The Man Behind Google’s AI Machine | Demis Hassabis Interview (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) Hello and welcome to the Tech Download, (00:00:03) a new CNBC original podcast where we (00:00:05) unpack the tech stories that matter (00:00:07) most. Each season, we dive into one big (00:00:10) theme and what it means for your money (00:00:12) with insights from the industry's most (00:00:14) influential voices. I've always thought (00:00:16) that in the end, it would be the most (00:00:18) important technology we'll ever invent. (00:00:21) And it's sort of the natural progression (00:00:23) really of the computer age. This season, (00:00:25) we're looking at Google DeepMind, the (00:00:27) powerhouse driving the tech giant's AI (00:00:30) push. We've been given rare access to (00:00:32) key figures at the company, including (00:00:34) our guest for this episode, DeepMind (00:00:37) co-founder and CEO Deis. I think it's (00:00:41) going to be like the industrial (00:00:42) revolution, but maybe 10 times bigger, (00:00:45) 10 times faster. So, it's incredible (00:00:47) amount of transformation, but also (00:00:49) disruption that's going to happen. (00:00:54) Hey everyone and welcome to the tech (00:00:56) download. Allow me to reintroduce (00:00:58) myself. I'm Arjent Karple, senior (00:01:00) technology correspondent at CNBC based (00:01:02) in London. Um, and I've got a very (00:01:05) special new co-host with me. Hey there, (00:01:07) Arjent. Yeah, Steve Kovac here. Um, I (00:01:09) cover tech over here in New York. I (00:01:12) mostly focus on Apple and Microsoft, but (00:01:14) look, I've been covering the tech (00:01:16) industry for over 15 years now. I kind (00:01:18) of have a good grasp on everything and (00:01:20) I'm so excited to be here with you Arjun (00:01:22) because I've just admired your work from (00:01:24) across the ocean for so long and now we (00:01:27) actually get to kind of collaborate and (00:01:28) do this thing together. Uh I think it's (00:01:30) going to be a good time. It's going to (00:01:31) be so fun Steve. So between us, we think (00:01:33) we've got nearly three decades of (00:01:35) experience covering tech. And the crazy (00:01:37) thing is we've got so much to learn. And (00:01:39) I think over the course of us doing this (00:01:41) podcast, we're going to learn so much, (00:01:42) speak to so many uh interesting people. (00:01:45) I'm so excited that this first series (00:01:47) we're kicking off with an insight into (00:01:50) Google DeepMind, one of the world's (00:01:51) leading AI labs as well. Um and just for (00:01:54) our listeners and our viewers, a a quick (00:01:56) intro, I guess, to uh Google DeepMind. (00:01:59) It was a company founded in 2010 here in (00:02:01) London where where I sit as well. Very (00:02:03) small company founded by three people. (00:02:05) Deisabis, Shane Leg, and Mustafa Sleman (00:02:08) who who's at Microsoft now, right? (00:02:10) >> Yeah. And in fact, I interviewed him uh (00:02:12) god nearly a year ago now. Uh Mustafa (00:02:14) Sullean. Uh he's basically doing what (00:02:17) Deis is doing over at Google. And it's (00:02:19) just kind of interesting to see how (00:02:21) Google was like this incubator, so to (00:02:23) speak, for all of this top AI talent (00:02:25) around the world. Demis obviously stuck (00:02:27) around. He's running deep mind over (00:02:29) there. Uh what I also think is really (00:02:31) interesting though is just this AI (00:02:33) moment Arjun we've been living through (00:02:35) for the last three years and how three (00:02:38) years ago chat GPG comes on the scene (00:02:40) and Google was kind of seen as under (00:02:42) threat. They went through this code red. (00:02:44) They had to go through a bunch of (00:02:45) reorganizations internally. Eventually (00:02:48) Demis came out on top as the leader of (00:02:50) AI. And guess what it 2025 was a really (00:02:54) interesting year for AI over at Google. (00:02:56) they kind of caught up and in some ways (00:02:58) even surpassed what chatbt was already (00:03:01) doing. And this is really interesting (00:03:03) because the fundamental technology for (00:03:06) all this these large language models (00:03:07) we've been talking about for so many (00:03:09) years started at Google and the (00:03:11) perception was Google let chatbt kind of (00:03:14) take that technology and run away with (00:03:16) it. But now in my view at least Gemini (00:03:19) is pretty much on par if not better than (00:03:21) Chat GBT (00:03:22) >> and Google DeepMind is integral for (00:03:25) this. I mentioned it was found in 2010. (00:03:27) Google actually acquired DeepMind in (00:03:30) 2014. I was very new into my career as a (00:03:33) tech reporter as well. Google paid (00:03:34) around 400 million pounds uh for Deep (00:03:37) Mind at the time in 2014. About $540 (00:03:40) million. It's a stake this day that (00:03:43) could be worth tens of billions, maybe (00:03:45) hundreds of billions of dollars (00:03:47) according to some estimates today. And (00:03:49) DeepMind really is very much responsible (00:03:52) for Google's AI. We talk about Gemini, (00:03:54) the the the