[0:00:00] Thank you very much, Big Finish. I'm Mike Allen, co-founder of Axios. On behalf of my co-founders, Roy Schwartz, Jim Van Hai, thank you to all of you who for coming up nine years now have been fans of Axios. And thank you for turning out here in San Francisco in this historic bank, this very cool setting for this Axios AI Plus SF Summit. Welcome to all of you around the world for our Big Finish, Demis Hassabis, PhD,
[0:00:28] Co-founder and CEO of Google DeepMind. He's a neuroscientist, an entrepreneur, an AI pioneer. Demis was a chess prodigy at five, a Nobel laureate at 48. He's a British-born genius. He's been knighted. Demis Hassavas, welcome to Axios. Thank you so much. Thanks for having me. We've been looking forward to this. We appreciate it. Great to be here. It was just over 400 days ago that you found out
[0:01:05] You were a Nobel laureate. And you said in that moment, you said this is surreal. This is the big one. What has changed since then about your life and work? What has it made possible? Well, look, it's still pretty surreal, actually. It still hasn't fully sunk in.
[0:01:22] But it has made quite a big difference. The thing it makes a difference to is when you speak to people not in your field, including big government people, things like that, who maybe don't know that much about AI. If you have the Nobel Prize, it's a sort of shortcut to almost anyone to know that you're expert in your field. So it's going to be useful, I think, in the future. And you had endless resources at your...
[0:01:50] At your disposal, are there new resources that you have or that you think you can tap now? Not really. I mean, you're right. We're lucky at Google and at DeepMind. We have a lot of resources. They're not endless. We always need more, more compute, no matter how much compute you have. But we have a lot of great things, which is why we're able to do such a broad portfolio of things. But it's mostly, again, this platform it gives you to basically speak out about things that you care about.
[0:02:19] And I haven't done a lot of that yet, but I think it will be important. Maybe we're going to talk about AI safety and other things. I think the Nobel and the platform that gives you could be useful for that. And what's on the short list of, in addition to AI safety, you think you'll be talking more about using your platform? Yeah, well, it's not just about safety in the long term, AGI safety. Obviously, I think that we think a lot about that, but it's also about responsible use of AI today.
[0:02:47] What are the kinds of things we should be using AI to improve and to power up and to accelerate and maybe what sorts of things we should be careful about even in the near term. So I think that's one thing. I think also just getting society ready for what's to come.
[0:03:04] AGI, probably the most transformative moment in human history, is on the horizon. And we need to get prepared as a society and as a species. And I think, of course, governments and other important people, other important leaders are going to be critical in that. And I think having something like the Nobel platform opens pretty much any door. One of the things that distinguishes you is you're deep in the science.
[0:03:33] And yet you also are on the front line of this fight and this race among companies, hyperscalers, superpowers, and you sort of in the mold of Steve Jobs, you also have a product mind. You want to create delightful things for people, but you always say you're a scientist first.
[0:03:54] Yeah, science. I'm a scientist first. The reason I say that is that's the sort of default approach I take to everything. And what I mean by that is the scientific method, really, that way of thinking. I really love the, I mean, I think it's the most, the scientific method is the most important maybe idea humanity's ever had. You know, it created the Enlightenment and then modern science. So basically modern civilization depends on this idea of
[0:04:22] Scientific method and experimentation and then updating your hypotheses and so on. And I think it's an incredibly powerful method, but I think it can be applied to more than just science. I think it can be applied to everyday living and indeed business. And that's what I've tried to do is sort of take that to its limit.
[0:04:41] I think that's what gives us an advantage in some ways as a research organization, as an engineering organization. Yes, we're in the middle of this ferocious, probably the most ferocious competitive battle maybe tech has ever seen. But one of the things that I think gives us an edge is the rigor and precision we bring to our work.
[0:05:02] Because we have a scientific method sort of at the heart of it. And we blend world-class research with world-class engineering with world-class infrastructure. And I think you need all three of those things to be at the frontier of something like AI. And I think we're sort of pretty unique in having world-class capabilities in all those areas. Yeah. So in Axios fashion, we're going to divide our conversation between Zoom Out and Zoom In. So Zoom Out.
[0:05:30] Getting your priceless mind on the state of AI. So we're going to talk about the blunt state of AI. And what I'm going to ask from you is, given the known knowns today, be blunt, clinical, no hype, no soft selling. Can we do that? I'll do my best. All right. What does the next 12 months of progress look like? What do you believe that if we sit here a year from today, and I would love to, what will have changed in the world?
