Google DeepMind 首席执行官 Demis Hassabis 谈生活在 AI 未来｜EP 137

This year, Google talked about AI very differently. This time, they want you to sit up. They want you to lean in. They want you to pay them $250. And they want you to get to work.

I've been working every hour there is for the last 20 years because I've felt how important and momentous this technology would be, whether it's 5 years, 10 years, or 2 years. They're all actually quite short timelines when you're discussing the enormity of the transformation this technology is going to bring.

When I see a Van Gogh, the hairs go up the back of my spine because I remember what he went through and the struggle to produce that in every one of Van Gogh's brushstrokes. Even if AI mimicked that and you were told that it was—so what?

Now, there's a very large, what looks like a circus tent, over there. What do you think's going on in there?

That is Shoreline Amphitheatre.

Oh, that's the amphitheater under that tent. Yesterday, I thought that was just some carnival they were setting up for employees.

Okay, my mistake. I thought Ringling Brothers had entered into a partnership with Google. It's a revival tent. They're bringing Christianity back.

I'm Kevin Roose, a tech columnist at the New York Times. I'm Casey Newton from Platformer and this is Hardfork. This week, we're taking a field trip to Google. We'll tell you all about everything the company announced at its biggest show of the year. Then Google DeepMind CEO Demis Hassabis returns to the show to discuss the road to AGI, the future of education, and what life could look like in 2030.

Kevin being very old, for starters.

Somebody did text me to ask why I freaking yell the name of the show every episode.

And did you say it's because I started yelling my name?

I said it's because of the cold brew. Well, Casey, our decor is a little different this week.

Mm-hmm. I'll say it: it looks better.

Yes. We are not in our normal studio in San Francisco. We are down in Mountain View, California, where we are inside Google's headquarters.

I'm just thrilled to be sitting here surrounded by so much training data. That's what they call books here at Google.

So, we are here because this week is Google's annual developer conference, Google I/O. There were many, many announcements from a parade of Google executives about all the AI stuff they have coming. We're going to talk in a little bit with Demis Hassabis, who is the CEO of Google DeepMind, essentially its AI division, and who's been driving a lot of these AI projects forward. But first, let's set the scene for people, because I don't think we have ever been together at an I/O before. So, what is it like?

Google I/O has a bit of a festival atmosphere. It takes place at the Shoreline Amphitheatre, which is a concert venue, but once a year it gets transformed into a sort of nerd concert where, instead of seeing musicians perform, you see Google employees vibe-coding on stage.

Yes, there was a vibe-coding demo. There were many other things. I did actually see Googapella, Google's a cappella group, warming up in anticipation of a concert. So, you've got some old-school Google vibes here, but also a lot of excitement around all the AI stuff.

I didn't see Googapella perform. Where was this performance?

I didn't see them perform either. I just saw them warming up. They were sort of doing their scales. They sounded great. You know what? I bet it was a classic a cappella situation where they warmed up and someone came up to them and said, “Please don't perform.”

All right, Kevin. Before we get into it, shall we say our disclosures?

Yes. I work for The New York Times, which is suing OpenAI and Microsoft over copyright violations related to the training of AI systems. And my boyfriend works at Anthropic, a Google investment.

Oh, that's right.

Yeah. So, let's talk about some of what was announced this week. There was so much. We can't get to all of it, but what were the highlights from your perspective?

Well, look, I wrote a column about this, Kevin. I felt a little bit like I was in a fever dream at this conference. I think often it is the case at a developer conference that they'll try to break it out into 1, 2, 3 big bullet points. This one felt a little bit like a fire hose of stuff.

By the end, I'm looking at my notes saying, “Okay, so email's going to start writing in my voice, and I can turn my PDFs into video TED Talks. Sure, why not?” So, I had a little bit of fever-dream mentality. What was your feeling?

Yeah. I told someone yesterday that I thought the name of the event should have been “Everything Everywhere All at Once.” That did actually feel like what they were saying: every Google product that you use is going to have more AI, that AI is going to be better, and it's all going to make your life better in various ways. But it was a lot to keep track of.

Yeah. I mean, look, if we were going to try to pull out one very obvious theme from everything that we saw, it was: AI is coming to all of the things. And it's probably worth drilling down a little bit into what some of those things are.

Yeah. So, the thing that got my attention—and I was sitting right next to you—the one time when I really noticed you perking up was when they started talking about this new AI Mode in Google Search, their core search product. So, talk about AI Mode and what they announced yesterday.

So, Kevin, this gets a little confusing because there are now 3 different kinds of major Google searches. I would say there is the normal Google Search, which is now augmented in many cases by what they call AI Overviews, which is sort of an AI answer at the top.

Yeah, that's the little thing that will tell you the meaning of phrases like “You can't lick a badger twice,” right?

That's right. And if you don't know the meaning of that, Google it. That's thing 1. Thing 2 is the Gemini app, which is kind of a 1-for-1 ChatGPT competitor. That's in its own standalone app, standalone website. And then the big thing that they announced this week was AI Mode, which has been in testing for a little while.

I think this sort of lands in between the first 2 things, right? It is a tab now within Search. This is rolling out to everybody in the United States and a few other countries. You can have the sort of longer, multi-step questions that you might have with Gemini or ChatGPT, but you can do it right from the Google Search interface.

Yeah, and I've been playing with this feature for a few weeks now. It was in their Labs section, so you could try it out if you were enrolled in that. And it's really nice. It's a very clean thing. There's no ads yet. They will probably appear soon.

It does this thing called the fan-out, which is very funny to me. You ask it a question, and it dispatches a bunch of different Google searches to crawl a bunch of different web pages and bring you back the answer. It actually tells you how many searches it is doing and how many different websites it is searching. So, I asked it, for example, “How much does a Costco membership cost?” It searched 72 websites for the answer to that question. So, AI Mode is very eager to answer your question, even if it does verge on overkill sometimes.

Yeah. Well, you and I had a chance to meet with Robby Stein, who is one of the people leading AI Mode, and I was surprised by how enthusiastic you were. Like you said, you've really found this quite useful in a way that I think I have not so far. So, what are you noticing about this?

The main thing is that it's just such a clean experience. On a regular Google Search results page—we've talked about this—it has just gotten very cluttered. There's a lot of stuff there: ads, carousels of images, sometimes a shopping module, sometimes a Maps module. It's just hard to actually find the blue links sometimes.

I imagine that AI Mode will become more cluttered as they try to make more money off of it. But right now, if you go to it, it's a much simpler experience. It's much easier to find what you're looking for.

Yeah. And at the same time, they're also trying to do some really interestingly complex stuff. One of the things they showed off during the keynote was somebody asking a question about baseball statistics that required finding 3 or 4 different tricky-to-locate stats and then combining them all together in an interactive chart. That was just a demo. We don't have access to that yet. But that is one of those things where it's like, well, if that works, that could be a meaningful improvement to search.