chatbot, the the AI um that (00:03:57) that Google's released to consumers. (00:03:59) This is powered so much by the (00:04:01) technology coming out of DeepMind. But (00:04:03) even before all of this, DeepMind was (00:04:05) having some big breakthroughs. There was (00:04:07) a a big moment a few years ago when they (00:04:10) released a system called Alph Go. This (00:04:12) was the first computer program that was (00:04:14) able to defeat a world champion in a (00:04:16) game called Go. This is a very complex (00:04:19) uh game and it was seen at the time as (00:04:21) one of the grand challenges uh of AI (00:04:23) because it was such a complex game with (00:04:25) so many different combinations (00:04:27) available. The other big breakthrough of (00:04:28) course was was something called (00:04:30) Alphafold. This was another AI system (00:04:32) developed at deep mind that could (00:04:34) accurately predict 3D models of protein (00:04:36) structures and the idea is here is if (00:04:39) you could do that this may lead to some (00:04:41) medical breakthroughs. So this (00:04:42) advancement of science uh has been (00:04:45) pretty core to what uh DeepMind's been (00:04:47) up to and clearly it was a significant (00:04:50) bet from Google more than 10 years ago (00:04:52) because it's helped turn Google into an (00:04:54) AI world leader today. (00:04:56) >> Yeah and that that's exactly right and (00:04:58) what really struck me about Deep Mind (00:05:00) having watched them for so many years is (00:05:02) how rooted in science they were. They (00:05:04) weren't necessarily trying to build (00:05:06) consumer products like they do now. They (00:05:08) were really trying to solve fundamental (00:05:10) problems in science and really usher in (00:05:13) this era of AI powered drug discovery of (00:05:16) other big complex problems like climate (00:05:18) change. I know Demis talks about that a (00:05:20) lot and he's going to talk about that in (00:05:22) your conversation as well Arjun. (00:05:23) >> Absolutely Steve um look it's a great (00:05:26) scene setter for Deep Mind. So let's get (00:05:28) into the conversation with its CEO (00:05:30) Demis. (00:05:33) >> Demis thanks for joining me on the tech (00:05:35) download. Appreciate it. Thanks for (00:05:37) having us. (00:05:37) >> Uh Dis, we're going to try to get (00:05:39) through a lot in our time here, but I (00:05:40) want to start first with the technology (00:05:42) itself. And we've been talking about AI (00:05:44) and we've been talking about the the (00:05:46) capabilities and how they've been (00:05:47) continuously improving um as well. Now, (00:05:50) in the tech world, I know there's a lot (00:05:51) of conversations about how good can (00:05:54) these models get, how good can these (00:05:55) systems get, and there's a lot of debate (00:05:57) around this idea of of scaling laws. Um (00:06:00) for our for our listeners, you know, (00:06:01) it's this idea of of more compute, more (00:06:04) data, bigger models. uh eventually will (00:06:06) lead to bigger systems as well. You said (00:06:08) we need to push scaling laws to the (00:06:10) maximum. Um (00:06:12) >> there's questions over now. Are we (00:06:14) hitting any kind of walls in terms of (00:06:16) progress of those scaling laws in terms (00:06:17) of the ability for these models to get (00:06:19) better? And just from you know what (00:06:21) you've been developing here at DeepMine, (00:06:22) what are you seeing? (00:06:24) >> Well, look, I think scaling laws um are (00:06:28) going very well. So we're definitely (00:06:30) seeing increased capabilities by putting (00:06:33) in more compute, more data, uh, and (00:06:35) making these models generally larger. So (00:06:37) that trend is continuing. Um, may be not (00:06:40) as fast as it was a couple of years ago. (00:06:42) So um, there's some talk of diminishing (00:06:45) returns. Uh, and and but but there's a (00:06:47) big difference between sort of no (00:06:49) returns and exponential. And I think (00:06:51) we're somewhere in the middle where (00:06:52) there's very good returns and that's (00:06:54) worth doing. Um on top of that if I to (00:06:56) you know in terms of like getting all (00:06:58) the way to AGI artificial general (00:07:00) intelligence um you know maybe that (00:07:02) there's one or two uh big innovations (00:07:05) still needed as well and maybe missing (00:07:07) in addition to the scaling up of um kind (00:07:10) of the existing ideas. (00:07:11) >> We'll get on to AGI very shortly but (00:07:13) what what are missing in your view? (00:07:15) Well, if you look at I mean we've all (00:07:17) you know played around with different (00:07:18) chat bots and you can see that uh you (00:07:21) know they can do very impressive things (00:07:24) um in some dimensions but they're kind (00:07:26) of like jagged intelligences I like (00:07:27) calling them in the sense of like (00:07:29) they're very good at certain things but (00:07:30) there are other things that they don't (00:07:32) do they're not capable of at all and um (00:07:34) and if you pose a question in a certain (00:07:36) way you you know you find that they're (00:07:38) flawed