0:00
Sundar, uh, we're here talking about a week or so after the wider release of Gemini 4 Argon.
Search this video
0:05
Um, and I'd love to hear from you what the early reception has been to the new model. >> Pretty excited.
0:12
You know, we had been using it internally for a while.
0:14
Uh, it's definitely a a world-class frontier model.
0:18
Nowadays, when you look at the frontier, the breadth and depth of that uh frontier intelligence is pretty comprehensive.
0:27
So if I were to mentally visualize a heat map across agentic tasks, knowledge work, coding etc.
0:32
I think I think it is very very strong on many dimensions and and I think it competes very well.
0:39
The reception has been um you know in the first 72 hours since Argon launched you know we've had about almost 25 trillion tokens which is about 3x what 3.
0:50
0 had in the first 72 hours.
0:54
So you know it's definitely and we are capacity constraint as we work through this launches.
0:58
So really good feedback um uh particularly from our customers uh but it's clearly areas to uh do better uh you know but it's I think the frontier is constantly you ship you get feedback and you're on to the next thing and it's it's as dynamic a moment as I've ever seen.
1:18
>> What are the areas um that you feel like you want to do better on still?
1:20
you want to do better on still? you know we want to continue on coding u it's a very good model uh our engineers have loved using it uh internally but there are areas where we can see we can do better at any given moment in time when we have shipped we internally are
1:37
continuing to make progress on and and so you have behavioral refinements based on the feedback you're getting from people using it um so there are areas we're going to do better there >> you were mentioning the compute constraints it's interesting interesting that even a company like Google has those for a new model like this? I mean,
1:54
I mean, Argon's a large model.
1:55
It's obviously a very expensive model to serve.
1:57
Can you walk me through the decision of how you have to now gate this?
2:01
It's I believe right now it's accessible through the API AI Pro and Ultra subscribers.
2:07
Obviously, not to the widest possible user base yet, and I imagine that's challenging for a model of this size.
2:13
So, how do you think through that gating that and releasing it?
2:16
>> One of the things we've been very good at, uh, if you, you know, if you go back to the 3.
2:20
0 models we can distill of these models and and get versions which are nearly as capable but much more efficient.
2:28
So we working on those models as well u so which will help us give wide access to as many people as possible.
2:35
So those efforts are underway.
2:39
We also with this model you know we are definitely at a moment as we are working through frontier you want to really think about safety of these models.
2:46
So we've invested a lot more uh effort both in terms of how we tested it externally uh through our fair program with the US government um giving early access to cyber uh researchers uh internally widely testing it and we've obviously brought it out externally.
3:06
You're right it's a it's a large model.
3:09
Uh we are serving it pretty fast given the size of the model.
3:11
So I'm excited what we are doing but you know we'll we'll continuously expand access from now on and so you saw the launch time we had made made it available to all ultra tiers uh protier in the Gemini app we followed it with protier and the anti-gravity app and we are working on uh you know distilled versions which which we can serve to more people with the 4. 0 series.
3:37
Is this the carbon model that there's rumors about?
3:41
>> At any given moment in time, we have internal versions of our Frontier models.
3:45
Uh, and so some of these are code names of those versions.
3:47
That's different from us getting a more efficient version of Argon available more broadly.
3:56
>> So more efficient version coming.
3:56
Um, the process of testing it for safety.
4:01
I'm curious to hear you talk about that.
4:02
It's it's interesting to think about how the private frontier labs open anthropic talk about the security and privacy of their new models versus a Google or a meta or you know a larger hyperscaler.
4:17
It sounds like you guys are being very methodical with the safety testing for this one and and it's maybe taken longer than previous models. Is that fair to say? >> Uh definitely.
4:25
>> Uh definitely. I think I think with the with the agentic capabilities and and and its ability to code so well uh particularly when you're deploying it uh you know we we are serving many enterprise customers at scale you want
4:39
to make sure so for example the work we've done obviously the work you do to align these models well but more importantly both the compute and the systems you build uh to monitor them while they are working right and that is that can be comput intensive too
4:57
>> but that real time monitoring as these models are deployed I think ends up being critical uh so that you can you you know you you get as early signals as possible if something is not working as intended and you can intervene so we've taken time to build those systems the
5:13
sandboxing you would need to do the real-time monitoring you would need to do and and and with that also for customers giving them the protections they need so you exactly know where it's running what access it has, uh, all the telemetry you need, all of that. >> This interview is presented by Hyper
5:30
>> This interview is presented by Hyper Agent.
5:32
Hire AI co-workers that take on real roles, collaborate with your team, and help run your business.
5:37
Learn more at hyperagent. com.
5:41
Thanks also to Granola, the AI notepad for people in back-toback meetings.
5:44
It works everywhere you do and lets you focus on what matters. Try it at granola.
5:49
ai/sources and use the code sources for 3 months off.
5:53
This episode is also brought to you by Jira by Atlassian, where teams and agents get the context, coordination, and control to move work forward. Try it free at jira. com. That's ji. com.
6:08
This episode is also brought to you by Mercury, AI native banking that's loved by more than 300,000 entrepreneurs, including me. Visit mercury. com to learn more.
6:16
Mercury is a fintech, not a bank.
6:18
Check the show notes for details.
6:18
How closely do you follow the rat race of new models?
6:22
I mean, I'm sure you do, but OnX and everything of like who's up, who's down, and you know, I think you all really had a moment with Gemini 3 last year where it was like, wow, Google's back at the frontier.
6:31
I think everyone obviously sees the benchmarks and Argon is is there for sure, but um it feels like the energy is constantly moving to who is ahead in the pole position.
6:42
And how much do you focus on that?
6:44
How much do you encourage, you know, your team to focus on that versus not?
6:48
I'd be curious to hear right now how you're thinking about it.
6:51
>> It's a great question.
6:51
Look, I think there's so much.
6:53
It's such a dynamic moment.
6:55
There's so much noise coming through.
6:56
I obviously pay attention. You know, I'm on X. I I pay attention.
