Sundar, we're here talking about a week or so after the wider release of Gemini 4 Argon. And I'd love to hear from you what the early reception has been to the new model.
Pretty excited. You know, we had been using it internally for a while. It's definitely a world-class frontier model. Nowadays, when you look at the frontier, the breadth and depth of that frontier intelligence is pretty comprehensive. So if I were to mentally visualize a heat map across agentic tasks, knowledge work, coding, etc., I think it is very, very strong on many dimensions, and I think it competes very well. The reception has been — you know, in the first 72 hours since Argon launched, we've had almost 25 trillion tokens, which is about 3x what 3.0 had in the first 72 hours. So, you know, it's definitely — and we are capacity constrained as we work through these launches. So really good feedback, particularly from our customers, but there are clearly areas to do better, you know. But I think the frontier is constantly — you ship, you get feedback, and you're on to the next thing, and it's as dynamic a moment as I've ever seen.
What are the areas that you feel like you want to do better on still?
You know, we want to continue on coding. It's a very good model. Our engineers have loved using it 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 continuing to make progress on — and so you have behavioral refinements based on the feedback you're getting from people using it. So there are areas we're going to do better there.
You were mentioning the compute constraints. It's interesting that even a company like Google has those for a new model like this? I mean, Argon's a large model. It's obviously a very expensive model to serve. Can you walk me through the decision of how you have to now gate this? It's, I believe, right now it's accessible through the API, AI Pro and Ultra subscribers. Obviously not to the widest possible user base yet, and I imagine that's challenging for a model of this size. So how do you think through gating that and releasing it?
One of the things we've been very good at — if you go back to the 3.0 models, we can distill these models and get versions which are nearly as capable but much more efficient. So we're working on those models as well, which will help us give wide access to as many people as possible. So those efforts are underway. We also — with this model, you know, we are definitely at a moment as we are working through the frontier, you want to really think about safety of these models. So we've invested a lot more effort, both in terms of how we tested it externally through our fair program with the U.S. government, giving early access to cyber researchers, internally widely testing it, and we've obviously brought it out externally. You're right, it's a large model. We are serving it pretty fast given the size of the model. So I'm excited what we are doing, but you know, we'll continuously expand access from now on. And so you saw at launch time we had made it available to all Ultra tiers — Pro tier in the Gemini app; we followed it with Pro tier in the Anti-Gravity app, and we are working on distilled versions which we can serve to more people with the 4.0 series.
Is this the Carbon model that there's rumors about?
At any given moment in time, we have internal versions of our frontier models. And so some of these are code names of those versions. That's different from us getting a more efficient version of Argon available more broadly.
So, more efficient version coming. The process of testing it for safety — I'm curious to hear you talk about that. It's interesting to think about how the private frontier labs, OpenAI, Anthropic, talk about the security and privacy of their new models versus a Google or a Meta or, you know, a larger hyperscaler. It sounds like you guys are being very methodical with the safety testing for this one, and it's maybe taken longer than previous models. Is that fair to say?
Definitely. I think with the agentic capabilities and its ability to code so well, particularly when you're deploying it — you know, we are serving many enterprise customers at scale, you want 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 to monitor them while they are working, right? And that can be compute intensive too. But that real-time monitoring as these models are deployed, I think ends up being critical, so that 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 sandboxing you would need to do, the real-time monitoring you would need to do — and with that also for customers giving them the protections they need, so you exactly know where it's running, what access it has, all the telemetry you need, all of that.
This interview is presented by Hyper Agent. Hire AI co-workers that take on real roles, collaborate with your team, and help run your business. Learn more at hyperagent.com. Thanks also to Granola, the AI notepad for people in back-to-back meetings. It works everywhere you do and lets you focus on what matters. Try it at granola.ai/sources and use the code SOURCES for 3 months off. 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 J-I-R-A.com. 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. Mercury is a fintech, not a bank. Check the show notes for details.