um and they they can't do some (00:07:41) relatively simple things and so for a (00:07:43) true general intelligence you shouldn't (00:07:45) see that inconsistency should be (00:07:47) consistent across the board. Um, and (00:07:48) also there are things like it can't (00:07:50) continually learn. It can't learn new (00:07:51) things online. It can't truly create (00:07:54) original things. So there's quite a few (00:07:56) capabilities that you would like to see (00:07:58) and you would need for general (00:07:59) intelligence that are missing from (00:08:00) today's systems. That's really (00:08:02) interesting. So what what would be the (00:08:03) sort of unlock to get to those (00:08:05) intelligent systems? I just want to (00:08:07) quickly discuss a conversation I had (00:08:08) with with Thomas Wolf who's the (00:08:10) co-founder over at Hugging Face and he (00:08:12) was talking to me um a few months back (00:08:14) about his view on LLMs in particular (00:08:16) large language models and just saying (00:08:17) they're really great and you know you (00:08:19) use these chat bots and the chat bots (00:08:20) say hey great question great idea um and (00:08:23) here's here's all the information you (00:08:24) need to know but what's missing is the (00:08:27) ability for these systems to come up (00:08:29) with new and novel ideas perhaps and (00:08:31) particularly I know you're so interested (00:08:33) in science and what AI could do to (00:08:36) unlock new drugs or discover new (00:08:37) diseases etc. Um that actually maybe the (00:08:40) LLM's limitations are there that you (00:08:42) can't come up with these Nobel Prize (00:08:44) winning ideas, these novel ideas. (00:08:46) >> So perhaps there needs to be some sort (00:08:48) of new architecture. What's your what's (00:08:50) your thinking on that at the moment? (00:08:51) >> Yeah. Well, look, my passion for and my (00:08:53) whole reason I I've spent my whole (00:08:54) career on AI is I think eventually it (00:08:57) will be the ultimate tool for science. (00:08:59) And of course, we've shown that with (00:09:00) things like Alphafold and all of the (00:09:02) science work we've been doing over the (00:09:03) last decade, but there's still a long (00:09:05) way to go uh in terms of uh can an AI (00:09:08) actually come up with a new hypothesis (00:09:10) itself? not just solve a conjecture that (00:09:13) is already out there which would be (00:09:14) already useful and impressive but can it (00:09:16) actually come up with a new conjecture a (00:09:18) new a new idea about how the world might (00:09:21) work and so far um these systems can't (00:09:24) do that they don't really have the (00:09:25) capability to do that so there seems to (00:09:27) be something missing um I think uh uh uh (00:09:29) some of the capabilities that are (00:09:30) required are kind of long-term planning (00:09:32) better reasoning maybe also the idea of (00:09:36) a world model uh this idea of like you (00:09:38) know the system actually understanding (00:09:40) better the physics of the world so that (00:09:42) it can run simulations um you know kind (00:09:45) of in its mind uh to test its own (00:09:48) hypothesis you know these are uh things (00:09:50) that you know the best scientists do (00:09:52) human scientists do uh and so far our AI (00:09:55) systems you know are not able to do that (00:09:57) >> can you just help us understand a bit (00:09:58) more of this idea of world models (00:10:00) because it may be a term people are (00:10:01) hearing for the first time you know how (00:10:02) that guess they differ from LLMs (00:10:05) language (00:10:05) >> so LLMs and and the models we use at the (00:10:07) moment are you know mostly around text (00:10:09) um Of course, things like Gemini, our (00:10:11) our foundation model can also cope with (00:10:14) images and video and audio. So, (00:10:16) different modalities. Um, but it's still (00:10:20) actually understanding the physics of (00:10:21) the world, the causality of the world. (00:10:23) You know, how one thing affects another (00:10:25) thing. Um, can you plan a long time into (00:10:28) the future? These are all related (00:10:29) concepts. And if you really want to (00:10:31) understand how the world works so that (00:10:33) maybe you can invent something new in (00:10:35) the world or explain something about the (00:10:37) world that was not known before which is (00:10:39) basically what a scientific theory does (00:10:41) then you have to have uh this this (00:10:44) accurate model of how the world works. (00:10:46) Um you know starting with intuitive (00:10:48) physics and and and how the physics of (00:10:50) the world works but all the way up to (00:10:51) biology you know and and and economics. (00:10:55) >> Yeah. And and do you envision the world (00:10:57) if we get to this idea of artificial (00:10:58) general intelligence this sort of human (00:11:00) level of intelligence that that there (00:11:02) will be a combination of LLMs and world (00:11:03) models working together or will sort of (00:11:05) world models supersede in some sense (00:11:07) LLMs? (00:11:08) >> No, I think there will be some (00:11:09) convergence