7:01
>> It's important to get the pulse externally because you get a sense of where you're doing well, where you're not.
7:06
So, I observe all that internally, but more importantly, internally, we now have, you know, we have our own evals, uh, which are pretty robust evaluate SSR models.
7:16
We can see the capability curve they are on.
7:18
So we get very high quality signals internally as well.
7:20
I think what I encourage the teams is we are very committed to being at the AI frontier for the very long term.
7:28
We didn't get on this train recently.
7:30
We've been uh you know focused on this for a long time and we want to be at that intelligence frontier a decade from now and so on.
7:36
So it is doing everything you need to including safety by the way.
7:42
I think that's going to be critical to develop uh intelligence at the frontier and but I think one of the things for us through this year is we needed to pick up the cadence at which we are shipping things right I mean the industry has been dynamic and you know the frontier models are evolving pretty fast too.
7:59
So some of what we had to adapt this year was to uh make make deeper progress on agentic coding particularly and then and then use that to improve our release cycles.
8:15
Uh we were able to start doing that you know with 3. 5 to 3. 6 to 3. 7 to 3. 8 flash.
8:22
>> We kind of in about 3 months went through that sequence uh sequence and so you know obviously with 4.
8:28
0 until we expect to have uh uh improved release cadence as well.
8:35
>> There's aspects of Argon I'm interested in.
8:37
One thing I saw was that it beat Minecraft uh on its own in a little over a day with I think only less than 2,000 in compute, which is pretty impressive.
8:47
I'm curious what the implications of this level of agentic capability.
8:49
I mean, a video game is one thing, but you could extrapolate that to potentially a lot of things.
8:55
That's advanced decision-m uh long horizon task, right?
8:59
Minecraft when you see that what do you think and how does that impact how you plan ahead >> we've had internal versions where for example with Argon we have seen this like step change in internal productivity ever since Argon has been rolled out internally a version of internally there are for example there
9:19
are large code bases on which we have pointed it to and it's gone let's say we've had problems which we know we need to fix it's millions of lines of codeed Yeah, >> you know, in Chromium, it's you know where if you want to create an experience where you can drag out a tab, it's uh move it around, put it in an AI experience. How do you keep it live
9:38
How do you keep it live without reloading?
9:40
That's the kind of like you can get excited about this feature in your head, but historically for teams, it cuts through all layers of Chromium, right?
9:49
So, but you can point Argon at it and it can persistently go and work through it and get it done.
9:55
So we are seeing examples like that like that everywhere.
10:00
So what that excites me is I think if if in some ways I view 27 2027 as a golden age of that uh productivity frontier for us because of the new class of models and how well we are integrating it with our internal environments.
10:18
You can also apply these models to deeply uh improve model development itself. Right?
10:24
and you know what what we in the industry talk about as uh recursive self-improvement.
10:28
So those are all you know practically I think areas where you will see continued progress uh which is exciting to me.
10:37
Look I think what you saw in Minecraft you know from now on I think we are the capability frontier is there where you can in a dedicated way apply assuming you have the compute and resources deeper progress on many areas than we have before >> any verifiable domain really right.
10:54
Yeah, and you know for sure and even in other areas you can always have humans in the loop to give feedback and keep making progress right and so >> it feels like infinite possibilities it's one of those moments where you think about you know if you can think about a thing you know you can apply it
11:11
and make progress and you've seen that with mathematics recently you know and so things which you would think would take a long time you you you can point and put effort into it and show dramatic progress like OpenAI did with uh Mathematics with their uh with the recent model. So uh I think we're going
11:27
So uh I think we're going to see a lot more of that in uh 2027.
11:34
>> Yeah, I've heard Sergey also talking about RSI recently and pushing the teams on that and I'm curious like where is Google at on RSI internally?
11:40
How much of current research and development is I assume most of it's being assisted by the models but where are we at on that continuum of the models really pushing?
11:52
There's a lot of humans in the loop on these things and uh given given the uh uh safety aspects of it but we are definitely at a stage where AI researchers are co-working with the models in a deep way right and you know the models are involved and maybe
12:08
suggesting recipes automatically evaluing next checkpoints and you know so so there's a healthy amount of uh models and researchers co-working there are aspects of the pipeline mine which are like probably closer to being like
12:24
you know what I would say in this continuum more RSI like some of them are lesser but the whole continuum is shifting so you know these are the reasons I think you know it's important to think about how we develop the
12:37
frontier as we go through it um but you know I would expect the next 12 months to see continued acceleration because of RSI >> quickly while we're on just the model front here uh I was at the last IO and you guys talked about Gemini 3.5 Pro uh
12:51
5 Pro uh and there's still some questions out there of like whatever happened to that model is that did that just become Argon or what happened there?
12:59
Now when we were uh training 3.
13:02
5 series we had flash and pro which we were working on.
13:06
An area where we internalized through the middle of this year is that there were gaps in um both agentic capabilities particularly around coding and they're related and and how fundamental it is to overall progress.
13:20
So we wanted to definitely hill climb there and and the team felt the fastest way to hill climb was using flash.
13:28
So it's a smaller model, it's more efficient, >> so you can really get uh a lot more iterative loops there.
13:37
>> And so that's what led to 3. 5 to 3. 6 to 3. 7 to 3. 8 flash.
13:40
By the time we got that done, we had a chance to apply it to uh Pro, but Argon pre-training was done. >> Got it.
13:52
So it was just realizing that 3.
13:54
5 Pro would have had a very short shelf life and so we decided to take all the learnings and apply it in Oregon post training and uh and and and so it was a you know in a constrained environment you know if there were things like compute time uh etc you would do all of it. >> Yeah.
14:15
>> But I think it gave us a chance to kind of skate to where the future needs to be. >> Yeah.
14:20
>> And so obviously with 4.
14:20
0 You will have again different model versions in 4. 0.
14:25
Argon is our largest model and what we have outside but we will ship more models there.
14:29
But we are also making argon available to the pro tier uh in the meantime as well.
14:35
>> Where is the difficulty still in coding?
