by Reed Albergotti: It has been a pretty crazy week with all the product announcements, the quantum computing announcement. You have Nobel Prizes. You have new products, Gemini 2.0. Did you plan it for all this week, or did that just happen because of the crazy nature of AI?
Sundar Pichai: In 2015, I set the company [in] this AI-first direction. As part of that, we said we would do a deep, full-stack approach to AI all the way from world-class research, building the infrastructure. And then building models, both for us to developers, and putting it in our product. That kind of deep investment — and I felt AI is very profound — it cuts across everything we do as a company, across Google and Alphabet. That’s the foundation, taking a deep technical innovation approach, a full-stack approach. And then in the current genAI moment, sometimes you invest to get things right upfront. For me, that was getting Google DeepMind set up from the ground up, start Gemini, and build it to be natively multimodal and with long context, and now at 2.0.
So it’s getting that foundation right, and aligning teams, setting up the company. And I think you’re seeing… the benefits of all that beginning to ship, get out in the hands of people. I expect our pace to be higher because there’s always a fixed cost. We have to get our TPUs ready at scale to do the kind of models we want to build, build out our data centers, get the right teams. The innovation pipeline feels very, very strong. Some of these are long-term bets, and it takes time to play out. Quantum, to me, looks like where AI was in the 2010s. Few people know about it, but you’re working on it methodically. It’s the same. Waymo, we’ve been in the journey for over 15 years now, and it’s an exciting moment, so it’s probably a combination of all of that. The Nobel Prize wasn’t planned for, but I said this when it happened. I felt watching Demis [Hassabis] and John [Jumper], the team, work on AlphaFold… I was privileged to see a Nobel work from inception to finish. And so that’s been icing on the cake.
A year or two ago, the narrative was, really, Google got caught off guard by ChatGPT. This week is a good time to look at how far you’ve come in terms of the company and the way people look at Google. Do you feel a difference now?
It’s an exciting time. Internally, I had a palpable sense of the progress we were making. When you’re working on AI models, you’re looking at all these loss curves, and you’re looking at the capabilities of the models. You’re looking at various benchmarks. We are world-class people with access to state-of-the-art resources. The combination of Google DeepMind and Google Research, they’re the most cited in this genAI field. We’re responsible for many of the breakthroughs on which this revolution is happening. It’s definitely very satisfying to see the momentum, but we plan to do a lot more. We’re just getting started.
You said at DealBook last week that progress is getting harder, that the low hanging fruit has been picked. What does that mean exactly? What is the low-hanging fruit that’s been picked? And what is that really hard thing?
To be very clear about my answer, I said it there. I’m actually very excited about the progress ahead. What I meant by that is, I think in this field you can throw compute at it and make that initial progress. But then, it’s not just a question of scaling alone, it’s achieving breakthroughs. So getting our models to work with up to 2 million tokens as input length, that’s long context. That’s an example of a breakthrough. And with Gemini 2.0, we have a multimodal live API, so now it does native image and audio output. You can stream inputs into the model, get output — those are all breakthroughs. As we go to this next level, you need more insightful breakthroughs. The caliber of the work, I think, will be pretty high. All I meant to say is, I think it’ll help distinguish the really elite teams. And it’s not just us, there are a few other teams out there, but that’s what makes 2025 exciting.
I think people took that wrong. If I looked at some of the content after that, it was like, ‘Whoa, we’ve hit the plateau.’ What you’re saying is, at Google, when it gets hard, that’s where we shine?
Precisely. I see parallels with Waymo too. Many people were working on the problem, but then it got harder. When it gets harder, being able to work through that, to get to that next level, it’s important.
There’s this tendency to look at the last two years, from November 22 until now, as the curve. If you look at it that way, it does look like we’re hitting a plateau because you had this huge, out of nowhere — at least from the outside perspective — this new thing. If you zoom out and you look at that curve, I imagine there are little plateaus along the way that you’ve seen. Where do you see it going from here? Is it the same trajectory?