of these technologies. (00:11:10) That's at least my betting is is um uh (00:11:13) there will be these LLMs or foundation (00:11:16) models you know like Gemini under the (00:11:18) hood that will be a key component. I (00:11:20) think the question I think there's (00:11:21) almost no doubt about that in my mind (00:11:23) which is why uh we must try and scale (00:11:25) those systems as as as big and as as (00:11:27) powerful as we can but the question is (00:11:30) is it the only component that's needed (00:11:32) for an AGI and um that's where I think I (00:11:36) suspect uh other types of technologies (00:11:38) and other types of capabilities will be (00:11:40) needed and I think these world model (00:11:42) capabilities and we're working on our (00:11:45) versions called Genie uh and uh and we (00:11:48) have video models like VO (00:11:50) state-of-the-art video models that you (00:11:51) can generate videos from from text and (00:11:54) you can think of video models and and (00:11:56) interactive models like Genie as kind of (00:11:58) uh you know early embionic uh world (00:12:01) models where if you can generate (00:12:03) something that's realistic about the (00:12:05) world then in a sense your model (00:12:07) understands that about the world (00:12:08) otherwise how could it have generated it (00:12:10) demos you mentioned this AGI artificial (00:12:12) general intelligence I know there's (00:12:13) various definitions of it floating (00:12:15) around you've previously said you (00:12:17) believe that reaching AGI could (00:12:19) somewhere in the in the realm of 5 to 10 (00:12:21) years away. Um, is this still your view (00:12:24) given, I guess, some of the profound (00:12:26) developments we've seen in 2025? (00:12:28) >> Yes, I think we're right on track from (00:12:29) that. Actually, when we started Deep (00:12:30) Mind back in 2010, we thought this would (00:12:32) be a 20-year kind of mission to to build (00:12:35) AGI, uh, you know, a system that's (00:12:37) capable of exhibiting all the cognitive (00:12:39) capabilities we we we have, including, (00:12:41) you know, things like, uh, uh, uh, true (00:12:43) innovation and creativity, um, and (00:12:46) planning and reasoning and things like (00:12:48) that. And I think we're about 5 to 10 (00:12:50) years away from that. Um, but that's, (00:12:52) you know, pretty incredible if you think (00:12:54) about how transformative a technology (00:12:56) this is. (00:12:56) >> You mentioned there might need to be (00:12:57) some more technology breakthroughs. (00:12:59) We're seeing things like the models (00:13:00) advancing. We're seeing the (00:13:01) semiconductors advancing rapidly as (00:13:03) well. Are there any currently (00:13:05) bottlenecks and things you need to (00:13:07) figure out? I know energy is something (00:13:08) that's been bought up so much saying, (00:13:10) well, look, we can keep advancing chips. (00:13:11) we can keep advancing models, but at (00:13:13) some point (00:13:14) >> we're just not going to have enough (00:13:15) energy to run these data centers, to run (00:13:17) these AI models. (00:13:19) >> Um, (00:13:19) >> well, look, look, there's there's lots (00:13:21) of physical constraints. So, um, of (00:13:23) course, there's, you know, no one ever (00:13:24) has enough chips and, you know, we're (00:13:26) lucky that we have, you know, our own (00:13:27) TPU range in addition to GPUs and um, (00:13:30) but there just aren't enough uh, compute (00:13:32) chips in the world really for the (00:13:33) demand. Uh, and of course, in the end, (00:13:35) that comes down to energy as well. (00:13:36) there's this idea of energy will be (00:13:38) effectively is synonymous with (00:13:40) intelligence as we get into the era (00:13:41) towards AGI. Um now the interesting (00:13:44) thing is I think that AI itself will (00:13:47) help here in the sense of getting more (00:13:49) efficiencies out of existing (00:13:50) infrastructure but helping with things (00:13:52) like material design better better solar (00:13:54) materials but it could also help with (00:13:56) new breakthrough technologies like (00:13:58) fusion. you know, we have a (00:13:59) collaboration with Commonwealth Fusion (00:14:01) uh in the US to help contain plasma and (00:14:03) fusion reactors and um one of my pet (00:14:06) projects is can we come up with a room (00:14:08) temperature superconductor uh material (00:14:10) using AI. So I think there are multiple (00:14:12) breakthroughs that AI could come up with (00:14:14) and help uh us come up with that would (00:14:17) help with the energy uh situation. In (00:14:19) fact, indeed that's I think that's one (00:14:20) of the most promising use cases of AI. (00:14:22) Um and then the other thing is as these (00:14:24) systems are getting better, they're also (00:14:26) getting, you know, 10x more efficient (00:14:28) per year. So if you look at our range of (00:14:31) models, we have our kind of lighthouse (00:14:32) model, our pro versions of Gemini, but (00:14:34) then we have our flash versions which (00:14:36) are way more efficient and the sort of (00:14:38) workhorse models that are used for (00:14:39) everything. And