[0:00] We effectively you can think of as 50% of our effort is on scaling, 50% of it
[0:04] is on innovation. My betting is you're going to need both to get to AGI. I've
[0:08] always felt this that if we build AGI and then use that as a simulation of the
[0:13] mind and then compare that to the real mind, we will then see what the
[0:18] differences are and uh potentially what's special um and remaining about
[0:22] the human mind, right? Maybe that's creativity, maybe it's emotions, maybe
[0:26] it's dreaming. There's a lot of consciousness. There's a lot of um
[0:30] hypotheses out there about what may or may not be computable. And this comes
[0:35] back to the chewing machine question of like what is the limit of a chewing machine.
[0:39] >> So, there's, nothing, that, cannot, be, done within the sort of computational.
[0:43] >> Well,, no, one's, put, it, this, way., Nobody's found anything in the universe that's
[0:46] that's non-computable >> so, far. >> So, far. Welcome to Google Deep Mind the podcast
[0:55] with me, Professor Hannah Fry. It has been an extraordinary year for AI. We
[1:00] have seen the center of gravity shift from large language models to agentic
[1:04] AI. We've seen AI accelerate drug discovery and multimodal models
[1:09] integrated into robotics and driverless cars. Now, these are all topics that
[1:13] we've explored in detail on this podcast. But for the final episode of
[1:17] this year, we wanted to take a broader view, something beyond the headlines and
[1:20] product launches to consider a much bigger question. Where is all this
[1:25] heading really? What are the scientific and technological questions that will
[1:30] define the next phase? And someone who spends quite a lot of their time
[1:34] thinking about that is Demis, CEO and co-founder of Google DeepMind. Welcome
[1:39] back to the podcast, Deis. >> Great, to, be, back. >> I, mean,, quite, a, lot's, happened, in, the
[1:43] last year. >> Yes., What's, what, sort, of, the, biggest shift do you think?
[1:47] >> Oh, wow., I, mean, um, it's, just, so, much, has happened as you said it's just it feels
[1:52] like we've packed in 10 years in one year. I think a lot's happened. I mean
[1:56] certainly for us uh the the progress of the models um we've just released Gemini
[2:00] 3 which we're really happy with. Um the the cap multimodal capabilities all of
[2:05] those things have just advanced really well. And then probably the thing I I
[2:08] guess over the summer that I'm uh very excited about is world models being
[2:12] advanced. I'm sure we're going to talk about that. Yeah, absolutely. We will
[2:15] get on to all of that stuff in a bit more detail in a moment. Um, I remember
[2:18] the very first time that I interviewed you for this podcast and you were
[2:21] talking about the root node problems about this idea that you can use AI to
[2:24] kind of unlock these downstream benefits. Um, and you've made pretty
[2:28] good on your promise, I have to say. And do you want to give us an update on on
[2:32] where we are with those where what are the things that are just around the
[2:35] corner and and the things that we've that you've sort of solved or near solved?
[2:38] >> Yeah., Well,, of, course,, obviously, the, big proof point was was AlphaFold and sort
[2:42] of crazy to think we're coming up to like 5year sort of anniversary of of
[2:45] AlphaFold being sort of announced to the world, Alpha Fold 2 at least. So that
[2:49] was the proof I guess that it was possible to do these root node type of
[2:53] problems. And we're looking we're exploring all the other ones now. I
[2:56] think material science uh I'd love to do a room temperature superconductor um and
[3:02] uh you know better batteries these kinds of things. I think that's that's on the
[3:06] cards. uh better materials of all sorts. We're also working on fusion um
[3:11] >> because, there's, a, new, partnership, that's been announced fusion.
[3:13] >> Yeah,, we've, just, announced, partnership with a deep one. We we already were
[3:16] collaborating with them, but it's a much deeper one now with Commonwealth Fusion
[3:19] who, you know, I think are probably the best startup uh uh working on at least
[3:22] traditional TOKAC uh reactors. So they're probably closest to to having
[3:28] something uh uh viable and we want to help accelerate that uh you know helping
[3:32] them contain the plasma in the magnets and maybe even some material design
[3:36] there as well. So that's exciting. And then we're collaborating also with our
[3:40] quantum colleagues which they're doing amazing work uh at the at the quantum AI
[3:45] team at Google and we're helping them with error correction codes uh where
[3:49] we're using our machine learning to help them and then maybe one day they'll help us.
[3:53] >> That's, perfect., Exactly., The, fusion, one is particularly I mean the difference
[3:58] that that would make to the world that would be unlocked by that is gigantic.
[4:01] >> Yeah., I, mean, fusion's, always, been, the holy grail. Of course I think solar is
[4:04] very promising too, right? Effectively using the fusion fusion reactor in the
[4:08] in the in the clouds in the sky. But um I think if we could have uh modular
[4:14] fusion reactors, you know, this promise of uh almost unlimited renewable clean
[4:20] uh energy uh would be obviously transform everything. And that's the
[4:23] holy grail. Of course, that's one of the ways we could we could um uh help with climate
[4:29] >> does, make, a, lot, of, our, existing, problems sort of disappear if we can if we can
[4:33] >> definitely, I, mean, it, opens, up, many, this is why we think of as a root node. Of
[4:36] course it helps directly with energy and and pollution and and so on um and helps
[4:42] with the with the climate crisis. But also, if energy really was renewable and
[4:46] clean and and and super cheap or almost free, then many other things would
[4:50] become viable. Um, like, you know, water access cuz we could have desalination
[4:54] plants pretty much everywhere. Uh, even making rocket fuel. Uh, you know, it's
[4:59] just there's lots of seawater that contains hydrogen and oxygen. That's
[5:02] basically rocket fuel, but it just takes a lot of energy to split it out into
[5:05] hydrogen and oxygen. But if energy is cheap, uh, and and renewable and sort of
[5:10] clean, then why not do that? you know you could have that producing 247.
[5:14] >> You're, also, seeing, a, lot, of, change, in the uh the AI that is applying itself to
[5:18] mathematics, right? That you know winning medals in the International Math
[5:21] Olympiad and yet at the same time, these models can make quite basic mistakes in
[5:26] high school math. Why is there that paradox? >> Yeah,, I, think, it's, fascinating, actually.
[5:30] One, of the, most, fascinating, things, and probably that needs to be fixed uh as
[5:34] one of the key things while we're not at AGI yet. Um, as you said, we've had a
[5:38] lot of success in other groups on getting like gold medals at the
[5:40] International Mass Olymp. You look at those questions and they're they're
[5:43] super hard questions that only the top students in the world can can do. And on
[5:48] the other hand, if you pose a question in a certain way, we've all seen that
[5:51] with with experimenting with chat bots ourselves uh in our daily lives that it
[5:55] can make some fairly trivial mistakes on logic problems. They can't really play
[6:00] decent games of chess yet, which um is surprising. So there's something missing
[6:05] um uh still from these systems in terms of their consistency. And I think that's
[6:09] one, of the, things, that's, you, would expect from a general intelligence and
[6:13] art, you know, an AGI system is that it would be consistent across the board.
