Google Cloud's CTO on their agentic AI plans | TechCrunch Disrupt 2025

[Music]

Nice moves.

>> Oh, thank you. Um,

yeah. So, this was I think when we

started talking about this panel, I

didn't even realize how much work you've

been doing on the Agentic Cloud. I think

it's sort of if it works, nothing

breaks. You don't hear about it. Um, but

so I guess let's start with that. You

were talking about the building blocks.

I think if people aren't building

>> cloud-based agents, they may not know

about everything that's out there to

build on. So, so what are the building

blocks that people in the audience could

use right now?

>> Sure. Sure. So, there's two paths. If

you want to go code first and uh you

know you want to really get into the

nuts and bolts of building um agents uh

you can take a path through like agent

development kit and more fine grain

control. This will give you like the

observability, the debugging, the u you

know tool management, tool creation, you

know, kind of all of those things uh

bound together in a nice little package

to help you get started if you're you're

really after like refining and building

agents themselves for the sake of

getting value from those uh first coded

agents. If you're more of like I want to

build an app and the agent flow is like

how I want to build an app, then we have

a completely you know separate track

which is Gemini Enterprise. And with

Gemini Enterprise, you can basically

just go into UI and you're like, "Hey, I

want to uh look ahead my calendar this

week and I want to prep for any

meeting." This is obviously one that I

use all the time and I want to prep for

any meetings that I have. And because it

has connectors to your calendar, you

know, built in because it has connectors

to thirdparty data sources already built

in like your CRM, you don't have to fuss

with all like the data engineering and

the data management. You can just define

an objective for your agent. You have

got the connectors built in. the agent

now has context. It'll run ahead. It'll

look at your calendar and it will uh and

it it will actually propose like, hey,

based on who you're meeting with and the

nature of those meetings. Here's some

ideas also for research. Uh here's some

ideas for background that uh may be

helpful to you in those meetings. So, uh

really depends, you know, what you're

building. You know, code first and agent

kind of optimization ADK uh and you and

that takes you down a path of also

orchestrating multiple agents. And we'll

get into I think probably agent, you

know, commerce as well. If you're trying

to build apps, you know, go to Gemini

Enterprise and take the easy button.

>> Yeah. Well, so this is commerce. I'm

going to I'm going to pick that right

up.

I got really interested in this because

in the last few months you've rolled out

this new payments protocol that

basically is allowing agents to maybe

just put things in your cart for you to

buy or maybe actually buy things for

you. Really handing over the credit card

to the to the intern so to speak. I

guess h how do you build up to that?

>> Yeah. So it really uh depends on your

use case but there are there are kind of

two main um paths for using this agent

payments protocol and again you know

we'll we'll think about this in like

terms of a stack. So first we had to you

know get agents up and running and give

you the tools get agents up and running

quickly. So that's the ADK. Then on top

of that you know a lot of these

workflows are going to be multiple

agents. So we have the agentto agent

protocol. And then like once you're

transacting and you're interacting

across multiple agents you want to

exchange value. And so that's the agent

payment protocol. And there are kind of

two main use cases. One is I'm going to

give you some parameters and I want you

to go and like if if you're a sneaker

head, you know, and you always miss out

because the bots are getting you, you

know, because they're getting to the

sneaker sales before you're able to get

to the sneaker sales, you can define an

agent that says, you know, I I want this

type of sneaker and kind of these cost

parameters, these features, and when

it's available, I want you to let me

know. And that's what we call a mandate.