um they use techniques (00:14:41) like distillation where you have a big (00:14:43) model that teaches a smaller model and (00:14:45) the smaller model is really really (00:14:46) efficient. And I think there are more (00:14:48) and more innovations and techniques like (00:14:49) that that will keep bringing the (00:14:51) efficiency curve uh down and so you get (00:14:54) you know much better performance per per (00:14:56) watt. We hear a lot about sort of AGI (00:14:59) and I think there's a lot of people (00:15:00) wondering technology sounds amazing (00:15:02) sounds great but there's al also a lot (00:15:03) of fear right around uh the (00:15:05) proliferation of this technology and the (00:15:06) impact it's going to have on on people (00:15:08) every day and their lives. Um I guess (00:15:10) for you what what are some of the the (00:15:11) things we need to consider? Yeah. From (00:15:13) from that perspective in terms of the (00:15:14) impact on society, whether it's around (00:15:16) jobs, whether it's around kind of what (00:15:18) we're going to do with our time if if we (00:15:19) reach this goal versus I guess the (00:15:22) benefits that you believe this (00:15:24) technology is going to bring for (00:15:25) humanity. (00:15:26) >> Well, of course, you know, I I believe (00:15:27) that overall AI is going to be one of (00:15:29) the most beneficial technologies and (00:15:30) humanity's ever uh uh invented. Uh (00:15:33) that's why I spent my whole career (00:15:35) working on it. But it's only, you know, (00:15:36) it's not a given. It's a dual-purpose (00:15:38) technology. Um, I dream about using AI (00:15:41) for things like curing diseases. We have (00:15:43) a spin out called isomorphic that builds (00:15:45) on on alphafold work on protein folding (00:15:48) work that we did a few years ago to (00:15:50) accelerate drug discovery and try and (00:15:52) solve all disease. I think that's now (00:15:54) you know within reach that type of thing (00:15:56) in the next decade or two. Um, we've (00:15:58) discussed energy. There's many benefits (00:16:01) I think AI is going incredible benefits (00:16:02) AI is going to bring. Um, but there are (00:16:04) also risks. Obviously there's kind of (00:16:06) economic disruption. Um and I think (00:16:09) there it's going to be like the (00:16:10) industrial revolution but maybe 10 times (00:16:13) bigger 10 times faster. So you know it's (00:16:16) incredible amount of transformation but (00:16:18) also disruption that's going to happen. (00:16:20) And you know we need some uh new (00:16:22) economic models probably for that. Um (00:16:24) and then on terms of the the worries (00:16:26) about the usage of AI I have two which I (00:16:29) think are are worth worrying about. one (00:16:30) is bad actors repurposing these general (00:16:33) purpose technologies AI technologies for (00:16:35) harmful ends. Um and then the second one (00:16:37) is AI itself as it get we get towards (00:16:39) AGI and agentbased systems. So these are (00:16:42) systems that are able to do things more (00:16:44) autonomously than than today's systems. (00:16:46) Um they can you know what are the guard (00:16:49) rails around that? How do we make sure (00:16:51) we can keep them uh uh doing the things (00:16:53) that we want them to do and not veer off (00:16:55) into uh something that we didn't expect. (00:16:59) And so those are the two kind of risks (00:17:00) that are kind of uh that I foresee. (00:17:03) >> Do you feel that you're developing (00:17:05) systems that you can be in control of? (00:17:07) >> I think we're we're very confident about (00:17:09) that. You know, we we've had uh and (00:17:11) thought about responsibility and safety (00:17:13) and security of these systems from the (00:17:15) very beginning. Um you know, we started (00:17:17) Demi back in 2010. Almost no one was (00:17:19) working on AI back then, but we planned (00:17:21) for success and we knew success would (00:17:23) mean these extremely powerful systems. (00:17:25) So we also understood the the the the (00:17:28) other side of the coin of that. So from (00:17:30) the very beginning we've tried to be (00:17:31) very thoughtful use the scientific (00:17:33) method and scientific approach to try (00:17:34) and understand as much about our systems (00:17:36) we're building before we deploy them. Um (00:17:38) of course that doesn't mean we won't (00:17:39) make any mistakes. There's too it's too (00:17:41) it's it's such a incredible and (00:17:43) fastmoving technology. But I think with (00:17:45) with something like AI we need to be you (00:17:48) know I call myself a kind of cautious (00:17:50) optimist. I'm I'm I'm very uh big (00:17:53) believer in human ingenuity. I think (00:17:54) given enough time and care, we'll get (00:17:56) this right as scientists and as a (00:17:58) society, but it's it's not a given. And (00:18:01) so, um, we shouldn't be sort of rushing (00:18:03) into this. Um, and and we need to go (00:18:05) into it with our eyes open because I I (00:18:07) guess the reason I ask it because I know (00:18:08) you've spoken to people like Joshua (00:18:09) Benjio and Max Tegmark and and these