How closely do you follow the rat race of new models? I mean, I'm sure you do, but on X 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. I think everyone obviously sees the benchmarks, and Argon is there for sure, but it feels like the energy is constantly moving to who is ahead in the pole position. And how much do you focus on that? How much do you encourage your team to focus on that versus not? I'd be curious to hear right now how you're thinking about it.
It's a great question. Look, I think there's so much — it's such a dynamic moment. There's so much noise coming through. I obviously pay attention. You know, I'm on X. I pay attention. It's important to get the pulse externally because you get a sense of where you're doing well, where you're not. So I observe all that. But more importantly, internally we now have, you know, we have our own evals, which are pretty robust at evaluating our models. We can see the capability curve they are on. So we get very high quality signals internally as well. I think what I encourage the teams is: we are very committed to being at the AI frontier for the very long term. We didn't get on this train recently. We've been focused on this for a long time, and we want to be at that intelligence frontier a decade from now and so on. So it is doing everything you need to, including safety, by the way. I think that's going to be critical to develop intelligence at the frontier. 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. So some of what we had to adapt this year was to make deeper progress on agentic coding particularly, and then use that to improve our release cycles. We were able to start doing that, you know, with 3.5 to 3.6 to 3.7 to 3.8 Flash. We kind of, in about 3 months, went through that sequence, and so, you know, obviously with 4.0 we expect to have an improved release cadence as well.
There's aspects of Argon I'm interested in. One thing I saw was that it beat Minecraft on its own in a little over a day with, I think, less than 2,000 in compute, which is pretty impressive. I'm curious what the implications of this level of agentic capability are. I mean, a video game is one thing, but you could extrapolate that to potentially a lot of things. That's advanced decision-making, long-horizon tasks, right? 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. There are, for example, 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 code. You know, in Chromium, it's where if you want to create an experience where you can drag out a tab, move it around, put it in an AI experience — how do you keep it live without reloading? That's the kind of thing where you can get excited about this feature in your head, but historically for teams, it cuts through all layers of Chromium, right? But you can point Argon at it and it can persistently go and work through it and get it done. So we are seeing examples like that everywhere. What excites me is I think in some ways I view 2027 as a golden age of that productivity frontier for us because of the new class of models and how well we are integrating it with our internal environments. You can also apply these models to deeply improve model development itself, right? And what we in the industry talk about as recursive self-improvement. So those are all, practically, areas where you will see continued progress, which is exciting to me. Look, I think what you saw in Minecraft — from now on I think 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?
Yeah, and 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 can apply it and make progress, and you've seen that with mathematics recently. And so things which you would think would take a long time, you can point and put effort into it and show dramatic progress, like OpenAI did with mathematics with their recent model. So I think we're going to see a lot more of that in 2027.
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? 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?
There's a lot of humans in the loop on these things, given the 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 suggesting recipes, automatically evaluating next checkpoints, and so there's a healthy amount of models and researchers co-working. There are aspects of the pipeline which are probably closer to being, what I would say in this continuum, more RSI-like; some of them are lesser, but the whole continuum is shifting. So these are the reasons I think it's important to think about how we develop the frontier as we go through it. But I would expect the next 12 months to see continued acceleration because of RSI.
Quickly, while we're on just the model front here — I was at the last I/O and you guys talked about Gemini 3.5 Pro, and there's still some questions out there of like, whatever happened to that model? Did that just become Argon, or what happened there?
Now, when we were training the 3.5 series we had Flash and Pro which we were working on. An area where we internalized through the middle of this year is that there were gaps in both agentic capabilities, particularly around coding — and they're related — and how fundamental it is to overall progress. So we wanted to definitely hill climb there, and the team felt the fastest way to hill climb was using Flash. So it's a smaller model, it's more efficient, so you can really get a lot more iterative loops there. And so that's what led to 3.5 to 3.6 to 3.7 to 3.8 Flash. By the time we got that done, we had a chance to apply it to Pro, but Argon pre-training was done.
