Hey everyone, welcome back to the Asian

Factory, the podcast where we go beyond

the hype and dive into some of the more

technical AI details in the world of AI

right now. Today we have on the show

Paige Bailey and I'm super excited

because she's from the DeepMind team.

Paige, do you want to give a quick

intro?

>> Yeah, so greetings everyone. My name is

Paige. I'm the UTL for our developer

relations team at DeepMind. uh which

means that I get to work on the Gemini

APIs as well as things like Bio, Nano

Banana, Gemini, uh Gemma, uh all of the

things that we've released for AI

developers to use in their projects.

>> So, I know Google's been up to a lot

recently and last week Google just

launched a bunch of new tech. Do you

want to give us a quick rundown?

>> Yeah, sure. So, so I think I have a

slide actually to show kind of the uh

how much we've been shipping over the

course of the last while. Um the and

this is just going up to May 2025. So

obviously like shipping continues uh

shipping continues until models reach

AGI I suppose or ASI. Um but we've been

releasing a new model or a new feature

about every single day. Um sometimes

even more than one a day over the course

of the last year. Um the latest model

releases are really Gemini uh 3. So we

released our Gemini 3 Pro model which

everybody was very excited about last

Tuesday. Um and then also uh Nano Banana

Pro which is kind of the next iteration

of our NanoBanana series of models. Um

really really excited about both of

these. And then we also have a new IDE

for developers called anti-gravity which

is released by uh by some of the folks

that we have in deep mind who had uh

previously been working on the wind surf

team. So lots of goodness coming out for

every single part of the stack for

developers and for people who care about

gen media.

>> Yeah, I can't imagine how busy the deep

mind team has been up to trying to

release models. It seems like what you

like you said once a day. Um, and

today's November 24th, so if you're

listening to this after, of course,

these models and releases have come out

in the middle of November. Uh, and it's

been kind of taking the AI world by

storm a little bit. Uh, I remember last

week when Gemini 3 was announced, I was

really excited to try it out. And then

they launched anti-gravity. I saw the

blog post right after

>> and then the day after that was Nano

Banana Pro and it was a lot. It was it

almost felt like Christmas Day. So,

>> yep.

>> Yeah,

>> strong agree. And so, so I guess it

might be good to talk a little bit about

Gemini 3, what it is, was it, what it

isn't, um, and some of the things that

folks can can build with it. Uh, does

that sound good?

>> Yeah, I would love to hear.

>> Heck yeah. So, so Gemini 3 is really

intended to be the next iteration in our

Gemini series of models. Um, I I think

the I've certainly been excited about

it, especially for doing things like app

development, which we'll see in a

second. Um, and it's available in all of

the places. So, you can use it in the

Gemini APIs and the Vert.Ex AI APIs. Um,

places like Gemini CLI, Android Studio,

the Gemini app, AI mode, jewels, they've

all been kind of replatted to use this

third series of models. Um, and I I

think it's the natural evolution for for

Gemini at large. like the first Gemini.

Um, so I was I was on the original

modeling team for both Palm 2 and for

Gemini. And I think that model was

really all about being able to

understand a lot of different kinds of

content. So being unable to understand

video and images and audio and text and

code and all of the above all at once.

Um, Gemini 2 was really about thinking.

So we introduced this concept of

reasoning of being able to plan of being

able to kind of think step by step

through processes and have these

detailed thought traces and thinking

tokens. And then the third iteration of

Gemini is all about like getting much

much better at code and also being able

to act on things. Um uh so we've made

huge improvements to tools use function

calling um which are great for agent

style workloads but also being able to

incorporate things like tools into the

reasoning process. Um so you can have um

these kind of more composite

architectures of different models

working together, different models

interacting with systems to accomplish

tasks um as opposed to just one model

trying to do all of the work all by

itself. If that makes sense.

>> Yeah, this is really well explained. It

almost like you said sounds like the

natural evolution between 1, 2, and 3.