you give this agent a mandate and it'll

go out and it'll be uh monitoring the

web and once that condition those

guardrails are met it'll come and prompt

you and then you have the choice to go

and put something in the card and to

finish the transaction. There's another

one which is you know what uh I trust

the agent to go and kind of curate all

these things to put it in the cart and

all I really want to do is finish the

transaction and so this has kind of two

levels of mandate there's like this

initial you know set of guardrails and

then there's a trust mandate and uh

think of like travel you know going you

want to go somewhere you want to have uh

you know a bunch of options put together

for uh airfare hotel experiences you're

going to need multiple agents going out

bringing those things back putting them

into the cart and saying like, "Okay,

it's time to check out." And then the

human can just check out. One really

important thing about this um as a

technologist is that because we're in

the early days of agents, um agent

payment protocol is is like a whisper of

the future, which is how do you automate

the exchange of value? And this is

relevant to, you know, if you're uh a

startup and you're trying to advertise

your inventory, you know, the future is

you're going to be advertising inventory

to a bunch of agents and you want to get

those transactions done in as much an

automated fashion as quickly as

possible. And actually like human

slowness and human weirdness about, you

know, like they'll go and scan your

website, they'll look at your inventory,

they'll leave it, they'll abandon the

cart, they'll do all this stuff. You

know, in a way, we're also building, you

know, one this kind of new commerce of

the future and more of a consumer

version, but also in a B2B scenario, you

can imagine if you're like a large

shipping company and you get millions of

invoices every day, you want to be able

to complete those transactions in the

background, you know, receive invoices,

pay valid invoices, flag ones that might

be, you know, a little bit out of band,

um, and do that in a highly scaled and

automated fashion. So, you know, expect

to see more of these multi-agent

automated, you know, real-time

fulfillment, but we had, you know, you

had to start somewhere. So, you have to

start with the foundations just like the

web, you have to start with the

protocols and the rules of engagement

for how you uh transact.

>> Yeah. So, I'm glad you, you know, I

think this idea of the transition to a

commerce is so interesting and so kind

of fraught. I want to pick on one thing

you said. You were sort of saying, "Oh,

okay. If you're a sneaker head and those

terrible bots keep snapping up the early

sneakers,

but like what you've built here is a bot

for buying sneakers.

>> Yes. Yes. But it is a precision bot that

you control and now you are on a level

playing field with these aggregators

that would go and take all this

inventory and then they'd resell it and

they jack the prices up. So in a way

we're also you know fine-tuning these

agent flows so that you can get more

value out of you know a workflow and as

an individual consumer or as a business

maybe you don't have to pay those

intermediaries anymore. Maybe now you

can operate with like the power that

only you know these kind of these

aggregators had in the past.

>> Yeah. I just I I think it kind of gets

to the social complexity of it where as

as we're in this world where maybe some

people are using agents to buy things

and other people are buying things on

the internet. Maybe some people are

still buying things in stores. Like how

do you think about the perspective of

the store where they say, you know, are

they going to be glad when the agentic

buyers show up

>> eventually? Uh

>> eventually. Yeah.

>> And the reason I say that is because uh

if how you know how many of us have been

designing against like robots.ext text

and like these you know automated bots

that come into your uh website or into

your business. Uh and in the past we

might call these malware or we might

call these you know cyber security

incidents. So if you're the CISO or if

you're the head of security for any

organization of a reasonable size right

now you have really had to flip your

thinking from you know the the bots are

bad to the bots are your customer. So

that's why I say eventually because you

know there right now and it's kind of

like this AI play counterplay. So now

the CISOs and there's a huge opportunity

in cyber security where if you can build

tools that can reason and and

rationalize this kind of inbound swarm

or flood of agents into your

organization into your website into your

uh commerce platforms then you're going

to be able to take advantage of this

kind of early mover. And if you aren't

designing AI systems to kind of reason

about these new behaviors, uh, then

you're going to get left behind. And as

a again, you know, as a technologist,

one of the things I always think about

is how do I get learning loops as

quickly as possible. And so a lot of the

leading edge companies right now,

they're, you know, they're creating

perimeters or they're creating kind of

sandboxes. They're creating ways that

they can characterize the flood of these

agents coming in and they can learn how

to instrument against it faster. So it's

kind of like you have to let them in a

little bit so you can get the learning

curve. Um but a lot of that's being

designed right now.

>> Yeah. Yeah. So one of the alternate

visions of this the other you know major

uh agentic commerce idea is of course

OpenAI's instant checkout where we're

seeing these

sort of company by company launching

with Stripe. Okay. Now we're partnering

with PayPal so PayPal can do it too.

It's not the open protocol approach, but

we're kind of back in the classic

technologist struggle of okay, do we

want the walled garden? Do we want the

open protocol? I guess h how do you

think about the consequences of that?