[0:06:00] I think the things that we're pressing hard on are the convergence of modalities. So Gemini, which is our main foundation model, has always been multimodal from the beginning. It takes images, video, text, audio, and then can now increasingly produce those types of outputs as well. And I think we're getting some really interesting...
[0:06:24] cross-pollination by being multimodal. The best example of that is our latest image model, Namno Banana Pro, which I think shows some astonishing sort of understanding of visuals. And it can kind of create infographics that are really accurate and so on. So I think over the next year, you're going to see that progress a lot. And I think, for example, in video, when that converges with the language models, you're going to see some very interesting combinations of capabilities there.
[0:06:51] I think the other things we're going to see over the next year and I'm personally working on is world models. So we have this system called Genie, Genie 3, which is like an interactive video model you can think about. So you can sort of generate a video, but then you can start walking around it like you're in a game or simulation and it stays coherent for a minute. I think that's very exciting. And then, you know, maybe the other thing is agent-based systems. So I think the field's been talking a lot about agents, but then they're not reliable yet enough to...
[0:07:21] We've heard a lot about that today here on the Axia stage. What would you say a year from now, how will agents have progressed? What's an example of how it will work in everyday life a year from now? Well, look, we have this concept of a universal assistant that we want Gemini eventually to become. I think this is also you're going to see from us over the next year. This will be on...
[0:07:43] More devices as well. By universal, we mean it's not just on your computer or your laptop or your phone, but maybe it comes around with you on glasses or other devices. And I think we want to create something that is useful to you in your everyday life that you consult many times a day. It becomes a part of the fabric of your life and it just improves your productivity, but also your personal life. Recommendations for books and films and other activities that you'd like.
[0:08:12] But agents at the moment, you can't delegate to them a whole task and be sure they're going to complete that entire task completely reliably. But a year from now, you think they will? I think a year from now, we'll start having agents that are close to doing that. And bull case, bear case. What is the best case for what AI can do for the world? And what do you fear most?
[0:08:40] Well, look, the best case scenario that I've always dreamed about and why I've worked my whole life on AI and getting closer to this moment we've been working towards for decades now, many of us, is something that's called radical abundance. So this idea that we solved a lot of the biggest issues confronting society and humanity today.
[0:09:03] Free, renewable, clean energy. Maybe we solve fusion or better battery, optimal batteries and solar materials, semiconductors, you know, material science. We've solved a lot of diseases. So then we're in a situation where, you know, we're in this new era, post-scarcity era, and where potentially, you know, humanity is flourishing and traveling to the stars and spreading consciousness to the galaxy. And what do you fear most?
[0:09:33] Well, even that utopian kind of view has some questions around it about what will be our purpose as humans if there are these technologies that are out there that are solving all these problems. What will be left to solve? I worry about that as a scientist and the scientific method even. So there's that. But there's also obviously the well-known challenges and risks with AI.
[0:09:57] Well, twofold. One is bad actors using AI for harmful ends, or the AI itself, as it gets closer to AGI and becomes more gentic, it goes off the rails in some way that harms humanity. So you've mentioned going off the rails. Yes. How worried are you about these catastrophic outcomes, your level of concern? I'm just going to rattle them off. One, pathogens created by the evil actor using AI.
[0:10:26] I think that's definitely one of the bad use case scenarios that we have to guard against, for sure. Energy or water cyber terror using AI by a foreign actor? Yeah, that's probably almost already happening now, I would say. Maybe not with very sophisticated AI yet, but I think that's the most obvious vulnerable vector.
[0:10:46] And which is why we focus quite a lot, and we are focusing quite a lot as Google and as DeepMind on AI for cybersecurity. So to power up the defensive side of that equation. AI operating outside human control on its own.
[0:11:01] Well, this goes back to the agentic stuff where I think as that becomes more sophisticated and it's clear why the industry will build those things because they'll be more useful as things like assistance. So they're definitely going to happen. But the more agentic and autonomous they are, the more room there is for these things to deviate from what you maybe had intended when you gave the initial instruction or the initial goal. So this is a very active area of research, which is to how to make sure that systems that
[0:11:30] Maybe a capable offer of continual learning or online learning stay within the guardrails that you set. I mean, I think the good news is because AI has become so big commercially and for enterprises, if you think about
[0:11:46] Renting or selling one of your agents as a model provider, leading model provider to another big business. Those businesses will want guarantees around the agent's behavior, what it does with their data, what it does with their customers. And if those things go wrong, they're not going to be existential in any way, but you'll lose the business for sure.