And I would imagine the Gemini 3 so good

at being able to do things like coding

because of the data set maybe that

you're training it on and its ability to

action on whatever it's been reasoning

on which a lot of coding is, you know,

basically what it is, right? Action,

taking actions, writing code, changing

functions, so on so forth.

>> Absolutely. And and to your point, like

there there are really two parts of the

model building process, right? there's

pre-training and then there's

post-training. Um, and pre-training is

all about giving Gemini as many many

tokens as it can possibly have access

to. So, so getting it uh like the

entirety of the internet, like we're

creating a lot of synthetic data sets um

that uh that the model can use to to

help accomplish its work. So things like

video game footage, um things like uh a

lot of synthetically generated code that

get then gets run and so you can kind of

get an idea that it's it's actually

doing something useful with descriptions

of what it does. um all of that kind of

getting put into post-training and then

um or or pre-training and then

post-training is really all about giving

um much more uh kind of u sort of

specific examples of kinds of use cases

that the user would be expected to take.

Um so having really really handcurated

examples of people using a whole bunch

of tools to accomplish a task. um lots

of multi-turn conversations uh that

include edits of websites um lots and

lots of those kinds of things getting

added into our into our post- training

mixture in addition to techniques like

reinforcement learning. Um so so I think

that uh there's a lot of really really

hard work that's been done across tens

of thousands of people across Google um

to get the models into a much better

state and it's it's been really

inspiring to see honestly.

>> Yeah. Interesting. So this pre and post

training process is almost like sending

Gemini to school to learn the basics and

then the post training is almost like on

the job experience where they get to

learn what the workflow looks like so

they can then replicate that. Very cool.

>> Yeah. And then the the other thing these

are just kind of the benchmarks um for

for the model itself. My favorite one

though is uh my favorite one is

something called vending bench um which

is vending bench for folks who are

uninitiated is gauging the model's

ability to run a passive business. Uh so

so uh h I I'll add a link to the vending

bench 2 website um so everybody can see

but but basically you're saying hey

model your job is you run this vending

machine you need to figure out what we

should stock it with you need to uh

whenever you realize that something's

about to go out of stock like you should

reorder it. Um you should deploy a

person to restock the machine whenever

the order comes in. And you should also

do long range planning to make sure that

as you're like handling all of these

component parts of the business um that

uh you know it's able to have as much

uptime as possible. So like I don't come

to the to the vending machine looking

for pretzels and find pretzels are

missing. Um but uh but I'm able to to

kind of have those expectations when the

when the vending machine is running. And

then the way that the models are gauged

in terms of performance is how much

money they make over the course of a

year. Um, and which which is awesome.

It's like the models ability to run a

passive business. Of course, you would

you would gauge it by how much money it

actually makes over time. Um, and right

now Gemini 3 Pro is coming in at around

$5462

per vending machine, which, you know,

like, um, given that we're operating on

this exponential trajectory, maybe we

can all retire and run passive

businesses at some point in the future.

But it's it's been really exciting to

see how the models have improved on

these on these things that require a lot

more strategy and a lot more decision

making, especially long-term decision

making um for these kinds of use cases.

>> Yeah, I just sent you an invite for a

meeting after this uh for a vending

machine entrepreneurial sync you and I

can

>> Excellent. I I really so so like like I

really really want uh ju just kind of

like as a segue I really really want

somebody to create an eval for uh like a

laundromat bench

>> um because I I do feel like these these

kinds of passive businesses

um have immense potential and a lot of

people who are who have built um AI

startups AI companies just recently they

have a few operating just kind of more

or less autonomously they won't have to

step in very often. Um, but there's

still like pretty significant revenue

streams every single month. It might not

be a billion dollars, but it's certainly

enough to uh to kind of meet the needs

for users and to to kind of take care of

take care of some of these tasks.

>> Yeah, it really allows people to bridge

the gap, I find, and unsolicited Google

non-certified Google advice, but I think

laundromats and vending machines tend to

be some of the most secure and good

entrepreneurial beginning ventures, I

guess. So,

>> absolutely. Maybe we should get a

vending machine. That would be kind of

fun.

>> Wait, put it full of Google swag.