>> Uh I bet on open. I mean, if you look at

uh the foundations of the web, if you

look at the foundations of commerce

today, underneath all of them are

standards and protocols that were built

in communities. And that goes from you

know how packets transit a network to

how you trust that a site is exactly the

site that it should be um to authorities

around security uh starting with open is

absolutely critical and that's act and

that's why like in ADK that I mentioned

earlier you can bring models other than

Gemini you know into ADK and build

agents using other people's models. uh

this has been an approach that we have

taken and you know I've been at Google

cloud now for 11 or 12 years and you

know since the very beginning it was

kubernetes it was tensorflow it was open

standards it's how do we give you as

builders an option where if you want to

compose and manage yourself you know you

have these open versions and you have

lots of choice if you want to take

advantage of some of the managed bits

you know we can do that for you as well

and so uh right now like scaled use of

AI infrastructure it turns out that uh

customers and businesses and builders

don't want to pay reseller fees to get

AI models and AI uh infrastructure these

days. So they like coming to a place

like Google Cloud where we have our own

firstparty models, we have our own

firstparty AI infrastructure and we're

not just reselling somebody else's

stuff. And we also learn how to optimize

because that's true because we have

these things. Uh we learn to optimize.

So if you want scaled AI infrastructure,

we can give you managed scaled tensor

processing units through Google

Kubernetes Engine GKE in a managed way.

Or if you want to roll your own like

abstractions and you want to manage them

yourself, you know, like uh Anthropic

and others, customers have been learning

how to build and go from GPUs to TPUs

along the way, too.

>> Yeah. Please be careful rolling your own

abstractions, though. That's

>> Yeah, I would agree with that. You're in

for a p in for a pound, I guess, is uh

is way to think about it.

>> So, I mean, we're describing a kind of

world of agent commerce that is not

really here yet, but I guess how do you

how fast do you see this coming down the

road?

Well, I mean, I work with a lot of uh

startups, a lot of uh enterprises, and I

can tell you that the first iterations

of these agentic workflows, generally

speaking, it's kind of small uhish like

numbers of agents. So, I would say like,

you know, the swarms of hundreds or

thousands aren't here yet, but in the

tens to, you know, getting to hundreds,

the orchestration of them is a really uh

important thing to think about. Um

but there are quite a few deployments of

um AI agents right now. One for example

uh Sully AI which is a startup that

provides um in healthcare basically like

healthcare task automation agents as

like a service to healthcare workers. So

like there are a lot of administrative

tasks you take care of every single day

in healthcare and Sully AI has basically

created agents that allow them you know

using Google cloud that allow healthcare

workers to take a lot of that

administrative burden and minimize it so

they can focus on patient care. Um so

they are out in the real world and they

are you know creating value today. I

think it's more of the multi- aent

interactions, you know, complex

multi-agent orchestration that is the

frontier.

>> And so, do we think when are we going to

start to see that? I mean, is that like

>> five years out? Is it five months out?

>> You're ask you're asking a technologist

and a CTO to give a roadmap timeline

with assurityity in front of a large

crowd. So,

>> yeah,

>> it's not my first rodeo, Russell. Um,

>> fair. But what I will say is I think you

can bet on like waves of innovation. So

the the first wave which is just b um

building singular task agents to

accomplish singular function is here now

and it's growing rapidly. So you could

also kind of this is semi close to RPA

and kind of the first you know

iterations of uh you know automation but

we're rapidly getting to uh multi- aent

uh orchestrated workflows within I'd say

like one use case. So for example,

marketing uh generating uh creative

right now you can uh use you know Gemini

as a front end and you can say hey I I

really have an idea for this interesting

creative thing and using you know

imagine under the hood or nano banana

under the hood it'll give you a first

order like this is what it could look

like and then you can send it to VO and

it can uh turn it into a short form

video but then you have other things

that you need like brand guidelines and

you need you know AI as a judge to say

like Is this content even something I

would put out to my users? And so in

those workflows, a lot of the

innovations happening right now is like

the AI being capable of evaluating the

outputs. So that um uh in my team we

talk about like if you're a manager now,

if if you're in the in the flow of of

any work, basically AI has made your

life a nightmare because uh it's

stacking up all of these things that

need validation or evaluation. And as a

human, you know, I think of it like

folders on a desk. That's how old I am.

