Demis Hassabis：为什么 AGI 的影响将超过工业革命，以及 AI 的瓶颈在哪里

Demis, I'm so excited to be doing this. Thank you so much for joining me today.

Great to be here.

1. What Actually Counts as AGI & Where Are We Today?

There are many places we could have stopped, but I was watching the documentary that you did, which was fantastic, and I wanted to start with AGI. Definitions vary widely, and you've been very thoughtful about what it means to you. Can you explain how you think about it today, so we have that as a kind of ground center?

We've always defined AGI in a very consistent way: as a system that exhibits all the cognitive capabilities the human mind has. That's important because the brain is the only existence proof we have, that we know of—maybe in the universe—that general intelligence is possible. For me, that's the bar for what AGI should be.

How close are we? It's the worst question. Everyone says different things, and it's very difficult when you have prominent figures saying it could be as early as 2026 or 2027.

I've got a probability distribution around the timings, but I would say there's a very good chance of it being within the next 5 years. That's not long at all.

Is that closer than you thought? Has that changed over time?

Not really. Actually, when we started DeepMind back in 2010, my co-founder Shane Legg, who's chief scientist here, used to write blog posts predicting when AGI would happen. Bear in mind that in 2010, when we started, almost nobody was working in AI, and everyone thought AI basically didn't work. It was a dead end.

Those posts are still on the internet for people to check. We used to extrapolate compute and algorithmic progress, and we basically predicted it would take around 20 years from when we started out. I think we're pretty much on track.

2. What Are the Biggest Bottlenecks Holding AI Back Today?

What are the biggest bottlenecks when you look at where we are today? In the documentary, you said you just never have enough compute. What are the biggest bottlenecks?

I think compute is the big one, not just for the obvious reason of scaling up your ideas and your systems. As the scaling laws, as they're called, predict, you keep building bigger and bigger architectures with more and more parameters, and as you do that, you get more intelligent systems.

3. Have We Hit the Limits of Scaling Laws?

The other thing you need a lot of compute for is doing experiments. The computer—the cloud—is our workbench, basically. If you have a new algorithmic idea and you want to test it, you've got to test it at a reasonable scale; otherwise, it won't hold when you actually put it into the main system. You need quite a lot of compute if you have a lot of researchers with lots of new ideas.

You mentioned the word “scaling laws.” A lot of people suggest that we're hitting scaling laws and starting to see that plateauing effect. Do you think that's true?

No, I don't think so. I think it's a bit more nuanced than that. When the leading companies all started building these large language models, you were getting enormous jumps with each generation of a new system. Maybe they were almost doubling in performance.

At some point, that had to slow down, so it's not continuing to be exponential. But that doesn't mean there aren't still great returns from scaling the existing systems up further. We and the other frontier labs are getting a lot of great returns on that kind of compute expansion.

I would say the returns are still very substantial, although they're obviously a bit less than they were at the start of all of this scaling.

4. Where Is AI Ahead of Expectations & What's Still Missing?

Where are we behind where you thought we would be?

I think, actually, in most areas we're ahead of where I thought we would be. If you think about things like video models, or even our newest systems like Genie, which are interactive world models, I think that's incredible if you step back and think about it. If you'd shown me that 5–10 years ago, I would have been pretty amazed.

5. Why Can't AI Systems Learn Continuously Like Humans?

In most domains, we're ahead of where the field thought we would be. There are still some big things missing, though, like continual learning. These systems don't learn after you finish training them, after you put them out into the world. They're not very good at learning further things, and I think some critical capabilities are lacking.

I'm sorry to ask—I'm asking some basic questions. Why don't we have continuous learning?

People haven't quite figured it out yet, and all the leading labs are working on how to integrate new learning into existing systems that you've spent months training. The brain does this very elegantly, probably through things like sleep and reinforcement learning.

6. How Did DeepMind Go from Behind to Leading the Pack?

You get consolidation, as it's called in the brain, where your memories during the day are replayed, and then some of that information is elegantly incorporated into your existing knowledge base. Perhaps we need something like that to incorporate new information along with the existing information base.