14:37
Where is the the area you're still trying to hill climb?
14:40
>> It's not that difficulty.
14:40
It's just the frontier is constantly evolving. Right.
14:44
And um you know on every aspect of coding, Argon would be the frontier if you were to look two months ago. >> Right. Right?
14:50
So it's not I won't say it's a difficulty as in you're constantly giving it more complex end to end agentic tasks and you want the model to do incredibly well in a way that uh you know worldass team programming would accomplish it with good taste uh uh great code reviewability uh it's very safe etc etc so it's just a dynamic frontier.
15:20
Uh there are times, you know, all our models, you know, for example, take hallucination.
15:24
It's a it's a never- ending frontier to keep making progress on.
15:29
It's an inherent part of the current state-of-the-art technology, but you're constantly working to reduce that.
15:37
So, it's both an opportunity and a difficulty.
15:39
To give one example of that, >> we haven't talked since um there were some leadership changes that I wanted to ask you about at Google Deepmind.
15:44
Uh Demisabus, the CEO, stepping back uh becoming chief scientist and also focusing more on isomorphic which I'm very curious about and then Corey you know being elevated running Deep Mind Jeff Dean Google Legend leaving to start Discovery Loop um and you know when that happened there were there was a lot of press about that and and intrigue about what was going on behind the scenes.
16:09
Uh I haven't heard you or anyone named there speak on it since then.
16:14
So I'd be curious to hear from you.
16:15
Maybe you can put some of the rumors to rest.
16:17
Um what was the thinking there? What motivated that? Why did that happen?
16:23
>> You know, you're seeing some uh natural evolutions as people spend their time doing things for a while.
16:28
Um >> Jeff was 27 years, I believe. >> 27 years. You know, I feel proud.
16:32
I've worked together with Jeff for a long time. Uh a long time.
16:35
You know, obviously, uh he was seinal in his contributions here and we are excited uh to see what he does with Discovery Loop.
16:47
You know, I would start with saying he definitely wanted a a I think moment in time where he saw an opportunity to go do something which he was excited by with a set of people.
16:55
The the change internally I would I would just say you're going through a moment in the industry where the frontier model development is an allconsuming thing, right?
17:05
You know, you know, we've had broad AI research, AGI research and so on, but frontier model development is all consuming. It's moving.
17:13
It's picking up pace, right?
17:17
And it's a very scaled operation >> all the way from data to compute to the algorithmic work you do to managing the teams and your model agents to uh do that.
17:31
So I think I think it's become an important job on its whole.
17:36
Uh Demesis he's obviously been thinking about this moment for a long time.
17:39
He's particularly interested in both how do you uh go through this frontier moment safely and how do you apply it to solve other problems.
17:50
It's part of the work that he did with alpha fold and hence what we are doing now with isomorphic are examples of how you can apply uh you know AGI to solve other problems >> medicine drug discovery. Yeah. >> Yeah.
18:05
Robotics and so on like you know it'll apply to everything.
18:08
So I think he wants to spend more of his time.
18:11
wants to spend more of his time. So he's very leaned in into those things >> and >> which if you know Demis you know he's always been passionate about that >> passionate about it and he has been at it uh early on and I think in some ways both seeing that the model development is kind of a wellestablished process and Korea has been a deep mine for you know
18:34
well over a decade too and so we we saw this opportunity for Koreai to step up do that allows Demis Demis is still involved as chair of deep mind but uh he's also spending a lot more time you've seen him more involved with uh AI security conversations but as well as thinking about how to apply AI in a more practical way be it science be it
18:57
healthcare and there'll be other areas as well >> does this signal any kind of shift with how deep mind relates to the rest of Google is it merging in closer yeah what what are the natural next order effects of this >> there's been a natural process over many years including when Demis was running it uh where Deepmine um the Gemini app gets built within DeepMP right so as an
19:21
as an example um Corey had moved here to the Bay Area almost uh you know maybe 18 months ago there's been close integration makes sense for compute for >> deep mind to impact the rest of the Google so there's been deep collaboration which has already been underway a lot so I don't think any of this uh in the continuum wasn't maybe externally people noticed that as a moment. >> Yeah. >> Yeah.
19:48
>> But for us it's been a deep coupling which has been underway for a long time.
19:52
How do you manage and are managing allocating compute towards long-term research bets?
19:57
research bets? I think what Deep Mind was historically known for kind of preacht and the current AI boom we're in with the near-term the the frontier models that you talked about the Gemini app the products the commercialization how do you allocate that has that shifted at all in in the last year or two >> you know the problem is always uh you
20:16
know stakes are high and you know and I think we are fortunate to see extraordinary opportunities at the being at the frontier of uh intelligence having a very successful consumer business across many many products and services and being you know one of the top cloud hyperscaler companies in the world and the opportunities to provide the technology. So I think the
20:37
So I think the opportunity space is large and so obviously we we we have to plan ahead but I think we have a a pretty rigorous approach to plan ahead.
20:46
You know these teams plan independently underlying all of that we've been AI first company for a long time and so with that comes an commitment to invest in the compute that you need to be at the frontier both both for research and developing the models.
21:02
So that is ground zero with which we plan for the future.
21:03
plan for the future. Beyond that there are these businesses for example Thomas and the cloud team they have their own long range plans you know with their revenue projections the opportunities and we use that to plan compute for them you can always have moments when you
21:20
look back in hindsight there's no company in the world now sitting here to give you a sense of how fast the technology frontier is moving same time last year this agent decoding liftoff hadn't happened >> and remember we build our own compute right uh And that has long lead times. >> Yeah. >> Yeah.
21:38
>> So at any given moment in time, you're living with all the compute decisions you made 2 years ago. >> Yeah.
21:45
>> But people are assessing you based on the computer decision.
21:47
They they have the perfect information today, right?
21:51
And so >> you had to make the call two years ago. >> Two years ago.
21:53
And and you make the call based on the signals.
21:55
You know, you're confident.
21:57
You you take a long-term optimistic view and you're aggressive in your investments. >> Yeah.
