI: Yeah, I want to see if it installs the latest version. Allow. Okay... seems

version. Allow. Okay... seems

to be installing.

And then hopefully any second now on AI Studio we will see the new build complete.

PAIGE BAILEY: There it goes!

AMMAAR RESHI: So it's installed on advice, and you can see I have this piano app now on my foldable phone.

[Applause] PAIGE BAILEY: Amazing!

AMMAAR RESHI: And super fun, because I think you're just going to see so many new developers come through who just wanted to make apps for their phone or the web, and all they had to do was describe in AI Studio.

PAIGE BAILEY: Do you want to start it a little?

AMMAAR RESHI: Yeah, should I play it? We'll see how it goes.

play it? We'll see how it goes.

PAIGE BAILEY: I'm not sure if you can hear here, but it looks like it's popped up on the screen as well. Maybe we could also try in the Android

emulator.

AMMAAR RESHI: The emulator is getting audio soon.

PAIGE BAILEY: Gotcha.

AMMAAR RESHI: Soon we will have Play store distribution too. So you'll be animal to

too. So you'll be animal to just submit an app from the play to the play store from here, and then your publish to the world.

PAIGE BAILEY: Amazing.

AMMAAR RESHI: Yeah. And

rolling out today is workspace apps as well.

PAIGE BAILEY: Which means you can connect gmail and Calendar and all of the other workspace products into AI Studio. Is

that right just for prompting AMMAAR RESHI: Just through prompting, yeah. Do you have an

prompting, yeah. Do you have an idea?

PAIGE BAILEY: I have an idea... so I know we've been

idea... so I know we've been talking about our meeting invites, and our calendars are a little bit overwhelmed. Wasn't

there an idea that you had around sort of calendar roulette?

AMMAAR RESHI: I did. I did.

You know, I've already built that one. Should I show you?

that one. Should I show you?

PAIGE BAILEY: Yep.

AMMAAR RESHI: It's a pretty fun one. It's a slightly scary

fun one. It's a slightly scary app for those who might not want to trust a model to do this, but I was basically like, hey, just look at my calendar, find any meeting and then just delete it.

[chuckles] But definitely ask me before you do that, because I need to check PAIGE BAILEY: And make sure

it's not a meeting with Logan.

AMMAAR RESHI: A little risky It picked up some meetings.

It's got that, I can spin it and then it's going to go and figure out what meeting to delete.

PAIGE BAILEY: I love thought's got all of the I/O keynote, launch Kaggle, all of these things. And it's

these things. And it's automatically used OAuth to log you in with your Google account based on our Firebase integrations. Anid Exactly.

integrations. Anid Exactly.

An AMMAAR RESHI: Exactly. So

you can see it even knows there are 15 events later today so I definitely want to get out a few of those I'll be using this app to make that happen.

PAIGE BAILEY: Excellent.

Sounds good to me. So cool. So

this is a great overview of Playground and build some of the newest features in both.

AMMAAR RESHI: Yeah.

PAIGE BAILEY: And I think we're really, really excited to see what you all can create with AI Studio and with all of these tools. To access it, all people

tools. To access it, all people have to do is go to ai.dev, right?

AMMAAR RESHI: Yeah. Somebody

is telling me to spin the wheel and I'm too afraid to see what it's going to do live. So we'll

save that for later, you and me.

PAIGE BAILEY: Cool, cool.

So let's go back... this shah

brief taste of all of the things we've been building in the AI Studio team. If you're curious

Studio team. If you're curious about AI Studio, you're in luck, because the next section is all about AI Studio, Antigravity, and how you can incorporate those workflows into your own

projects. We've also released

projects. We've also released something called the interactions API within AI Studio. So you can create an

Studio. So you can create an API key and get started with agents and build new agentic

experiments with the With the.

Gemini APIs If you're curious about that about our managed agent platform, we talked a little bit about it in the keynotes earlier this morning.

But it's basically giving you the ability to that same vibe coding style that you just saw where you can describe a natural language anything. It will

language anything. It will generate a managed agent for you connecting to services and running on a cloud VM.

AMMAAR RESHI: Yeah, and it's the Antigravity agent powering all of it, which is really awesome. So you have the exact

awesome. So you have the exact same power and capabilities working in an environment with just one API call.