are (00:18:12) people I've also spoken to and and (00:18:13) they're of this cohort that believes do (00:18:15) do we need to be rushing so quickly into (00:18:17) a world of AGI and agentic systems? (00:18:20) Maybe we need more uh toolbased uh AI, (00:18:23) AI to solve specific things rather than (00:18:25) these allpurpose or general purpose kind (00:18:27) of systems and I know they've called for (00:18:30) for perhaps a slowdown to the (00:18:32) development of of these AGI systems. (00:18:36) >> In your view, do you think you should be (00:18:38) slowing down? Well, I've I've had lots (00:18:40) of you know, I know them very well. (00:18:41) Yoshua and Max we've had many (00:18:42) discussions and many others and and (00:18:44) actually I have some sympathy for that (00:18:45) view that you know building a tool based (00:18:48) AI is you know thinking of AI as a tool (00:18:51) or the ultimate tool for say science is (00:18:53) the right way to build AI in the initial (00:18:55) stages um and uh certainly that's the (00:18:58) way we're viewing it and the kinds of (00:19:00) things we apply AI to uh like AlphaFold (00:19:03) but um the thing is you know it's a very (00:19:06) complex geopolitical and corporate uh (00:19:09) system that we're in And it isn't just (00:19:11) about you know there are many companies (00:19:13) trying to build this there also many (00:19:15) nations trying to build it and um it's (00:19:18) there's a sort of race dynamic which I (00:19:20) ideally wouldn't be there. So in an (00:19:22) ideal case this would be a scientific (00:19:23) endeavor and it would be very carefully (00:19:26) uh each step would be carefully (00:19:27) considered but unfortunately the the the (00:19:29) pra the real world isn't isn't like that (00:19:32) and we have to kind of be pragmatic (00:19:34) about uh where we are. So what we're (00:19:36) trying to do is be good role models for (00:19:38) um yes being on the frontier pushing (00:19:40) that uh the benefits of that as quickly (00:19:42) as we can and as broadly as we can um (00:19:45) but also try and be as responsible as (00:19:48) possible with that along the way and (00:19:49) thoughtful as possible and I think we've (00:19:51) got that balance pretty pretty good (00:19:53) right now and hopefully that's a bit of (00:19:54) a role model uh to the rest of the field (00:19:56) in the industry too. Yeah, I want to (00:19:57) address some of those dynamics as well, (00:19:58) but just just first I guess just from a (00:20:00) personal point of view, have you ever (00:20:01) you said you sort of started this (00:20:03) mission of deep mind, you know, you (00:20:04) believe in the technology, but has there (00:20:06) ever been any moments in your career (00:20:07) when you g like should we be doing this? (00:20:09) Um, look, you when you look at how (00:20:12) powerful the technology is. Um, I really (00:20:15) think there that there are so many (00:20:18) challenges confronting society today, (00:20:20) not to do with AI, climate, poverty, you (00:20:23) know, the access to water. There's a (00:20:25) there's just so many uh issues um health (00:20:28) uh aging, population, uh disease. So (00:20:31) like uh uh um you know energy we talked (00:20:34) about earlier. So if a some if I if (00:20:37) there wasn't a technology transformative (00:20:39) as AI coming down the road, I'd be (00:20:41) really worried about uh society's (00:20:43) ability to deal with these challenges. (00:20:45) So, interestingly, AI itself is one of (00:20:48) those challenges, maybe one of the (00:20:49) greatest ones, but it's also one which (00:20:51) can help us um cope with and resolve and (00:20:55) solve some of these other big grand (00:20:57) challenges. So, it's a very interesting (00:20:58) one, right? It's it's it's sort of (00:21:00) double-edged and I've always believed in (00:21:03) that. I've always thought that um uh in (00:21:06) the end it would be the the the the most (00:21:07) important technology uh we'll ever (00:21:09) invent. And um I think it's sort of the (00:21:12) natural progression really of the (00:21:14) computer age. (00:21:15) >> Dennis, you just just a quick aside, you (00:21:18) started uh life in gaming, which is (00:21:20) amazing. Co-developing theme park. (00:21:22) Fantastic. (00:21:23) >> Fantastic game as well. Um did you ever (00:21:25) do you still play games? (00:21:26) >> Yes, I love games. It's my main and only (00:21:28) hobby really. Well, like these days like (00:21:31) League of Legends with my two two boys (00:21:33) and my brother and we have a little (00:21:34) team. We've done it since lockdown. Um (00:21:37) but yeah, I love games in all its forms (00:21:39) from from football to (00:21:40) >> such a high impact stressful role as you (00:21:42) have potentially. Is that your unwind? (00:21:45) >> It is. It is. I would say so. And it's (00:21:47) also, you know, it's a it's a kind of in (00:21:50) the past as well as being a great (00:21:51) creative endeavor for me, you know, and (00:21:53) it's how I learned programming and other (00:21:55) things was was through making games. (00:21:57) >> I have nowhere near as a stressful a job (00:21:59) as you, but that's my unwind, too. (00:22:00) >> Yes, for sure. (00:22:01) >> Get home, turn the console on. (00:22:03) >> Exactly. Exactly. (00:22:07) Just in that small segment alone, Steve, (00:22:09) there's so much to unpack and I want to (00:22:11) focus on on two kind of big buzzwords (00:22:14) right now. The first is artificial (00:22:16) general intelligence or AGI. This idea, (00:22:19) and I know there's so many different (00:22:20) definitions of it, but broadly speaking, (00:22:22) this idea of AI that that is as smart or (00:22:25) smarter than humans. And I think that so (00:22:27) many of these big AI labs including Open (00:22:29) AI, including Deep Mind, are pushing and (00:22:32) hoping to get to this stage of AGI. A (00:22:36) and so far they've approached this with (00:22:38) a a technique called large language (00:22:40) models. These AI models that are trained (00:22:42) on huge amounts of data, but mainly (00:22:44) text. But there's this other buzzword, (00:22:46) right? World models. This idea of these (00:22:49) AI models that understand the physical (00:22:52) world. And this is this this buzzword is (00:22:55) really growing in popularity, right? (00:22:56) >> Yeah. And I think this is going to be a (00:22:58) big theme of AI going into uh the rest (00:23:01) of 2026 and even into next year because (00:23:04) the idea here is that LLMs, sure, we got (00:23:06) the language part down. It can mimic the (00:23:09) way humans talk and and speak and and (00:23:11) write and things like that. Uh but when (00:23:13) it comes to the physical world, you (00:23:14) know, we talk so much about robotics and (00:23:16) AI and physical AI. Well, they need to (00:23:19) understand how the physical world works, (00:23:21) how water flows, how air moves, and (00:23:23) things of that nature. And what really (00:23:25) struck out to me when you brought this (00:23:26) up to Demis, he he said, "Yeah, we do (00:23:29) need to start exploring that more." And (00:23:31) in fact, he sees a world in which the (00:23:33) LLM and those world models start to (00:23:36) converge. I think that was the word he (00:23:37) used, converge uh into something uh more (00:23:40) unique and and more powerful and (00:23:42) capable. This is also a debate that's (00:23:44) been playing out among AI leaders like (00:23:46) on social media. You could fire up X or (00:23:49) your favorite social media site and uh (00:23:51) what really struck out to me is Yan Lun. (00:23:53) He was the head of AI for many years (00:23:56) over at Meta. He recently left uh to (00:23:58) start his own thing because he kind of (00:24:00) got superseded by Alexander Wang and (00:24:02) that whole big talent wars that happened (00:24:04) over last summer. um he had a really (00:24:07) interesting interview in the Financial (00:24:09) Times. He doesn't think LLMs are what's (00:24:12) going to get us to AGI. To your point, (00:24:14) that's what everyone's chasing, the (00:24:15) super intelligence, AGI, whatever you (00:24:18) want to call it. His thing is LLMs can (00:24:20) only get you part of the way. You need (00:24:22) world models and all sorts of other (00:24:24) things. And he kind of uh harshly (00:24:26) criticized Meta for not thinking beyond (00:24:28) the LLM. Um and that seems to be part of (00:24:31) the reason why he left to do his own (00:24:32) thing. And it's really interesting to (00:24:34) see one of Meta's big competitors, (00:24:36) Gemini, just talk openly about it and (00:24:38) say them is saying, "Yeah, we we need to (00:24:40) do this. We need to start thinking about (00:24:42) this." Uh it enables so many things from (00:24:44) robotics, autonomous driving um and just (00:24:47) a better understanding for these AI (00:24:49) models and um intelligent systems that (00:24:51) we're chatting with to get you that (00:24:53) right answer. (00:24:54) >> Steve, do you ever use a a chatbot um (00:24:57) and you put something in and it will (00:24:58) say, "Hey, Steve, great question. That's (00:25:00) a really clever thought." (00:25:01) >> All the time. That's the sickop fancy of (00:25:03) all these chatbots, right? Where they're (00:25:04) like, "Oh, you're so smart and great at (00:25:06) asking me these questions." Yeah. All (00:25:08) the time. (00:25:08) >> Exactly. Because the reason I bring that (00:25:10) up is is partly to this point, this (00:25:11) growing criticism of LLMs is that (00:25:13) actually, yes, they're great and they'll (00:25:15) give you the information and but (00:25:17) actually when it comes to LLM as a (00:25:20) foundation for being able to create new (00:25:22) ideas, novel ideas, there's limitations (00:25:24) there. And I think that's partly what uh (00:25:26) Demis was speaking to and why this idea (00:25:28) of world models is really growing in (00:25:30) popularity. Um it's going to be (00:25:32) interesting to see how this plays out as (00:25:34) you mentioned into this next phase of AI (00:25:36) where uh it's key for things like (00:25:38) robotics, driverless cars, and many (00:25:39) other use cases, too. (00:25:40) >> Yeah. And I'm you'll you'll notice as we (00:25:42) continue this podcast, I'm