Got it. So it was just realizing that 3.5 Pro would have had a very short shelf life.
And so we decided to take all the learnings and apply it in Argon post-training, and so it was — you know, in a constrained environment, if there were things like compute time etc., you would do all of it. But I think it gave us a chance to kind of skate to where the future needs to be.
And so obviously with 4.0 you will have again different model versions in 4.0. Argon is our largest model and what we have outside, but we will ship more models there. But we are also making Argon available to the Pro tier in the meantime as well.
Where is the difficulty still in coding? Where is the area you're still trying to hill climb?
It's not that difficulty. It's just the frontier is constantly evolving, right? And you know, on every aspect of coding, Argon would be the frontier if you were to look two months ago. 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, you know, a world-class team programming would accomplish it with good taste, great code reviewability, it's very safe, etc. So it's just a dynamic frontier. There are times, you know, all our models — for example, take hallucination. It's a never-ending frontier to keep making progress on. It's an inherent part of the current state-of-the-art technology, but you're constantly working to reduce that. So it's both an opportunity and a difficulty. To give one example of that—
We haven't talked since there were some leadership changes that I wanted to ask you about at Google DeepMind. Demis Hassabis, the CEO, stepping back, becoming chief scientist, and also focusing more on Isomorphic, which I'm very curious about. And then Koray, you know, being elevated, running DeepMind. Jeff Dean, Google legend, leaving to start Discovery Loop. And you know, when that happened, there was a lot of press about that and intrigue about what was going on behind the scenes. I haven't heard you or anyone named there speak on it since then. So I'd be curious to hear from you. Maybe you can put some of the rumors to rest. What was the thinking there? What motivated that? Why did that happen?
You know, you're seeing some natural evolutions as people spend their time doing things for a while.
Jeff was 27 years, I believe.
27 years. You know, I feel proud. I've worked together with Jeff for a long time, a long time. You know, obviously he was seminal in his contributions here, and we are excited to see what he does with Discovery Loop. You know, I would start with saying he definitely wanted a moment in time where he saw an opportunity to go do something which he was excited by with a set of people. The change internally — I would just say you're going through a moment in the industry where the frontier model development is an all-consuming thing, right? You know, we've had broad AI research, AGI research and so on, but frontier model development is all-consuming. It's moving, it's picking up pace, right? 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 do that. So I think it's become an important job on its own. Demis is — he's obviously been thinking about this moment for a long time. He's particularly interested in both how do you go through this frontier moment safely, and how do you apply it to solve other problems. It's part of the work that he did with AlphaFold, and hence what we are doing now with Isomorphic — are examples of how you can apply AGI to solve other problems—
Medicine, drug discovery. Yeah.
Yeah. Robotics and so on — like you know, it'll apply to everything. So I think he 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 early on. And I think in some ways, both seeing that the model development is kind of a well-established process, and Koray has been at DeepMind for, you know, well over a decade too, and so we saw this opportunity for Koray to step up and do that — allows Demis... Demis is still involved as chair of DeepMind, but he's also spending a lot more time — you've seen him more involved with AI security conversations, but as well as thinking about how to apply AI in a more practical way, be it science, be it healthcare, and there'll be other areas as well.
Does this signal any kind of shift with how DeepMind relates to the rest of Google? Is it merging in closer? What are the natural next-order effects of this?
There's been a natural process over many years, including when Demis was running it, where DeepMind — the Gemini app gets built within DeepMind, right? So as an example, Koray had moved here to the Bay Area almost maybe 18 months ago. There's been close integration — makes sense for compute, for DeepMind to impact the rest of Google. So there's been deep collaboration which has already been underway a lot. So I don't think any of this in the continuum — maybe externally people noticed that as a moment, but for us it's been a deep coupling which has been underway for a long time.
How do you manage and are managing allocating compute towards long-term research bets? I think what DeepMind was historically known for kind of predates the current AI boom we're in, with the near-term — 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 the last year or two?