You mentioned video models, media, and images. It seems that DeepMind has progressed very quickly and caught up with or overtaken other providers. I tweeted about what I used and how it had changed over time—I think you liked it—and DeepMind is now my number one for research for new shows. It wasn't that way before.

What has led to the acceleration and progression of DeepMind in a way that it wasn't there 2–3 years ago?

We made some organizational changes. I think we've always had the deepest and broadest research bench at Google and at DeepMind. If you look at the last decade or more—15 years—I would say about 90% of the breakthroughs that underpin the modern AI industry were done either by Google Brain, Google Research, or DeepMind, so by one of our groups.

If you think of things like AlphaGo, reinforcement learning, and, of course, transformers, these are all the key breakthroughs. I would back us to make those breakthroughs in the future if there are any missing ones.

I think we've basically helped put together all the talent from around the company, pushing in one direction. We talked earlier about compute resources. It was also about combining all of our resources so we could build the biggest models, rather than having 2 or 3 versions around the company.

A lot of it was assembling all the ingredients we already had and then pushing with relentless focus and pace, acting almost like a startup, really, to get back to the frontier and be ahead in many areas.

You say that if anyone's going to make the breakthrough, it could and should be us. When you think about that, is continual learning the next breakthrough you're most excited about?

There are quite a few things that are missing. I think there's a lot of mileage in looking at different memory systems. At the moment, we have these long context windows, which are a bit brute force. You just put everything in them.

Then there's long-term planning and hierarchical planning. These systems are not very good at planning over long time horizons—many years into the future—which we can do with our minds. There's quite a lot of problems still left to overcome.

Maybe one of the biggest is consistency. I sometimes call these systems “jagged intelligences” because they're really amazing at certain things when you pose the question in a certain way. But if you pose a question in a slightly different way, they can still fail at quite elementary things.

A general intelligence shouldn't be that sort of jagged.

Sure. One hundred percent. That's a disaster.

The general intelligence, if you think about how our minds work, shouldn't have those kinds of holes in it.

7. Are We Heading Toward Model Commoditization?

We talked about a plateauing of scaling laws.

Everyone talks about a commoditization of models in terms of capabilities. Do you think we'll see that, or do you think we'll see ones that continuously accelerate ahead of the others?

Yeah, I feel like maybe the 3 or 4 leading labs now, of which we're 1, I think the gap is starting to pull away because a lot of these tools also, of course, help you build the next generation—things like coding tools and math tools. It's getting harder and harder, I would say, to eke out the same gains from just the same ideas.

So I think those labs that have the capability to invent new algorithmic ideas are going to start having a bigger advantage over the next few years, as all the juice has been wrung out of the set of ideas.

8. What Does the Future of Open Source Really Look Like?

I'm intrigued—you were very open with a lot of your research for years, and we see many very good-quality open models. How do you think about the future of open? I have many portfolio companies that use frontier models, and then they use that to set a benchmark, and they use open models to get as close as possible, but with more cost-effectiveness. What does that future look like?

Yeah, I think it's probably similar to what we're seeing today. We're big supporters of open science and open models, and we've done many things, obviously, from the original Transformers to AlphaFold. These are all things we've given out into the world to help the research community, and we plan to continue to do that, especially in applied domains—scientific domains, applying AI to science, which is obviously my passion.

But I think increasingly what you're going to see is that the open-source models are probably 1 step back from the absolute frontier. It usually takes about 6 months for the open-source community to reimplement and figure out what those ideas are.

But we are also pushing hard on a suite of open-source models called Gemma, which we're determined to make best-in-class for their sizes. Specifically for small developers or academics, or the beginnings of a startup, I think they're perfect for that, and also edge computing too. So we're very interested in open-source models for certain types of applications.

9. What Does a Post LLM World Look Like?

How do you think about a world post-LLMs? Different people have different views. You have Yann LeCun with very different views.

For me, I kind of disagree with Yann on a few things. I think there's a 50/50 chance there are some things missing that we still need to make breakthroughs in, perhaps world models or these kinds of approaches. But my bet is pretty strong that we've seen how successful these foundation models have been. They can do incredibly impressive things, and I don't think that's going to go away. We're still seeing gains from the returns from the scaling laws.