[6:17] And so sometimes people call it jagged intelligences. So they're really good at
[6:22] certain things, maybe even like PhD level, but then other things they're
[6:26] like not even high school level. So it's very uneven still the performances of
[6:30] these systems. They're very very impressive in certain dimensions. Um but
[6:34] they're still pretty uh uh basic in others and we've got to close those
[6:38] gaps. And you know there are theories reasons to why and depending on the
[6:41] situation it could even be uh the way that an image is is perceived and
[6:46] tokenized. So sometimes actually it doesn't even get all the letters that
[6:50] you you know so when you count letters in words um it sometimes gets that wrong
[6:54] but but it may not be seeing that each individual letter. So there's sort of
[6:57] different reasons for some of these things. uh and each one of those can be
[7:01] fixed and then you can see what's left. Um but I think consistency I think
[7:05] another thing is reasoning and thinking. So uh we have thinking systems now that
[7:11] at inference time they spend more time thinking and they're better they're
[7:14] better at outputting their answers. Um but it's not sort of super consistent
[7:19] yet in terms of like is it using that thinking time in a useful way um to
[7:24] actually double check and use tools to double check what it's outputting. I
[7:28] think we're we're on the way, but maybe we're only 50% of the way there.
[7:32] >> I, also, wonder, about, that, that, story, of of Alph Go and then Alpha Zero where you
[7:36] sort of took away all of the human experience and found that the model
[7:39] actually improved. Is there a sort of is there a scientific or a maths version of
[7:44] that in in the the models that you're creating? >> I, think, what, we're, trying, to, build
[7:47] today, it's more like Alph Go. So, you know, you effectively these these large
[7:52] language models, these foundation models, they're starting with all of
[7:55] human knowledge. you know, what we put on the internet, which is pretty much
[7:57] everything these days, and um compressing that into some useful
[8:02] artifact, right, which they can look up and and generalize from. But I do think
[8:06] we're we're still in the in the early days of having this uh uh search or
[8:11] thinking on top like AlphaGo had to kind of uh use that model to direct in useful
[8:16] reasoning traces, useful planning uh ideas. Um and and then come up with the
[8:22] best, you know, solution to whatever the problem is at that point in time. So I I
[8:27] don't feel like we're constrained at the moment with the kind of limit of human
[8:31] knowledge like the internet. I think the main issue at the moment is we don't
[8:34] know how to use those systems in a reliable way fully yet in the way we did
[8:38] with Alph Go. Um, but of course that was a lot easier because it was just it was
[8:42] a game. I think once you have Alph Go there, uh, you could go back just like
[8:47] we did with the the Alpha series and do an Alpha Zero where it starts sort of
[8:52] discovering knowledge for itself. I think that would be the next step, but I
[8:55] I that's obviously harder and so I think it's good to try and create the first
[9:00] step first with some kind of alphike system and then we can think about an
[9:04] alpha zero like system. But that is also one, of the, things, missing, from, today's
[9:09] systems is the ability to online learn and continually learn. So you know we
[9:13] train these systems, we balance them, we postrain them and then they're out in
[9:17] the world but they don't do they don't continue to learn out in the world like
[9:21] like we would. Um, and I think that's another missing critical missing piece
[9:25] from from these systems from from, you know, that will be needed before AGI.
[9:29] >> In, terms, of, all, of, those, missing, pieces, I mean, I know that there's this big
[9:32] race at the moment to release commercial products, but but I also know that that
[9:36] Google Deep Mind's roots really lie in in that idea of scientific research. And
[9:40] I I found a quote from you where you recently said, "If I had had my way, we
[9:43] would have left AI in the lab for longer and done more things like AlphaFold
[9:48] maybe cured cancer or something like that." Mhm. >> Do, do, you, think, that, we, lost, something
[9:52] by not taking that slower route? >> Um, I, think, we, lost, and, gained, something.
[9:57] So I feel like that would have been the more, pure, scientific, approach., At least
[10:02] that was my original plan say 15 20 years ago that you know when almost no
[10:07] one was working on AI. We just started we were just about to start Deep Mind.
[10:10] People thought it was a crazy thing to work on. Um but we believed in it and
[10:15] and and and I think that the idea was if we would make progress we would continue
[10:19] to sort of um incrementally build towards AGI be very careful about what
[10:24] each step was and and the safety aspects of it and so on analyze what the system
[10:29] was doing and so on. But in the meantime you wouldn't have to wait till AGI
[10:33] arrived before it was useful. You could branch off that technology and use it in
[10:37] really beneficial ways to society namely advancing science and medicine. So
[10:42] exactly what we did with AlphaFold actually which um it's not it's not a
[10:46] foundation model itself general model but it uses the same techniques you know
[10:51] transformers and other things and then blends it with um uh more specific
[10:55] things to that domain. So I imagined a whole bunch of those things getting done
[11:00] while which would be hugely you know you'd release to the world for just like
[11:03] we did with AlphaFold and indeed do things like cure cancer and so on. um
[11:08] whilst we were working on the sort of more the AGI track in the lab. Now it's
[11:13] turned out uh that chat bots were possible at scale and people find them
[11:18] useful and then they've now morphed into these foundation models that can do more
[11:22] than chat and text obviously including Gemini. Uh they can do images and video
[11:26] and all sorts of things and um that's also been very successful commercially
[11:33] and in terms of a product and I love that too. Like I've always dreamed of
[11:36] having the ultimate assistant that would help you in everyday life, make it more
[11:40] productive, maybe even protect your brain space a bit as well from an
[11:43] attention so that you can focus and be in flow and so on cuz you know today
[11:47] with social media it's just noise noise and I think AI can actually that works
[11:52] for you could help us with that. Um, so I think that's good, but it has created
[11:57] this pretty crazy race condition where there's many commercial organizations
[12:01] and even nation states all rushing to you know, improve and overtake each
[12:06] other and that makes it hard uh to do sort of rigorous science at the same
[12:11] time. We try to do both and I think we're getting that balance right. On the
[12:15] other hand, there are lots of pros of the way it's happened which is of course
[12:18] there's a lot more resources coming uh into the area. So that's definitely
[12:22] accelerated progress. Um and also um I think the general public are actually
[12:27] interestingly only a couple of months behind the absolute frontier in terms of
[12:32] what they can use. So everyone gets the chance to sort of feel for themselves
[12:36] what, AI, is, going to, be, like., And, I, think I think that's a good thing and then
[12:39] governments sort of understanding this better. The thing that's strange is that
[12:42] I mean this time last year I think there was a lot of talk about um you know
[12:46] scaling eventually hitting a wall about us running out of data and yet you know
[12:51] we're recording now Gemini 3 has just been released and it's leading on this
[12:56] whole range of different benchmarks. Um how how has that been possible? Like
[13:01] wasn't there supposed to be a problem with scaling hitting a wall? I think a
[13:04] lot of people thought that especially as other companies have sort of had slower
[13:08] progress should we say but I think we've never really seen any wall as such like
[13:13] what I would say is um maybe there's like diminishing returns and people when
[13:18] I say that people think only think like oh so there's no returns like it's zero