22:02
>> Yeah. I don't think any company in the world today would sit and tell you like you know today they got their decision right two years ago right and so you're working through those moments and I think that comes with the territory um uh comes with the territory we plan to
22:17
be state-of-the-art in terms of our compute investments on an ongoing basis to develop uh uh to both drive research and develop frontier models right and u um and beyond that you know we'll have opportunities to uh apply in many other parts and I think uh you know we plan to do it to the best of our ability. >> So you're that means you're planning
22:38
>> So you're that means you're planning right now for 2028 2029 right?
22:39
Yeah, depending on, you know, yeah, >> uh yeah, if you're if you're planning for land, power, yeah, etc.
22:46
, you're probably in the outer years, >> right?
22:51
And there was a big moment over the summer where you all went cash flow negative for the first time since the company IPOed.
22:57
And I think a lot of people in Wall Street free cash flow. Yes, of course.
23:01
And and people saw that and go, "Wow, Google's continuing to invest more aggressively."
23:05
I think you guided up cap capex for for 2027.
23:11
you're thinking about 2028, 2029.
23:11
What are you seeing that's giving you the resolve?
23:16
Especially the moment in the summer where Google goes negative on on free cash flow.
23:20
I assume that was a big decision and and how do you think about that and what what is giving you that resolve?
23:27
>> The strength and durability of the demand signals we are seeing.
23:30
So it's probably you to do something like this you need to be confident at multiple levels.
23:36
levels. One is I think we are fortunate um to have early signals from such a diversified um uh opportunity set right you know all the way from consumer to enterprise and and so we are seeing that robust demand signals >> you probably have more signal than any
23:54
company on earth actually >> yeah I haven't thought about it that way but you could arguably make a case for it and so you have that robust demand signals we have the demand signal for Google itself which is a cutting edge
24:06
enterprise of how we are using AI and adopting it >> and you're seeing the capability frontier then you're able to independently outside of the demand signal in some objective way actually assess the capability frontier and see how it's progressing right and seeing
24:23
how that curve looks like and be the scaling loss or like how our new models are doing compared to our old models the the complexity of tasks we are able to tackle our internal eval benchmarks and so So you're getting a firsthand sense of it. You're also getting sense of the
24:38
You're also getting sense of the breakthroughs from research which give you a sense of okay will Gemini 5.
24:41
0 be much better than where for argon is and the answer uh you know is a yes right so you're getting all those confidence signals across that layer of the stack and we also have durable long-term contracts with committed revenue which we are signing.
25:03
So it's it's not just hope, right?
25:05
You're planning to uh uh book of business as well.
25:08
So you take all that into account and you know and and and uh you know uh you lean in in a responsible disciplined way and which is what we're doing.
25:19
what we're doing. It's interesting to hear you talk with confidence about that those signals which obviously I believe and then I think people are also trying to grapple with for example like core reeve was the latest one but there's fears about the neoclouds there's feel
25:34
you know fears about the financing of these things the brittleleness of it when you look out there even beyond Google but Google's involved in a lot of these things do you sense anything that gives you worry about a near-term not not even Google specific but just that the markets maybe over at skis. >> So I was obviously answering the earlier
25:50
>> So I was obviously answering the earlier question from a Google standpoint right like looking at it etc.
25:53
you know the broader market I mean you have a anytime you work through these moments where the opportunity is uh uh you know massive and the rewards are high and you have a lot of players coming in you are going to have dynamism and volatility associated with it and and I think that is inherent part of a moment like that.
26:15
This is why I try to focus very hard on that signals which I I have control over and I can see right and the demand signals the business clear line of sight to the business opportunities we have in terms of contracts backlogs and so on and you plan for that right and and I think you you design the company in a
26:36
way in which you can work through moments where even if there's some variation the long-term arc of it is good for the company right and so that's why your ability to take a long-term view uh helps and even if you look through the last 5 years the pandemic pandemic drove a big shift towards digital. >> Mhm. >> Mhm.
26:55
>> Then some of it corrected. >> Yeah.
26:58
>> But it made sense to invest to to drive that shift forward for most companies which did did so have done better. Right.
27:06
So so you have to understand long-term secular trends and invest to uh you know capture the transition.
27:11
In fact, we invested a lot of our headcount growth as a company through those years was for Google cloud and we are benefiting on the other side of having scaled up.
27:23
We were relatively later to the cloud market but we you know took a a very aggressive investment uh posture into that for a uh for a period of 6 to 8 years and and you know and that helps us uh through a moment like that.
27:41
I think that's how I think about the AI moment as well.
27:47
>> Can you look out anywhere in the future and see where you're not compute constrained right now?
27:50
Is there any moment in time you can look out and see?
27:53
I hesitate to answer that way because um >> yeah, you got to be careful because everyone will freak out if it's like >> No, what I mean is it's more that you know we live in a moment in time in which the global economy is so intermingled in
28:11
such complex ways and you can have the butterfly effect of like some you know one incident which can have ripple effects everywhere right and so I'm trying to decouple two things I'm not saying from now on everything will be a smooth no you know life doesn't work
28:27
that way right you know economies grow there are hiccups we work through there are business cycles >> all that are inevitable parts of how the world works >> right and I think no one thinks this is like the past bubble like this is not fiber yeah right we're using all this
28:42
>> we're using all this we're super constrained and and and more importantly we're using it >> for real today in practical valuer ways and And we see opportunities, we see clear line of sight in which all of this will work better in the future. But that
28:57
But that doesn't mean as a collective industry, you know, or or an economy, you overshoot or, you know, or you have shocks to the system which are out of your control and and and all that.
29:09
So those are two separate things, right?
29:13
And which is why part of it is I focus on the things that are in our control are our ability to take a long-term view, take our years of experience to better forecast, better model, uh better invest and apply it to this uh wide opportunity set we have in a pretty disciplined way.
29:32
And I I so I put my energy, time and attention into that and less thinking about, you know, is there too much exuberance or not.
29:43
Employ AI staff to run your business with Hyper Agent.