So, you know, when you were a manager

back in the day, it's like you have a

desk and people bring you their stuff

and it's in a folder and you have to

like approve it. And in and in, you

know, the old days, you know, it was

gated by what humans could put on your

desk, how fast they could get the work

done. So, it's like, okay, cool. Like

every single day, I get like 20 to 100

folders. I burn down through as many as

I can. By the end of the week, as long

as it's down to zero folders or a low

number, it's great. Problem is now is

that AI can create like infinite folders

almost instantly. So one of the big

unlocks for the you know in the shorter

run is going to be AI as a judge AI

evals grounding in like brand guidelines

or grounding in enterprise data that

allows the outputs of these things to be

contextual to be useful and to be things

that companies are proud of and can

actually use. Um whereas like I think

these big agent swarms I think there's a

lot of fragility right now in for

example um like in financial services

say you want to build uh I want to build

the the um like financial adviser bot

well there's no financial adviser bot

because financial advising is a series

of tasks there's like 10 to 14 tasks are

involved in that and trying to get 10 or

14 agents to reconcile their

perspectives and synthesize that into

like one output that can be delivered to

a human today is a very complex process.

and is very much on the frontier and I

would say there's still room to go.

>> Okay. Yeah. This idea of like stacking

folders is so visceral but it's also I

mean right now we're still trying to

evaluate I mean AI evals I when I hear

it I think oh you're talking about

benchmarking and humans evaluating AI

and it's like no we need AI systems that

can evaluate AI outputs absolutely and

start clearing the folders out. we need

to move at the same p. The thing is that

like human cognition and human, you

know, like uh evaluation runs at human

speed, but like AI can generate content

way faster than humans are able to go

like that's good, that's bad. And so you

have to that is like one of the key

problems to solve in almost any workflow

is how fast the decisions how fast the

content how fast the analysis the

synthesis the reports depending on which

use case you're in how fast you can get

AI to kind of sort through take out the

stuff that's obviously bad uh and really

queue up the human for the highest value

like Eval's most complex edge cases uh

you know that's super critical.

>> Yeah. So that gets to this question that

I think keeps coming up when we talk

about these obviously AI agents are

making huge technical progress.

I wonder if sort of are they there yet?

Is is the missing piece really kind of

more capabilities from the underlying

agentic tech itself?

>> I think it's um it's multiple things. So

for example uh you know Gemini today you

know leading frontier model it's been

that way for you know kind of six months

plus which in this world feels like an

eternity and that has to continue. So we

definitely need innovation we need more

like reasoning capabilities we need more

um kind of navigating and and kind of

like ambient capabilities inside the

front end or like a frontier model like

Gemini. So you need that but you also

need grounding tools. uh you need the

ability to bring in enterprise data more

easily. You know, some of this is going

to be through connectors and you know

creating tools and kind of like the old

school publish and subscribe. If

anybody's been around for a while like I

have, you know, this is concept of like

publish and subscribe where you have

data sources and they publish their

availability and then you have consumers

that subscribe to topics. We're back

again. You know, we just use different

words. And so now, you know, there's

also a need for companies and

organizations if they haven't

externalized their data through an API

or through a tool, you know, they're at

risk of like getting left behind or

becoming irrelevant in a world like

this. So, you need the, you know, you

need the connectors, you need the tools

created, you need this kind of rich

ecosystem of data sources available and

APIs, you need API management, you know,

you need a way to sort through and

navigate against all of these different

um options. So, and that's just like

three layers of the stack. If we had,

you know, another like three days, we

could go through kind of all the layers

of it because there's also a lot of

bindings in between that make these

workflows useful like in the real world.

But you have to you can't just choose

one area like as as Google Cloud like we

can't just choose one area to innovate.

We have to provide and that's why we do

a complete end-to-end AI stack that is

both first party but also has uh you

know is built kind of with open in mind.

So, you know, we have model garden, we

have, you know, agent gardens, we have

these places where if what we have isn't

at that point in time exactly what you

need, you can go get it from somebody

else. Those are the table stakes to be,

you know, on the frontier of technology

today.

>> Yeah. Well, so I want to ask about one

specific layer that I think gets

overlooked a lot, but these are

incredibly powerful tools. Not everyone

on the internet is as nice and friendly

as we are. I guess I hear about, you

know, generating these endless folders,

doing all of this work at scale,

particularly once payments are in the

mix. Yeah. And I worry about the

security aspect of do do we need better

tools for identifying bad actors, for

validating agents so that we can sort of

smoke out bad activity?