[0:12:07] Because why would that business enterprise go with that provider? They would choose a different provider that was more responsible and had better guarantees. So I think what's great about that is that capitalism will reward naturally, ideally, more responsible actors. But it's possible that the AI could jump the guardrail.
[0:12:29] Potentially, if done wrong. I mean, there was always a possibility where nobody really knows what the, that's one of the big unknowns. I think it's non-zero, that potential. So it's worth very seriously considering and mitigating against. But, you know, I hear people talk, you know, give very precise percentages about what the chances of these. You have to put a PDoom. A PDoom, which I think is kind of nonsense because no one knows what it is. What I know is it's not. So you don't quantify it, but you say it's. It's non-zero.
[0:12:59] So clearly, if your PDOM is non-zero, then you must put significant resources and attention on that. Where is the US winning the AI race against China and where are we losing? I think that we're still in the US and in the West in the lead. If you look at the latest benchmarks and...
[0:13:22] The latest systems, but China is not far behind. If you look at the latest DeepSeq or the latest models, they're very good. And there are some very capable teams there. So maybe the lead is only a matter of months as opposed to years at this point. Because when you put chips aside, AI, China probably is winning. No, I think chips is one thing, but I think algorithmically, innovation-wise, I think the West still has the edge.
[0:13:49] I don't think any of the Chinese models or companies have shown they can innovate on algorithmically something new beyond the state of the art. They've been very good at fast sort of following the current state of the art. Our last zoom out question, and you're going to like this one. What's the most astonishing thing about AI that you think gets shockingly little attention?
[0:14:17] The most astonishing thing about AI that gets shocking little attention. Yeah, if I think of the things we're working on and already have working, it's the multimodal understanding these models have. Multimodal is video. Yes, video, image, and I mean audio. But I'm thinking specifically video, actually. So if you give Gemini a YouTube video to process, you can ask it all sorts of incredible things about...
[0:14:45] The video, it's just sort of mind-blowing to me that it can understand sort of conceptually in a lot of cases, like not always, but in many really impressive cases, what's happening. What's an example of a question like that? Well, I've asked questions on like, I mean, this was just something I tested Gemini on the other day, was I love the film Fight Club. And there's some scene in it, I think, where Brad Pitt, or maybe it's Ed Norton, I can't remember, takes off his ring before...
[0:15:14] Having a fight. And I asked Gemini, what's the significance of that action? And it came up with an interesting philosophical point about leaving behind everyday life and just symbolically showing that. It was a very interesting kind of meta-insight that these systems have now. And I think if you use it, the other thing that's not appreciated is we have this thing called Gemini Live where you can...
[0:15:42] Point your phone at something and say you're a mechanic. It can actually just help you with whatever task you have in front of you. Ideally, that should be glasses, because you want to have your hands free, really, for that. But I think people don't realize how powerful that multimodality capability is yet. All right, you've given us the perfect bridge and transition to zooming in. Congratulations on Gemini 3. Thank you. Last month, your game-changing model, you say it reasons with unprecedented
[0:16:11] Depth and nuance. Tell us what's unique about the nuance part of Gemini 3. Yeah, I think it's just...
[0:16:18] We're really pleased with almost the personality of it, the style of it, as well as its capability. I like the way that it answers succinctly. It pushes back a little bit. It doesn't just agree with whatever you're saying. It pushes back gently on some ideas that if they don't make sense. And I think people are appreciating. I feel like you can feel it's a bit of a step change in its kind of intelligence and therefore usefulness.
[0:16:47] And what's something that Gemini has answered or produced where you said, I didn't know it could do that or I didn't know it would do that?
[0:16:56] Well, actually, this is the amazing thing of why we love what we're doing so much is that this era we're now in with research connected to product, the great thing about that is that you get millions and potentially at Google, billions of users immediately taking advantage of the new technology you put out there. And we're continually surprised by the cool things that people figure out very quickly to use these models for. And a lot of those things sort of tend to go viral.
[0:17:25] The other thing I most enjoyed with Gemini 3 was one-shotting games. So back to my very first career of making AI for games, I think we're very close now with these models, maybe the next version of models where you could start really creating perhaps commercial-grade games, you know, vibe-coding them with, you know, in a few hours, which used to take years. And that shows nuance. What does that show about the model?
[0:17:49] Well, I think it's just incredible depth and capability of these models to understand very high-level instructions and produce very detailed outputs. And the other things that Gemini 3 specifically is good at is front-end work and developing websites. And it's pretty good aesthetically and creatively as well as technically.