I still vividly recall the early 2010s and just understanding that this model can barely recognize images, and getting excited about it. Progress has been relentless over the last decade.It’s definitely broadened the wider public into the world now, so it’s mainstream. But when I look ahead into 2025, I definitely think we already have capable models enough now (that) we can build many, many use cases on top of it. That progress is going to be very real. With Gemini 2.0, we are laying the foundation for it to be more agentic. While it is still within the realm of research, we’re putting it out in the hands of the trusted testers, things like Project Mariner, to work on Chrome from inception. To watch a model being able to use the browser is pretty incredible to see, but we have to break through some barriers because we are in these fields where we have to do it safely, reliably. The saying goes, ‘the final 20% takes 80% of the effort.’ In this case, the last 10% may take 90% of the effort. But that’s why we have benchmarks. We’re making progress. We’re putting it out in the hands of trusted testers. That way we can responsibly test, get feedback, then we’ll give it to more people, and so on. But think about all the workflows in the world which AI can begin to influence. We may actually see dramatically more progress than what we have seen. Both are simultaneously true.
Speaking of the long-term approach, going multimodal with Gemini from the beginning seemed to sacrifice some other capabilities, maybe on the language and coding-specific benchmarks, for this multimodal approach. Is that right?
When we launched Gemini 1.2 we really wanted it from the ground up to be multimodal. I think our models were almost state-of-the-art in multimodality, but we hadn’t exposed the capabilities of the model. There was no native image out or audio out. With 2.0 we are unlocking those capabilities, but at the same time, we are becoming state-of-the-art in all the coding or reasoning, or so on. On the SWE-bench, which is a popular benchmark, our models are state-of-the-art now. We’ve thrown out other experimental models, which aren’t released yet, which have shown even more capabilities. We are definitely pushing the frontier, but we’ll do it responsibly, which is why you see some of it being in trusted tester mode, some of it only as experimental APIs for developers, but we’ll work hard, get feedback, and then take it to the next level.
The people of DeepMind have said this and theorized that the multimodel approach is the path to AGI because you need to have this world model, and maybe you need to have it an embodied AI, to be able to really reason and understand. Are you finding, as you progress, that’s true? Do you think that’s the right approach?
As humans, our experience with the world is incredibly multimodal, so it’s always made sense. This is why we did Google Lens for search. You shouldn’t always have to type if you could point your phone to something you’re looking at and ask the question. Lens gets billions of queries every month for us, it’s one of our fastest-growing use cases. It’s always been clear to me that that is the future of where things will go. Demis and (the) team, they’ve always had the strong vision. One of the things we launched as part of all this is, you can use it to help in games. Project, NaVi. If you’re a new gamer, it’s looking at what you’re doing and talking to you to give feedback. I think that’s the foundation of it. And down the line, when you look at things like robotics and stuff, it’s going to be incredibly important. It’s important in Waymo. Waymo is all about seeing the world around you and making decisions. Our work, which we are doing with these natively multimodal models, will intersect with Waymo and make Waymo even better over time.
When you get these multimodal products into the hands of billions of people, and you look at Astra, when that becomes widely available, does that become a really good source of training data? Is that an advantage for Google?
There’s nothing like real-world feedback across everything we do. People using Google Lens in search, people as they use Astra. I think that virtuous cycle becomes super important for our products. I think all that makes our products better. If you look at Waymo, for example, we simulated a lot, and then we drove in the real world. But now we are in the real world, deploying in cities and we are doing 175,000 rides per week, or a million miles. I think that’s the best way you can in creatively improve your product.
And 10 cities next year, that’s fast. Do you think that’s going to become a real source of revenue? And what’s the metric we should look at? Is it cost per mile?
The metric for us right now is making sure we’re building a generalized Waymo driver. The more situations we can take that, scale it, make it work safely with a very high bar, in urban environments, in highways, in all weather conditions. And then deliver it in various scenarios — in our cars, working with partners, like they’re doing with either Uber or with other players — and getting it to scale, and having a great user experience is what we are thinking about as we’re working through it.
One thing I did not think would happen is that I’d be referring to data centers by name, like Colossus, Rainier. You have this billion-dollar data center in Kansas City. Are you going to build one of these massive clusters, or have you already? And is there going to be a name?
We should have Gemini take over our naming from now on. Look, we are constantly pushing the state of the art in our data centers, and I think we have some of the most powerful clusters in the world. Things I’m proud of, most of the world is adopting liquid cooling. We’ve had widely deployed liquid cooling in our data centers for a while. We will be one of the first customers, not just TPUs, but working with Nvidia to get GB200s in our data centers. We already have a data center partially powered by geothermal. And our top, top data centers, many of them operate 90% carbon-free basis in terms of its energy use. So we are cutting edge, and we are scaling it up. And everything I see, everything we benchmark against, I think we are at the frontier there as well.