[13:22] or one it's either exponential or or it's asmtopic no actually there's a lot
[13:27] of room between those two regimes and I think we're in in between those so it's
[13:32] not like you're going double the performance on all the benchmarks every
[13:35] time you release a new iteration. Maybe that's what was happening in the early
[13:40] very early days, you know, three four years ago. But you are getting
[13:43] significant improvements like we've seen with Gemini 3 that are well worth the
[13:48] investment and the return on that investment and doing. So I that we
[13:51] haven't seen any slowdown on. There are issues like are we running out of just
[13:56] available data but there are ways to get around that you know synthetic data uh
[14:01] generating your you know these systems are good enough they can start
[14:03] generating their own data especially in certain domains like coding and math
[14:07] where you can verify the answer in some sense you could produce unlimited data
[14:11] so all of these things though are research questions and I think that's
[14:16] the advantage that we've always had is that um we've we've always been sort of
[14:21] research first and We I think we have the broadest and deepest research bench
[14:25] always have done. Um and if you look back at the last decade of advances
[14:29] whether that's transformers or alpha zero any of the things we just discussed
[14:33] that they all came out of Google or deep mind. So I've always said like if if
[14:38] more innovations are needed uh scientific ones then I would back us to
[14:43] be the place to do it just like we were you know in the previous sort of 15
[14:47] years for a lot of the big breakthroughs. So I think that's just
[14:50] what's transpiring and I actually really like it when the terrain gets harder
[14:54] because then it's not just worldass engineering you need which is already
[14:58] hard enough um but you have to ally that with worldclass research and science
[15:03] which is what we specialize in uh and on top of that we also have the advantage
[15:07] of world-class infrastructure with our TPUs and and other things that we've
[15:10] invested in a lot for a long time um and so that combination I think allows us to
[15:17] uh uh sort be at the frontier of the innovations as well as the scaling part
[15:22] and we effectively you can think of as 50 50% of effort is on scaling 50% of it
[15:27] is on innovation and I think my betting is you're going to need both to get to AGI
[15:31] >> I, mean, one, thing, that, we, are, still seeing even in Gemini 3 which is an
[15:35] exceptional model is uh this idea of hallucinations so I think um there was
[15:39] one metric that said uh it can still give an answer when actually it should
[15:44] decline um I mean could you build the system where Gemini gives a confidence
[15:50] score in the same way that Alpha Fold does. >> Yeah,, I, think, so., And, I, think, we, need
[15:54] that actually. And I think that's sort of one of the missing things. I think
[15:57] we're getting close. I think the better the models get, the more they know about
[16:01] what they know, if that makes sense. And so, and I think the more reliable we
[16:05] could sort of rely on them to actually introspect in some way or do more
[16:10] thinking and actually realize for themselves that they're uncertain or
[16:14] there's there's there's uncertainty over this answer. Uh, and then we've got to
[16:18] sort of work out how to train it in a way that where it can it can output that
[16:23] as a as a reasonable answer. Um, we're getting better at it, but it still sometimes, you know
[16:30] it sort of forces itself to answer when it probably shouldn't. Um, and then that
[16:34] can lead to a hallucination. So, I think, you know, a lot of the hallucinations are of that type
[16:39] currently. So, there's a missing piece there that that sort of has to be
[16:43] solved. And you're right, as we did solve it with Alpha Fold, but in in
[16:46] obviously a much more limited way >> cuz, presumably, behind, the, scenes, there
[16:50] is some sort of measure of probability of whatever the next token might be.
[16:54] >> Yes,, there, is, of, the, next, token., That's how it all works. But that doesn't tell
[16:58] you the overall arching piece is this is you know how confident are you about
[17:02] this entire fact or this entire um statement. And I think that's why you'll
[17:08] need this. I I think we'll need to use the thinking steps and the planning
[17:11] steps to go back over what you just output. At the moment, it's a little bit
[17:15] like the systems are just it's like talking to some a person and they just
[17:19] you know, when when they're in on a bad day, they're just literally telling you
[17:21] the first thing that comes to their mind., Most, most, of the, time, that, would
[17:24] be okay. But then sometimes when it's very difficult thing, uh you'd want to
[17:28] like stop pause for a moment and maybe go over what you were about to say and
[17:32] adjust what you were about to say. But perhaps that's happening less and less
[17:35] in the world these days, but um that's still the better way of having a
[17:39] discourse. So, you know, I think you can think of it like that. These models need to do that better.
[17:43] >> I, also, really, want, to, talk, to, you, about um the the simulated worlds and putting
[17:47] agents in them because we got to talk to your genie team earlier today.
[17:51] >> Tell, me, why, you, care, about, simulation. What What can a world model do that that
[17:55] a language model can't? >> Well,, look,, I, it's, it's, actually, been
[17:59] it's probably my longest standing passion is world models and simulations.
[18:04] uh in addition to AI and of course it's all coming together in our most recent
[18:07] work like Genie and I think um language models are able to understand a lot
[18:13] about the world I think actually more than we expected more than I expected
[18:17] because language is actually probably richer than we thought it contains more
[18:21] about the world than we maybe even even linguist maybe imagined and that's you
[18:24] know proven now with these new systems but there's still a lot about the the
[18:28] spatial dynamics of the world you know how spatial awareness um and the cont
[18:33] the physical context we're in um and how that works mechanically that um isn't is
[18:38] hard to describe in words and isn't generally described in in corpuses of of
[18:43] words and a lot of this is allied to learning from experience online
[18:47] experience there's a lot of things which you can't really describe something you
[18:50] have to just experience it um maybe the sensors and so on are very hard to put
[18:55] into words you know whether that's you know motor angles and smell and you know
[19:00] these kind of sensors it's very difficult to describe that in any kind
[19:03] of language. So I think there's a whole set of things around that and I think if
[19:07] we want robotics to work or a universal assistant that maybe comes along with
[19:11] you in your daily life maybe on glasses or you know on your phone um and helps
[19:17] you in your everyday life not just on your computer um you're going to need
[19:21] this kind of world understanding and uh world models are at the core of that. So
[19:26] this what we mean by a world model is this this sort of model that understands
[19:29] the causitative and effect of of the mechanics of the world right intuitive
[19:34] physics but um how things move how things behave. Um now we're seeing a lot
[19:39] of that in our video models actually and one way to show how do you test you have
[19:44] that kind of understanding well can you generate realistic worlds cuz if you can
[19:49] generate it then in a sense you must have understood uh uh the system must
[19:53] have encapsulated a lot of the mechanics of the world so that's why Genie and VO
[19:57] and these models are our video models and our sort of interactive world models
[20:02] are really uh impressive but also important steps towards showing we have
[20:07] generalized models and then hopefully some point we can apply it to, you know
[20:11] robotics and and and universal assistance. And then of course, one of
[20:15] my favorite things I'm definitely going to have to do at some point is
[20:17] reapplying it back to games and and uh you know, game simulations and create
[20:22] the ultimate games, which of course was maybe always my subconscious plan. >> All, of, this.