PAIGE BAILEY: Amazing. And

if you're curious about that, make sure to go see the build agents with Gemini API session with Phillip and Thor later this week. They'll also be touching

week. They'll also be touching a little bit on the Gemini Live API, around our text to speech models, and all of the things that you can do with those.

We've not only been shipping Frontier Models.

AMMAAR RESHI: Right.

PAIGE BAILEY: And something that's really near and dear to my heart is Open Models.

AMMAAR RESHI: Open Models open source models, And I think to see how far gemma four has come and, what's possible with it and especially the demos we've been seeing this week on how it's running on device and pulling off some incredible things.

PAIGE BAILEY: Absolutely.

And Gemma 4 is available to test out in the playground in AI Studio as well for folks who want to get started, and under known fact I think is that you get a non-zero number of gemma

calls and the gemini APIs for free. So if you wanted to

free. So if you wanted to incorporate it into your app, it's great at understanding video audio images text code, kind of all of the above.

And you can kind of compare and contrast with the Gemini models and other models that you might have incorporated into your apps as well.

We've got a huge, huge Gemma 4 community that is pulling down many, many downloads across the

world. Gemma supports over 140

world. Gemma supports over 140 different languages. You can

different languages. You can find it today on hug Hugging Fa It has a 256000 token context window, and if you want to learn more about the gemma open model

family make sure to check out this session with Olivier, gus, and ian also later this week.

We've also been really blessed to have strong partnerships with the Unsloth community with Ll Llama. And many others like

Llama. And many others like Kaggle to bring these Gemma models into the world. And it's

been, it's been, you know only the beginning for what's possible with Gemma 4. So this

is for folks who haven't tried running models at the Edge before. I'm curious. How many

before. I'm curious. How many

folks have used Open Models before? Are you using them in

before? Are you using them in your projects? Few hands. If

your projects? Few hands. If

not, definitely make sure to try out the Google AI Edge Gallery.

It's a great way to bring down costs for your projects, especially as the models that you can use on mobile devices, on laptops, they are getting more and more powerful.

AMMAAR RESHI: Yeah, especially love the offline usage. Because I was on a plane

usage. Because I was on a plane ride, about eight hours, and Gemma 4 five' coding with it and it actually worked really well.

Amazing. So it's amazing to see that all run locally.

PAIGE BAILEY: Yeah. And

Gemma 4 can run on your laptop gemma two can even run on a mobile device so I think we've added it to the Pixel 10 so that it can power some of the on device agentic use cases that folks might want on On android

and on pixel devices. And,

again, if you're curious about on device ml Sachin and Aaron will tell you all about it later this week. Google is kind of

this week. Google is kind of special though and the in the sense that we aren't just a company that has you know models that are available via rest APIs, models that are available to download and to fine tune and to use for your projects. We

also have our open source TPU software stack.

AMMAAR RESHI: Yeah.

PAIGE BAILEY: So everything from inference to post-training pre-training and model building, kind of the usual suspects, like vLLM for TPU inference, for both

JAX and PyTorch. Tunix for

Post-training with RL and agentic workflows. MaxText.

agentic workflows. MaxText.

And then also JAX and TorchTPU for making TPUs run really, really fast and in a really performant way. It's been

performant way. It's been amazing to see how much JAX has been able to push the boundaries for every single model that we build at Google, and our DeepMind team is exclusively using JAX to build out all of

its all models and infrastructure. And if you're

infrastructure. And if you're curious about that, make sure to go to Josh's session later this week where he will be talking about how you can scale AI with the TPU software stack.

This full stack approach is really unique. Everything from

really unique. Everything from AI infrastructure to data, to models, to our AI platform on Cloud. To things like agents

Cloud. To things like agents and apps via Gemini Spark via AI Studio, and more.

And also all of this being done in a very secure way. I was

really excited to see the Synth ID announcements earlier.

AMMAAR RESHI: Yeah, with the partners.

PAIGE BAILEY: Partnering with even model service providers across the industry, and attempting to do this in a way that really is kind of putting safety and responsibility...

AMMAAR RESHI: I think we'll need to give the Calendar Roulette app to everybody here to clear their schedule for all these sessions.