Yeah, it could be a meaningful improvement to search. And we should also say that it's a big unknown how all of this will affect the main Google Search product, right? For now, it's a tab. They have not merged it into the main, core Google Search, in part because it's not monetized yet. It costs a lot more to serve those results than a traditional Google search.

But I imagine over time these things will kind of merge, which will have lots of implications for publishers, people who make things on the internet, and the whole economic model of the internet.

Before we get dragged down that rabbit hole, let's just talk about a few other things that they said on stage at Google I/O. So, I was really struck by the usage numbers that they trotted out for their products. Gemini, according to them, the app now has 400 million monthly users. That is a lot. That is not quite as many as ChatGPT, but it is a lot more than products like Claude and other AI chatbots.

They said that the number of tokens being output by Gemini has increased 50 times since last year. People are really using this stuff. In other words, this is not just some feature that Google is shoving into these products that people are trying to navigate around. People are really using Gemini.

I think that’s right. And I think it’s the Gemini number in particular that struck me. Four hundred million is a lot of people, and I don’t see many obvious ways that Google could be faking that statistic. In contrast, for example, they said 1.5 billion people see AI Overviews every month. It’s like, well, yeah, you just put them in Google search results. That’s an entirely passive phenomenon. But with Gemini, you have to go to the website or download the app, so that tells me that people are actually finding real utility there. So that’s Gemini, but they also released a bunch of other stuff, like new image and video models. Do you want to talk about those?

Yeah. Like the other companies, they’re working on text-to-image and text-to-video, and while OpenAI’s models have gotten most of the attention in this regard, Google’s are quite good. I think the marquee feature for this year’s I/O is that the video-generating model Veo 3 can also generate sound. They showed us a demo, for example, of an owl flapping its wings. You hear the wings flap, and it comes down to the ground. There’s this sort of nervous badger character, and they exchange some dialogue that was basically incomprehensible, just pure sloth, but they were able to generate that from scratch.

The dialogue was: “They left behind a ball today. It bounced higher than I can jump. What manner of magic is that?” I guess that’s something.

Yeah. They also announced a new Ultra subscription to Google’s AI products. Now, if you want to be on the bleeding edge of Google’s AI offerings, you can pay $250 a month for Gemini Ultra. And Casey, I thought to myself, no one is going to do this. Who is going to pay $250 a month? That’s a fortune for access to Google’s leading AI products. And then I look over to my right, and there’s Casey Newton in the middle of the keynote, pulling out his credit card from his wallet and entering it to buy a subscription to this extremely expensive AI product. So, you might have been the first customer of this product. Why?

Well, I hope that they don’t forget that when it comes time to feed me into the large language model. Look, I want to be able to have the latest models. And I think one clever thing that these AI companies are doing is saying, “We will give you the latest and greatest before everyone else, but you have to pay us a ridiculous amount of money.” If you’re a reporter and you’re reporting about this stuff every day, I do think you sort of want to be in that camp.

Is it true that I now spend more on monthly AI subscriptions than I paid for my apartment in Phoenix in 2010? Yes. And I don’t feel great about it, but I’m trying to be a good journalist. Kevin, please. Your family is dying.

Another thing that made me perk up was that they talked a lot about personalization, right? This is something we’ve been talking about for years. Basically, Google has billions of people’s emails, search histories, calendars and all their personal information, and we’ve been waiting for them to start weaving that stuff in so that you can use Gemini to do things in those products. That has been slow, but they are taking baby steps. They showed off a few things, including this new personalized Smart Replies feature that will be available for subscribers later this year in Gmail. Instead of just getting the formulaic suggested replies at the bottom of an email, it’ll actually learn from how you write. Maybe it can access some things in your calendar or your documents and suggest a better reply. You’ll still have to hit send, but it’ll prepopulate a message for you.

Yeah, I have to say I’m somewhat bearish on this one, Kevin, only because I think that if this were easy, it would just be here already, right? When you think about how formulaic so much email is, it doesn’t seem to me like it should be that hard to figure out what kind of emailer you are. I’m basically a 2-sentence emailer. That doesn’t seem like it should be hard to mimic. So that’s an area where I’ve been a little surprised and disappointed.

We also know that large language models do not have large memories. One thing that I would love for Gmail to do, but it cannot, is understand all of my email and use that to inform the tone of my voice. But it can’t do that. It can only take a much more limited subset. Is that going to make it difficult to accurately mimic my tone? I don’t know. What I’m trying to say here is that I think there are a lot of problems here, and my expectations are pretty low on this one.

Yeah, that was the part where I was like, I will believe that this exists and is good when I can use it. But as with other companies, like Apple, which demoed a bunch of AI features at its developer conference last year and then never launched half of them, I have become a little skeptical until I can actually use the thing myself.

Yeah, it really is amazing how, looking back, last year’s WWDC was just a movie about what a competent AI company might have done in an alternate future. It had very little bearing on our reality, but it was admittedly an interesting set of proposals.

Okay, so that is the software AI portion of I/O. There was also a demo of a new hardware product that Google is working on: Android XR glasses. Basically, they’re Google’s version of what Meta has been showing off with its Orion glasses. You have a pair of glasses with chunky black frames and a holographic lens in them, and you can actually see a little thing overlaid on your vision telling you what the weather is, what time it is or that you have a new message.

They showed off an integration with Google Maps where it’ll show you a little miniature Google Maps inside your glasses. It’ll turn as you turn and tell you where to go. They did say this is a prototype, but what did you make of this?

Well, I think a lot of it looked really cool. Probably my favorite part of the demo was when the person demonstrating it looked down at her feet because she was getting ready to walk to a coffee shop, and the Google map was actually projected at her feet. She knew, “Okay, go to the left, go to the right.” If you’ve ever been walking around a foreign city and desperately wanted this feature, I think you would see that and be pretty excited. What did you think?

Yeah, I thought to myself, Google Glass is back. It was away for so long in the wilderness, and now it’s back. It might actually work this time.

Absolutely. I did get to try the glasses. There was a very long line for the demo, but—

Let me guess. You said, “I’m Kevin Roose. Let me in the front of the line.”

No, they made me wait for 2 hours. I mean, I didn’t literally wait for 2 hours. I went and did some stuff and then came back. But I got my demo. It was 5 minutes long. It was pretty basic, but it is cool. You can look around and say, “Hey, what’s this plant?” Gemini will look at what you’re seeing and tell you what the plant is.

Totally. I did a demo a few months back and also really enjoyed it. So I think there’s something here.

And I think, more importantly, Kevin, consumers now, when they look at Google and Meta, finally have a choice: Whose advertising monopoly do I want to feed with my personal data? You have consumer choice now, and I think that’s beautiful. That’s what capitalism is all about.

So, those are some of the announcements. What did you make of the overall tenor of the event? What stuck out to you as far as the vibe?

The thing that stuck out to me the most was contrasting it with last year’s event. Last year, they had this phrase that they kept repeating: “Let Google do the Googling for you.” To me, that put me in the mind of somebody leaning back into a floating chair from the movie WALL-E and just letting the AI run roughshod over your life.