>> Cool. And then the uh so additional

additional kind of benchmark things.

It's the first model to break 1500 on

Elmarina. Right now we come in at around

151. It's also doing really really good

at webdev arena. Um, so the new feature

that you might have seen in Replet,

that's their design feature uh to build

beautiful UIs is powered by Gemini 3.

Um, and you can really really see a

difference in the quality of the apps

that get generated with build and AI

studio. Um, and then it's also kind of

state-of-the-art at reasoning and

multimodality at tools use and planning.

Um, if you're a Minecraft enthusiast,

it's much better at these kind of voxal

art experiences. So you can see the

difference between Gemini 2.5 Pro and

Gemini 3 Pro um for for uh these these

kinds of voxal art um demonstrations.

Um really really good at making games

including games with interesting game

mechanics and kind of beautiful

beautiful um beautiful UI.

Um and it feels like so it feels like

we've been talking about it for a while.

Would you be down for me to show how to

use it in in AI Studio as part of our

build feature?

>> Let's build.

>> Excellent. Let's do it. And I I think

both of us, we were having a

conversation at the very beginning like

both of us less of a fan of slides and

more of a fan of like showing stuff in

action. So like let's show some stuff in

action. I also like one last thing

though before I pop over. Um it's

currently available in the API. So you

just have to change out the model doc

string. And then also if you've been

using uh if you've been using OpenAI as

part of your projects um we have

something called the OpenAI

compatibility layer um which allows you

to swap out a few lines of code. So this

API key and base URL um you would just

modify it to be your Gemini API key and

you would swap out the model name. Um

but then you would be able to run it

with your OpenAI uh your OpenAI code

hopefully without any changes. Okay,

cool. I've uh taken up enough of your

time, but I want to take up even more

now. Super excited for building.

>> Cool, cool, cool. Let's let's do a

thing. So, let me zoom over to AI

Studio. Um this is the build feature in

AI Studio. If you expand out this

section on the left, you can see that

there there's a lefth hand pane. Um you

can select the build feature. Um, and

then it immediately puts you into kind

of this build uh this build environment.

I'm going to collapse this so there's a

little bit more real estate for us to

see what's going on. Um, and then

basically you can just describe anything

you would like to create. Um, you can

also see uh kind of an app gallery with

remixable uh remixable apps, things that

the team has created, including a comic

book style experience, a product mockup

visualization. Um, lots of games and

immersive 3D worlds, including like

simulation games, which are pretty cool.

Um, and you can fork these, remix these,

uh, share them with your friends. Um,

we'll also show how to uh how to share

them with Google Cloud um in just a

second. Um, and then also if you can't

think of something to build, if you

click I'm feeling lucky, it gives you

kind of a out ofthe-box ready-made app

idea that you can incorporate into your

projects. Um, but what I've been trying

to do is I've been trying to like build

um increasingly complex apps uh with the

with the goal of trying to make it so

that Gemini can't build it and it has to

to kind of fix itself along the way. Um

uh because we do have an autofix uh an

autofix step within AI Studio. It's just

that it's triggering much less

frequently now because the models are

getting so good. Um, so I'm going to I'm

going to try to see if we can build

something that um builds something that

is sufficiently complex that Gemini will

have to fix itself. So let's let's try

this out. Um

uh cool. Create an app that uses the

webcam and the microphone

um to have a conversation with a user.

Um the app should uh the app is intended

to be kind of a um what it should be is

kind of like an insurance

um an insurance app. So the user will

show them uh the user will show a bunch

of different objects. The app will uh

sort of catalog all of the objects, like

describe them, put down their name, um

do an assessment of what kind of wear

the object might have, and then also use

Google search grounding to try to ground

um with how much that uh that item might

be worth. The app should put that in a

table format um so the user can see all

of the all of the objects that they've

been cataloging um through the duration

of the video conversation. Um, uh, the

app should also be really well-designed,

should show some stats about all of the

objects,

um, uh, that the user that the user has

put down for insurance purposes. Um, and

the app should also have,

um, the app should also have like a

Nordic theme, so it should feel like

very IKEA vibes. Um,

and uh, for the video conversation, the

app should be using using Gemini Live

along the way.