PAIGE BAILEY: Yeah. This is

true. Amazing!

So what's next? Like I guess that's the general question.

AMMAAR RESHI: Always.

PAIGE BAILEY: Given all of these things that exist today, that were announced today, this overwhelming onslaught of AI that's even just from Google and the DeepMind team that's being released into the world, what are you most excited about

that's coming next?

AMMAAR RESHI: Well, I love all the robotics stuff. It's

always super exciting. I know

you love the robots too.

PAIGE BAILEY: I really love the robotics stuff, especially since because of the Gemini APIs, they can be put pretty much anywhere.

AMMAAR RESHI: Anywhere.

PAIGE BAILEY: Any device that can call a Rest API is capable of using our robotics.

AMMAAR RESHI: Yeah, and there was a great demo of Reshi mini from Hugging Face, and it was using Gemini robotics, and that's open source as well for that project. But yeah, we

that project. But yeah, we recently announced Gemini Robotics 1.6 and it just incredible to see that out in the wild now.

PAIGE BAILEY: Yeah, and I think it's also really powerful in the sense that, you know, you mentioned the Reshi Mini.

There's also a beautiful open source project from Stanford called "the Pupper," which is used for their CS curriculum.

It has a Raspberry Pi powering it behind the scenes. All of

the parts are 3D-printed. But

since it's just kind of using this commodity hardware, it can call the Gemini Live API, use Gemini robotics ER to kind of manipulate this little robotic

dog, and you can ask it to do anything from like "follow me," and it will follow you around the stage.

AMMAAR RESHI: So cool.

PAIGE BAILEY: Yeah.

AMMAAR RESHI: How much training is needed for that bun?

PAIGE BAILEY: Not any training required. Batteries

training required. Batteries included, the first dog you don't have to train, but you just have a conversation with Pupper, and it's capable of having this dynamic interop with

you across a broad spectrum of use cases.

AMMAAR RESHI: That's awesome.

PAIGE BAILEY: Yeah. In

addition to that we also have our world Model family, Genie 3, three which was touched on a little bit today in the keynote.

AMMAAR RESHI: Yeah, part of the Omni model. And it gives it the understanding of the physics of the world and how things actually work. And that's why

actually work. And that's why you see the video coherence is so good, because it actually understands how things actually need to work in the real world.

PAIGE BAILEY: Absolutely.

And you can describe any world that you would like to experience, and then navigate around it using your arrow keys

in order to create a short video. So as an example, this

video. So as an example, this is a prompt that, you know, a tranquil waterfall cliff area featuring dynamic water physics And then a character, which is a high-speed paper airplane. And

you can kind of see that in action along the way. If they

could play the video for us...

I'm not sure the clicker is

capable of playing the video.

So if you go back a slide, there should be a video embedded within the slide.

Yep, there we go. I can see it on the screen. And then all you should have to do is click that

play button in the center.

Nice. There we go. Awesome.

Thank you all. Yeah, and so you can see similar to what Ammaar was just mentioning the physics of the airplane, as it's kind of interacting with some of the splashes the physics of the currents and the river, how this is moving and banking.

AMMAAR RESHI: And then the reflections of the light on the water.

PAIGE BAILEY: It's so nice.

Could you play it again, please?

Yeah, awesome. Thank you. And

so you can see even the splashes kind of going down in the water as it jumps off the waterfall.

So I strongly, strongly encourage you all to play around with Genie 3. It's available

today if you have a Google AI Ultra subscription, in addition to many of the other products that were featured in the keynotes this morning.

Awesome!

And then so I think the most important thing to call out is that we're building kind of the first generation of tools for people who are AI native, and to

really don't have to ask for permission in order to create anything that they've been imagining.

AMMAAR RESHI: I keep telling people, all you literally have to do is just go to that prompt box and describe your idea. And

people have asked me, what is the tip you have, or something, for folks. And I usually just

for folks. And I usually just go in assuming that the idea I have is going to work. And

you'll realize how far you can actually push the model to make your idea very real.

And if it doesn't work in this iteration, I guarantee you Another model iteration, and you'll be able to make that idea come to life.