[0:18:14] Something we've written a fair amount about at Axios is that even the authors, creators of these models, don't totally understand them. What's something about Gemini 3 that you feel like you don't totally get? Well, actually, I feel like with all these models, and maybe all of the audience are feeling this too, is that
[0:18:33] There's such a fast pace of innovation and improvement. We're spending almost all of our time building these things. I have to have this feeling every time we release a new version that I haven't even explored a tenth, had time to even explore a tenth of probably what the existing systems can do. Because, of course, we're on to immediately, you know, we're referencing back to the ferocious race and competition we're in. We're immediately focusing on the next innovation and obviously making sure it's safe and reliable.
[0:19:02] So again, our users end up taking them much further than often we've tried internally. And one more question on Gemini 3. A little back story. And you had a number of irons in the fire. But LLMs, the text-based large language models, you didn't necessarily go all in on that as the Holy Grail, something that Walter Isaacson
[0:19:31] The great author and thinker and your friend said to me is that when you saw the power of the LLM, you did a pivot, a pirouette, as Walter said it, and were able to leapfrog to great success. And Walter's point was that most business people would have been stubborn, might have doubled, tripled down on their other bets. How did you make this decision to go all in?
[0:19:56] On your LLM. Well, I think this is, again, the beauty and the strength of the scientific method. If you're a true scientist, you can't get too dogmatic about some idea you have. You need to go with where the empirical evidence is taking you. So first of all, Walter is probably referring back to the 2017, 2018 era. So there we had a lot of irons in the fire, as we said. We had our own very capable language models that were called Chinchilla and then Sparrow. We had these various different code names for them.
[0:20:25] They weren't publicly released, but they were internal. In fact, some of the scaling laws were originally figured out by our team. They're called the Chinchilla scaling laws. But we also had other...
[0:20:34] Types of programs, AlphaZero, things that were building on AlphaGo, pure RL systems. And we also had some cognitive science, more neuroscience-inspired architectures as well. And at the time, my job was to make sure we build AGI first, fast, and safely. That's always been our sole intelligence, our mission at DeepMind. And so I'm kind of agnostic, actually, to the approach that's taken. I'm pretty pragmatic on that. That's maybe my engineering side of me is I have some theories.
[0:21:03] As a good scientist would, but at the end of the day, it's got to pragmatically work. And so when we started seeing the beginnings of scaling working, then we increasingly put more and more resources onto that branch of the research tree. So that's refreshing about your approaches with artificial general intelligence, human capable AI. You don't shy away from it. Some other people say, well, we won't know or we're already there or it doesn't matter. You say that it does matter and we will.
[0:21:33] No, and you say it's not far off. Yeah, we're definitely not there now.
[0:21:37] Actually, quite close is how you say it. Yes, quite close. I think we're like five to 10 years away, if you were to ask me. I'm sorry, say that again. Five to 10 years away. I think my bar, though, is quite high. So this is the, we define AGI as the system that exhibits all the cognitive capabilities we have. And that includes inventive and creative capabilities. I think there are missing, as all of you have used the current LLMs, they're amazing in some ways. They're really impressive in some senses.
[0:22:06] They've got incredible, almost PhD-level skills in some areas, IMO, gold medals, and so on. But in other areas, they're very flawed still. And so they're these sort of jagged intelligences. So you would expect across-the-board consistency from a true AGI. And they're missing other capabilities like continual learning, online learning, long-term planning and reasoning. They can't do any of these things currently.
[0:22:29] I think they will be able to, but maybe one or two more breakthroughs are going to be required. And a question from the great Ina Fried, who we've seen today and whose coverage from Day Zero of Axios has helped make Axios what it is. She says you're obviously, you've said that AI might be one advance, two advances away from AGI. Yes.
[0:22:53] Will we get there just by improving LLMs and generative AI, or do you think that there might be a different approach that's needed to hit AGI in your five to 10 years? I think, again, this is an empirical question, but what I do know, this would be my best guess, is the scaling of the current systems, we must push that to the maximum, because at the minimum, it will be a key component of the final AGI system. It could be...
[0:23:21] The entirety of the Agile system. There's a chance that just scaling will get you there. But my guess is, if I was to guess from my advance point now, is that one or two more big breakthroughs, when I mean, there's innovation going on all the time, by the way, even including in scaling, existing techniques. But I'm talking like a transformer level or AlphaGo level type of breakthrough. I think we might, I suspect, when we look back,
[0:23:43] Once AGI is done, that one or two of those things were still required in addition to scaling. We're about to get the hook, so a super rapid round. Another question from Ina. You obviously are a big believer in AI, but if you look at what's being spent, that doesn't mean that there might not be a big enough bubble to rattle the economy. How worried are you about that?