[20:27] >> Yeah., All, of, the, time., Exactly. >> What, about, science, too, though?, Could, you
[20:30] use it in that in that domain? >> Yes,, you, could., So, uh, science, you, know
[20:34] again I think building models of scientifically complex uh uh domains uh
[20:40] whether that's materials on atomic level um you know in biology uh but also like
[20:47] some physical things as well like weather one way to um understand those
[20:52] systems is to build simula learn simulations of those systems from the
[20:56] raw data right so you have a bunch of raw data let's say it's about the
[20:59] weather and obviously we have some amazing weather projects going on um and
[21:04] then you have a model that kind of learns those dynamics and can recreate
[21:08] those dynamics uh more efficiently. So uh than doing it by brute force. So I
[21:14] think there's huge potential for simulations and uh kind of world models
[21:18] maybe specialized ones for aspects of of science and mathematics.
[21:22] >> But, then, also, I, mean, you, can, drop, an agent into that simulated world too right?
[21:27] >> Yes., your, Genie, 3, team,, they, had, this really lovely quote which was almost no
[21:32] prerequisite to any major invention was made with that invention in mind. And
[21:36] they were talking about dropping agents into these simulated environments and
[21:39] allowing them to explore with sort of curiosity being their main motivator
[21:43] >> right?, And, so, that's, that's, another really exciting use of these these uh
[21:47] world models is you can we have another project called Simma. We just we just
[21:51] released Simma 2. sim, you know simulated agents where you have an
[21:54] avatar or an agent and you put it down into a virtual world. It can be a
[21:58] normal, it can be a kind of actual commercial game or something like that
[22:02] very complex one like No Man's Sky, kind of open world space game. Uh, and then
[22:07] you can you can instruct it with because it's got Gemini under the hood, you can
[22:11] just talk to the agent and and give it give it tasks. But then we thought
[22:16] well, wouldn't it be fun if we plug Genie into Simma and sort of drop
[22:20] Simmer, a Simma agent into a another AI that was creating the world on the fly.
[22:25] So now the the two AIs are kind of interacting in the minds of each other.
[22:29] So Simmer's, you know, the Simma agents trying to navigate this world and Genie
[22:33] is, as far as Genie is concerned, that's just a player and uh an avatar doesn't
[22:38] care. There's another AI. So it's just generating the world around whatever Sim
[22:41] is trying to do. So, so that it's kind of amazing to see them both uh
[22:46] interacting together. And I think this could be the beginning of an interesting
[22:49] training loop where uh you almost have infinite training uh examples because uh
[22:55] whatever the simmer agent's trying to learn, Genie can basically create on the
[22:59] fly. So, I think that you could imagine a whole world of like uh setting and
[23:03] solving tasks, just millions of tasks automatically, and they're just getting
[23:07] increasingly more difficult. So, we might try to set up a kind of loop like
[23:11] that. Um, as well as obviously those simmer agents could be great as game
[23:16] companions. Um, also some of the things that they learn could be useful also for robotics.
[23:21] >> Yeah., The, end, of, boring, NPCs, basically. >> Exactly., It's, going, to, be, amazing, for
[23:25] these games. Yeah. >> Those, worlds, that, you're, creating though, how do you make sure that they
[23:30] really are realistic? I mean, how do you ensure that you don't end up with
[23:32] physics that looks plausible but is actually wrong? >> Yeah,, it, that's, that's, a, great, question
[23:38] and and and and can be an issue. It's basically hallucinations again. So some
[23:42] hallucinations are good cuz cuz you it also means you you might create
[23:45] something interesting and new. So in fact sometimes if you're trying to do
[23:48] create creative things or trying to get your system to create new things, novel
[23:52] things, um a bit of hallucination might be good, but you want it to be
[23:55] intentional, right? So not uh so you kind of switch on the hallucinations
[24:00] now, right? Or the the creative um exploration. But yes, with with the with
[24:05] when you're trying to train a simmer agent, you don't want genie
[24:08] hallucinating kind of physics that are wrong. So actually what we're doing now
[24:11] is we're almost creating a physi physics benchmark where um we can use game
[24:16] engines which are very accurate with physics to create lots of like um fairly
[24:22] simple like the sorts of things you would do in your physics A level uh lab
[24:26] uh lessons, right? like you know rolling little balls down different tracks and
[24:30] seeing how fast they go and so like really teasing a part on a very basic uh
[24:35] level like Newton's three laws of motion has it encapsulated it um whether that's
[24:41] VO or Genie have these models encapsulated the physics of that 100%
[24:45] accurately and right now they're not they're kind of approximations and they
[24:49] look um realistic when you just casually look at them but they're not uh they're
[24:54] not accurate enough yet to rely on for say robotics So now we've got these really
[25:00] interesting models.
[25:08] Um, and with physics, I think that's going to probably involve generating
[25:12] loads and loads of ground truth. Simple videos of pendulums, you know, what
[25:16] happens when two pendulums go around each other, but then very quickly you
[25:19] get to like three body problems which are not solvable anyway. So I think it's
[25:23] going to be interesting. But what's amazing already is when you look at the
[25:26] the video models like VO and just the way it treats reflections and liquids
[25:32] it's pretty unbelievably accurate already,, at least, to, the, naked, eye., So
[25:36] the next step is actually going beyond what a human can amateur can perceive
[25:41] and uh would it really hold up to a proper physicsgrade experiment? I know
[25:46] you've been thinking about these simulated worlds for a really long time
[25:49] and uh I went back to the transcript of our first interview and in it you said
[25:53] that you really like the theory that consciousness was this consequence of evolution
[25:58] >> um, that, you, know, at, some, point, in, our evolutionary past there was like an
[26:01] advantage to understanding the internal state of another and then we sort of
[26:04] turned it in on ourselves. >> Does, that, make, you, curious, about, running
[26:08] sort of an a agent in evolution inside of a simulation? Sure. Um, I mean, I'd
[26:13] love to run that experiment at some point. Kind of re rerun evolution, rerun
[26:19] um almost social dynamics as well. Like the the Santa Fe used to run lots of
[26:25] cool experiments on little grid worlds. I used to love some of these, but
[26:28] they're mostly economists and they were trying to like, you know, run like
[26:31] little uh artificial societies and they found that things all sorts of
[26:35] interesting things got invented like that uh if you let agents run around for
[26:39] long enough with the right incentive structures. markets and banks and all
[26:42] sorts of crazy things. So I think it would be really cool and also just to
[26:46] understand the origin of life and the origin of consciousness. And I think
[26:50] that is the one of the big passions I had for for working on AI from the
[26:54] beginning was I think you're going to need these kinds of tools to really
[26:57] understand where we came from and what these phenomena are. Um, and I think
[27:02] simulations is is is one of the most powerful tools to do that because you
[27:06] can then do it statistically because you can run the simulation many times with
[27:11] control slightly different initial starting conditions and then um maybe
[27:15] run it millions of times and then understand what the slight differences
[27:18] are in a very uh controlled experiment sort of way which of course is you know
[27:24] very difficult to do in the real world for any of the really interesting
[27:27] questions we want to answer. So I think accurate simulations will be an unbelievable boon to science
[27:32] >> given, you, know, what, we've, discovered about sort of emergent properties of
[27:36] these models right having sort of conceptual understanding that we weren't
[27:39] expecting do you also have to be quite careful about running this sort of simulation
[27:42] >> uh, I, think, you, would, have, to, be, yes, but that that's the other nice thing about
[27:45] simulations you can run them in you know pretty safe sandboxes maybe eventually
[27:50] you want to air gap them uh and you can of course monitor what's happening in
[27:55] the in the in the simulation 24/7 uh and you have access to all the data.