PAIGE BAILEY: Yeah, and since we've been adding databases to ai studio authentication, even things like these workspace integrations...

integrations...

AMMAAR RESHI: Yeah.

PAIGE BAILEY: It really does feel like you can go very, very far without ever needing to leave the browser and popping into an IDE.

AMMAAR RESHI: Yeah. And soon

on your phone too with the AI Studio mobile app coming very soon.

PAIGE BAILEY: Amazing. And I

can't wait for this world where I'm just walking around Describe an idea that I have and then by the time I get home it's already landed in my inbox.

AMMAAR RESHI: Yeah. Just a

few weeks away from that.

PAIGE BAILEY: Very cool.

Amazing. So for anyone who is curious, excited, who wants to try out and test more wants to try our models, make sure to go to ai.studio. And we also just

to ai.studio. And we also just want to say thank you so much for coming to I/O this year, for your energy, for your excitement.

AMMAAR RESHI: Yeah.

PAIGE BAILEY: We really appreciate you all and can't wait to see what you build.

And we also have time for questions. If anyone has

questions. If anyone has questions or curious about things, happy to answer or attempt to find answers. There

should be a couple of microphones set up on either corner of the room, if folks are interested.

AMMAAR RESHI: We have a question over there. Do you

want to walk to that mic over there?

PAIGE BAILEY: For folks, if you're unable to walk to the microphone, if you ask your question, I'll also repeat it, so that everyone can hear.

AUDIENCE MEMBER: Hi. So

thank you for that demo. So,

the Antigravity is like an inbuilt in that studio, or you have to just do something else to get it like the Antigravity?

PAIGE BAILEY: So there is export the Antigravity available in AI Studio. But to use Antigravity, you would need to download it Antigravity 2.0, and

Anshul and Joana's section just after this should talk all about Antigravity and how you can use it and install it on your local

laptop. Awesome. Excellent.

laptop. Awesome. Excellent.

AUDIENCE MEMBER: A quick question on packaging. So is it really AI ultra to get access to flow and all these tools, or is it going to be available in Gemini Enterprise?

PAIGE BAILEY: That's a great question. I know that right

question. I know that right now, you know, some of the features around Gemini Flash and others are in the process of being added to Gemini Enterprise, but at the moment

Google AI Ultra is the best place to go to get access to them.n

them.n AUDIENCE MEMBER: Hey, thank you for taking the question, and as code is becoming so disposable, you could say, you

could build things at the speed of thought, what should people be building so that it doesn't get obsolete by tomorrow?

AMMAAR RESHI: Sorry, what was the last part?

AUDIENCE MEMBER: What should people be building, or how should they be thinking about building things, so that they don't get obsolete, whatever they build today tomorrow?

AMMAAR RESHI: That's a great question. I mean, I think, you

question. I mean, I think, you know, we're building a lot of our stuff internally at Google now with agents as well. I

don't think we think of it as, like, that thing is going to be disposable. Because I think

disposable. Because I think what we see is that you still need a lot of supervision around the agent to check their work.

With that said, I think things that allow us to get more monitoring on what agents are up to, more transparency around the process, more verification flows around what they're building, a lot of the robust testing,

things We're going to start also building into our vibe coding platforms as well. So when you go ship something, we'll do the security checks. We'll check if

security checks. We'll check if the thing is safe to deploy and all of that stuff. Really

building for this new partner that you have is how we're thinking about it.

But yeah, I wouldn't say disposable yet. I would just

disposable yet. I would just say it's more that it's a lot faster than the barrier to entry has lowered for a lot of people, which is pretty exciting.

PAIGE BAILEY: Yeah. And I

see a lot of people... if

they're building companies, they're building around kind of the limitation of the models or the platforms where they exist today. The best thing to expect

today. The best thing to expect is that the models eventually will be perfect, and they will be super fast, and they will be,

you know, pretty close to $0 in terms of you know your ability to run these inference calls for next to nothing. And so if you're building, imagining that world...

world...

AMMAAR RESHI: That's such a great tip. I think we were

great tip. I think we were building something where we knew that for the experience to be just right, it had to be a bit faster, but we still built most of the experience knowing that

speed was coming. Right? And I

think you can... like Paige

said, just assume that the ground from beneath you will rise with the model, right? So

just build for that experience that you intend to give to people.