So I think the only question really is, when you think about a future AGI system, is an LLM foundation model going to be the key component only, or is it the total system? I just think it's a question of whether there's anything else needed. I don't think it's going to get replaced. I think it's going to get built on top of these foundation models, just like the way we do with our world models.

When we think about that future 5 years out, as you said, potentially with AGI, what does that world look like? Many people have different concerns.

Yeah.

If we just start generally, what does that world look like to you?

Well, I think on the positive side—and the things I've obviously spent my whole career and life building towards AGI—is that I think it will be the ultimate tool for science and medicine. In terms of advancing scientific discovery and finding cures to diseases, I think we need that kind of technology, and so I'm hoping in 5-plus years' time we'll be entering a new golden age of scientific discovery.

10. Can AI Really Fix Drug Discovery?

My mother's got multiple sclerosis, so it's something—it's the thing that I'm always most excited about. The thing I worry about is actually drug discovery: the process of getting it through all the trials and knowing that it takes a decade before my mother will actually get any benefits from it. How do we solve that? I think we'll get to that point soon.

First of all, what we're doing is, after we did the AlphaFold project on protein folding, we spun out a company called Isomorphic Labs, which is doing extremely well. The idea there is that we're focusing on solving the rest of the drug-discovery process, which is a lot of chemistry: designing the compounds, checking that they're not toxic, and all the different properties you need for drugs to be safe. I think we'll have that whole drug-design engine ready in the next 5 to 10 years.

Then you're right. The next problem is that clinical trials still take many, many years. But I think AI can help there in terms of maybe simulating parts of human metabolism, and also stratifying patients to make sure that certain patients get exactly the right type of drug that's suitable for their genomic makeup. So I think AI can help there too.

But I think the real revolution will come when a few, maybe a dozen or so, AI drugs get through the whole process, and then the government and the regulatory bodies see that and have enough data to back-test the predictions of those models. Then maybe, a further 10 years out, we can really just trust the predictions that the models are making and perhaps skip some steps, like animal testing is no longer needed. Maybe we can go up the dosage ladder quicker because you can rely on these models.

So I think we've got to do it in 2 steps: solve the drug-design problem first, and then look at the regulatory length of time it takes.

11. What Does "Good" AI Regulation Actually Look Like?

Speaking of regulation, AI safety is a big topic and a big concern. I watched it last night over dinner, which was a great watch—obviously a documentary. I think it was Stephen Hawking who said, “We must get it right because we might not get another chance.” Do you think that's right?

Yeah, I do think that's right. I think those are the stakes that we have to deal with, and there are 2 things I worry about.

One is the misuse of these systems by bad actors. They can be repurposed. These are dual-purpose technologies. They can be used for incredible good in science and health, as we just discussed, but they can also be repurposed for harmful ends by a bad actor. So that's 1 issue.

The second issue is a technical one: making sure these systems, as they get more powerful—not today's systems, but maybe in a year or 2's time, when they become more agentic, more autonomous, as we get towards AGI—can be kept on the guardrails that we want.

I think regulation, the right kind of regulation, could help here in terms of making sure there's at least a set of minimum standards from all of the leading providers, but ideally it needs to be an international standard.

What is the right kind of regulation? And again, I'm kind of quoting you back from this documentary. You're like, “I think we need more global coordination,” which worries me because we're getting worse at it.

Yes, which I think would be an unwavering truth. Yes, for sure. I mean, that's sort of crazy, the timing that we're in, right? With this most consequential, maybe, technology the world has ever seen, at the same time as a very fragmented international system. It's not ideal, but I think we're going to have to try and do the best we can to at least come up with a set of minimum standards—some benchmarks that test for undesirable properties, for example, deception.

Nobody should be building systems that are capable of deception, because then they could be getting around other safeguards. And then I imagine, if things go well, some kind of certification process that basically—it's almost like a Kitemark of quality—that this model has certain safeguards and certain guarantees, and therefore consumers and companies can safely build on top of it.