29:48
Agents can find new customers, run campaigns, and handle your business operations.
29:52
They work in the cloud, collaborate with everyone on your team, and pass work to each other like real co-workers.
29:58
Hire your AI staff at hyperagent. com.
30:02
I spend a lot of time context switching between meetings, often with no time to process one before the next starts.
30:08
Thankfully, Granola runs in the background the whole time.
30:10
It's an easytouse AI notepad for meetings that works everywhere, even on phone calls.
30:13
I use Granola to recall what was said in meetings and create helpful summaries.
30:17
I use it every day to stay on top of what I need to get done with my team.
30:21
It connects to my email and suggests follow-ups for me to quickly review and send, saving me valuable time.
30:26
Granola isn't just a core part of my workflow.
30:31
It's basically my second brain. Try Granola at granola.
30:33
ai/sources and use the promo code sources for 3 months off.
30:38
AI is only as useful as the context it has.
30:41
But when that context is scattered across tools, threads, and DMs, your team and your AI agents are flying blind.
30:47
That's the problem Jira by Atlassian solves.
30:50
What's the goal tied to your project?
30:52
What got decided last week in Slack DMs?
30:54
Atlassian's teamwork graph pulls all of the valuable pieces together from Jira, Confluence, GitHub, Slack, and more.
31:01
So nothing falls through the cracks.
31:03
You get 44% more accurate results with 48% less token usage.
31:08
With Jira, you can easily share your work context with the AI agents you already love, like Claude, Cursor, and GitHub Copilot.
31:14
Assign them work directly or connect your tools through MCP.
31:18
All of this lets you spend less time digging through endless links and messages, chasing down what got decided and by who, and spend more time actually shipping. Learn more at jira. com. That's jir. com.
31:32
One of the biggest pains I've had running sources has been managing my books.
31:35
I'm not an accountant, and all of the software out there is unintuitive, clunky, and detached from where my transactions actually live until now.
31:43
Mercury recently launched Mercury Books.
31:46
It's AI powered accounting that works with your Mercury banking and credit card transactions, plus external cards and payroll systems as well.
31:51
You can choose between cash and acrruel accounting and quickly generate reports for your cash flow, P&L, and balance sheet.
32:00
Like everything Mercury does, the design of Mercury Books is super approachable and clean.
32:03
I finally don't feel overwhelmed when I'm trying to understand the big picture of my business.
32:09
I love that AI does most of the heavy lifting, including for things like autoc categorizing transactions, and that Mercury's command AI agent can handle tasks end to end.
32:17
You don't need to be paying for separate accounting software anymore.
32:21
And you can even invite your accountant or CPA to work alongside your Mercury books with you.
32:27
This doesn't cost extra, and Mercury will even help you find a human accountant if you need one. Visit mercury.
32:32
com/books to learn more and apply online in minutes.
32:36
Mercury is a fintech company, not an FDIC insured bank.
32:38
Banking services provided through Choice Financial Group and column NA members FDIC.
32:43
Framer is the AI website builder that powers the sources podcast website at podcast. sources. news.
32:50
Framer brings AI agents into the same canvas where your website is designed, managed, and published so you can move faster without giving up your taste or control.
32:58
I use Framer to make the sources podcast website be the destination for everywhere you can find the show.
33:04
Plus, you can also see recent issues of my newsletter.
33:06
Start building with agents for free today at framer.
33:09
com/sources for 30% off a framer pro annual plan. framer. com/sources.
33:16
Rules and restrictions may apply.
33:16
I've heard you say that you would love for AI to help you more with compute allocation inside Google.
33:22
I'm curious, even the last few months, uh, is it involved at all?
33:27
Are you are you finding that you're able to make better decisions or that you would even maybe hand off certain decisions uh to a model on that something as big and important as that?
33:38
>> Two ways I would answer the question.
33:40
one is given the stakes it's it's squarely uh you know uh the decisions are being made by uh humans in the loop >> including me and you know you know directly involved but I would say u even for me one of the use cases you know I have agents running internally for me you agents have deployed both to help me understand you know simple thing it's not super fancy but you know at our scale even to understand the model development arc.
34:12
You know, there are all these chat groups, there are pre-training chat groups, mid-training chat groups, you know, post- training data, etc.
34:20
It could easily take me an hour a day to kind of catch up what's happened in the last 24 hours, right?
34:24
But now I have agents running which summarize and bring that information to me.
34:31
I can still go deeper and I do so often.
34:36
Similarly, there are agents with telemetry into our compute usage byproduct area, efficiency, utilization and is giving me inputs and insights, you know.
34:46
So, those are the ways we are weaving AI into all this.
34:48
We have extraordinary people working with all of this as well.
34:52
I think it's made all of us u uh uh you know be better at what we do and so that's how I'm using it.
35:01
>> So, no future near-term where AI is making the compute decisions.
35:03
I mean even five years ago you know we used AI to make power consumption efficient in our data centers.
35:12
So um we do have when we're making computer allocation decisions AI is a core part of the input it's providing right.
35:20
So um >> I'm trying to get you off the hook here like oh the model said it not me.
35:26
>> Yeah I mean you know it's a it's a continuum there.
35:28
By the time it's doing that you would be applying yourself to the next problem and so on.
35:32
So that's how I I think about it. >> Yeah.
35:35
I'm curious what you make of this personal agent race that the industry is now in with Muse, which is a great product, Instinct.
35:41
I remember at the last IO you guys showed Gemini Spark.
35:46
>> And I'm curious, you know, do you think Google needs to do more here on consumer?
35:50
I know you just had the Gemini workspace where you're giving an agent an email, a workspace identity, all that, but in the consumer race when you look at it, um it's early, but I imagine you want Google to have even more of a stake in that.
36:04
>> Oh, you know, it's a huge area of focus.
36:06
In fact, you know, you know, the goal we set out for Gemini, how we call it, we wanted to be a uh a personal proactive and powerful assistant, right?
36:12
So, in >> it's a core part of what we're trying to do with Gemini.