>> Yeah. Well, the answer to that is is

absolutely yes. Um and we're just at the

beginning of this. So you know within um

like Google cloud one of the things that

we focus a lot on is uh governance

security um trust and quite often trust

you know it's interesting because in the

generative AI world you know trust has

this kind of new layer of do I it's not

deterministic you know and so it's like

do I really trust the outputs and how do

I create systems of trust and I find

that's as much a barrier or something to

solve as you know like classic

identifying potential you know malware

um you injection threats. Um, so there's

like a there's a track where the

fundamental technology will continue to

become, you know, more secure, but then

there's like this trust tier where, you

know, it's really about making sure that

the AI is doing what it's supposed to be

doing. So, you know, observability,

telemetry eval

um, and you know, believe it or not, not

every company has exactly the same

sensibilities about how much risk

they're willing to take.

>> I've heard that. Yeah. So, you know,

there's also like configurations of this

and one of the things we learned very

early on, you know, uh you could say

we've been very deliberate about um you

know, safety and responsibility. One of

the things that we discovered very early

on is uh some companies would take our

filtering and our settings uh you know

of kind of like do you want it to be

creative or do you want it to be you

know definitive or or you know

authoritative and so many people would

actually set it to creative

um way more than we thought. And so, you

know, for us, it's about surfacing a

platform that you can configure your own

filters and add your own governance to

whatever like your brand, you know, how

you want to take on risk. And that's way

more important than us coming up with

some singular vision of like how do you

do risk management and we operate in so

many constituencies in in in Europe,

AMIA, JPAC in um the Americas and there

are subtle differences in all of those.

So we also have we're building you know

aenta capabilities being able to deploy

in um you know infrastructure that's

connected or sometimes disconnected. You

know one of the projects that my team is

working on right now in the CTO group is

how do you give signals of intent of

what of what's happening but it can

cross trust boundaries that are actually

airgapped physically separated because

you can't share the data itself and

obviously like AI hasn't figured out a

way to like come out of the machine walk

over and like walk into the other

machine and figure you know

>> not yet. Yeah.

>> Yeah. not yet. Um, but there are ways

there are emerging techniques to

understand the intent of a workflow or

kind of early telemetry and then ways to

like bridge the gap. Uh, and so that's

also really really exciting in terms of

security is being able to bring more AI

and more capabilities of agents to

scenarios that typically would have been

closed off to AI, which is smaller

compute shapes, uh, you know, uh,

disconnected infrastructure. Uh, that's

really a frontier that I'm excited

about.

>> Yeah. So we're sort of talking about

this complex agentic world. Obviously

that's going to involve consuming a lot

of cloud services I think is a safe

assumption.

>> That'd be great.

>> But we also don't know how fast it's

coming, how soon it's getting here, how

quickly we're going to solve all of

these problems that that you know you

identify that we're actively working on.

At the same time, we're in the middle of

this unprecedented infrastructure

buildout. You know, I have Meta has said

they're going to spend 600 billion

dollars, you know, in the next few

years. We're getting 500 billion from

Stargate.

There are all of these sort of

uncertainties about how fast all of the

stuff we're describing will get built. I

guess do you worry that we're

overbuilding as a kind of society and

that there's going to be some harsh

correction that that gets in the middle

of this? I mean to be able to answer

that as you know as like society build I

would have to have a lot more data on

the de on the demand side from other

firms. I'd have to I'd have to see

demand signals from other firms to know

um I can at least tell you from the

Google perspective uh one I think AI has

definitely redefined uh and kind of

restarted the cloud market itself and

and infrastructure being a large part of

that. So we're in an entirely new game

today when it comes to cloud and

infrastructure. Um you know second thing

that is important about the scaled

infrastructure moment is that uh

there is so much demand. There is so

much demand Russell um and I think part

of that is that uh we're in a place now

where the um the kind of the frontiers

are being built and so you have multiple

frontier model providers. Um and as we

transition into inference, there's also

a recognition, you know, that we need to

make sure that compute is uh readily

available for inference. And so we're

kind of building we're kind of building

across segments right now. So like

there's a segment of like the large AI

labs and they need uh you know

availability to massive amounts of

infrastructure uh for training and to

and this continuous innovation

development. But there's also another

track which is as more of these become

real um inference becomes more and more

important and that's that's been a big

driver at least for us on you know our

tensor processing units and going from

really a training mindset in the past in

the architecture to now uh with Ironwood

which I'm happy to report is in the

hands of our first set of trusted

testers through GKE um you know Ironwood

is was built for a world where yes

training is going to be there but

inference is going to start to take, you

know, the preponderance of of workloads

over time and uh so, you know, 10x

performance improvement year-over-year,

TW uh 2x improvement in um efficiency,

power efficiency year-over-year. So,

those are the things that we're

interested in is, you know, we've done

infrastructure at scale a bit at Google.