This year, Google talked about AI very differently. This time, they want you to sit up. They want you to lean in. They want you to pay them $250, and they want you to get to work. AI is your superpower. It’s your bionic arm, and you’re going to use it to get further and faster than ever before.

But even while presenting that vision, Kevin, they were also very much like, “But it’s going to be normal. It’s going to be chill. It’s going to be kind of like your life is now.” You’re still going to be in the backyard with your kids doing science experiments. You’re still going to be planning a girls’ weekend in Nashville, right? There wasn’t really a lot of science fiction here. There was just a little bit of, “Oh, we put a little bit of AI in this.” So that was interesting to me.

Yeah, I had a slightly different take, which is that I think Google is being AGI-pilled. For years now, Google has distanced itself from the conversation about AGI. It had DeepMind, which was sort of its AGI division, but they were over in London and were a separate thing. People at Google would not exactly laugh, but kind of chuckle when you asked them about AGI.

It just didn't seem real to them, or it was so remote that it wasn't worth considering. They would say, “What does this have to do with search advertising?” Exactly. So now, it's still the case that this is a company that wants you to think about it as a product company, a search company. They're not going all in on AGI, but once you start looking for it, you do see that the culture of AI, and how people at Google talk about AI, has really been shifting.

It is starting to seep into the conversation here in a way that I think is unusual and maybe indicative that the technology is just getting better faster than even a lot of people at Google were thinking it would.

So, I don't totally agree with you, Kevin, because while I'm sure that they're having more conversations about AGI here than they were a year ago, when you look at what they're building, it doesn't seem like there's been a lot of “rip it up and start again.” It seems a lot like, “How do we plug AI systems into Google-shaped holes?” Maybe that will eventually ladder up to something like AGI, but I don't think we've seen it quite yet.

The other observation I would make is that I think the Google of 2025 has a lot more swagger and confidence when it comes to AI than the Google of 2024 or 2023. Two years ago, Google was still trying to make Bard a thing, and I think they were feeling very insecure that OpenAI had beaten them to a consumer chatbot that had found some mass adoption.

They were just playing catch-up. I don't think anyone would have said that Google was in the lead when it came to generative AI just a few years ago. But now they feel like there is a race and that they are in a good position to win it. They were talking about how Gemini stacks up well against all these other models. It's at the top of this leaderboard, LMArena, for all these different categories.

I don't love the way that AI is sometimes covered as if it were sports—who's up, who's down, who's winning, who's losing. But I do feel like Google has the confidence now when it comes to AI of a team that knows it's going to be in the playoffs, at least. And that was evident.

Oh, yeah. Well, when you look at the competition—just what's happened over the past year—you have Apple doing a bunch of essentially fictional demos at WWDC, and you have Meta cheating to win at LMArena, making 27 different versions of a model just to come up with one that would be good at one thing, right? So I think if you're Google, you're looking at that and thinking, “I could be those guys.”

So that is what it felt like inside Google I/O. What was the reaction from outside? I noticed, for example, that the company's stock actually fell—not by a lot, but to a degree that suggested that Wall Street was kind of meh on a lot of what was announced. What was the reaction like outside of Google?

I think the external reaction that I saw was struggling a little bit to connect the dots, right? That is the issue with announcing so many things during a 2-hour period: Sometimes people don't have that one thing they're taking away and saying, “I can't wait to try that.” When you're just looking at a bunch of Google products that you're already using, I think if you're an investor, it's probably hard to understand. “I don't understand why this is unlocking so much more value at Google.”

Maybe millions of people are going to spend $250 a month on Gemini Ultra, but unless that happens, I can understand why some people feel like, “This feels a little like the status quo.”

Yeah, I see that. I also think there are many unanswered questions about how all of this will be monetized. Google has built one of the most profitable products in the history of capitalism in the Google search engine and the advertising business that supports it.

It is not clear to me that whatever AI Mode becomes, or whatever AI features it can jam into search—if search as a category is just declining across the board, if people are not going to Google.com to look things up the way they were a few years ago—I think it's an open question what the next thing is and whether Google can seize on it as effectively as it did with search.

Well, I think they gave us one vision of what that might be, and that is shopping. A significant portion of the keynote was devoted to one executive talking about a new shopping experience inside Google where you can take a picture of yourself, upload it, and then virtually try things on. It will use AI to understand your proportions and accurately map a garment onto you.

There was a lot of stuff in there that would just let Google take a cut, right? Obviously, you can advertise the individual thing to buy. Maybe you're taking some sort of cut of the payment. There's an affiliate fee in there somewhere.

One of the things I'm trying to do as I cover Google going forward is understand that, yes, search is the core, but Gemini could be a springboard to build a lot of other really valuable businesses.

An important question I know that I always ask you when I go to these things: How was the food?

Let's see. I think the food was really nice. Here's the thing: Last year it was a purely savory experience at breakfast, and I am, shamefully, an American who likes a little sweet treat when I wake up. This year they had both bagels and an apple-cinnamon coffee cake, so when I was heading into that keynote, I was in a pretty good mood.

I had some of them. They have little bottles of cold brew, and I'm a huge caffeine addict, so I took 2 of them. Boy, I was on rocket fuel all day. I was just humming around. I was bouncing off the wall. I was doing parkour. I was feeling great.

I thought I saw you warming up with the a cappella team. Now it all makes sense.

Demis, welcome back to Hard Fork.

Thanks for having me again.

A lot has happened since the last time you were on the show. Most notably, you won a Nobel Prize. Congrats on that. Ours must still be in the mail. Can you put in a good word for next year with the committee?

I will do. I will do.

I imagine it's very exciting to win a Nobel Prize. I know that has been a goal of yours for a long time. I imagine it also leads to a lot of people giving you crap during everyday activities, like if you're struggling to work the printer and people are just like, “Oh, Mr. Nobel Laureate, does that happen?”

A little bit. I tried to say, “Look, I can't.” Maybe it's a good excuse not to have to fix those kinds of things, right? So it's more of a shield.

You just had Google I/O, and it was really the Gemini show. I think Gemini's name was mentioned something like 95 times in the keynote. Of all the stuff that was announced, what do you think will be the biggest deal for the average user?

Wow. Well, we did announce a lot of things. I think, for the average user, it's the new powerful models, and I hope it's this Astra-type technology coming into Gemini Live. I think it's really magical, actually, when people use it for the first time and realize that AI is already capable today of doing much more than they thought.

Then I guess Veo 3 was the biggest announcement of the show, probably, and seems to be going viral now. That's pretty exciting as well, I think.

One thing that struck me about I/O this year compared to previous years is that it seems like Google is getting AGI-pilled, as they say. I remember interviewing researchers at Google even a couple of years ago, and there was a little taboo about talking about AGI. They would sort of be like, “Oh, that's like Demis and his DeepMind people in London. That's sort of their crazy thing that they're excited about, but here we're doing real research.”

But now you've got senior Google executives talking openly about it. What explains that shift?