So, so that was just like spitballing a

whole bunch of information. Um, I'm I'm

not sure uh I'm not sure how well it'll

do, but let's try.

>> Um, yeah. And then, uh, so what's

happening here is you immediately get

put into something that looks like an

IDE. Um, if I expand out this thinking

section, you can see that Gemini is

figuring out what it should be doing.

It's figuring out how it should be um,

doing function calls. I said the Gemini

live API and it's figuring out what

model name that should be, figuring out

what the structure of the data should

be, if it's supposed to be cataloging

all of these items. You also see uh,

kind of some recommendations about what

to do with the app. here in the center.

Um there's a place up in the right where

you can save to a public or a private

repo on GitHub. Um and then there's also

a place where you can add files if you

would like. Um add files. So add things

like images or text files or PDFs. Um

and as Gemini is working, you can see

all of the code getting getting created

along the way as well as the prompts

that are created to give to each one of

the each one of the models. Um the I I

love asking Gemini to write the prompts

for Gemini just because usually it does

a much much better job than I can do. Um

all of the files are getting put into

this really nice directory structure

here in the center. Um and the nice

thing too is that since this experience

is baked into AI Studio, it has insight

into the latest models, the latest

features, um what's possible with the

APIs. Um right now it's generating a

React app. You could have also specified

Angular. Um, it's using React Native, so

you can access all of these apps via

mobile phones as well. Um, uh, but this

is, uh, this is really really cool to to

kind of see and to just have an idea,

um, and to figure out, uh, to figure out

how to get it get it started without you

necessarily having to wade through

documentation or figuring out which

models to use for which experiences. I

also quite Yeah, go for it.

>> No, I was just going to say so you threw

your messy thoughts into this thing and

obviously it's very impressive that it's

able to not only format but format your

thoughts into something that you know is

a bit more coherent but also write the

code. I think the most impressive part

is the name that it came up with because

I know that you didn't I know that you

didn't specify that.

>> Yeah. Nordic Nordic Shield. So So what

we we see an app uh we see an app. So,

I'm going to try to see um how well how

well this does just out of the box. Um

hey there.

Let's see.

>> So, yeah, that's that little autofix

button, I guess, probably.

>> Yeah. So, so it looks like uh just

taking a look at the app itself.

Um, it t taking a look at the app

itself, it was able to hook in the

webcam.

>> Um, it even gave me the options of

turning on or turning off my video and

my camera or uh my camera and my audio

rather. Um, it created the inventory

list which is pretty cool. Um, it has

some condition breakdowns, shows uh

shows connections. Um, but then it

didn't pick up on my audio which often

happens whenever you have multiple audio

sessions simultaneously. Um, so I think

what it's doing is it's figuring out

where um it would uh like which section

it would need to uh need to figure out

to attend to. Um, and then that would

trigger the Gemini live call. And so

there's a big reasoning trace on the

left um which shows kind of like all

right, is the mic on? Is it enabled? How

do I toggle the mic? How do I understand

the media stream? Um, and then it should

be able to understand that as part of

the website as well. Um, yeah, I also

like I I feel like it's it's a much

better design than I would have been

able to come up with myself. I also

really like that it automatically

included instructions. Apparently, this

person that I or the the entity that I'm

talking to will be called Sven, which is

also very Nordic. Um, and uh and we'll

see we'll see how well it's able to make

the changes. I mean you you did ask for

IKEA. So

>> this is this is true you know like it's

the and and I think the one of the nice

things too is um you know I started

doing machine learning around 2009 so a

really long time ago. Um but I've always

been more of like a Python human than a

JavaScript or TypeScript human. I I

built um I worked on part of the VS Code

team for a while. So I I've built like

VS Code extensions in Typescript. Like

I've built small websites. Um but I

would not call myself a TypeScript like

expert by any stretch of the

imagination. So having this help um

figure out where the problems might be,

where to debug it, how to debug it, like

what libraries to include to be part of

a modern stack is really really helpful

just for me personally. Yeah, I was

about to say like it's it's quite

impressive that it not only recognized

that there was an error,

>> it parsed that error, made a reasoning

state for you and and allowed you to

kind of have some level of

interpretability into its thought

process, which I think is like so

important because obviously if it's

doing something that's kind of not what

you intended, you can easily like pull

the plug if you will, which is doesn't

seem like it's the case here, but you

never know, right?