[0:24:05] I think it's not a binary. I think some parts of the AI industry are probably in a bubble. I don't know, like the seed rounds of $50 billion seed rounds and things like that seems a little bit unsustainable. But on the other hand, of course, I, or more than anyone, believes that AI is the most transformative technology ever. So I think in the fullness of time, this is all going to be more than justified. And my job...
[0:24:32] As head of Google DeepMind and the engineer with Google is to make sure we win either way. If the bubble, the so-called bubble bursts or if things continue to be good like they are now. We're in a strong position. The AI recruiting wars. What's the end state of this competition for talent?
[0:24:48] Well, look, it's gone pretty crazy recently. Things like what Meta have been doing, and everyone's got to do what makes sense for them. What we found for us is that we want people who are mission-driven. We have, I think, the best mission. We have the full stack. So I think if you want to do the most impactful work and have the most positive impact on the world, then I think there's nowhere better than at Google DeepMind. And in the end, I think the best scientists, the best researchers, the best engineers, they want to work on the most cutting-edge stuff.
[0:25:18] So if you're the sort of top of the leaderboards with the best systems, that's sort of a self-fueling. This is a question from James Vande Hei in the Entrepreneurial Young Mind at High Point University in North Carolina. He says there's a lot of conversation about AI gaining a mind of its own. Is there a scenario where AI could act in its self-interest?
[0:25:40] Well, it's a great question, and it's related to some of the more sort of catastrophic outcomes. If that went wrong, that would be one of the issues that with agent-based systems or very autonomous systems, if somehow they developed a self-interest that was in some sense conflicting with what the designers or even perhaps humanity wanted it to do. And finishing with a fun thing, you're still...
[0:26:06] What does gaming teach us about the world and what does gaming teach us about where these machines are headed?
[0:26:14] My chess background and my training in that and then other games subsequently is being critical to how I do my work, both in business and in science. I think the thing I love about games, and there's many things I've loved about them, but I love the creativity of making them, but I also just playing them, I think, is the best way to train your mind because the best games, whether that's chess or Go or whatever or poker, they're microcosms of something in the real world, right? But in general, you don't get...
[0:26:42] In the real world, to have several practice goes at making the decision correctly in that moment. Maybe in the real life, you only get a dozen of those critical moments. But you can practice your decision-making capabilities as much as you want within almost a simulation really of the world with games. And as long as you take the games very seriously, so you put a lot of thought into your decision-making, it really does train your decision-making and planning capabilities, in my opinion.
[0:27:10] Now, you pointed out that our squishy brains evolved to be hunter-gatherers, and yet we're facing a disruption that, as you put it to The Guardian, will be 10 times bigger than the Industrial Revolution and maybe 10 times faster. Are we facing a situation where most humans can't keep up and maybe no human, including you, can keep up? Well, the good news is, and I think my point on the hunter-gatherer was, look how adaptive
[0:27:38] Our brains have been. We evolved to be hunter-gatherers, and yet here we are sitting in our modern cities, modern civilization with all the technology around us. And the same human brain pretty much has been able to adapt to that. So I'm a really big believer in human ingenuity, and I think we're infinitely adaptable. We are the only existence proof. Our brains are the only existence proof of general intelligence.
[0:28:05] Perhaps in the known universe that we know of so far. So we are general intelligences ourselves. And so we should be able to infinitely adapt. There is a question about when post-AGI, what kinds of technologies can we create? Brain computer interfaces, other things that some of us may choose to use in addition to our existing technologies. And that could be one way for us to keep up. And as we say goodbye, you're a lifelong Liverpool fan. You've helped them with their analytics.
[0:28:34] How will AI affect and inform the World Cup here in North America? Well, we've had a lot of teams approach us for help, too. And I have to try and be equal with that. But it's hard having a lifelong supporter of Liverpool. But I'm looking forward to trying to make it out here, maybe at least for the World Cup final. But let's be serious.
[0:28:54] What will it change between now and then? It's a lifetime in AI between now and then, right? Yeah. Well, what in AI or AI for sport or just in... Yes. Yeah. Well, I mean, look, sport has immense amount of data and it's all about extreme elite performance. So it's actually a natural bedfellow for AI to come in and help optimize that process even further. And without giving away a trade secret, what will it be able to do for a World Cup team?
[0:29:20] Maybe score more headers from corners. That's one of the things I think our system found out. Precise positioning of the players.