I think it's the AI part of the equation becoming more and more central. I sometimes describe Google DeepMind now as the engine room of Google, and I think you saw that probably in the keynote yesterday.

If you take a step back, then it's very clear. You could say AGI is maybe the right word: We're quite close to this human-level general intelligence, maybe closer than people thought even a couple of years ago. It's going to have broad, cross-cutting impact, and I think that's another thing that you saw at the keynote. It's literally popping up everywhere because it's this horizontal layer that's going to underpin everything.

I think everyone is starting to understand that, and maybe a bit of the DeepMind ethos is bleeding into general Google, which is great.

You mentioned that Project Astra is powering some things that maybe people don't even realize AI can do yet. I think this speaks to a real challenge in the AI business right now, which is that the models have these pretty amazing capabilities, but either the products aren't selling them or the users just haven't figured them out yet. How are you thinking about that challenge, and how much do you bring yourself to the product question as opposed to the research question?

Yeah, it's a great question. One of the challenges of this space is that the underlying technology is moving unbelievably fast, and I think that's quite different even from the other big revolutionary technologies, the internet and mobile. At some point, you get some stabilization of the tech stack so that the focus can be on the product, right, or on exploiting that tech stack.

What we've got here, which I think is very unusual but also quite exciting from a researcher perspective, is that the tech stack itself is evolving incredibly fast, as you guys know.

So I think that makes it uniquely challenging, actually, on the product side. Not just for us at Google DeepMind, but for startups, for anyone really—for any company, small and large. Where do you bet right now when that could be 100 percent better in a year, as we've seen? And so you've got this interesting thing where you need fairly deeply technical product people, product designers and managers, in order to anticipate where the technology may be in a year.

There are things it can't do today, and you want to design a product that's going to come out in a year. So you've got to have a pretty deep understanding of the technology and where it might go to work out what features you can rely on. It's an interesting one. I think that's why you're seeing so many different things being tried out, and then, if something works, we've got to really double down quickly on that.

During your keynote, you talked about Gemini powering both productivity-assistant-style stuff and fundamental science and research challenges, and I wonder: In your mind, is that the same problem that one great model can solve, or are those very different problems that just require different approaches?

I think, when you look at it, it looks like an incredible breadth of things, which is true. How are these things related, other than the fact that I'm interested in all of them? That was always the idea with building general intelligence—truly general intelligence—and in this way that we're doing it, it should be applicable to almost anything, right? Whether that's productivity, which is very exciting and could help billions of people in their everyday lives, or cracking some of the biggest problems in science.

Ninety percent of it, I would say, is the underlying core general models—in our case, Gemini, especially Gemini 2.5. In most of these areas, you still need additional applied research or a little bit of special casing from the domain. Maybe it's special data or whatever to tackle that problem. We might work with domain experts in the scientific areas, but underlying it all, when you crack one of those areas, you can also put those learnings back into the general model, and then the general model gets better and better.

So it's a very interesting flywheel, and it's great fun for someone like me who's very interested in many things. You get to use this technology and go into almost any field that you find interesting.

A thing that a lot of AI companies are wrestling with right now is how many resources to devote to the core AI push on foundation models—making the models better at the basic level—versus how much time, energy and money you spend trying to spin out parts of that, commercialize them and turn them into products.

I imagine this is both a resources challenge and a personnel challenge. Say you join DeepMind as an engineer and you want to build AGI, and then someone from Google comes to you and says, “We actually want your help building the shopping thing that's going to let people try on clothes.” Is that a challenging conversation to have with people who joined for one reason and maybe are asked to work on something else?

Well, we don't have to—one advantage of being quite large is that there are enough engineers on the product teams and in the product areas who can deal with product development. And the researchers, if they want to stay in core research, absolutely—that's fine, and we need that.

But actually, you'll find a lot of researchers are quite motivated by real-world impact, be that in medicine, obviously, and things like Isomorphic Labs, but also having billions of people use their research. It's actually really motivating, and there are plenty of people who like to do both. So we don't need to pivot people to certain things.

You did a panel yesterday with Sergey Brin, Google's co-founder, who has been working on this stuff back in the office. Interestingly, he has shorter AGI timelines than you. He thought AGI would arrive before 2030, and you said just after. He actually accused you of sandbagging—basically, of artificially pushing out your estimates so that you could underpromise and overdeliver.

I'm curious about that because you will often hear people at different AI companies arguing about when the timelines are. Presumably, you and Sergey have access to all the same information and road maps, and you understand what's possible and what's not. So what is he seeing that you're not, or vice versa, that leads you to different conclusions about when AGI is going to arrive?

Look, first of all, there wasn't that much difference in our timelines. If he's just before 2030 and I'm just after, our timelines are pretty similar. My timeline has been pretty consistent since the start of DeepMind in 2010. We thought it was roughly a 20-year mission, and amazingly, we're on track. It's somewhere around then, I would think.

I have, obviously, a probability distribution of where the most mass of that is, between 5 and 10 years from now. Partly, it's to do with the fact that predicting anything precisely 5 to 10 years out is very difficult, so there are uncertainty bars around that. There's also uncertainty about how many more breakthroughs are required and about the definition of AGI.

I have quite a high bar, which I've always had. It should be able to do all of the things that the human brain can do, even theoretically. That's a higher bar than what a typical individual human could do, which is obviously very economically important and would be a big milestone, but in my view, it wouldn't be enough to call it AGI.

We talked onstage a little bit about what is missing from today's systems: true out-of-the-box invention and thinking. Inventing a conjecture rather than just solving a math conjecture—solving one is pretty good, but actually inventing the Riemann hypothesis, or something as significant as that that mathematicians agree is really important, is much harder.

Consistency is also a requirement of generality, really. It should be very, very difficult for even top experts to find flaws, especially trivial flaws, in these systems, which we can easily find today. The average person can do that. So there's a capabilities gap and a consistency gap before we get to what I would consider AGI.

And when you think about closing that gap, do you think it arrives via incremental 2 to 5 percent improvements in each successive model, stacked up over a long period of time? Or do you think it's more likely that we'll hit some sort of technological breakthrough, and then all of a sudden there's liftoff and we hit some sort of intelligence explosion?

I think it could be both, and for sure, both are going to be useful. That's why we push unbelievably hard on scaling and on what you would call incremental improvements—although there's a lot of innovation even in that—to keep moving things forward: pre-training, post-training, inference-time compute, all of that stack. There's actually lots of exciting research, and we showed some of that—the diffusion model, the Deep Think model.

So we're innovating at all parts of that traditional stack, should we call it. And on top of that, we're doing more greenfield things, more blue-sky things, like AlphaEvolve, maybe you could include in that.

Is there a difference between a greenfield thing and a blue-sky thing?

I'm not sure. Maybe they're pretty similar. So, some new area, let's call it. And then that could come back into the main branch, right?

As you both know, I've been a fundamental believer in foundational research. We've always had the broadest, deepest research bench, I think, of any lab out there. That's what allowed us to do past big breakthroughs: obviously, Transformers, but also AlphaGo, AlphaZero, distillation—all of these things.

