Google CEO Sundar Pichai on Agents, Open Source, Race to AGI (Transcript)

How do we account for that?

SUNDAR PICHAI: People in our experience with Search and in YouTube — YouTube is a great example — people have this sense of connection with the creators they like and follow. And so that’s a big part of what they’re looking for. So I think people are in different mindsets. There are times they want to discover content on the web. Shopping is delightful for a lot of people in a lot of moments, and so they’re not trying to fully outsource that. That’s why I try to distinguish between what feels like something which may be a chore at times and what feels like something which is delightful.

Similarly, it could be in news — people have their trusted sources. At least at Google through Search and YouTube and through agents, we think there will always be an incredible value from the ecosystem which users want to connect to, and the agents should be in the job of doing that.

MATTHEW BERMAN: Yeah.

SUNDAR PICHAI: But you are right, there are times the agents are playing a role in the middle. Sometimes it’s very good because it helps improve user satisfaction because the users are getting to what they want in a better way. But there’s a layer of abstraction too, and that’s what you’re talking about. But I think it’s always been true with technology a bit. And at the same time, the tools with which you can create content are also exploding. So people will also be creating more content. So I think there’ll be a new balance which we will find. But yeah, it is an interesting moment of evolution.

MATTHEW BERMAN: Yeah, I like that. I think you’re saying there’s definitely going to be a strong place for agents to do that curation on our behalf, especially in the era of complete slop domination. But also that feeling of exploration can still be there.

SUNDAR PICHAI: Yes. Because it’s a fundamental human need that doesn’t go away.

Cybersecurity and AI-Enhanced Cyberattacks

MATTHEW BERMAN: Yeah, exactly. Okay, so speaking of the Wild West — and I know you’re probably feeling this pretty strongly at Google — there have been increasing cybersecurity, cyberattacks. The models are getting better at cyber. Obviously Google has been thinking about cybersecurity for decades. Are you seeing cyberattacks, especially AI-enhanced cyberattacks, ramp up at Google?

SUNDAR PICHAI: Look, we’ve been seeing — to be very clear, we’ve deeply cared about cyber because we’ve been working on frontier technologies for a while. Google pioneered many important security frontiers like Zero Trust and so on. We have worked hard to keep the company at the frontier, and also we operate many products and platforms around the world that touch billions of people.

We have been pretty aggressive in deploying agentic workflows. Our internal security teams use agentic workflows to help detect vulnerabilities. And then how do you work to patch them? We have steadily seen over the last 2 years, as the model capabilities have progressed, we are able to detect more vulnerabilities and we’ve been working very hard to patch them.

I think Metos was a point of inflection of capturing that moment in time. They put out a model which is really well-built for that particular task and it was frontier there. What we are excited about — and this is part of the reason we are sharing maybe one of the undermentioned announcements today at I/O — is CodeMender. CodeMender is a product which we use internally and which we are building to share externally. It not only helps you identify the vulnerabilities, generate patches, test and verify that they work, and deploy them.

MATTHEW BERMAN: And it’s running 24/7.

SUNDAR PICHAI: That’s right.

MATTHEW BERMAN: Yeah.

SUNDAR PICHAI: And in real-time. We completed our recent acquisition of Wiz, so this is state-of-the-art in being able to do this real-time monitoring of vulnerabilities, et cetera. So I think the combination of what we have with Wiz and CodeMender — we are using it internally to stay at the frontier. And I think it’s an important moment for the industry.

I have to say I’m heartened by the cross-industry collaboration going on at this moment. I think one of the examples I would call out today, be it SynthID or watermarking, companies coming together around cyber — that is so important for this industry with this technology. So I’m encouraged by those trends as well.

Model Release Strategy: Open vs. Restricted

MATTHEW BERMAN: So you mentioned Mythos, so I want to talk about that for a moment. Obviously Anthropic decided not to release Mythos publicly, just to a handful of companies. We have OpenAI releasing GPT-5.5 Cyber. Which approach do you think is more appropriate? Which is right for Google? Is there some model that is just too good and you’re going to hold it back? Or is the more iterative deployment strategy that OpenAI takes more aligned with what Google believes?

SUNDAR PICHAI: Look, I think it depends on where you feel — it depends on the model capability. If it is not fundamentally changing what’s out there already in terms of the state of the art, I think it’s definitely okay to put it out. But in the security world, there’s a well-established practice. Google has done Project Zero for a long time. We have teams of people who find vulnerabilities, then we notify the vendor, give them 90 days to patch it before we acknowledge the vulnerability in the wild. And so there are well-established practices in the security industry around how to do it.