But you don’t talk about it as much as some of the others. Amazon announced hundreds of thousands of their Trainium2 chips in it. You trained Gemini 2.0 exclusively on your TPUs. Why don’t you go out there and brag about it?
I remember in the year 2017 or 2018 in Google IO talking about building AI-first data centers and showing our TPU parts. Maybe we’ve been doing it for a while. We are proud of what we’re doing, but I’ll take your suggestion and make sure we talk about it more.
Maybe it’s a trade secret; maybe if you’re in the lead, you don’t have to talk about it. Can you compare to some of these cluster that are considered among the biggest in the world?
To be very clear, there are a few companies doing it. What matters is, for generating your cutting-edge models for pre-training, you need these large clusters, ideally located concurrently. We are definitely at the cutting edge of it. I think we have some of the largest compute clusters available for Google DeepMind and for our cloud customers too. We’ll do a lot more there.
The incoming Trump administration has this “Manhattan Project” for AI idea. Have you gotten any insight into what that is going to look like and what part in it Alphabet will play?
It’s early days. The transition team is underway there, but the president has been very clear that he wants to invest in American technological leadership and critical technologies. From my standpoint, we’ve always done it over the years, but we want to help. You saw our announcement on quantum computing or AI, we’ve announced teams for building small modular nuclear reactors with our partners. Some of these are big, physical infrastructure projects. And I think there is a chance for us to work as a country together, to take these big, ambitious projects and go back to that. There’s no one in the world who wasn’t impressed by watching the SpaceX booster come back and land back that way. I think setting a high bar and pursuing these big, physical infrastructure projects and doing it well and fast. Tax rate progress is something we would be very excited by and happy to play any part we can.
Have you spoken with the new AI czar?
No, I’m looking forward to meeting David [Sacks], clearly there are people coming in who are experts in these areas, in the technology sector. I think that’ll be very, very helpful. We look forward to engaging there.
So no specifics yet, you don’t know exactly what that’ll look like?
That’s correct, other than early indications that they are definitely interested in driving innovation at scale. We are looking forward to those conversations.
There’s also export controls that limit where and how many chips can leave the country. You’re building an AI hub in Saudi Arabia. Do you see anything changing with regard to export controls in the new administration?
AI is a critical technology, so I think from a national security standpoint, there’ll be frameworks associated with it. We are committed to working with the right people. But what you’ve seen over the last couple of years is consumers, enterprises, governments — people are excited about the possible use cases. So getting AI deployed in various useful scenarios is really important. It can help drive productivity, it can drive economic growth. And above all, as a society, we need to learn to use the technology, adapt, and begin conversations.
You had this big breakthrough from the Quantum AI team on error correction. Did you expect that? What was your reaction?
One of the exciting things is our quantum team, with Hartmut [Neven]and team, they’ve always had a very rigorous framework, and they’ve defined progress in clear milestones. Every time we get a milestone, I’m like, these are ambitious projects. This one has definitely been one of the more positive surprises. This is definitely a deeper breakthrough, tackling error correction while you’re scaling up in your quantum computer. It’s definitely been one of the tougher challenges in the field. I couldn’t be more pleased with it. But for us now, these are all milestones in the way we are focused on developing practically usable quantum computers, which we can apply to new novel use cases. That’s the goal. I would compare it to our journey on Waymo or AI in the sense that it’s going to take time. But I think progress is inevitable if you put your mind to it.
You said quantum is like AI in 2010. That means that pretty soon it’s going to start having a real impact. What does achieving an at-scale quantum computer mean for Alphabet?
To your earlier question, classical computing, or super computers are getting more and more powerful. But I think quantum will, for certain types of use cases, end up playing a powerful role. It’ll be an important tool in our arsenal. And down the line, the intersection of quantum and AI is very exciting to us. We published the state-of-the-art weather forecasting models with GenCast. But in a future when we can use quantum computing, you shouldn’t underestimate our ability to predict these things on a much deeper, better scale. These are profound implications on a practical basis. Some things like, what AlphaFold did, what more can you do to understand nature, simulate nature, and all of that will have practical applications. And then there’s always this deeper, both through AI and quantum, to the extent we are more deeply trying to understand the nature and fabric of the universe we live in. I think it gives us the best shot because the universe is fundamentally quantum. And so there are deeper implications of making progress as well.