To the extent any of those things are needed again—another big breakthrough of that level—I would back us to do that. We're pursuing lots of very exciting avenues that could bring that sort of step change, as well as the incremental improvements. They also interact, because the better your base models are, the more things you can try on top of them. Again, with AlphaEvolve, you can add evolutionary programming, in that case, on top of the LLMs.

We recently talked to Karen Hao, who's a journalist who just wrote a book about AI, and she was making an argument essentially against scale. You don't need these big general models that are incredibly energy-intensive and compute-intensive and require billions of dollars, new data centers and all kinds of resources to make happen. Instead, you could build smaller models. You could build narrower models. You could have a model like AlphaFold that's just designed to predict the 3D structures of proteins. You don't need a huge behemoth of a model to accomplish that. What's your response to that?

I think you need those big models. We love big and small models. You need the big models often to train the smaller models. So we're very proud of our Flash models, which are the most—we call them our workhorse models.

They're really efficient and some of the most popular models. We use a ton of those types and sizes of models internally. But you can't build those kinds of models without distilling from the larger teacher models. And even with things like AlphaFold, which obviously I'm a huge advocate of, more of those types of models can tackle really important problems in science and medicine today. We don't have to wait for AGI.

That will require taking the general techniques and then potentially specializing them—in that case, around protein structure prediction. I think there's huge potential for doing more of those things, and we're largely in our AI-for-science work. I think we're producing something pretty cool on that pretty much every month these days.

I think there should be a lot more exploration on that. Probably a lot of startups could be built combining some kind of general model that exists today with some domain specificity. But if you're interested in AGI, you've got to push, again, both sides of that. It's not an either-or in my mind. I'm an “and,” right? Let's scale, let's look at specialized techniques, combining that and hybrid systems—sometimes they're called that—and let's look at new blue-sky research that could deliver the next transformers. We're betting on all of those things.

You mentioned AlphaEvolve, something that Kevin and I were both really fascinated by. Tell us what AlphaEvolve is.

At a high level, it's basically taking our latest Gemini models—actually, 2 different ones—to generate ideas and hypotheses about programs and other mathematical functions. Then those go into a sort of evolutionary programming process to decide which of those are most promising, and then that gets ported into the next step.

Tell us a little bit about what evolutionary programming is. It sounds very exciting.

It's basically a way for systems to explore new space, right? In genetics, what things should we mutate to give you a kind of new organism? You can think about it the same way in programming or mathematics: You change the program in some way, and then you compare it to some answer you're trying to get. Then the ones that fit best, according to some evaluation function, you put back into the next round of generating new ideas.

We have our most efficient model, a sort of Flash model, generating possibilities, and then we have the Pro model critiquing that and deciding which one of those is most promising to be selected for the next round of evolution.

So it's sort of like an autonomous AI research organization, almost, where you have some AI coming up with hypotheses, other AI testing them and supervising them. The goal, as I understand it, is to have an AI that can improve itself over time or suggest improvements to existing problems.

Yes. I think it's the beginning of a kind of automated process. That's why people are so excited about it, and why we're excited about it. It's still not fully automated, and it's still relatively narrow.

We've applied it to many things, like chip design; scheduling AI tasks on our data centers more efficiently; even improving matrix multiplication, one of the most fundamental units of training algorithms. It's amazingly useful already, but it's still constrained to domains that are provably correct, right? Obviously, maths and coding are, but we need to fully generalize that.

It's interesting because I think for a lot of people, the knock they have on LLMs in general is, well, all you can really give me is the statistical median of your training data. But what you're saying is we now have a way of going beyond that to potentially generate novel ideas that are actually useful in advancing the state of the art.

That's right. But we already had this type of evidence. This is another approach, AlphaEvolve, using evolutionary methods, but we already had evidence of that even way back in the AlphaGo days.

AlphaGo came up with new Go strategies, most famously Move 37 in Game 2 of our big Lee Sedol world championship match. It was limited to a game, but it was a genuinely new strategy that had never been seen before, even though we've played Go for hundreds of years.

That's what kicked off our AlphaFold projects and science projects, because I was waiting to see evidence of that kind of spark of creativity, you could call it, or originality, at least, within the domain of what we know. But there's still a lot further that has to go.

We know that these kinds of models, paired with things like Monte Carlo tree search or reinforcement learning and planning techniques, can get you to new regions of search space to explore. Evolutionary methods are another way of going beyond what the current model knows, to explore—to force it into a new regime where it hasn't seen it before.

I've been looking for a good Monte Carlo tree search for so long now. If you could help me find one, it would honestly be a huge help. One of these things could help.

I read the AlphaEvolve paper—or, to be more precise, I fed it into NotebookLM and had it make a podcast that I could then listen to that would explain it to me at a slightly more elementary level. One fascinating thing that stuck out to me is a detail about how you were able to make AlphaEvolve more creative.

One of the ways that you did it was by essentially forcing the model to hallucinate. So many people right now are obsessed with eliminating hallucinations, but it seemed to me like one way to read that paper is that there's actually a scenario in which you want models to hallucinate, or be creative—whatever you want to call it.

Yes. I think that's right. Hallucination, when you want factual things, is obviously something you don't want. But in creative situations, you can think of it as a little bit like lateral thinking in an MBA course or something, right? Just create some crazy ideas. Most of them don't make sense, but the odd 1 or 2 may get you to a region of the search space that is actually quite valuable, it turns out, once you evaluate it afterward.

You can substitute the word hallucination, maybe, for imagination at that point, right? They're obviously 2 sides of the same coin.

I did talk to one AI safety person who was a little bit worried about AlphaEvolve, not because of the actual technology and the experiments, which this person said are fascinating, but because of the way it was rolled out.

Google DeepMind created AlphaEvolve and then used it to optimize some systems inside Google and kept it sort of hidden for a number of months, and only then released it to the public. This person was saying, well, if we really are getting to the point where these AI systems are starting to become recursively self-improving and they can sort of build a better AI, doesn't that imply that if Google DeepMind does build AGI or even superintelligence, it's going to keep it to itself for a while rather than doing the responsible thing and informing the public?

Well, I think it's a bit of both, actually. First of all, AlphaEvolve is a very nascent self-improvement thing, right? It's still got a human in the loop, and it's only shaving off, albeit important, percentage points from already existing tasks. That's valuable, but it's not creating any kind of step changes.

There's a trade-off between carefully evaluating things internally before you release them to the public, out into the world, and also getting the extra critique back, which is also very useful from the academic community and so on. We have a lot of trusted-tester-type programs where people get early access to these things and then give us feedback and stress-test them, including sometimes the safety institutes as well.

My understanding was you weren't just red-teaming this internally within Google. You were actually using it to make the data centers more efficient, using it to make the kernels that train the AI models more efficient.

I guess what this person is saying is, we want to start getting good habits around these things now, before they become something like AGI. They were just a little worried that maybe this is going to be something that stays hidden for longer than it needs to. I'd love to hear your response to that.