>> 100%. I also really really like if you

zoom in to the bottom um you can see

like warnings uh insights into the build

process. So rendering start rendering

end um uh recommendations on what to do

to change like the ergonomics and the

accessibility of the of the website

itself. So like hey this formatting

might not show up as well on like a

mobile device. Um and then our design

team has also implemented a feature

called annotate. So if you click

annotate app, what you can do is you can

circle features, you can add comments.

So say things like um I don't uh um make

this cataloged items

uh section more

attractive or similar. Um, and it's able

to to kind of take the comments and make

changes and modifications to the

website. Um, the same way as you would

work with a designer to kind of do red

lining against the uh against the UI

itself.

>> Um,

>> this is really nice too, I think,

because like it looks like it's still

working on trying to fix the bug while

you're able to annotate. So, you can

almost do two things at once. have it

like work

>> um and then also have it like um you

know annotate your designs or review

your designs.

>> Yeah. And I love I love also like there

there are a couple of different agents

that we have or I suppose a few

different agents that we have um as part

of our our Google AI stacks right now.

So we have Jules which is an

asynchronous agent, Gemini CLI which is

kind of a synchronous agent that you can

operate with uh within a terminal. Um,

but it's really really nice to have, you

know, a team member go and attempt to

accomplish a task and debug it along the

way. Um, so let's see. It looks like

it's made some changes. I'm going to go

ahead and try again.

Hey there.

Hello there. I am Sven, your insurance

cataloging assistant. Please show me the

items one by one and I will help you

catalog them.

>> What about this?

I see a dark colored smartphone in a

case with a ring grip. It looks to be in

good condition. Would you agree with

this assessment? Sure. It's a Pixel 7,

but you might also see that it has like

some scuffing and like a cracked screen

if I turn it around.

>> Ah, thank you for clarifying. I see the

scuffing and cracked screen now when you

turn it. Based on this, I would say it

has fair condition. Do you agree? Great.

I have cataloged the Pixel 7 smartphone.

Is there anything else?

>> What about this?

>> I see a Diet Dr. Pepper can. It looks

like it is in new condition. Shall I add

this to your catalog?

>> Well, how much do you think it would be

worth?

>> I can't estimate value myself. I can

only catalog the item with its name,

description, and condition. Would you

like to add the new Diet Dr. Pepper Can?

The Diet Dr.

>> Let's do it. And then also, how about uh

how about this computer screen that you

see in the back?

>> I see a large screen on the wall with

some images and text. It looks to be in

good condition. Would you agree?

Understood. I have cataloged the

computer screen. Is there anything else

you would like to catalog today?

>> Nope, that's it.

>> Thank you for using Nordic Shields

insurance cataloging service. Have a

good day.

>> Cool. And so you can see uh let me zoom

out a little bit so folks on the call

can see as well. Um it looks like there

were two phases. So as we were talking

about these different uh these different

items um the uh live API feature showed

them on the inventory list. So it gave

the item name, the description, the

condition, the estimated value. Um but

uh as part of the estimated value step

um there was a second kind of follow-up

action uh that did a search around

Google to pull in like what would the

estimated value of this thing be? Um so

so the live API uh the live API kind of

assistant was just responsible for

cataloging and then after she added uh

an item to the table to the inventory

list that kicked off a second step. So

almost like an agent to go search um

round up uh a bunch of different items

and then give us an estimate for how

much these things um how much these

things might be worth which is pretty

cool. Um yes. So it was uh and it also

added links um like citations for those

different uh for those different uh uh

items as well, which is pretty cool.

>> Yeah.