PAIGE BAILEY: I think when Build first started it was using either like the Gemini 1.5 or 2.0 series. So the number of

2.0 series. So the number of apps that people could create...

AMMAAR RESHI: Super low.

PAIGE BAILEY: Super low. And

like now much, much more expansive. And it's been really

expansive. And it's been really magical to see how that platform shift has evolved over just the span of a year.

AMMAAR RESHI: That's a great question.

AUDIENCE MEMBER: Hi, my name is John. I work in the public

is John. I work in the public sector, so we have a lot of legacy code, which is in COBOL, and how good are the Gemini models that could essentially help us in converting a lot of

backlog there?

PAIGE BAILEY: Yeah, oh, man, this is one of my favorite questions. So like COBOL,

questions. So like COBOL, legacy code, Fortran code, like you are not alone. Like NASA

also has a ton of legacy code.

So does so does the energy industry and banking and many others. I find that our models

others. I find that our models are getting better and better at even understanding legacy code, even though there isn't as much in the pre-training and post-training mixture, especially when you give it

access to invoke things like manuals as tool calls, or to invoke things like managed environments with those languages kind of configured, so

that you can do sandboxed code creation, and then execution to see if like all of your tests pass, and similar. Like there's

a great blog post that I can also point you towards, if that would be interesting, from our from one of our teams internally, that was talking about migrating a Google

specific internal-only language to a more well-known language externally. And then also one

externally. And then also one that was migrating TensorFlow models to JAX models, using like a special flavor of internal TensorFlow. So it's not

TensorFlow. So it's not impossible. It just takes some

impossible. It just takes some extra work. And the models are

extra work. And the models are getting better and better every day.

AUDIENCE MEMBER: Thank you.

PAIGE BAILEY: Yep.

AUDIENCE MEMBER: Hi. A

question... Gemini, only... is

it an evolution of Veo for a replacement, or is it a separate track?

AMMAAR RESHI: Yeah, it's definitely an evolution. It's

effectively the next direction of... like Paige said earlier,

of... like Paige said earlier, it's going to be the Nano Banana for video. So you can expect

for video. So you can expect the omni iterations to be where you'll find all the video generation and editing

capability. So it'll be Veo++,

capability. So it'll be Veo++, it's like the still owe generation plus the intelligence

plus the editing all in one.

PAIGE BAILEY: And that really is one of the most magical things I think about the Gemini model family is that you find many models in the industry can understand different modalities or a subset of different modalities but very few are capable of outputting different modalities, whether it's audio or images or text or audio and

text simultaneously, or whatever it might be abp This video output is something that the team has been waiting on for a while, and I think we're all really excited about.

AUDIENCE MEMBER: Thank you.

AMMAAR RESHI: Yeah, thank you.

AUDIENCE MEMBER: Hi, I have a quick question regarding the output rate of the Gemini Live audio. We have seen some issues

audio. We have seen some issues in carrier compatibility because of the downsampling of the

24kHz. I don't know if you are

24kHz. I don't know if you are working on an pd kilohertz output for the gemini live models.

PAIGE BAILEY: I could barely hear you. I think I heard 8

hear you. I think I heard 8 gigahertz. Is that right?

gigahertz. Is that right?

I'm not sure if we're working on that kind of support for the outputs. But we can definitely

outputs. But we can definitely ask. And if it's not, we can

ask. And if it's not, we can file a feature request.

AUDIENCE MEMBER: Yeah, because when you try to use the Gemini model in an actual carrier, that downsampling degrades the quality. And we

have to choose other models because of that. But we would prefer using Gemini.

AMMAAR RESHI: Yeah, we'll take the feedback for sure and take it to the team.

PAIGE BAILEY: Thank you.

AUDIENCE MEMBER: Hi. My

question is... let's say I use one of the latest models to create an agent. So is there any support for me to go and get the agent deployed, so it can be

used by other people, like my family, or somebody else?

AMMAAR RESHI: Yeah, you can deploy your agent. So once...

right now they spin up in a sandbox But if you're ready to deploy, there's a whole documentation guide that should be available today that'll walk you through how to put that out there.

AUDIENCE MEMBER: Thank you.