36:17
do with Gemini. Um and you are right in IO that's what you know with spark I think uh you know that's the direction we are doing we obviously spoke more about it in our enterprise event in the context of the Gemini agent >> to be very clear between enterprise and consumer >> it's the same model same harness same runtime it is just the context of what you have in a business or your personal
36:43
context you know is different and everything we announced in the event Because we're building it for enterprise has worldclass security governance every task there you know we provisioning essentially a microVM to go complete the task running securely 24/7 with realtime monitoring etc to pro and providing enterprises with the governance and control and auditability of those agents. We're bringing all that
37:10
We're bringing all that to consumers as well.
37:12
So already you saw you know you have chat and task now in Gemini.
37:17
Gemini. Um our vision you know in the near future is it's it's one box and you know there's it's a continental what is a chat what is a task you know which of those need to be persistent which of those need to be scheduled how does it come back to you uh I think I think we view it as a natural opportunity empirically when we look at what people
37:40
are doing in Gemini if you were to go back and classify it we see the trend already right So the percentage of if I were to call queries which are more agentic has been steadily increasing over time right and and we're going to make all of that more widely available as we go through the next few months and we mainly launched all of this to >> ultra and pro users. Yeah. And and uh so Yeah.
38:06
And and uh so we're going to make it more widely available.
38:10
The newer Argon models deliver uh a huge step change, but you will be able to again as a consumer, you know, talk to Gemini, not just in the app, but in the context of like, uh, you know, maybe from Gmail or you can call, you could, you know, so, so we are focused on it, uh, you know, from search and we're focused on it in a pretty deep way. >> Yeah.
38:37
A thing I've been dying to ask you, especially with your experience, you know, running Chrome before, like that the moment we're in, all these agents are using browsers like humans to complete tasks.
38:45
I mean, with Muse, you can literally see, right?
38:47
And >> you know, it feels when I look at it, I get why it's happening.
38:51
You want to get the task done and that's the way to do it.
38:54
But it feels like trying to fit a triangle into a square.
38:57
It doesn't feel quite right.
38:59
And to me, I'm it's making me go, does the web need to fundamentally change for these agents?
39:04
and how and I know there's protocols you guys have worked on them there's others um but I feel like Google's uniquely positioned to do something here bigger and I'm curious if you've been thinking about that >> no you know deeply you are right look it's a it's a complex moment because
39:21
you know when we say agents working at scale that means many many different parts of the economy interoperating with each other in some way in a way that has to make sense for all the players involved right and I think the technology frontier will ahead of the ecosystem uh how it comes to work together in this case. I definitely think a big part of
39:40
I definitely think a big part of this is why there's been a lot of momentum behind MCP or agent protocols or payment protocols like what we did with UCP along with a set of companies.
39:53
Uh and there's going to be a chaotic evolution there.
39:57
Uh hopefully working towards more interoperable standards, >> right?
40:02
And and I think that's a very elegant way to do it.
40:04
browser use and computer use is a very elegant way to work too because that's what you as a user would do and allows you to efficiently build controls uh for for for the users to be in the loop as well. >> Mh.
40:20
So I think you have both ways and they're going to develop and and uh but there are moments where anytime I think if you solve it with better protocols um you know I think it's a far more scalable approach and it it'll work anywhere and instantly.
40:37
So I think but that will take longer time I think. >> Yeah.
40:40
How do you think about protecting the interface in the way that you need to in this agent world? Right.
40:44
There's an argument you made there's a new kind of agentic Google search that could be invented.
40:50
Now agents don't need you know mu Amazon is blocked muse for example you have a very large ads business that requires engagement.
40:57
So what do agents do to those kinds of interfaces?
41:01
How do you protect or not from that?
41:04
>> The right answer longer term is you put users first and say what would users want and and you you want to try your best to be aligned in that direction.
41:15
work towards that direction, right?
41:15
And and I think I think that's what will drive these experiences.
41:21
But ultimately, it has to work for all the businesses who are providing value to so at any given point you're going to reach equilibrium and that equilibrium will shift.
41:30
There'll be new players who view each of these moments as disruptive ways to do things differently.
41:35
Like when the internet came, you know, you would sit and think, shouldn't the way real estate works, how people buy and sell houses overnight transform, right?
41:42
But, you know, it's kind of played out not quite what people thought it would happen that way, right?
41:49
And I would argue even today most of it happens the way it used to happen uh like you know uh 20 years ago.
41:59
>> So there are real constraints in the real world.
42:01
real world. I think AI agents and AI are even more uh they will will be even more disruptive but you know there'll be micro equilibriums which which will emerge for me from our standpoint I think you embrace the innovation you kind of put yourself in the shoes of users and you
42:21
try to innovate in that direction >> that's how I would approach our product development so I think that's what users would demand >> I've talked to Liz and some of your other colleagues over the last year or so about this idea of when do Google
42:34
search AI overviews and Gemini fully merge if ever because right now you've got these two largecaled you know AI products uh Gemini billion users search billions of users but relative you know different but overlapping in some cases AI experiences um do you ever see them
42:52
fully combining >> okay here's an example right like when we ship personal intelligence they work I mean you can turn on personal intelligence both in Gemini and in search and it has your context and so kind of works in a uh unified way. I
43:04
I expect our agents to provide that unifying experience across these products.
43:12
In fact agentic layer rather than try to converge or you know think about different products and how they come together I think uh the agentic layer naturally ends up doing that right so because your agents they're running in the cloud they're running 24/7 they have your context and so they can pick up things where you left off right so you can start a task in some place you go to the other place your agent has that context for you.
43:43
>> So I think the agent layer offers us a chance to bring a unified product experience better than ever before.
43:50
>> So that's how I uh I think about it.
43:52
>> You guys are also uh starting to send data centers to space.
43:54
Uh you've got very early you've got project suncatcher um your first pilot uh with planet labs and and SpaceX. Um how's it been going? I mean it went up. Is it is it still there? Is it burned up?
44:08
What's what's been the early feedback?
44:09
Uh it's been doing great.