[28:01] So we may need AI tools to help us monitor the simulations because um
[28:06] they'll be so complex they'll be and there'll be so much going on in them. If
[28:10] you imagine loads of AIs running around in a simulation uh uh it will be hard
[28:14] for any human scientist to keep up with it on but we could probably use other AI
[28:18] systems to help us analyze and flag anything interesting or worrying in
[28:23] those simulations uh automatically. I mean this I guess we're still talking
[28:27] sort of medium to long term in terms of this stuff. So so just going back to the
[28:31] trajectory that we're on at the moment. Um I also want to talk to you about the
[28:35] the impact that AI and AGI are going to have on on wider society.
[28:39] >> Um, and, last, time, we, spoke, you, said, that you thought AI was overhyped in the
[28:44] short term but underhyped in the long term. >> Um, and, I, know, that, this, year, there's
[28:48] been a lot of chatter about an AI bubble. >> Yes.
[28:50] >> What, happens, if, there, is, a, bubble, and, it bursts? What happens? Well, look, I I
[28:55] think yes that I still subscribe to it's overhyped in the short term still and
[28:59] still underappreciated in the in the medium to long term what's going to you
[29:03] know how transformative it's going to be. Um yeah, there is a lot of talk of
[29:07] course right now about AI bubbles. Um in my view uh I I think it there isn't it's
[29:14] not one thing binary thing are we or aren't we? I think there are parts of
[29:18] the AI ecosystem that are probably in bubbles. What one example would be, you
[29:23] know, just seed rounds for startups uh that basically haven't even got going
[29:27] yet and they're raising at tens of billions of dollars uh valuations just
[29:32] out of the gate. It's sort of interesting to see how how can that be
[29:35] sustainable? Um you know, my guess is probably, not, uh, at least, not, in, general.
[29:41] Um so there's that area. Then the people are worrying about obviously there's
[29:45] there's the big tech valuations and other things. I think there's a lot of
[29:48] real business underlying that. So um but it remains to be seen. I mean I think
[29:52] maybe for any any uh new unbelievably transformative and profound technology
[29:58] of which of course AI is probably the most profound. Uh you're going to get
[30:02] this uh overcorrection in a way. So when we started Deep Mind no one believed in
[30:06] it. No one thought it was possible. People were wondering what's AI for
[30:09] anyway. And then now fast forward 10 15 years and now obviously it seems to be
[30:15] the only thing people talk about in business. And um so it's a but you're
[30:19] sort of going to get it's almost an overreaction to the underreaction. Um so
[30:23] I think that's natural. I think we saw that with the internet. I think we saw
[30:26] with mobile and I think we're we're seeing or going to see it again with AI.
[30:30] Um I don't worry too much about are we in a bubble or not because from my
[30:34] perspective as you know leading Google deep mind and also obviously with Google
[30:38] as as and alphabet as a whole our job and my job is to make sure either way we
[30:45] uh are come out of it very strong and I think and we're very well positioned and
[30:49] I think we are tremendously well positioned either way. So if it continues going like it is now
[30:54] fantastic. we'll carry on, you know, all of these great things that we're doing
[30:57] and experiments and progress towards AGI. If there's a retrenchment, fine.
[31:02] Then also, I think we're in a great position because, uh, we have our own
[31:06] stack with TPUs. We also have, um, all these incredible Google products and
[31:11] you know, the profits that all makes to plug in our AI into. And we're doing
[31:15] that with search is totally revolutionized by AI overviews, AI mode
[31:20] with Gemini under the hood. We're looking at workspace at email, you know
[31:24] at YouTube. So, there's all these amazing things in Chrome. There's a lot
[31:27] of these amazing things that um AI we can see already are lowhanging fruit to
[31:32] apply uh Gemini 2 as well of course as Gemini app which is doing really well as
[31:37] well now and and and the idea of universal assistant. So, there's new
[31:42] products and I think they will in the fullness of time be super valuable, but
[31:46] we don't have to rely on that. we can just power up our existing uh ecosystem.
[31:51] Uh which is all sort of I think that's what's happened over the last year.
[31:54] We've got that really efficient now. >> In, terms, of, the, the, AI, that, people, have
[31:57] access to at the moment, I I know you said recently how important it is not to
[32:01] build AI to maximize user engagement just so we don't repeat the the mistakes
[32:05] of social media. But I but I also wonder whether we are already seeing this in a
[32:10] way. I mean people spending so much time talking to their chat bots that they end
[32:14] up kind of spiraling into self-radicalizing. Yeah. >> Um,, how, do, you, stop, that?, How, do, you
[32:19] build AI that that puts users at the center of their own universe, which is
[32:24] sort of the point of this in a lot of ways, but without creating echo chambers of one?
[32:28] >> Yeah,, it's, a, very,, you, know,, um,, careful balance that, you know, I think is one
[32:32] of the most important things that we as an industry have got to get right. So I
[32:37] think we've seen what happens with uh you know some systems that were overly
[32:41] syopantic or you know then you get these these sort of echo chamber
[32:45] reinforcements that are really bad for the person. So I think part of it is and
[32:49] actually what we want to build with with Gemini and I'm really pleased with the
[32:53] Gemini 3 persona that we had a great team working on and I helped with too
[32:57] personally is um just this sort of almost like a scientific uh personality
[33:03] that's um it's warm, it's helpful, it's light, but it's it's it's succinct to
[33:08] the point and it will push back on things in a friendly way that don't make
[33:13] sense. you know, rather than trying to reinforce you, you know, the idea that
[33:17] the earth's flat and you said it and it's like wonderful idea, you know, I
[33:20] don't think that's good in general for society if that were to happen. Um, but
[33:25] you got to balance it with what people want cuz people want uh these systems to
[33:28] be supportive um to be helpful with their with with their ideas and their
[33:34] brainstorming. So, you've got to get that balance right. And I think I think
[33:38] we are we're sort of developing a science of of personality and persona of
[33:42] like how to to to kind of measure what it's doing and where do we want it to be
[33:47] like on authenticity on humor you know these sorts of things. And then you can
[33:51] imagine there's a kind of base personality that it ships with. And then
[33:55] everyone has their own preferences. You know do you want it to be more humorous
[33:58] less humorous or or more succinct or more verbose? People like different
[34:02] things. So you add that additional personalization layer on it as well. But
[34:06] there's still the core base personality that everyone gets, right? Which is
[34:10] trying to try and adhere to the scientific method, which is the whole
[34:13] point of these. And we want people to use these for science and for medicine
[34:16] and health issues and so on. Uh and so um I think it's it's it's part of the
[34:22] science of getting these uh large language models right. And um I'm I'm
[34:28] quite happy with the direction we're going in currently. uh we got to talk to Shane Le a couple
[34:33] weeks ago um about uh AGI in in particular across everything that's
[34:38] happening in in AI at the moment the language models the world models you
[34:41] know and so on what's closest to your vision of AGI I think actually the
[34:47] combination of obviously there's Gemini 3 which I think is very capable but the
[34:52] Nano Banana Pro system we also launched last week which is an advanced version
[34:57] of our image creation tool what's really amazing about that it has also Gemini
[35:01] under the hood. So, it can understand not just images, it sort of understands
[35:05] uh what's going on semantically in those images. Uh and people have been only
[35:09] playing with it for a week now, but I've seen so much cool stuff on on social
[35:13] media about uh what people are using it for. So for example um you know you can
[35:18] give it a picture of a of of a of a complex plane or something like that and
[35:22] it can label all the diagrams of uh you know all the different parts of the
[35:26] plane and even visualize it in in a for like with all the different parts sort
[35:31] of exposed. Um so it has some kind of deep understanding of mechanics and and
[35:37] what make what you know makes up parts of objects what's materials. So it's a
[35:42] sort of um and it can you know render text really really uh accurately now. So
[35:46] I think that's sort of um it's getting towards a kind of AGI for imaging. Um I
[35:52] think it's uh a kind of general purpose system that can do anything across
[35:56] images. So I think that's very exciting. And then the advances in in world
[36:00] models, you know, Genie and Simma and what we're doing there. And then
[36:03] eventually we got to kind of converge all of those different they're kind of
[36:08] different projects at the moment and they're they're they're intertwined but
[36:11] we need to you know converge them all into one one big model and then that
[36:15] might be start becoming you know candidate for protoagi.