So I think it makes sense to me if you suddenly have something which dramatically changes the frontier — first of all, I think it’s important to work closely with the government on that. And you approach it in a responsible way. So I think that is consistent with how the security industry works. And I do think it’s important to also make sure enough people get access to it so they can patch their systems and so on. And they go hand in hand. So I think there’s validity to that approach.

MATTHEW BERMAN: Is there some threshold by which you would say, past that point, we can’t release it? Or is it more, let’s look at what the landscape of model capabilities are, let’s look at how cyber is right now, and let’s make the determination on a model-per-model basis?

SUNDAR PICHAI: That’s what I would say. Is the next one you’re introducing, does it dramatically change the frontier? Is it a 1 to 2% improvement over the current state of the art, or are you taking these 20% jumps? That’s where the judgment comes in, and I think that’s how I would change my approach.

Open Source Strategy at Google

MATTHEW BERMAN: So on the topic of model strategy, something near and dear to my heart is open source. Basically, Google and NVIDIA are the only companies with a real open-source model strategy nowadays. Google’s model is smaller, meant to run on edge devices. Why not release a large open-source frontier model?

SUNDAR PICHAI: Look, first of all, Google — we’ve been big fans of open source. Google was built on a lot of open-source systems. We have worked on many big things which are open source. I personally worked on Chromium and Android and so on, Kubernetes, and I can name many projects which Google has contributed pretty strongly to the world in open source.

In AI, we’ve been building Gemma models and we’ve been updating them year after year.

MATTHEW BERMAN: And they’re awesome.

SUNDAR PICHAI: And the recent release of Gemma 4 was a great release. So we are pushing it. I think all of us are trying to make sure the frontier takes a lot of investment to get the frontier done. You’ve seen our CapEx dollars, and you’re putting a lot of R&D dollars to generate those incremental frontier models. And you’re discovering new techniques as part of doing those models. So we all have to be mindful of that.

But I think we are also committed to making sure there’s an open-source ecosystem which is able to develop. And so we take a balanced approach there, and I think we’ll continue to take that balanced approach. That’s how I see it.

Open Source, China, and Model Strategy

MATTHEW BERMAN: And by the way, I do love the Gemma models. I run them locally at home. They are fantastic. So definitely, thank you. Obviously a company of the size and the resources of Google, if you’re making that decision of, okay, large closed source, large open source, we can’t do both. Probably most startups in the US also will struggle with that decision. What is your sense of the business model for open source in America? Is it viable right now?

SUNDAR PICHAI: Look, first of all, we not only do open source, part of the reason we invest so much in Flash and Flash models is so that we are giving a range of options, right? And those models are workhorses too to support. But your question on open source, it’s not that we don’t — we’ve been moving the open source frontier too. There are a lot of very, very good open source models, particularly from China, which startups are adopting too.

MATTHEW BERMAN: Yeah, we’re going to talk about that.

SUNDAR PICHAI: I think it depends, right? You go through moments in technology where the frontier moves so fast, maybe sometimes the open source may not be able to fully keep up with it. But then there are moments where open source will take leaps if the technology curve slows down or takes a break, right? It’s tough to fully predict the future, but I expect there to be a demand for a strong open source ecosystem, and I think we will definitely play a part in it, and I hope others do too.

MATTHEW BERMAN: There have been a number of open source ecosystems that have been really successful over many decades in technology, but I think just the upfront cost of baking a model makes it extremely difficult, especially if you’re putting out the model and then all of a sudden your competitors are serving inference at a higher margin than you. But I am hopeful. I do love open source, and we’re going to talk about the workhorse models in a moment, but I want to talk about China. They have been putting out incredible open source models. If you can put yourself in the shoes of another enterprise CEO and you’re looking at the landscape of which AI model to choose for your business, and you’re seeing DeepSeek at a fraction of the cost but still near the frontier, why wouldn’t America adopt Chinese open source AI? What’s the argument to go with American AI?

SUNDAR PICHAI: You’re saying why wouldn’t you just use the best open source models available? Look, at the end of the day, what are companies trying to do? They’re trying to solve problems. And they’re trying to solve a problem with a solution. So the question is, what are the solutions available? Let’s say you’re doing something in customer service — you want predictability, you want reliability, you want consistency, you want safety and security. So companies are optimizing for a lot of factors.