I think that is how it should go, ideally, but it does have to be international because, of course, these systems are cross-border and cross-territory.

12. Who Should Be the Ultimate Arbiter of Truth in an AI World?

Who is that ultimate verification system? You obviously started with Theme Park a long time ago.

Yes, a long time ago.

Yes, brilliant. Don't put the burgers down too close to the roller coaster.

But obviously, as a media company, I go through media platforms where I don't know what's real or fake. I'm always having to ask, “What's real or fake?” Who is that arbiter of verification?

Yeah. Well, I think there are—I mean, ultimately, it's got to be government, I think. But the kinds of technical bodies that would be able to do the technical work would be the AI Safety Institutes. There's a very good one in the UK that was set up under Prime Minister Sunak, and I think it's doing great work. Then there's one in the US, and maybe some of the leading countries that have the best research should also have an equivalent body that is staffed with high-quality researchers, too, that can actually evaluate and audit these kinds of systems against certain benchmarks and independently check whether they are meeting the right standards.

13. If Demis Had One Shot to Fix AI Safety, What Would He Do?

If I could give you a magic wand that was only applicable to AI safety, what would be your implementation idea or program that you would put in place with this magic wand?

Yeah, I think we need some kind of international body, maybe similar to the atomic agency, something like that, that perhaps the AI Safety Institutes sort of feed into. And the research community also has to do this and be involved in what the right set of benchmarks are to check what types of traits and what types of capabilities.

Maybe there are other safeguards too. It wouldn't be desirable to have AI systems output tokens that are not human-readable, in some kind of machine language that we couldn't understand. I think that would introduce a new vulnerability.

There are quite a few things like that which I think most of the leading labs would agree are probably not best to do. These institutions would test against those things, and I think that would give the public confidence. Academia could be involved, as well as civil society, to ensure that these systems—which are going to get incredibly powerful—have been independently audited.

14. Is This Time Different for Jobs or Will History Repeat Itself?

And so your magic wand's done now. I don't know. Maybe I used it on the wrong thing. Time will tell.

Yes, exactly.

You said that science would be one of the most exciting areas over the next 5 years. I have to ask about it because one of the biggest concerns is the labor displacement problem. I just had Mark Andreessen on the show, actually, and he said that I was a Marxist, which I was like, “Bollocks.” Mark's wonderful, so I'm not blaming him, but he was like, “It's completely rubbish.”

Yeah, I don't agree with that at all. We've always overcome it.

How do you think about the labor displacement problem when you look at how truly capable these systems are?

And what that does to labor markets? Well, certainly, in the past, with every new revolutionary technology, there's been a lot of job disruption. That's for sure, and I think that's definitely happened. A lot of old jobs go away or aren't viable anymore, but the history of it is that a whole set of new jobs arrive that one perhaps couldn't even imagine before, and those are high-quality, higher-paying jobs. That's the normal course.

Of course, you have to be very careful not to say that this time is different. I guess that's what people like Mark are claiming: that it's the same as the last 10 massive technological breakthroughs, like the internet, mobile, and so on. I do think this is going to be bigger than all of those previous technological breakthroughs.

I sometimes quantify the coming of AGI as 10 times the Industrial Revolution at 10 times the speed, unfolding over a decade instead of a century. I've been reading a lot about the Industrial Revolution. There are a lot of great books about it, and that caused a huge amount of upheaval as well as a lot of advances.

We wouldn't have modern medicine today. Child mortality was at 40% back in pre-industrial-revolution times. These are things you wouldn't want not to have happened, but ideally, this time around, we mitigate some of the downsides a bit better than we did during the Industrial Revolution.

I often listen to amazing voices like yours, and I get very excited by how fast it's coming. Then I try and stop myself from being too youthful and think I should be wiser. I'm told that we always overestimate what can be done in a year and underestimate what can be done in 10. Is that the truth here, or is it actually coming faster than we think?

No, I think that's still the truth. Maybe both time scales—the short term and the long term—are nearer together than they are with other technologies. I do think that, literally today and over the next year, things are a bit overhyped in AI. They couldn't be any more hyped in some ways.