You know, the problem is always — you know, stakes are high, and I think we are fortunate to see extraordinary opportunities at being at the frontier of 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 opportunity space is large, and so obviously we have to plan ahead, but I think we have a pretty rigorous approach to plan ahead. You know, these teams plan independently. Underlying all of that, we've been an AI-first company for a long time, and so with that comes a commitment to invest in the compute that you need to be at the frontier, both for research and developing the models. So that is ground zero with which we 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 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 agentic coding liftoff hadn't happened. And remember, we build our own compute, right? And that has long lead times. So at any given moment in time, you're living with all the compute decisions you made 2 years ago. But people are assessing you based on the compute decision — they have the perfect information today, right? And so—
You had to make the call two years ago.
Two years ago. And you make the call based on the signals. You know, you're confident. You take a long-term optimistic view and you're aggressive in your investments. I don't think any company in the world today would sit and tell you that 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 — comes with the territory. We plan to be state-of-the-art in terms of our compute investments on an ongoing basis, to both drive research and develop frontier models, right? And beyond that, you know, we'll have opportunities to apply in many other parts, and I think we plan to do it to the best of our ability.
So that means you're planning right now for 2028, 2029, right?
Yeah, depending on, you know — yeah. If you're planning for land, power, etc., you're probably in the outer years, right?
And there was a big moment over the summer where you all went cash flow negative for the first time since the company IPO'd. And I think a lot of people in Wall Street—
Yes, of course. And people saw that and go, 'Wow, Google's continuing to invest more aggressively.' I think you guided up capex for 2027. You're thinking about 2028, 2029. What are you seeing that's giving you the resolve? Especially the moment in the summer where Google goes negative on free cash flow. I assume that was a big decision, and how do you think about that, and what is giving you that resolve?
The strength and durability of the demand signals we are seeing. So it's probably — to do something like this, you need to be confident at multiple levels. One is, I think we are fortunate to have early signals from such a diversified opportunity set, right? You know, all the way from consumer to enterprise, and so we are seeing robust demand signals.
You probably have more signal than any 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 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 how that curve looks like, be it the scaling laws, or like how our new models are doing compared to our old models, 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 a sense of the breakthroughs from research, which give you a sense of, okay, will Gemini 5.0 be much better than where Argon is? And the answer, 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. So it's not just hope, right? You're planning to a book of business as well. So you take all that into account, and you know, you lean in in a responsible, disciplined way, and which is what we're doing.
It's interesting to hear you talk with confidence about those signals, which obviously I believe. And then I think people are also trying to grapple with, for example, like CoreWeave was the latest one, but there's fears about the neoclouds, there's fears about the financing of these things, the brittleness 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 even Google-specific, but just that the markets maybe are over their skis?
So I was obviously answering the earlier question from a Google standpoint, right? Looking at it etc. The broader market — I mean, anytime you work through these moments where the opportunity is 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 I think that is an inherent part of a moment like that. This is why I try to focus very hard on those signals which I have control over and I can see, right? 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 I think you design the company in a 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 helps. And even if you look through the last 5 years — the pandemic drove a big shift towards digital.
Then some of it corrected. But it made sense to invest to drive that shift forward. For most companies which did so, have done better, right? So you have to understand long-term secular trends and invest to capture the transition. In fact, 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. We were relatively later to the cloud market, but we took a very aggressive investment posture into that for a period of 6 to 8 years, and that helps us through a moment like that. I think that's how I think about the AI moment as well.
Can you look out anywhere in the future and see where you're not compute constrained right now? Is there any moment in time you can look out and see?
I hesitate to answer that way because—
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 such complex ways, and you can have the butterfly effect of like, 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 smooth — no, you know, life doesn't work 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.