Yeah. Well, look, I think that system is not anything really that I would say has any risk on the AGI-type front. I think today's systems, although very impressive, still are not that powerful from any kind of AGI-risk standpoint that maybe this person was talking about.

I think you need to have both. You need to have incredibly rigorous internal tests of these things, and then you need to also get collaborative input from external sources. So I think it's a bit of both.

I actually don't know the details of the AlphaEvolve process for the first few months. It was just function search before, and then it became more general. So it's evolved over the last year in terms of becoming this general-purpose tool.

It still has a long way to go before we can actually use it in our main branch, which is, at that point, I think, when it becomes more serious. Like with Gemini, it's sort of separate from that currently.

Let’s talk about AI safety a little more broadly. It’s been my observation that the further back in time you go, and the less powerful the AI systems are, the more everyone seemed to talk about the safety risk. It seems like now, as the models improve, we hear about it less and less, including at the keynote yesterday.

I’m curious what you make of this moment in AI safety, whether you feel like you’re paying enough attention to the risks that could be created by the systems you have, and whether you’re as committed to it as you were, say, 3 or 4 years ago.

Yeah. We’re just as committed as we’ve ever been. From the beginning of DeepMind, we planned for success. Success meant something looking like this. This is what we imagined, and it’s still sort of unbelievable that it’s actually happened, but it is in the Overton window of what we thought was going to happen if these technologies really did develop the way we thought they were going to.

The risk, and attending to mitigating those risks, was part of that. We do a huge amount of work on our systems. I think we have very robust red-teaming processes, both pre- and post-launch, and we’ve learned a lot.

I think that’s the difference now: having these systems have, albeit early, contact with the real world. I’m persuaded now that that has been a useful thing overall. I wasn’t sure. Five years ago, or 10 years ago, I may have thought maybe it’s better to stay in a research lab and collaborate with academia, but there are a lot of things you don’t get to see or understand unless millions of people try it.

It’s this weird trade-off again. You can only do it when millions of smart people try your technology, and then you find all these edge cases. However big your testing team is, it’s only going to be 100 people or 1,000 people or something, so it’s not comparable to tens of millions of people using your systems. On the other hand, you want to know as much as possible ahead of time so you can mitigate the risks before they happen.

This is interesting, and it’s good learning. I think what’s happened in the industry in the last 2 or 3 years has been great, because we’ve been learning when the systems are not that powerful or risky, as you were saying earlier. I think things are going to get very serious in 2 or 3 years’ time, when these agent systems start becoming really capable.

We’re only seeing the beginnings of the agent era, let’s call it, but you can imagine—and I hope you understood from the keynote—what the ingredients are and how they’re going to come together. I think we really need a step change in research on the analysis and understanding of controllability.

The other key thing is that it’s got to be international. That’s pretty difficult, and I’ve been very consistent on that, because it’s a technology that’s going to affect everyone in the world. It’s been built by different countries and different companies in different countries, so you’ve got to get some kind of international norm around what we want to use these systems for and what kinds of benchmarks we want to test safety and reliability on.

There’s plenty of work to get on with now. We don’t have those benchmarks. We, the industry and academia, should be agreeing to a consensus on what those are.

What role do you want to see export controls play in doing what you just said?

Export controls are a very complicated issue, and geopolitics today is extremely complicated. I can see both sides of the arguments on that. There’s proliferation, uncontrolled proliferation, of these technologies. Do you want different places to have frontier model training capability? I’m not sure that’s a good idea.

On the other hand, you want Western technology to be the thing that’s adopted around the world. It’s a complicated trade-off. If there were an easy answer, I think we’d all—I would be shouting it from the rooftops. But it’s nuanced, like most real-world problems are.

Do you think we’re heading into a bipolar conflict with China over AI, if we aren’t in one already? I recently saw the Trump administration making a big push to make countries in the Middle East and the Gulf, like Saudi Arabia and the UAE, into AI powerhouses, having them use American chips to train models that will not be accessible to China and its AI powers. Do you see that becoming the foundation of a new global conflict?

Well, I hope not, but I think, in the short term, AI is getting caught up in the bigger geopolitical shifts that are going on. I think it’s just part of that, and it happens to be one of the most topical new things that’s appearing.

On the other hand, what I’m hoping is that, as these technologies get more and more powerful, the world will realize we’re all in this together, because we are. The last few steps toward AGI—hopefully we’re on the longer timelines, actually, right? The longer the timelines I’m thinking about, the more time we get to sort of get the collaboration we need, at least on a scientific level, before then. That would be good.

Do you feel like you’re in the final home stretch to AGI? Sergey Brin, Google’s co-founder, had a memo that was reported on by my colleague at The New York Times earlier this year. It went out to Google employees and said, “We’re in the home stretch,” and that everyone needed to get back to the office and be working all the time, because this is when it really matters.

Do you have that sense of finality, or of entering a new phase or an endgame?

I think we are past the middle game, that’s for sure. But I’ve been working every hour there is for the last 20 years, because I felt how important and momentous this technology would be. We’ve thought it was possible for 20 years, and I think it’s coming into view now.

I agree with that. Whether it’s 5 years, 10 years or 2 years, they’re all actually quite short timelines when you’re discussing the enormity of the transformation this technology is going to bring. None of those timelines are very long.

We’re going to switch to some more general questions about the AI future. A lot of people are now starting to think about what the world might look like after AGI, at least in conversations that I’m involved in. The context in which I hear the most about this is from parents who want to know what their kids should be studying and whether they’ll go to college. You have kids who are older than my kid. How are you thinking about that?

When it comes to kids—and I get asked this quite a lot—or university students, I wouldn’t dramatically change some of the basic advice on STEM. Getting good at things like coding, I would still recommend, because whatever happens with these AI tools, you’ll be better off understanding how they work and how they function, and what you can do with them.

I would also say immerse yourself now. That’s what I would be doing as a teenager today: trying to become a sort of ninja at using the latest tools. You can almost be superhuman in some ways if you get really good at using all the latest, coolest AI tools.

But don’t neglect the basics, too, because you need the fundamentals. I think you should teach meta-skills, really, like learning to learn. The only thing we know for sure is that there’s going to be a lot of change over the next 10 years, so how does one get ready for that? What kinds of skills are useful for that? Creativity, adaptability and resilience. I think all of these meta-skills are what will be important for the next generation.

It’ll be very interesting to see what they do, because they’re going to grow up AI-native, just like the last generation grew up mobile- and iPad-native, and sort of tablet-native. Before that, it was the internet and computers, which was my era. The kids of each era always seem to adapt to make use of the latest, coolest tools.

I think there’s more we can do on the AI side to make the tools actually useful. If people are going to use them for school and education, let’s make them really good for that and sort of provably good. I’m very excited about bringing them to education in a big way, and also, if you had an AI tutor, bringing it to poor parts of the world that don’t have good educational systems. I think there’s a lot of upside there, too.