It’s transformational for the world, but potentially, it could really feed a lot of these other projects.
My goal is, in a five-year time frame, we are commercially applying quantum to tackle some use cases, and then from there on, you build on out.
You mentioned that AI search is going to be a bigger deal next year. AI Overview is already becoming very useful for a lot of my searches. Can you go into more detail on that?
We’ve been very excited with the evolution of search with AI Overviews. I’ve been using AI overviews with Gemini 2.0 flash, and I can already see improvements in it. And that’s something we’re going to get out to more people. But we are also going to do more with it. Our AI models will help us in search, build experiences for more complex, deeper queries where you have to break it down and help the user iterate, and get to deeper answers. In 2025 we’ll definitely innovate rapidly where search will do things which you couldn’t do in 2024. That’s the goal I have for the team: A class of questions which search noticeably improves in 2025 over 2024, and I think we’ll deliver that. That’s exciting because that means you’re pushing the frontier of knowledge and information. I’m looking forward to all that getting out in the hands of users.
AI safety is another question that a lot of people bring up. Demis said maybe a year ago that one of the big risks as the race heats up, is you start to take resources away from the safety effort because you need as many resources as you can to win this race. Is that materializing, or can you talk specifically about how many people are working on safety, or how many compute resources? Is there a way to measure that?
We’ve always felt this is an area where it’s an add where you drive innovation, but what will help you drive progress is incorporating safety from the very beginning. I think one of the advantages of being able to think long term and invest for the long term is we are investing as much in the foundational safety of these models, the underlying research you need to do to drive that safety. This is why, for example, building synthetic, open sourcing aspects of it — these are all ways by which we are pushing the boundary on safety. We are all investing in safety frameworks as we make these models more agentic. But I think safety and innovation go hand in hand. It’s what has helped us make more progress in Waymo. Because from day one, we deeply incorporated safety in our innovation and development practices, and they go hand in hand. Similarly in AI, we have more people working on AI safety than before, including access to more compute. It is something we’ll always be very, very committed to.
One question on this antitrust I thought was interesting. When I did that demo of Mariner, it’s on Chrome, and that’s what the US government wants to force Google to spin off. If you look at the administration’s appointments, it doesn’t look like that’s going to go away. I just have to ask, what does Google look like without Chrome?
This is an important process, we’ll participate constructively. I do think judges acknowledge that we’ve been innovating and we build the best products. I think some of the remedy proposals are far-reaching in scope. We plan to make a strong case. You’ve just seen the innovation coming out. All of this benefits consumers and at the end of the day, that has to be the foundation. That’s what our laws are based on. As long as we continue staying true to that approach, to bring beneficial things for our users, I think we’ll end up doing well.
You talked a little about geothermal. I know you spun out this direct air capture company. I’m just wondering if there’s anything that excites you on that front, what type of energy is going to fuel this? And are there going to be new renewable innovations that come out of this?
We are tremendously undertapped on the potential of solar. There’s a lot more opportunity to scale up solar. Thinking about it from a physics and engineering perspective, we have so many options. Nuclear is proven, there are countries which have proven it and it works today. I think there are safer options which are constantly being worked on. There is so much energy inside the Earth too. We only live on the surface of the Earth, and there’s volumetric energy within the planet, which we barely tapped. I’ve always felt if you put your mind to it, we should be dealing with the energy surplus. Energy should be an accelerant, not a constraint. It’s only our imagination and result that’s in the way.
I was looking at a map of geothermal potential sites and it’s huge, the whole western US. You have one, but it’s still pretty low. It’s not gigawatts, it’s megawatts. Do you just have to dig deeper?
The demand for energy is so great, I think we’ll be able to meet it. But we’ll need more R&D dollars. We’ll need more favorable permitting to go actually work on these things. And this is an area where I think there’s a real opportunity for the new administration. They’ve signaled their commitment to making that possible. It’s something we all can make a lot of progress on.