On the other hand, I still think it's underappreciated how revolutionary this is going to be on a time scale of about 10 years. We could call that long term. There's still that dichotomy, even today, with AI.

With the concern around labor markets, there's also a concern around income inequality and the concentration of wealth among a few players. How do you see that shaping up, given what you said about the Industrial Revolution and what happened there?

Well, I think there are different ways that could play out. Maybe pension funds should be buying into all the big AI companies and making sure that everyone has a piece of that, or sovereign funds. Maybe every country should have a sovereign wealth fund that does that. That would be the sort of investment way of doing it.

I think there also needs to be thought about what happens if there is this massive productivity gain, but it's narrow where that accrues. How do we redistribute and distribute that so that everyone benefits from these huge gains? I can see all sorts of ways that could be done, including providing infrastructure and other things with that additional productivity gain.

There could be unbelievable things happening on a 5-to-10-year time scale, including a breakthrough in some kind of renewable, free energy. Maybe we solve fusion. We're working on that, right, with our partners at Commonwealth Fusion.

I think AI is going to usher in all sorts of new breakthrough technologies. Maybe we'll have amazing new superconductors, better batteries, and advances in materials science. There are all sorts of ways I could see that completely changing the nature of the economy.

15. How Do We Solve the Energy Crisis Created by AI?

How do we solve the energy crisis that comes with an AI revolution? What it means in terms of energy requirements is unprecedented. I know it's an incredibly hard question—I'm delving from really hard question to really hard question—but how do we solve that unprecedented need for new energy?

Well, I think AI will, in the medium to long run, more than pay for itself in terms of energy costs. We work on all these projects, like optimizing existing infrastructure and optimizing the grid. I think we could probably get 30–40% more efficiency out of our national grids.

Then there's modeling the climate and weather, and we have some of the best weather-modeling systems in the world. That helps us work out where the effects are really happening so we can mitigate them.

Finally, perhaps the most exciting are these new breakthrough technologies, like fusion, new batteries, and superconductors, which I think AI will be essential for helping us reach. I think we'll be in a completely new energy situation compared with anything we've ever experienced as humanity. That will, of course, help with things like the climate and environment, and eventually also help us get into space much more cheaply.

If you have an incredible energy source like fusion, then you have effectively unlimited rocket fuel because you can just electrolyze seawater.

I'm not going to ask you to solve space. Don't worry about that.

16. Why Stay in the UK Instead of Moving to Silicon Valley?

My question was about being in the UK. You're in London. I'm in London, and I'm very proud to be in the UK. You've been, I'm sure, pushed and prodded at every turn to move to the US. Why have you stayed?

Well, I should ask you that question too, but I think I saw London, when we started DeepMind, as a place—and the UK in general, and Europe to some degree—with incredible talent. We've always had, I don't know, 3 or 4 of the top 10 universities in the world, with Cambridge, Oxford, Imperial, UCL—these kinds of universities.

We're producing the envy of the world, really: these amazing graduates and PhD students. We have incredible scientists here, and we've got a rich heritage of that, all the way from Turing and Hawking to Darwin and Newton. We have this incredible history of scientific breakthroughs and great thinkers.

I felt we had all the ingredients—the talent and great engineers here—but it just hadn't been galvanized into an ambitious deep-tech startup idea. I felt that was possible, and I felt there was actually less competition here for that sort of talent. We could even draw in the best talent from the top European universities. That's what it was like in the early days of DeepMind, so I think it was a huge structural advantage for us.

The final thing is that maybe being a bit away from the Valley has some disadvantages. You're not plugged into the network, the gossip, the latest trends, and the vibes. We're a little bit out of it here, but I think it is very conducive to thinking deeply about things and being more original in how you think.

I think that's great for things like deep tech, where you don't want to be distracted by the latest fad. You want to know that it's going to be a 20-year mission, which is what we knew at the beginning of DeepMind. Being a little bit away from that maelstrom is quite good.

Palmer Lucky and Mark Andreessen often talk about being 400 miles away from the Valley. It's core to Palmer's kind of innovative thinking.