Another thing that kids are doing with AI is chatting a lot with digital companions. Google DeepMind doesn’t make any of these companions yet. Some of what I’ve seen so far seems pretty worrying. It seems pretty easy to create a chatbot that does nothing but tell you how wonderful you are, and that can lead into some dark and weird places.

I’m curious what observations you’ve had as you look at this market for AI companions, and whether you think you might want to build this someday or whether you’re going to leave that to other people.

Yeah, I think we’ve got to be very careful as we start entering that domain, and that’s why we haven’t yet. We’re being very thoughtful about that. My view on this is more through the lens of the universal assistant that we talked about yesterday, which is something that’s incredibly useful for your everyday productivity. It gets rid of the boring, mundane tasks that we all hate doing to give you more time to do the things that you love doing.

I also really hope that these systems are going to enrich your lives by giving you incredible recommendations, for example, on all sorts of amazing things that you didn’t realize you would enjoy. They could delight you with surprising things. I think these are the ways I’m hoping that these systems will go.

On the positive side, I feel like if this assistant becomes really useful and knows you well, you could program it, obviously with natural language, to protect your attention. You could almost think of it as a system that works for you as an individual. It’s yours, and it protects your attention from being assaulted by other algorithms that want your attention, which is actually nothing to do with AI. Most social media sites—that’s what they’re doing, effectively. Their algorithms are trying to gain your attention, and I think that’s actually the worst thing. It would be great to protect that so we can be more in creative flow, or whatever it is that you want to do.

That’s how I would want these systems to be useful to people. If you could build a system like that, I think people would be so incredibly happy. I think right now people feel assailed by the algorithms in their life, and they don’t know what to do about it.

Well, the reason is because you have to use your one brain. Let’s say it’s a social media stream: you have to dip into that torrent to get the piece of information you want. But then you’re doing it with the same brain, so you’ve already affected your mind and your mood and other things by dipping into that torrent to find the valuable piece of information that you wanted.

But if a digital assistant did that for you, you’d only get the useful nugget, and you wouldn’t need to break your mood or whatever it is that you’re doing that day, or your concentration with your family. I think that would be wonderful.

Yeah, Casey loves that idea. You love that idea. I love this idea of an AI agent that protects your attention from all the forces trying to assault it. I’m not sure how the ads team at Google is going to feel about this, but we can ask them when the show comes.

Some people are starting to look at the job market, especially for recent college graduates, and worry that we’re already starting to see signs of AI-powered job loss. Anecdotally, I talk to young people who, a couple of years ago, might have been interested in going into fields like tech, consulting, finance or law, who are just saying, “I don’t know that these jobs are going to be around much longer.” A recent article in The Atlantic wondered if we’re starting to see AI competing with college graduates for these entry-level positions. Do you have a view on that?

I haven’t looked at that. I don’t know. I haven’t seen the studies on that, but maybe it’s starting to appear now. I don’t think there are any hard numbers on that yet. At least I haven’t seen them.

For now, I mostly see these as tools that augment what you can do and what you can achieve. I think, like with most big new technology shifts, maybe after AGI things will be different again, but over the next 5 to 10 years, we’re going to find what normally happens with big new technology shifts: Some jobs get disrupted, but then new, more valuable, usually more interesting jobs get created. I do think that’s what’s going to happen in the nearer term.

Today’s graduates and the next 5 years, let’s say, I think it’s very difficult to predict after that. That’s part of this more societal change that we need to get ready for.

I think the tension there is that you’re right: These tools do give people so much more leverage. But they also reduce the need for big teams of people doing certain things. I was talking to someone recently who said they had been at a data science company in their previous job that had 75 people working on some kind of data science tasks, and now they’re at a startup that has 1 person doing the work that used to require 75 people.

So I guess the question I’d be curious to get your view on is: What are the other 74 people supposed to do?

Well, I think these tools are going to unlock the ability to create things much more quickly. I think there’ll be more people who do startup things. There’s a lot more surface area that one could attack and try with these tools that wasn’t possible before.

Let’s take programming, for example. Obviously, these systems are getting better at coding, but I think the best coders are getting differential value out of them because they still understand how to pose the question, architect the whole codebase and check what the coding does. Simultaneously, at the hobbyist end, it’s allowing designers and maybe nontechnical people to vibe-code some things, whether that’s prototyping games or websites or movie ideas.

In theory, those 70 people or whatever could be creating new startup ideas. Maybe it’s going to be fewer big teams and more small teams, or teams very empowered by AI tools. But that goes back to the education thing: Which skills are now important? It might be different skills, like creativity, vision and design sensibility, that could become increasingly important.

Do you think you’ll hire as many engineers next year as you hire this year?

I think so. Yeah, there’s no plan to hire fewer, but we have to see how fast the coding agents improve. Today, they can’t do things on their own. They need a human. They’re just helpful for the best human coders.

Last time we talked to you, we asked you about some of the more pessimistic views about AI in the public. One of the things you said to us was that the field needed to demonstrate concrete use cases that were clearly beneficial to people to kind of shift this.

My observation is that I think there are even more people now who are actively antagonistic toward AI. I think maybe one reason is they hear folks at the big labs saying pretty loudly, “Eventually this is going to replace your job.” Most people just think, “Well, I don’t want that.” So I’m curious, looking back on that past conversation, if you feel like we’ve seen enough use cases to start to shift public opinion, or, if not, what some of those things might be that actually change views here.

I think we’re working on those things. They take time to develop. I think a kind of universal assistant would be one of those things if it was really yours and working for you effectively—technology that works for you.

I think this is what economists and other experts should be working on: Does everyone have or manage a suite, a fleet of agents that are doing things for you, including potentially earning you money or building you things? Does that become part of the normal job process? I could imagine that in the next 4 or 5 years.

I also think that as we get closer to AGI and make breakthroughs in— we probably talked about this last time—materials science, energy, fusion, these sorts of things helped by AI, we should start getting to a position in society where we’re getting toward what I would call radical abundance, where there are a lot of resources to go around. Then again, it’s more of a political question: How would you distribute that in a fair way?

I’ve heard this term, like universal high income, something like that. I think it’s probably going to be good and necessary, but obviously there are a lot of complications there that need to be thought through.

In between, there’s this transition period between now and whenever we have that sort of situation. What do we do about the change in the interim? It depends on how long that is, too.

What part of the economy do you think AGI will transform last?

I think the parts of the economy that involve human-to-human interaction and emotion, and those things, will probably be the hardest things for AI to do.

Aren’t people already doing AI therapy and talking with chatbots for things that they might have paid someone $100 an hour for?

Therapy is a very narrow domain, and I’m not sure exactly—there’s a lot of hype about those things. I’m not actually sure how many of those things are really going on in terms of actually affecting the real economy, rather than just being more toy things. I don’t think the AI systems are capable of doing that properly yet.

The emotional connection that we get from talking to each other, and doing things in nature and in the real world—I don’t think AI can really replicate all of those things.

So if you lead hikes, that’d be a good job.