44:13
I think still there uh it's it's doing well and we are >> I think we'll get a chance to say hello to the TPUs but uh uh you know which I'm really excited by.
44:23
I had a chance to watch the launch here while it was happening.
44:27
Look, I think when we take these projects to me it's no different from starting Whimo.
44:32
I I still recall uh flying to the desert to watch the Stanley car race as part of the DARPA challenge.
44:42
And at that time, if you had tried to envision where Whimo would be today, you know, it would seem like, you know, you're talking about something futuristic sci-fi thing, right?
44:53
But, you know, 10 to 20 years is a long time, particularly the with the with the inflection points we are in.
45:00
So uh I think we are taking the early steps but we are uh deeply committed uh to the journey and uh the team is a world-class team and you know in a 10 to 20 year time frame they're going to have amazing milestones.
45:12
Uh so that's what I'm excited about.
45:15
>> What about like the desert, the ocean, all these places on planet earth where we could put them now.
45:19
I mean what why why invest in space right now?
45:21
why invest in space right now? If you think through from a first principle standpoint of energy availability um you know being in sun sun-synchronous orbits getting access to 24/7 energy kind of you know it's it's a pretty good
45:37
reason there'll be more reasons you know I think uh we have so many opportunities to use the physical space the constrained physical space on planet earth the data centers today look big to us but they are they are not just arbitrarily big to how they were like 10 years ago. So it is tough for us to
45:58
So it is tough for us to exponentially extrapolate our our compute needs 20 years out.
46:06
>> And so I don't think it earth would look nice if that's how you plan to do it.
46:09
Uh yeah, >> it's a good point. >> Yeah.
46:13
And so and we would have other constraints here and and so on.
46:17
But also by thinking about these technologies, it's like in the 60s going to the moon, going to space, you're pushing so many aspects of the technology curve.
46:28
That's how you drive R&D.
46:30
That's how you drive the future.
46:30
R&D. That's how you drive the future. So there will be uh you know it's the it's the the fundamental thing about a moonshot you realize is when you focus on something very ambitious you invent so many things along the way all of which end up being helpful >> what you found with whimo and
46:49
>> whimo are tpus and you know because we were >> trying to build AI so early >> right >> and the compute wasn't there is what led us to build TPUs right >> right and so and and I think so thinking about a technology curve early in pushing I think definitely gives you those opportunities >> and you said the word moonshot a couple times. Uh it's literally a moonshot but
47:07
Uh it's literally a moonshot but does is this I I saw that in the release too for it.
47:11
Is this the next Whimo in that sense is this the next Alphabet spin-off kind of new company Suncatcher?
47:18
>> Yeah, we are definitely in a uh comput is so integral to what we do as AI.
47:20
you know, we haven't thought through the struck me.
47:25
I I view it as our the future way we will scale up compute in a pretty fundamental way and so but it's definitely a moonshot and you know and uh and but we committed to doing that well.
47:38
>> I would love to hear the strategy for Whimo because I can see it going a lot of directions.
47:41
Um, and I'm curious, do you want to keep having these partnerships with the Ubers, etc.
47:46
, or will it be the fully, you know, vertically integrated Whimo fleet? That's one phase of it.
47:53
And then I have a question about just, you know, selling Whimo vehicles directly to consumers.
47:58
I'm curious if you've what those two buckets.
48:00
Can Whimo and Uber coexist?
48:00
Uh, and then could I just buy a Whimo car one day and maybe rent it out myself?
48:08
It's an area where we are so so far from a zero sum construct.
48:12
I think like you know just like uh it's so early.
48:18
The opportunity space seems so vast to me.
48:20
So many applications for autonomous uh uh driver technology.
48:23
I expect Whimo to work in a myriad of ways including with other partners.
48:28
I expect the other partners to do very well as as companies on their own too.
48:34
So all of that I'm very very bullish on the way team the you know they're very very focused on scaling up what they are doing.
48:42
Uh we just crossed um half a million uh paid rights per week and you know we are so the theme is razor focused on scaling that up.
48:56
>> Toronto launch is coming. >> Yeah.
48:57
10 new cities this year, right?
48:59
And so it's operationally, logistically >> scaling up the vehicles, scaling up in more environments, um scaling up the uh customer service, the operations, all with safety being the number one priority and I think that is what the next 12 months is going to be more of it.
49:18
international other weather conditions, highways and everywhere scaling up with this high safety bar we have right already driving with >> 95% fewer serious crashes uh compared to u you know regular driving and so that to me is where through it all the underlying technology we are building is so general it has all the option values you're talking about built on top of it.
49:47
I think everyone is like wondering where are you going to take that option value?
49:52
Some people are scared because they have businesses to protect and others are just going wow what is the expansion >> you know I would I would just argue historically if you've looked at looked at as a company I think you've looked us build Android mobile even in the AI ecosystem we partner with a wide variety
50:08
of companies so I think we will always work with the ecosystem and I think there'll be plenty of uh growth for everyone to do well >> will I ever be able to buy a Whimo car myself though >> it's not what we are working on at this moment Uh, but never say >> never say never. >> Never say never. >> Never say never. >> What about Wing?
50:26
This is one of your other bets that's really maturing and I've heard you say maybe 40 million American households will have access to Wing in the not too distant future. This is drone delivery. Yeah.
50:37
>> Um, what do you imagine that's going to unlock?
50:41
>> I just spent uh time with a wonderful Wing team on uh Friday uh uh Friday evening.
50:47
So, it was my last meeting of the day.
50:49
So look, I'm very excited because um they have a product which I think is very good.
50:58
Early on they've made the bet to design a lightweight product which is easy to make and easy to deploy >> and and and you can scale it, right?
51:06
So I think so they've gotten the primitives right.
51:11
I think they have ambitious expansion plans too.
51:13
uh so I would expect us to meaningfully increase the scale of deliveries in 27 and also bring it to newer categories than where we do today. Right?
51:25
So those are the growth opportunities >> and this is similar to Whimo through a partnershipsled approach.