>> Yeah. I uh I think like this is probably

one of the the the nice new things about

Gemini 3 or some of the APIs now that

allow for these types of models to

always be grounded in in search, right?

Um, I think something like this would be

helpful for like decluttering my space.

I sell a lot of stuff on Facebook

Marketplace. So,

>> yeah. Excellent. So, so like and I I

think it's it's also nice to help with

um kind of detailed descriptions of all

of the all of the different items. Um,

and then if you wanted to deploy this

app uh via Google Cloud, you could

select a different cloud project, um,

click deploy and it would give you a

unique URL with logging um, and and kind

of insights uh, into all of the the

different features that you might be

including um, as part of your project.

So, if you uh, like as an example, I'm

going to show my personal Google Cloud

um, my personal Google Cloud account.

There we go. Excellent. So, so this is

uh this is the logging feature in Google

AI Studio where you can see um uh if you

turn on logging all of the different

kind of JSON blobs that come in from

your your app interactions. You can see

things like usage and billing. So, the

number of API errors and requests that

you get per day. Um as well as kind of

any rate limits that you might be about

to hit. So, it looks like I've been

using VO3

um uh past my rate limit uh past my rate

limit options for at least the tier that

I'm on. Um and then you can also see the

different projects. Um and right now I

have around 85 keys created for

generative language client. Um but it's

it's cool to to be able to have a

one-stop shop where you can see all of

that usage, which is quite nice. And

just to give Nano Banana its time in the

spotlight and to show a couple of

examples that people have built with it,

um you can do everything from kind of

use a Pinterest style collage of items,

um compose them into a single image. Um

you can do things like build uh

orthographic blueprints for real life

places. So in this case, a castle um

being able to see No Feinstein castle

from different uh from different views

um with high reliability for all of the

text. Um you can create detailed physics

explainers and all of this is grounded

using tools including things like

grounding with Google search which is

pretty amazing. You can also change the

resolution, so 1K, 2K, or 4K. And then

also the aspect ratios as you use uh

these generative media models as part of

your projects. So, it's been really

exciting to see what people create.

Everything from uh everything from kind

of re-imagined game experiences to like

detailed posters and slides for their

papers and their projects at Nurips.

>> Yeah. Yeah. And I think a lot of the

creativity and power, I guess, gets

unlocked when you even combine all of

these things together. So, you can ask

Gemini to go and do research for you on,

I don't know, skiing areas in across

North America, across a train line or

something like that. Uh, pull all that

information, create an image, and then,

you know what I've seen people do? take

that image, put it into VO3.1 and they

get an animated explainer which then

they can use as a part of this slide or

presentation or something like that

which is very cool.

>> 100%. When you combine reasoning with

multimodal outputs like the the

possibilities are really endless and

it's so much easier to understand things

like images or video um I think than to

to have people kind of comb through just

raw text or raw code. Um so it's been

really exciting to see.

>> Wow. Thanks so much for showing all of

this to us, Paige. It's been really

insightful. I want to switch gears a

little bit and I want to show you one of

the projects that I've been working on

with Gemini 3. Maybe you can critique me

and let me know what I'm doing wrong or

what I'm doing right and what you think

about it. This is my website that I'm

using to share resources and articles

with people. As I've been making more

short form video content online through

Tik Tok, Instagram, YouTube Shorts, and

such, it's been kind of challenging to

find a way to share some of the

resources that I talk about with people

who are viewing my content. And so, you

can see here that I have a lot of URLs

that I often reference, including Nano

Banana Pro, Gemini 3, and so on. Uh, but

I don't know or have a good way of

sharing it. So, I created this website.

I threw these ideas that I had for a

more like doodly vibrant pastile neo

brutalistic, you know, notebookesque

theme into AI Studio and it created this

for me. It's a nice place I can use to

like manage all of my content, create

new assets for me, save it, and then

import it into my videos. The problem is

is that my previous website obviously

doesn't look anything like this. So I

thought maybe what I could do is combine

Google's new IDE anti-gravity and Gemini

3 to um inherently take advantage of

that multimodality that we spoke about

earlier. Take a screenshot of this uh

new website, feed it into the code for

my old website and see what it comes up

with. What do you think?