[36:19] >> I, know, you've, been, reading, quite, a, lot about the industrial revolution
[36:22] recently. Um, are there things that we can learn from what happened there to
[36:27] try and mitigate against the the sort of some of the disruption that that we can expect
[36:32] >> AGI, comes? >> I, think, there's, a, lot, we, can, learn., It's
[36:33] it's something you sort of study in school, at least, in, the, in, in, Britain,
[36:36] but but on a very superficial level like it was really interesting for me to look
[36:40] into how it how it all happened, what it started with, the reasons behind the
[36:45] economic reasons behind that which is like the textile industry and then the
[36:48] first computers were really the sewing machines, right? And then they became
[36:52] punch cards for the early forran computers, mainframes. And for a while
[36:56] it was very successful in Britain became like the center of the the textile world
[37:00] because they could make these amazingly high quality things for very cheap uh
[37:04] because of the automated systems. Um and then obviously the steam engines and all
[37:09] of those things came in. I think there's a lot of um incredible advances that
[37:13] came out of industrial revolution. So um child mortality went down and all of
[37:18] modern medicine uh and and um sanitary conditions, the kind of work life uh uh
[37:25] split and how that all worked was kind of worked out during the industrial
[37:28] revolution. But it also came with a lot of challenges like in it took quite a
[37:31] long time um roughly a century and um different parts of the labor force were
[37:38] dislocated at certain times and then new uh things had to be created new
[37:44] organizations like unions and other things had to be created in order to
[37:47] rebalance that. So like it was fascinating to see the whole of society
[37:51] sort of had to over time adapt and then you've got the modern world now. So
[37:56] there were I think there were lots of obviously pros and cons of the
[37:59] industrial revolution why it was happening but no one would want if you
[38:02] think about what it's done in total like abundance of you know people you know of
[38:06] food and in the western world and and modern medicine and all these things
[38:10] modern transport um that was all because of the industrial revolution. So, we
[38:14] wouldn't want to go back to pre-industrial revolution, but maybe we
[38:17] can figure out ahead of time by learning from it what those dislocations were and
[38:22] maybe mitigate those um earlier or more effectively this time. And we're
[38:27] probably going to have to because the difference this time is that it's
[38:30] probably going to be 10 times bigger than industrial revolution and it'll
[38:33] probably happen 10 times faster. So more like a decade then unfold over a
[38:37] decade than a century. One of the things that Shane told us was that that the
[38:40] kind of current economic system where you know you exchange your labor for
[38:44] resources effectively, it it just won't function the same way in a post AGI
[38:50] society. Do you have a vision of of how society should be reconfigured or might
[38:55] be reconfigured in a way that works? >> Yeah,, I'm, spending, more, time, thinking
[38:58] about this now and Shane's actually leading an effort here on that to sort
[39:01] of think about what a post AGI world might look like and what we need to
[39:05] prepare for. But I think society in general needs to spend more time
[39:08] thinking about that. Economists and social scientists and governments
[39:12] because, I, I I, as, with, the, industrial revolution you know the whole working
[39:17] world and working week and everything got changed from from pre-industrial
[39:20] revolution war agriculture and I think that's going to at least that level of
[39:25] change is going to happen again. So it's not surprising. I don't would not be
[39:29] surprised if we needed new economic systems, new economic models to uh to
[39:34] basically um help with that transformation and make sure for example
[39:38] the benefits are widely um distributed and maybe things like universal basic
[39:45] income and things like that are part of the solution. But I don't think that's
[39:47] the complete uh I think that's just what we can model out now, right? Because
[39:52] that would be a almost an add-on to what we have today. But I think there might
[39:56] be something way better systems where um more like direct democracy type systems
[40:01] where you can you know vote with a certain amount of of credits or
[40:05] something for what you want to see. It happens actually on local uh community
[40:09] level. You know here's a bunch of money. Do you want a playground or a tennis
[40:14] court or an extra classroom on the school? And then you let the community
[40:18] um sort of vote for it, right? So, and then and then maybe you could even
[40:22] measure the outcomes and then and then the people that sort of consistently
[40:26] vote for the for for things that that end up being um more wellreceived, they
[40:30] they have proportionally more influence for the next vote. So, there's there's a
[40:34] lot of interesting things I hear, you know, economist friends of mine who are
[40:37] are kind of brainstorming this and I think that would be great if we had a
[40:42] lot more work on that. And then there's the philosophical side of it of like
[40:46] okay so jobs will change and other things like that but then um but maybe
[40:50] we'll have fusion will have been solved and so we have this sort of abundant
[40:55] free energy so we're post scarcity so what happens to money um maybe
[41:00] everyone's better off but then what happens to purpose right because a lot
[41:03] of people get their purpose from you know their jobs and then providing for
[41:07] their families uh which is a very noble purpose so if that's you know so there's
[41:11] a lot of I I think some of these questions blend from economic questions
[41:15] into almost philosophical questions. >> Do, you, do, you, worry, that, people, don't
[41:19] seem to be paying attention sort of or moving as quickly as you'd like to see?