PAIGE BAILEY: And the agents you build in AI Studio can be used anywhere.

AMMAAR RESHI: Anywhere.

PAIGE BAILEY: Not just within ai Studio, but also across other platforms. AUDIENCE MEMBER: Hi, my question is about choosing the tooling. So we have Antig

tooling. So we have Antig Antigravity, but also AI Studio, which allow us to create applications from the prompt.

So when I should use the AI system, I should use the Antigravity, and the same question goes with Gemma 4 and

Gemini models. If I have the

Gemini models. If I have the necessary computing power locally, should I always use the Gemma models? Or are Gemini

Gemma models? Or are Gemini models more capable of Gemma models?

PAIGE BAILEY: I think... so

it's a great question. And if

I'm understanding correctly, you're asking when should I prefer open source models versus the proprietary models, and what are the trade-offs between the two?

I do think that for areas where you don't want data to leave devices, or you want to be able to fine tune and customize a model, or you want to be able to have a model that's capable of

running on something like a mobile device, then Open Models make a make a lot of sense.

If you're looking to rapidly prototype, and you don't necessarily have those constraints, then, you know, starting with something like a Gemini API might be... might

make a lot more sense. But from

a cost perspective, that's also something to balance. If you're

curious about any of this, the best people to ask would be Olivier and the Gemma team later this week. And they have I

this week. And they have I think like a solid 15 slides about the trade-offs between Open Models, proprietary models, and then also tools that you can use like Cactus Compute that

help you automatically move from one to the other based on any evals that you might have defined or created.

So, excellent question. I think

we're going to see more of that longer term, where like, you know, if you can get 15 model calls calling a local model, and then boost up to five calling a

proprietary model, you could save a ton of money.

Awesome. Any other questions?

AUDIENCE MEMBER: Just one more question. So, the

more question. So, the question, before vLLM world, actually we used to live in they have a small model which are more cost effective, and running in your environment, and you are

not actually worried about the cost so much, actually, you can run those models also on the devices also.

So any plans in the longer run that actually... we can actually

that actually... we can actually develop our models our use cases that are more cost effective?

PAIGE BAILEY: I'm very sorry I couldn't hear your question.

AUDIENCE MEMBER: So, my question is about the cost effective, actually. Because I

effective, actually. Because I have seen that actually the models which are running, they actually use so many tokens

suddenly, and before LLM, we used to run models a lot, but they were not that much cost issues. So I have seen that

issues. So I have seen that before, there were small model for your use cases, like actually Deeplab, all these models they work really well,

actually, and then not much cost actually issues. Now, the LLM

actually issues. Now, the LLM generic models what has happened is that actually lots of tokens are being used. So any plans that o, okay, small models

actually we can use it like proprietary and actually can run it and they are much more cost effective?

PAIGE BAILEY: So is the question that you're asking will Open Models be able to use the agent harnesses? I'm sorry,

agent harnesses? I'm sorry, it's really hard to hear from the stage.

AUDIENCE MEMBER: The question is more actually how we can actually make sure that we are not burning too many tokens in short.

PAIGE BAILEY: Gotcha. So,

token, like learning how to manage context in your token...

and like learning how to manage tokens in your context window.

Huge area of focus, like being able to figure out which kind of documents to call, if you need to use something like Retrieval, needing to understand, like, if

somebody asks you to only look at a certain section of a video, you should probably use a tool to segment the video instead of putting the entire thing into the context window. That's

definitely an area where we're hoping to improve. And one of our colleagues, Philipp Schmid, has an excellent blog post around managing context in your model windows. If you're

model windows. If you're curious about it, can send a link to you later. But Philipp

Schmid, "Managing Context Efficiently," if you do a Google search for it, it should get you to his blog post. Yeah, great

question.

Awesome.

And I think we are right at time. But the show is not yet

time. But the show is not yet over. There are more talks left

over. There are more talks left today. So if you're, again,

today. So if you're, again, curious about AI Studio, curious about Antigravity, stick around Anshul and Joanna will be showing you all of the above.

And we really appreciate you.

Thank you for coming out. And

looking forward to seeing you tomorrow!

Thank you! !

AMMAAR RESHI: Thank you!

[ Music ]