I think that gives a place for both open source models, and there’ll be providers who will take the open source models and build that ecosystem around it, which makes a lot of sense. There’ll be closed source models, and it’ll be an open marketplace, and people will have a lot of choice.

I am more okay if it is open source with the right licenses. It should matter less where it came from. I think over time there are good ways to inspect open source. I’m not saying that’s exactly true, particularly with how AI models are developed, but with open source comes a community which is responsible for it, cares about it. So if something wrong is happening in that software, it’s not like it’s going to go unnoticed. So I think that creates a level of trust for people to adopt that technology.

I worry less about whether we are adopting open source models from China, and more about whether we are doing enough in the US to make sure we are staying at the frontier. That’s how I think about it.

MATTHEW BERMAN: I know Google is big on co-design, full stack, and continuing on where the open source model is coming from. I’ve heard the argument that if it’s open source, it doesn’t necessarily matter — we’re going to fine-tune it or customize it for our needs. But ultimately, if we continue to build on top of China’s open source, there’s also an argument that they’re going to optimize their models for their own chips, and then all of a sudden we’re kind of built on another country’s technology. Is that an incorrect argument?

SUNDAR PICHAI: Look, I think the fundamentals of AI — the way people should be building use cases on top — is that because the models are changing so fast anyway, you need to build it in a way in which you’re able to evolve the models underneath, right? I think that has got to be the way you’re working through this moment. You have to be dynamic enough to adapt, because the model frontier and the model ecosystems are changing pretty sharply. That’s the way I would think about it now. And it’s too early to predict if this is a real concern or not.

The Workhorse Models: Google’s Flash Strategy

MATTHEW BERMAN: Okay, talking about model strategy — one of my favorite things is watching the frontier labs and seeing how their model strategies play out. Anthropic and OpenAI seem almost exclusively focused on the absolute frontier, and Google has that, but you guys also put a lot of emphasis on what I’ll call the workhorse class of models, the Flash class of models. Talk a little bit about why. Why is that such a big part of Google’s strategy?

SUNDAR PICHAI: Look, in our mission statement, we have this thing to make technology universally accessible and useful. We’ve always deeply cared that the most important technology in our lifetimes diffuses as broadly as possible. And we get really excited at driving efficiency and making sure the best models can work in the fastest possible way, cheaper, because we need to do it for Search — we have to give it to billions of people. We want to put it in Gemini. And so we want to give it to developers so that they can do powerful things with it.

We’ve had a lot of success with the strategy, and I think 3.5 Flash particularly — I made this point during the keynote — but I’ve heard anecdotally from a lot of CIOs who are so concerned about how much their companies are blowing through budgets.

MATTHEW BERMAN: Yeah.

SUNDAR PICHAI: And you can feel it talking to them. And I think the problem is going to get worse as we go through the year. That’s where I think the Flash model will really shine, because particularly in an agentic workflow where you need these things to be repeatedly used and used a lot of times, it’s so important to have a model which is very capable but is fast and efficient. Even accounting for token use, Flash is remarkably cost efficient. So I’m really excited. We are finding it internally — we are using it as a blend of Pro and Flash internally, and I think most companies should learn to use it that way.

To be very clear, we’re super committed to being at the frontier on every category. I’m excited for Pro, which we are working on, but I think Flash has a unique role to play in this compute-constrained time.

The Race to Self-Improving AI

MATTHEW BERMAN: Yeah, I agree completely, especially since most companies are not solving Math Olympiad problems. They’re not at the absolute cutting edge of science. They need real work done, and that is truly where the Flash model shines. Not everybody is token maxing or unconcerned with the budget. So I definitely appreciate that.

If you think about the future of AI, is it just truly a race to self-improving AI? Maybe just to play devil’s advocate for a second — the Flash family of models are great now, but ultimately whoever reaches self-improving AI first wins and then nothing else matters. Do you think about it like that?

SUNDAR PICHAI: Look, I think first of all, there’s a responsibility that comes with this technology. And I think we all need to be careful to avoid this race condition at all costs. We owe it to humanity to make sure we deploy this technology responsibly.

You are right in your question that there is this current moment where people feel like the curve is so steep and where you are in the curve matters. But just a few months ago when we launched 3.0, people were like, “Oh, we are so far at the frontier, no one will ever be able to catch up.” And I think at the frontier labs, it’s very dynamic. The competition is fierce. We all have our strengths and weaknesses. We all also have different cadences of our pre-training release cycles. So the peaks don’t exactly match. All this creates a perception gap, which shifts widely in like 4 to 6 weeks.