Yes. We're a few thousand miles away, but yeah.

17. Will Europe Ever Build a Trillion-Dollar Tech Giant?

Terrible question: will Europe have a trillion-dollar company? You see the Americans always bash us for our lack of large companies. I ping Daniel and say, “Come on, dude. Let's—”

Yes, exactly. But we don't have a trillion-dollar company.

Not yet. Daniel might well get there with one of his companies. Spotify and Helsing—I think those are 2 good options. I think there's no reason why we can't have that.

I'm going to try and do that with Isomorphic, which is headquartered here and I think has the potential to be that. But I think that's one of the disadvantages of Europe: obviously, we're a combination of smaller markets. That's one thing we have to overcome.

Maybe this EU Inc. thing could be a good innovation. I'm pulling out the magic wand again.

Yes. You can change the world.

Yeah, you've got the magic wand, but this time apply it to European technology. What would you do to implement a growth mindset and build the trillion-dollar company that we don't have today?

I think in the UK—I mean, this may apply to other European countries too—I think unlocking what pension funds can invest in for the growth stage would be important. I think we're brilliant at doing the startup idea and getting it to a certain level, like we did with DeepMind, but then if you really want to cross that sort of chasm into the trillion-dollar global player, where are the billion-dollar rounds going to come from, where you can really take on the existing incumbents?

I think that certainly was missing 10 years ago when I was doing fundraising for DeepMind, and I think it's still kind of missing today: just that kind of level of ambition and the amount the capital markets can support.

I remember some of your early rounds, raising in the sub-million-pound area.

Yeah.

It was quite hard work, actually. Families, kids, all that kind of—

Exactly.

18. Meeting Elon Musk for the First Time?

Okay, we're going to do a quick-fire round. Do you remember meeting Elon for the first time?

Oh, yeah, it was amazing. It was at Founders Fund, because both SpaceX and DeepMind were part of the same portfolio—kind of an amazing portfolio that Peter Thiel had at Founders Fund. I think we were both invited. I think I was invited to my first portfolio conference. It must have been back in 2011 or 2012, very early days.

We were the small, little upcoming thing, and I had a small speaking slot. Then Elon was the big thing in that portfolio, so he had the keynote. But then we met afterwards. I think, as Elon says, it was like we were passing each other in the bathroom or something.

We said hi, and we both hit it off immediately, as people who were almost too ambitious in their thinking, perhaps, and loved sci-fi. I really wanted to visit his rocket factory, so I was trying to get an invite to SpaceX in Los Angeles. He invited me at the end of that meeting, and that was our second meeting, in the SpaceX factory.

I love it. And now your speaking slots are as big as his.

I don't know about that.

What healthcare revolution or disease eradication are you most excited about?

Again, for me, it's specifically multiple sclerosis.

Yeah.

Well, look, I want to literally cure cancer. I know people say that's the cliché, but what we're building at Isomorphic is general-purpose. We're trying to build a platform—a drug-design platform—that will be applicable to any therapeutic area.

19. What Big Questions About AI Is No One Talking About?

Ideally, it will help with everything from neurodegeneration and cardiovascular disease to immunology and cancer. Those are the ones we're focusing on first, but eventually it should be applicable to every disease area.

What are you thinking about that you're not reading about or seeing anyone talk about?

I think it's more that a lot of people are worrying about the economic questions around AGI that we talked about earlier, but I worry a lot about the philosophical questions around it.

When it comes—let's assume we get the technical part right, and let's assume we get the economics part of it right; both of those are hard—then there's a philosophical question of: What is meaning? What is purpose? We'll find out maybe what consciousness is. What does it mean to be human?

I think that's what's coming down the road, and I think we need some great new philosophers to help us navigate that.

20. What Does Demis Want His Legacy to Be?

All right, final question. There are many different ways you could describe what you do. What would you most like to be remembered for? What would you like your legacy to be?

I would like my legacy to be remembered for advancing science and building technologies that bring incredible benefits into the world, like curing terrible diseases.

Demis, thank you so much for putting up with my meandering conversation. You've been fantastic. I really appreciate it.

Thank you very much.