We're using all this. We're super constrained, and more importantly, we're using it for real today in practical, valuable ways. And we see opportunities, we see clear line of sight in which all of this will work better in the future. But that doesn't mean, as a collective industry, you know, or an economy, you overshoot or, you know, or you have shocks to the system which are out of your control and all that. So those are two separate things, right? And which is why part of it is I focus on the things that are in our control — our ability to take a long-term view, take our years of experience to better forecast, better model, better invest, and apply it to this wide opportunity set we have in a pretty disciplined way. And so I put my energy, time and attention into that, and less thinking about, you know, is there too much exuberance or not.
Staff to run your business with Hyper Agent. Agents can find new customers, run campaigns, and handle your business operations. They work in the cloud, collaborate with everyone on your team, and pass work to each other like real co-workers. Hire your AI staff at hyperagent.com.
I spend a lot of time context switching between meetings, often with no time to process one before the next starts. Thankfully, Granola runs in the background the whole time. It's an easy-to-use AI notepad for meetings that works everywhere, even on phone calls. I use Granola to recall what was said in meetings and create helpful summaries. I use it every day to stay on top of what I need to get done with my team. It connects to my email and suggests follow-ups for me to quickly review and send, saving me valuable time. Granola isn't just a core part of my workflow, it's basically my second brain. Try Granola at granola.ai/sources and use the promo code sources for 3 months off.
AI is only as useful as the context it has. But when that context is scattered across tools, threads, and DMs, your team and your AI agents are flying blind. That's the problem Jira by Atlassian solves. What's the goal tied to your project? What got decided last week in Slack DMs? Atlassian's teamwork graph pulls all of the valuable pieces together from Jira, Confluence, GitHub, Slack, and more, so nothing falls through the cracks. You get 44% more accurate results with 48% less token usage. With Jira, you can easily share your work context with the AI agents you already love, like Claude, Cursor, and GitHub Copilot. Assign them work directly or connect your tools through MCP. 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 J-I-R-A.com.
One of the biggest pains I've had running sources has been managing my books. 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. Mercury recently launched Mercury Books. It's AI-powered accounting that works with your Mercury banking and credit card transactions, plus external cards and payroll systems as well. You can choose between cash and accrual accounting and quickly generate reports for your cash flow, P&L, and balance sheet. Like everything Mercury does, the design of Mercury Books is super approachable and clean. I finally don't feel overwhelmed when I'm trying to understand the big picture of my business. I love that AI does most of the heavy lifting, including for things like auto-categorizing transactions, and that Mercury's command AI agent can handle tasks end to end. You don't need to be paying for separate accounting software anymore, and you can even invite your accountant or CPA to work alongside your Mercury books with you. This doesn't cost extra, and Mercury will even help you find a human accountant if you need one. Visit mercury.com/books to learn more and apply online in minutes. Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column N.A., members FDIC.
Framer is the AI website builder that powers the sources podcast website at podcast.sources.news. 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. I use Framer to make the sources podcast website be the destination for everywhere you can find the show. Plus, you can also see recent issues of my newsletter. Start building with agents for free today at framer.com/sources for 30% off a Framer Pro annual plan. framer.com/sources. Rules and restrictions may apply.
I've heard you say that you would love for AI to help you more with compute allocation inside Google. I'm curious, even over the last few months, is it involved at all? Are you finding that you're able to make better decisions, or that you would even maybe hand off certain decisions to a model on something as big and important as that?
Two ways I would answer the question. One is, given the stakes, it's squarely, you know, the decisions are being made by humans in the loop, including me, and you know, directly involved. But I would say even for me, one of the use cases, you know, I have agents running internally for me, agents I 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. You know, there are all these chat groups, there are pre-training chat groups, mid-training chat groups, you know, post-training data, etc. It could easily take me an hour a day to kind of catch up on what's happened in the last 24 hours, right? But now I have agents running which summarize and bring that information to me. I can still go deeper, and I do so often. Similarly, there are agents with telemetry into our compute usage by product area, efficiency, utilization, and is giving me inputs and insights, you know. So those are the ways we are weaving AI into all this. We have extraordinary people working with all of this as well. I think it's made all of us be better at what we do, and so that's how I'm using it.