But I think a few labs are really at the frontier and then there’s a big gap. And I think there are scenarios in which things like recursive self-improvement come into play. But I think if they come into play, that’s no different from the cyber moment — we all have to handle those moments far more responsibly than today. So that goes hand in hand. The more AI becomes advanced, the more it’s a societal conversation versus a single company conversation.

Google’s Compute Challenges

MATTHEW BERMAN: Yeah, well said. So we know we’re talking a lot about models, but all of it is downstream from compute. I’m always both impressed and in awe of Google’s ability to serve — you’re serving your own models’ inference on the API, you’re also powering your suite of products with Gemini to literally billions of users, you’re also allowing your competitors to use your inference, and you’re also selling TPUs. I’ve heard the reason for this is that, as Thomas told me, “We planned really well,” which makes a lot of sense — you’ve been at this for 10+ years. I’ve also heard that Google’s revenue is literally constrained by compute. So I wanted to give you the opportunity: what is the state of compute at Google right now?

SUNDAR PICHAI: Look, all of us have made a set of bold, right decisions over the last few years to invest in compute and scale it up aggressively. But having said that, I don’t think any of us sit in the chair and say we’re satisfied. You look back and say, “I wish I had done a little bit more.” So we are living in one of those moments in time, and the costs are going up too. For a given budget, you may be getting less compute than you had previously planned for — memory prices, and what have you. So the costs are going up.

I think we plan well. We are able to plan long-term for Cloud separately from our own internal needs. We do long-range plans and we plan for it. And some of it is like when you’re in Google Cloud and you’re supporting customers, your customers may look at something and say, “I want access to that.” Like they may look at the demo of 3.5 Flash and say, “How are you exactly running it at 800 tokens per second? Could we get access to that?” So you’re also supporting customers through those journeys. But it is not an easy balancing act. We are constantly thinking as far ahead as possible and we are making trade-offs, like every other company right now.

MATTHEW BERMAN: And so do you have — maybe an obvious question — more demand than you have compute to serve it?

SUNDAR PICHAI: Absolutely.

MATTHEW BERMAN: What is the scale of that?

SUNDAR PICHAI: And hence the emphasis on something like 3.5 Flash even more, right? Could we have done an even better Omni model? Yes. But how can we do an Omni model which we can give to as many people as possible? So you’re constantly making trade-offs, including on whether you build a very large model — an ultra-sized model — that will increase the capability frontier, but then who all can you give it to? All of us are making these trade-offs, and sometimes you make a trade-off and it looks like, well, maybe you’ve done that trade-off a bit differently.

MATTHEW BERMAN: Yeah. So we only have a few seconds left. One last question for you: what is the main bottleneck for compute for Google right now? Is it land? Is it political will? What is it?

Bottlenecks in AI Infrastructure

SUNDAR PICHAI: I think the way it works is by definition bottlenecks work this way. If you think something is a bottleneck and you solve it, something else becomes a bottleneck, right? That’s the whole definition of bottlenecks.

I think at various times, you’re having a few areas — your ability to physically permit and construct data centers, the power they need, and then very quickly you get into the core components for these systems. They’re all the bottleneck. And it’s like you need all of that to work together. To get a square set of a chip you need, right? And so I feel like there are a few parallel bottlenecks going through and it almost doesn’t matter. There are a few of them which are bottlenecks. And at various times you may conclude it is memory, but then if everyone concludes it’s memory, tomorrow you turn around and say, okay, no, no, no, it’s actually this. But I think there are systemic bottlenecks across all layers of the stack now.

MATTHEW BERMAN: Yeah, whatever the bottleneck is in the moment, that’s the most compute you can have in that moment. Okay, everybody, please help me thank Sundar Pichai. Thank you so much.

SUNDAR PICHAI: Thank you, Berman.

Related Posts

• Full Transcript: Zanny Minton Beddoes Interviews Elon Musk
• Transcript of Daniel Kokotajlo Interview: Diary Of A CEO Podcast
• Transcript: Why a Social Media Ban Won’t Save Teens – Candice Odgers  
• Transcript of Elon Musk Interview: SXSW Live – March 9, 2013
• Digital Afterlife: Who Owns Your Data When You’re Gone? – Suresh Sankaran Srinivasan (Transcript)