[41:23] What would it take for for people to sort of recognize that we need
[41:26] international collaboration on this? I am worried about that and I wish that
[41:30] and and again in a sort of ideal world there would have been a lot more
[41:33] collaboration already and international specifically uh and a lot more research
[41:38] and and sort of um I guess exploration and discussion going on about these
[41:43] topics. I'm actually pretty surprised there isn't more of that being discussed
[41:47] given that you know even our timelines which were there are some very short
[41:50] timelines out there but even ours are 5 to 10 years which is not long for for
[41:54] for for institutions or things like that to be built to to handle this. Um and
[42:00] one of the worries I have is that the institutions that do exist they you know
[42:04] seem to be very fragmented and not very influential to to the level that you
[42:08] would need. Um, so it may be that that that that there are there aren't the
[42:13] right institutions to deal with this currently. And then of course if you add
[42:16] in the geopolitical tensions that are going on at the moment around the world
[42:19] it seems like collaboration, cooperation is harder than ever. Um, just look at
[42:23] climate change and and um how hard it is to get any agreement on anything to do
[42:29] with that. So um so we'll see. I think as the stakes get higher and as these
[42:35] systems get more powerful and maybe this is one of the benefits of them being in
[42:38] products is uh the the you know everyday uh person that's not working on this
[42:44] technology will get to feel the increase in the power of these things and the
[42:48] capability and so that will then reach government and then maybe um uh they'll
[42:53] see sense as we get closer to to AGI. >> Do, you, think, it, will, take, a, moment, an
[42:58] incident for everyone to sort of sit up and pay attention?
[43:01] >> I, don't, know., I, mean,, I, hope, not., Most of the main labs are pretty pretty
[43:05] responsible. We try to be as responsible as possible. You know, that's always
[43:09] something we've, as you know, if you followed us over the years, that's been
[43:12] at the heart of what everything we do. Doesn't mean we'll get everything right
[43:15] but we try to be as thoughtful and as scientific in our approach as possible.
[43:20] Um, I think most of the major labs are are trying to be responsible. Also
[43:24] there's good commercial pressure actually to be responsible. If you think
[43:28] about agents, uh, and you're renting an agent to another company, let's say, to
[43:32] do something, um, that other company is going to want to know what the limits
[43:38] are and the boundaries are and the guardrails are on those agents, you
[43:41] know, in terms of what they might do and not just mess up the data and all of
[43:44] this stuff. So, I think that's good because the pe the more kind of carboy
[43:48] operations, they won't um get the business because the enterprises won't
[43:52] choose them. So I think the kind of capitalist system will actually be useful here to reinforce
[43:58] responsible behavior which is good but then there will be rogue actors um maybe
[44:03] rogue nations maybe rogue organizations um maybe people building on top of open
[44:08] source I don't know like obviously it's very difficult to stop that then um
[44:13] something may go wrong and uh hopefully it's just sort of medium-sized and then
[44:19] that will be a kind of warning shot to to to humanity across the bow and then
[44:24] that might be the moment to kind of um advocate for uh international uh
[44:29] standards or international cooperation or collaboration at least on some the
[44:34] high level basic or you know kind of like what's the basic standards we we we
[44:38] would want and and and agree to I'm hopeful that that will be possible
[44:42] >> in, the, long, term, so, beyond, AGI, and, and towards ASI right artificial super
[44:47] intelligence do you think that there are some things that that humans can do that
[44:51] machines will ever be able to manage? >> Well,, I, think, that's, the, big, question
[44:55] and I feel like this is related to as you know, one of my favorite topics is
[44:59] cheuring machines. I've always felt this that if we build a GI and then use that
[45:04] as a simulation of the mind and then compare that to the real mind, we will
[45:09] then see what the differences are and uh potentially what's special um and
[45:14] remaining about the human mind, right? Maybe that's creativity, maybe it's
[45:17] emotions, maybe it's dreaming. There's a lot of consciousness. There's a lot of
[45:22] um hypotheses out there about what may or may not be computable. And this comes
[45:27] back to the chewing machine question of like what is the limit of a chewing
[45:31] machine? And I think that's the central question of my life really ever since I
[45:34] found out about chewing and chewing machines. And um you know I think that's
[45:39] that's I fell in love with that. That's my core passion. And I think um
[45:44] everything we've been doing is been sort of pushing the notion of what a cheuring
[45:49] machine can do to the limit including you know folding proteins right and so
[45:53] it turns out I'm not sure what the limit is maybe there isn't one right and of
[45:58] course the my quantum computing friends would would say there are limits and and
[46:02] you need quantum computers to do quantum systems but I'm really not so sure and
[46:07] I've actually you know discussed that with some some some of the quantum folks
[46:11] and it may that we need data from these quantum systems in order to create a
[46:16] classical simulation. Um and then that that comes back to the mind which is is
[46:20] it all classical computation or is there something else going on you know like
[46:25] Roger Penrose believes you know there's quantum effects in the brain. If there
[46:28] are then and that's what consciousness is do with then machines will never have
[46:32] that, at least, the, the, the, classical machines we'll have to wait for quantum
[46:35] computers. Um but if they if there isn't then there may not be any limit maybe in
[46:40] the universe everything is computationally tractable and therefore
[46:44] if you look at it in the right way and therefore chewing machines might be able
[46:47] to model everything in the universe I I'm currently if you were to get make me
[46:51] guess I would guess that and I'm working on that basis until physics um shows me otherwise
[46:57] >> so, there's, nothing, that, cannot, be, done within these sort of computational
[47:01] >> well, no, one's, put, it, this, way, nobody's found anything in the universe that's
[47:04] that's non-computable So far >> so, far, >> right?, And, I, think, we've, already, shown
[47:10] you can go way beyond the the usual complexity theorist P= MP view of like
[47:15] what a classical computer could do today. Things like protein folding and
[47:19] go and so on. So I don't think anyone knows what that limit is. And that's
[47:24] really if you boil down to what we're doing at Deep Mind and Google and what
[47:28] I'm trying to do is is find that limit. But then in the limit of that though
[47:32] right, is that in the limit of that idea is that, you know, we're sitting here
[47:35] sort of there's like the warmth of the lights on our face. We kind of hear the
[47:39] wear of the machine in the background. There's like the feel of the desk under our hands.
[47:42] >> All, of, that, could, be >> replicable, by, a, classical, computer.
[47:47] >> Yes., Well,, I, think, in, the, end,, my, view on this is why I love K as well is all
[47:51] all, of, all all, of, my, two, favorite philosophy
[47:58] is a construct of the mind. I think that's true. And so, yes, all of those
[48:01] things you mentioned, they're coming into our sensory apparatus and they feel
[48:05] different, right? The light, the warmth of the light, the feel, the touch of the
[48:08] table, but in the end, they're it's all information. And we're information
[48:12] processing systems. And I think that's what biology is. This is what we're
[48:15] trying to do with isomeorphic. That's how I think we'll end up curing all
[48:18] diseases is by thinking about biology um as an information processing system. And
[48:25] I think in the end that's going to be and I'm working on my spare time, my 2
[48:28] minutes of spare time, you know, physics theories about uh things like
[48:32] information being the most fundamental unit, should we say, of the universe
[48:36] not energy, not matter, but information. And um so it may be that these are all
[48:41] interchangeable in the end, right? But we just sense it. We feel it in a
[48:45] different way. Um but you know, as far as we know, this is still all these
[48:49] amazing sensors that we have, they're still computable by a chewing machine.
[48:53] But this is why your simulated world is so important, right?
[48:56] >> Yes., Exactly., Because, that, would, be, the way to get one of the ways to get to it.