51:32
>> Uh yeah that's that's how we will we will approach in 27. Yeah.
51:36
>> Um okay you've been spending a lot of time in DC recently I've noticed.
51:38
Um, and I haven't heard you recently talk about how you're feeling about the state of policym.
51:45
Um, Demis has obviously laid out his proposal for this kind of FINRA style body, which I think got a lot of interest.
51:53
Um, and I'm I'm sure you're aligned with that in a lot of ways, but but I'm I'd be curious, especially the time you've been spending recently, and you signed the the the thing that the White House did with all the other tech CEOs.
52:04
Um, what would you like to see happen?
52:08
You know, it feels like we're in this moment where everyone's trying to create the the framework to support regulation.
52:14
>> It is such a consequential uh technology with so many implications including extraordinary uh opportunity to make uh you know society uh better. >> Yeah.
52:28
>> Uh the economies better etc.
52:28
I think we are in a phase right now where there is growing realization about how powerful this technology is and that the companies which are at the forefront of this technology needing to operate with the right kind of safeguards >> and and approaches so that we deploy this technology as safely as possible.
52:52
So that's how I would capture the current moment.
52:53
current moment. you for us as a company this is uh so core to how we want to think and operate we've done that for a long time that's what we're doing with the argon launch as well so in some ways that that's what the current moment is I
53:08
think the next step is the more we can all standardize our testing approaches with better eval standards I think that will help us assess the frontier much more effectively if you look at the financial world or financial statements
53:22
there are clear standards accounting standards you operate to uh there are well defined independent evaluators who who do those things like the professional audit firms I think so those are all good good directions for us to evolve to there will be a time we
53:38
will need to do more >> uh I think it's but it's important to get it right right I think the technology is so dynamic one of the challenges here is uh anything you do might be too inadequate or just wrong uh 12 months from now Right. And so
53:53
And so >> I think building the competency, creating industry standards groups are all good ways to make progress.
54:00
So that's where at least I'm spending a lot of my time and effort on is as Google are we meaningfully increasing the level of investments the bar at which we are safely developing and deploying these models.
54:13
I think companies which do that well will have a competitive advantage in the future. Right?
54:18
So that's the way I think about it.
54:21
And so that's where a lot of my energy and It's interesting, you know, we started this with you talking about how you guys have intentionally kind of of your own accord been more slow in the release of Argon and testing it internally.
54:32
It was your widest internal test ever, the government review, all these things and yet people are saying the government should mandate maybe a slowdown and it sounds like market incentives are maybe working already based on what you're telling me of how you guys are working.
54:46
I get nervous about these things.
54:46
is I get nervous about startups uh giving startups room to still grow regulatory capture these kinds of things.
54:52
So I'm curious how you balance you know needing some regulation with that.
54:58
>> Look I in all scenarios I would the main thing I would advocate is a more uniform regulation than a patchwork of state-by-state regulations.
55:05
I think that really can set the industry back and actually make make it inherently unsafer because it's a more complicated frontier to work through.
55:14
I feel very confident in our ability to adapt to ongoing developments in this space.
55:21
>> Already one way you can think of all of us now if you're developing a model at the frontier even though it's voluntary we are submitting our models to the US government for testing. >> Mhm. >> Right.
55:31
So that's a change from a year ago. >> Yeah.
55:34
>> Whether it's codified as regulation or not you have these developments which have happened.
55:37
So I think it'll be an ongoing area where we are evolving based on what we are observing and I think you want to anticipate but I think it's important for a technology which has so much potential uh to make sure we are at the frontier as well uh as for us for me as a company uh for us as a country and so on.
55:56
I I don't understand quantum to be honest.
56:00
Um I know you guys have a deep bet here.
56:04
Um and quantum feels like this thing that exists kind of in a layer above everything we've been talking about that could influence a lot of things.
56:10
It's a hard question, but I'd be curious to hear you uh share what you think the opportunity of quantum actually is because it sounds like I've been seeing you post, you know, you guys are making some progress here. Yeah.
56:21
Uh and when you look at it like then when it hits what do you think u will happen?
56:27
>> First of all welcome to the club.
56:27
>> First of all welcome to the club. when you say you don't understand quantum okay >> you're you're joining each and every one of us right I think uh good >> it's a it's a technology the more you spend with it uh in some ways you understand it less uh right and uh >> it's like AI >> uh I think even more so I think um quantum is an area where we have clearly we had set internal milestones
56:54
and the team is tracking or exceeding to those milestones over almost now uh a 10-year time frame and in terms of robustly scaling up quantum systems uh with a view towards commercial applications right the future possibilities of quantum are both vast
57:14
in some ways uncertain etc but I genuinely see practical commercial applications in a in a 5year time frame which we would get excited by you know where we would be today like we provide TPU odds. You know, if I were sitting in
57:27
You know, if I were sitting in a chair 5 years out, our Google Cloud team would be negotiating deals on quantum deployments at some scale, right?
57:39
And and and so uh so it is incredibly exciting to me and uh and it's another technology we've been committed to for a long time and you'll continue to see us be committed and be at the frontier. >> Sar, thank you.
57:52
>> All right, Alex, real pleasure. Thank you.
57:57
This interview is presented by Hyper Agent.
57:58
Hire AI co-workers that take on real roles, collaborate with your team, and help run your business.
58:03
Learn more at hyperagent. com.
58:07
Framer is the AI native website builder that lets you build faster without giving up control. Visit framer. com/sources for 30% off.
58:14
Rules and restrictions may apply.
58:17
Granola is the best AI notepad I've tried.
58:19
It works everywhere on a video or phone call, in person, or an Apple Watch. Try it now at granola.
58:25
ai/sources and use the promo code sources at checkout for 3 months off.
58:29
Jiradlassing is where your team and your agents work from the same context. Try it free at jira. com. That's ji. com.
58:39
Banking should feel like modern software.
58:41
Get everything you need in one place. Visit mercury.
58:43
com to learn more and apply online in minutes.
58:45
Mercury is a fintech, not a bank.
58:47
Check the show notes for details.