And one of the things that we know is

that to, you know, build correctly and

build the right way is um constantly

looking for optimizations. So it may

gather a lot of headlines about, you

know, all the all the, you know, money

that's being spent and all the

infrastructure that's being proposed by

a bunch of these players, but underneath

the hood, we're busy making sure that

our TPUs, you know, we we have over

a,000x performance improvement since the

first generation of TPUs. We're now on

our seventh generation of TPUs, you

know, and efficiency. When you get a

year-over-year efficiency of 2x, I mean,

this is these are significant

engineering accomplishments. But we're

also down in the networking layer. You

know, we're also down in, you know, how

our subc cables route traffic and do

that in a highly performant way. And so

by, you know, our focus is optimizations

at every layer of the stack and those

benefits accreing to our customers. Uh,

and I'm not worried about demand um from

my chair right now.

>> Yeah. Although I want to unpack some of

what you said because you it does sound

like you see us turning a corner from

you know how do we get enough compute to

train these models

to how do we get enough compute to run

these models right that that that is a

fundamental turn and also that that

comes with if you're going to again try

to optimize at every level

>> you want different hardware

>> that is you you know if you're

optimizing for inference that goes down

to the chip level and and

>> these are choices that we're sort of

having to make

>> as we build out two years from now. You

know, looking at the world two years

from now, what are the compute needs

going to be? You do have to choose

between are we going to build out for

training or are we going to build out

for infrastructure?

>> Yeah. Well, and this is something again,

you know, uh you know, Google cloud part

of Google, part of Alphabet. you know,

one of the things that we've learned

over time in uh serving ML at uh web

scale is how to serve ML at webcale. Uh

you know, so this is not a new concept

for us and uh you know, probably the

thing that is most interesting for all

of you is that you know, we have as much

demand for ourselves and for the web.

And so from a like stability of demand

and stability of supply chain and

stability of research and development,

you know, Google cloud was built on

startups. 15 years ago or you know, 12

years ago roughly when I started at

Google cloud, you know, our largest

customers were startups and it was their

ability to scale rapidly kind of the

success disaster, you know, like when

Nano Banana hit, you know, our QPS

estimate was about two orders of

magnitude lower than what it ended up

being. And did it fall over? It

absolutely did not. Why? Because at

Google, we have been building, you know,

scaled infrastructure systems that are

highly reliable, you know, for decades.

And so when you hit the success disaster

with us, we understand like how to

dynamically, you know, drain shard jobs

across, you know, pools. We understand

how to do this inherently in our DNA.

whereas I think a lot of our competitors

are learning on the fly and have some

aspirations and infrastructure that uh

you know maybe exceed their track

record. Um so I'm feeling pretty bullish

about us.

>> Yeah. Yeah. Well, we're almost out of

time, but I did want to backstage when

we were talking, you had, you know,

there is so much uncertainty around just

the pace of progress has been so fast,

but you had this interesting analogy

between the moment we're at with AI and

the growth of cloud services themselves

where, you know, well, yeah, how did you

put it?

>> Yeah. was uh we tend to evaluate whether

something is good or not based on like

it's far in the future you know best

case scenario which in AI may be like

you know fully autonomous ambient AI

assistant that understands you from a

personal level also understands you at

work and can provide like very proactive

recommendations can run like in parallel

across a whole bunch of work tasks that

you've given it's super complex stuff

and then we go like what's available

today and we get really sad about the

state of things but if you go back to

like cloud. You know, at one point it

was just like I need an API. I need to

like store something outside of my data

center or I need a little bursty compute

for these jobs that have exceeded my

own, you know, envelope. And then over

time it rapidly became, you know, the

foundation for scaled data, scaled

analytics, you know, sovereign uh

compute, all these shapes, all of these

features, all these things emerged. And

we're going to see the exact same thing

in AI. We're in such early days uh in

how AI is going to like evolve and

manifest itself and uh so is a I I like

see the thing repeating and that's why

I'm so excited as you can probably tell

uh about the future and where we're

headed.

>> Well, great note to end on. Thank you so

much for coming down. This was such a

fascinating talk.

>> Thanks Russell.

>> All right. Cheers.

>> Thanks.