Yeah. Yeah, climb Everest.

My intuition on this is that it’s going to be some heavily regulated industry where there will just be a massive pushback on the use of AI to displace labor or take people’s jobs, like health care or education or something like that. But you think it’s going to be an easier lift in those heavily regulated industries?

I don’t know. It might be, but then we have to weigh that up as a society: whether we want all the positives of that—for example, curing all diseases or finding new energy sources. I think these things would be clearly very beneficial for society, and I think we need AI for our other big challenges. It’s not like there are no challenges in society other than AI, but I think AI can be a solution to a lot of those other challenges, be that energy, resource constraints, aging, disease—you name it—and water access, et cetera. There are a ton of problems facing us today. Climate—I think AI can potentially help with all of those.

I agree with you: Society will need to decide what it wants to use these technologies for. But then what’s also changing is what we discussed earlier with products: The technology is going to continue advancing, and that will open up new possibilities, like radical abundance and space travel. These things are a little bit out of scope today unless you read a lot of science fiction, but I think they’re rapidly becoming real.

During the Industrial Revolution, there were lots of people who embraced new technologies, moved from farms to cities to work in the new factories, and were sort of early adopters on that curve. But that was also when the Transcendentalists started retreating into nature and rejecting technology. That’s when Thoreau went to Walden Pond, and there was a big movement of Americans who just saw the new technology and said, “I don’t think so. Not for me.”

Do you think there will be a similar movement around rejection of AI? And if so, how big do you think it’ll be?

I don’t know. There could be a get-back-to-nature movement, and I think a lot of people will want to do that. I think this could potentially give them the room and space to do it, right? If you’re in a world of radical abundance, I fully expect that’s what a lot of us will want to do.

Again, I’m thinking about it as sort of spacefaring and maximum human flourishing, but I think those will be exactly some of the things that a lot of us will choose to do—and have the time, space and resources to do them.

Are there parts of your life where you say, “I’m not going to use AI for that,” even though it might be pretty good at it, for some sort of reason—wanting to protect your creativity or your thought process or something else?

I don’t think AI is good enough yet to have impinged on any of those sorts of areas. Mostly, I’m using it for things like you did with NotebookLM, which I find great: breaking the ice on a new topic, a scientific topic, and then deciding if I want to get more deeply into it. That’s one of my main use cases. Summarization, those sorts of things—I think those are all just helpful.

But we’ll see. I haven’t got any examples of what you suggested yet, but maybe as AI gets more powerful, there will be.

When we talked to Dario Amodei of Anthropic recently, he talked about this feeling of excitement mixed with a kind of melancholy about the progress that AI was making in domains where he had spent a lot of time trying to be very good, like coding. You see a new coding system that comes out, it’s better than you, and you think that’s amazing. Then your second thought is, “Ooh, that stings a little bit.” Have you had any experiences like that?

Maybe one reason it doesn’t sting me so much is that I’ve had that experience when I was very young with chess. Chess was going to be my first career, and I was playing pretty professionally when I was a kid for the England junior teams. Then Deep Blue came along, right? Clearly, the computers were going to be much more powerful than the world champion from then on.

Yet I still enjoy playing chess. People still do. It’s different, but it’s a bit like Usain Bolt: We celebrate him for running the 100 meters incredibly fast, but we’ve got cars and we don’t care about that, right? We’re interested in other humans doing it. I think that’ll be the same with robotic football and all of these other things.

That maybe goes back to what we discussed earlier about what I think, in the end, we’re interested in: other human beings. Even if AI could write a novel one day that was technically good, I don’t think it would have the same soul or connection to the reader if you knew it was written by AI, at least as far as I can see for now.

You mentioned robotic football. Is that a real thing? We’re not sports fans, so I just want to make sure I haven’t missed something.

I meant soccer. There are RoboCup-style little robots trying to kick balls and things. I’m not sure how serious it is, but there is a field of robotic football.

You mentioned that sometimes a novel written by a robot might not feel like it has a soul. I have to say, for as incredible as the technology is in Veo or Imagen, I sort of feel that way with it. It’s beautiful to look at, but I don’t know what to do with it. You know what I mean?

Exactly. That’s why we work with great artists like Darren Aronofsky and Shankar Mahadevan on the music. I totally agree. I think these are tools, and they can come up with technically good things. Veo 3 is unbelievable. I don’t know if you’ve seen some of the things that are going viral and being posted at the moment with the voices. I didn’t realize how big a difference audio was going to make to the video. I think it just really brings it to life.

But it’s still not, as Darren would say—we were discussing this in an interview yesterday—it doesn’t have deep storytelling like a master filmmaker or a master novelist at the top of their game. It might never do that, right? It’s just always going to feel like something’s missing. It’s a sort of soul, for lack of a better word, of the piece: the real humanity, the magic, if you like.

When I see a Van Gogh or a Rothko, why does that touch you? I feel the hairs on the back of my spine because I remember what they went through and the struggle to produce that, right? In every one of Van Gogh’s brushstrokes is his torture. I’m not sure what that would mean even if AI mimicked that and you were told it was AI. It would be like, “So what?”

I think that is the piece that, at least as far as I can see 5 to 10 years out, the top human creators will always bring. That’s why we’ve done all of our tools, Veo and Lyria, in collaboration with top creative artists.

The new pope, Pope Leo, is reportedly interested in AGI. I don’t know if he’s AGI-pilled or not, but that’s something that he’s spoken about before. Do you think we will have a religious revival or a renaissance of interest in faith and spirituality in a world where AGI is forcing us to think about what gives our lives meaning?

I think that potentially could be the case. I actually did speak to the last pope about that, and the Vatican has been interested in these matters even prior to this pope. I haven’t spoken to him yet, but the question is how AI and religion—and technology in general and religion—interact.

What’s interesting about the Catholic Church—and I’m a member of the Pontifical Academy of Sciences—is that they’ve always had, which is strange for a religious body, a scientific arm. They like to say Galileo was the founder. It’s actually really separate, and I always thought that was quite interesting. People like Stephen Hawking and avowed atheists were part of the academy, and that’s partly why I agreed to join it: It’s a fully scientific body.

I was fascinated that they’ve been interested in this for 10-plus years. They were on this early, in terms of how interesting this technology will be from a philosophical point of view. I actually think we need more of that type of thinking and work from philosophers and theologians. That would be really, really good. So I hope the new pope is genuinely interested.

We’ll close on a question that I recently heard Tyler Cowen ask Jack Clark from Anthropic, which I thought was so good that I decided to just steal it whole cloth. In the ongoing AI revolution, what is the worst age to be?

Oh, wow. I haven’t thought about that. I think any age where you can live to see it is a good age, because I think we are going to make some great strides with things like medicine. I think it’s going to be an incredible journey. None of us knows exactly how it’s going to transpire. It’s very difficult to say, but it’s going to be very interesting to find out.

Try to be young if you can.

Yes, young is always better. In general, young is always better.

All right, Demis, thanks so much for coming.

Thank you very much.