[49:00] What's the limits of what we can simulate? Because if you can simulate
[49:02] it, then in some sense, you've understood it. >> I, wanted, to, to, finish, with, some, personal
[49:07] reflections >> of, of, what, it's, like, to, be, at, the forefront of this. I mean, does the
[49:14] emotional weight of this ever sort of weigh you down? Does it ever feel quite isolating?
[49:18] >> Yes., Um,, look,, I, I, don't, sleep, very much, partly because it's too much work
[49:23] but also I have trouble sleeping. It's very complex emotions to deal with
[49:27] because it's unbelievably exciting. Um you know, I'm I'm basically doing
[49:32] everything I ever dreamed of. And we're at the absolute frontier of science on
[49:37] in so many ways, um, applied science as well as machine learning. And that's
[49:42] exhilarating as all scientists know that that feeling of being at the frontier
[49:46] and discovering something for the first time. And that's happening almost on a
[49:50] monthly basis for us. So, which is amazing. Um, but then of course we as as
[49:55] and Shane and I and others who've been doing this for a long time, we
[49:58] understand it better than anybody. Um the enormity of what's coming and this
[50:02] thing about is still under actually appreciated. In fact, what's going to
[50:06] happen in more of a 10-year time scale. Um, including to things like the the
[50:11] phil philosophical uh, you know, what it means to be human, what's important
[50:15] about that? all of these questions are going to come up. Um, and so it's it's
[50:20] it's a big responsibility. Um, but we have an amazing team thinking about
[50:25] these things. Um, but also it's something, I, guess, at least, myself, I've
[50:30] trained for my whole life. So, you know ever since my early days playing chess
[50:34] and and then working on computers and games and simulations and neuroscience
[50:39] it's all been for uh this kind of moment. Um, and it's roughly what I
[50:44] imagined it was going to be. So, that's partly how I cope with it is just training.
[50:48] >> Are, there, parts, of, it, that, have, hit, you harder than you expected though?
[50:52] >> Uh,, yes,, for, sure., On, the, way,, I, mean, even the Alpha Go match, right? Just
[50:56] seeing you know that how we managed to to crack Go, but Go was this beautiful
[51:02] mystery and it changed it. And so, that was that was interesting and kind of
[51:06] bittersweet. I think even the the more recent things of like language and then
[51:11] imaging and you know what does it mean for creativity uh I I'm you know have
[51:15] huge respect and passion for the creative arts and having done game
[51:19] design myself and you know I talked to film directors and it's it's an
[51:22] interesting dual moment for them too. There's like first on one hand they've
[51:25] got these amazing tools that speed up prototyping ideas by 10x but on the
[51:30] other hand um is it replacing certain creative skills? So I think there's
[51:36] there's sort of these trade-offs going on um all over the place which um I
[51:41] think is inevitable with something as uh a technology as powerful and as
[51:45] transformative as as AI is as in the past electricity was and internet and
[51:50] we've you know we've seen that that is the story of humanity is we are tool
[51:55] making uh animals and that's what we love to do and for some reason we also
[52:00] have a brain that can can understand science and do science which is amazing.
[52:05] but also sort of insatiably curious. I think that's the heart of what it means
[52:09] to be human. And I think I've just had that bug from the beginning. And my
[52:14] expression of trying to answer that is is to build AI.
[52:19] >> When, you, and, the, other, AI, leaders, are, in a room together, is there sort of sense
[52:22] of solidarity between you that that this is a group of people who all know the
[52:26] stakes, who all really understand their things, or or does the competition kind
[52:30] of keep you apart from one another? >> Well,, we, all, Yeah,, we, all, know, each
[52:33] other. I get on with pretty much all of them. Some of the others don't get on
[52:36] with each other. Uh and there is it's hard because that we're also in the most
[52:40] ferocious uh uh capitalist sort of competition there's ever been probably.
[52:45] You know, investor friends of mine and VC friends of mine who who who were
[52:50] around in the dotcom era say this is like 10x more ferocious and intense than
[52:55] that was. In many ways, I love that. I mean, I I live for competition. It's
[52:59] it's it's it's you know I've always loved that since my chess days but
[53:03] stepping back uh I understand I hope everyone understands that there's a much
[53:07] bigger thing at stake than just you know company successes and and and you know that type of thing
[53:13] >> when, it, comes, to, the, next, decade, when you think about it are there big moments
[53:18] coming up that you're personally most apprehensive about
[53:21] >> I, think, right, now, the, systems, are, you know I I call them passive systems you
[53:26] you put the energy in as the user you know the question or the what's the task
[53:30] and then they uh these systems kind of provide you with some summary or some
[53:34] answer. Um so very much it's it's human directed and human energy going in uh
[53:39] and human ideas going in. The next stage is agent-based systems, which I think
[53:43] we're going to start seeing. We're seeing now, but they're pretty
[53:46] primitive. Like in the next couple of years, I think we'll start seeing some
[53:49] really impressive reliable ones. And um I think those will be incredibly useful
[53:54] and capable if you think about them as an assistant or something like that, but
[53:58] also they'll be more autonomous. So I think the risks go up as well uh with
[54:03] those types of systems. So I'm I'm quite worried about uh what those sorts of
[54:08] systems will be able to do maybe in two, three years time, you know. So
[54:13] we're working on cyber defense in preparation for uh a world like that
[54:18] where maybe there's millions of agents you know, roaming around on the internet.
[54:22] >> And, what, about, what, you're, most, looking forward to? I mean, is there is there a
[54:25] day when you'll be able to retire sort of knowing that your work is done or or
[54:30] is there more than a lifetime's worth of work left to do? >> Yeah,, I, always, Well,, I, I, could
[54:35] definitely do with sabbatical um and I would spend it doing stuff. Yeah, a week
[54:40] off for even even a day would be good. Um, but look, I think my mission has
[54:45] always been to get to kind of help uh the world steward AGI safely over the
[54:50] line for all of humanity. So, I think when we get to that point, of course
[54:54] there's then there's super intelligence and there's post AGI and there's all the
[54:58] economic stuff we were discussing and societal stuff and maybe I can help in
[55:01] some way there. But I think um that will be my core part of my mission, my life
[55:07] mission uh will be done if it's a I mean it's only a small job, you know, just
[55:11] get that over the line or help the world get that over the line. You know, I
[55:15] think it's going to require collaboration like we talked earlier. Um
[55:18] and I'm quite a collaborative person. So I hope I can I can help with that from
[55:22] the position that I have. >> And, then, you, get, to, have, a, holiday
[55:25] >> and, then, I'll, get, I'll, have, the, Yeah, exactly. a well- earned sbatical.
[55:29] >> Yeah,, absolutely., Deis,, thank, you, so much. helpful as always.
[55:33] >> Well,, that, is, it, for, this, season, of, Goo Deep Mind the podcast with me, Professor
[55:37] Hannah Fry. But be sure to subscribe so you will be among the first to hear
[55:41] about our return in 2026. And in the meantime, why not revisit our vast
[55:46] episode library because we have covered so much this year. From driverless cars
[55:50] to robotics, world models to drug discovery. Plenty to keep you occupied. See you soon.