>> Yep, let's do it.

>> Okay, so Paige, here we are in

anti-gravity, which is meant to be an AI

native IDE from Google, and it has

Gemini 3 Pro baked right into it. So,

what I'll probably do is take those two

screenshots that I was um that I was

talking about before and feed them into

here and then just ask if if

Anti-gravity or Gemini 3 Pro can come up

with a way of redesigning my site to

adhere to this a little bit better. So,

I'm going to grab those two screenshots.

So, I'll ask it recreate this site to

adhere to my new design philosophy,

which you can see in the included

images. So, go ahead and run that. And

while that's running, I think it's worth

noting that obviously Gemini 3 Pro's big

job here is to not only interpret what's

inside of the image, but kind of break

it down into the most notable features

that it needs to then go and apply to

coding. within anti-gravity. So while

it's doing that, it obviously has

explored a little bit already and it's

created this task list here. So explore

the existing codebase and a little bit

more. It'll follow that task list to

continue ensuring that it stays grounded

and true to that kind of process of

development as time goes on. Uh so the

next thing it'll do is obviously analyze

the new design and then set up the the

new design system. Yeah. And so the next

thing I came up with here is this

implementation plan which I also find to

be really nice because through this

process it's creating a bit of

documentation which is nice because down

the road when the codebase becomes

larger and the AI has to continue

sifting through more and more code uh it

can continue to reference through things

it's done in the past. So you can see

this implementation plan it has these

actions like new modify and so on. Uh

and it keeps it from you know um not

going off the rails as much.

>> Awesome. Paige, have you created

anything cool with Nano Banana Pro?

>> Oh, yeah. The uh So, so with Nano Banana

Pro, I've been experimenting quite a bit

with different uh with different

outputs. Let me show

um let me show one from a little while

ago. The quality is just absurd.

Um, as a as an example,

I I created a bunch of uh let me share

screen really quick.

Um,

>> oh, nice.

>> I created a bunch of images with uh

popular like physicists and computer

science figures like Claude Shannon,

>> Richard Fineman, and Carl Sean. And then

also one of our winning hackathon

projects this past weekend was using was

using Nano Banana Pro for its game asset

generation. So I'm not

>> Yeah. So they used it for every single

one of the game assets as well as the

design patterns

>> for the games themselves which was quite

cool. This was at the AI engineering

world's fair. And how is your how is

your website coming along?

>> So it looks like we're done here. What

I'll probably do is pop open my terminal

and then run the app and then we can

take a look at what it looks like.

>> Yep.

>> Cool. Yeah. So, Paige, this is what it

came up with. You can see here that it

actually followed the brand guidelines

pretty well. It has these pills that

kind of jiggle when you hover over them

and it has like some more information

about my resources, prompt library,

agent blueprints, and things like that.

So, I found it did a pretty pretty good

job just like with one prompt. What do

you think?

>> Oh, it looks awesome. Thank you, Gemini

3 and anti-gravity. I can't wait to

continue playing around with this tool.

>> Very, very cool. It's I I'm really I I

love the the kind of attention to detail

and the creativity that it puts into

these UI designs. I feel like it's

better than ever at at kind of creating

these these kinds of experiences.

>> Paige, it was a pleasure having you on

the podcast today. Thanks so much for

coming and explaining a little bit more

about Gemini 3 and all of the great

things that Google launched last week.

>> Absolutely. Thank you so much for having

us. And I loved that we were able to

talk about not only just the models

themselves, but also how Google's entire

AI stack from the hardware to the

compiler to the machine learning

framework to our endto-end app

deployments. Um, even things like

anti-gravity and AI Studio and all of

the other surfaces that we have around

Google help make sure that we're

building the best possible models for

all of these different use cases. It was

really fun

>> and not to mention all the cool things

you can do and deploy them to Google

Cloud and have them ready in production.

So, thanks so much again. I'm EMTT

Mirage

>> and I'm Paige Bailey.

>> Powering down.