incredibly (00:25:44) cynical about the robotics angle of this (00:25:47) AI moment we're living in. All that so (00:25:49) many of the robots we're seeing, they're (00:25:51) literally puppets. They're teleoperated. (00:25:53) The best example, of course, is the (00:25:55) Tesla Optimus robot, uh which started (00:25:57) out as a man in a bodysuit dancing (00:26:00) around. Now it's a real robot, but again (00:26:02) it's tea operated. There are literally (00:26:04) people in a control room controlling it (00:26:07) over the internet and even using their (00:26:09) voice to talk to you and things like (00:26:10) that. So we are the robotics people I (00:26:13) talked to, we had one in the office just (00:26:14) a couple weeks ago and they said the (00:26:16) hardest part isn't building the actual (00:26:18) robot, it's training it and that's where (00:26:21) these world models are going to come in (00:26:22) so they can actually operate (00:26:23) autonomously like we've been promised. (00:26:29) Deus, you mentioned some of the dynamics (00:26:30) at play, right? And competition (00:26:33) >> commercially, of course, is one of (00:26:35) those. We've got Open AI, we've got (00:26:36) Anthropic, we've got all these different (00:26:38) AI labs um out there. It's intense. Uh (00:26:41) and Gemini 3 has had such good reception (00:26:43) uh so far. Um but there was a point (00:26:46) people were doubting (00:26:47) >> Google as a whole and its ability to (00:26:49) compete and I say a point it was at some (00:26:51) point in 2025 and it wasn't that long (00:26:53) ago and then you know, Gemini 3 really (00:26:55) came out and and impressed a lot of (00:26:57) people as well. Um but it's a space (00:26:59) that's ever changing. Uh so how are how (00:27:02) would you assess right now the (00:27:03) competitive environment? How do you feel (00:27:05) it? (00:27:05) >> Yeah. Well look it's a ferocious uh uh (00:27:07) competitive environment at the moment. I (00:27:09) mean many people who are telling me you (00:27:11) know been in tech for 20 30 years say (00:27:13) it's the it's the most intense (00:27:14) environment they've ever seen perhaps (00:27:16) you know ever in the technology (00:27:17) industry. and uh and and and you know (00:27:20) all the I guess most capable players (00:27:22) whether it's individual you know tech (00:27:24) titans or big tech companies or and or (00:27:27) the best startups they're all involved (00:27:29) in this space now cuz I think everyone (00:27:30) has understood what we've known for 20 (00:27:32) plus years now that this is really the (00:27:34) most important technology um so that's (00:27:37) sort of to be expected but it's tough (00:27:39) but it's also exciting and um you know (00:27:42) going back to games I uh I sort of I've (00:27:45) started playing chess when I was very (00:27:47) young for the England and junior chess (00:27:48) team. So I've kind of been brought up in (00:27:50) in competition. So you know I love (00:27:52) competition fortunately. In fact many (00:27:53) ways I live for competition. So a lot of (00:27:56) a big part of me sort of like likes to (00:27:58) lean into this. But on the other hand (00:27:59) the only thing I would say is at the (00:28:01) back of my mind I know there's something (00:28:02) much more important than individual (00:28:04) competition between companies or even (00:28:06) countries which is overall getting (00:28:08) stewarding AGI well for the world for (00:28:10) the whole you know for all of humanity. (00:28:12) And I think that's incumbent of all of (00:28:14) us who are leaders of the AI labs. um (00:28:18) and uh and and can have an influence (00:28:19) over this is to have that at the sort of (00:28:21) in the front of their minds in amongst (00:28:23) this sort of ferocious capitalist (00:28:25) competition that we're in as well. So (00:28:27) both are true at the same time. (00:28:29) >> I mentioned kind of the moment people (00:28:30) were questioning what Google was going (00:28:32) to do with with AI earlier in in the (00:28:34) year. (00:28:35) >> Did you do anything different? Yeah, I (00:28:37) think look I I feel like you know if we (00:28:40) go back over the last decade actually (00:28:42) you know Google Google brain (00:28:44) specifically uh the research division in (00:28:46) Google and deep mine as it was uh sort (00:28:48) of fairly independent we kind of (00:28:50) invented about 90% of the technologies (00:28:53) uh that everybody's using today you know (00:28:55) whether it's transformers of course most (00:28:56) famously the architecture behind all the (00:28:58) LLMs or AlphaGo you know sort of (00:29:00) introduced reinforcement learning at (00:29:02) scale uh on a really hard problem so (00:29:04) we've invented all this technology but (00:29:06) then um maybe we were in hindsight we (00:29:09) were a little bit slow to commercialize (00:29:10) it and scale it and um you know that's (00:29:13) what open and others did very well and (00:29:15) then the last 2 three years I think (00:29:17) we've had to come back to almost our (00:29:19) startup or entrepreneurial roots and um (00:29:23) be scrappier be faster ship things (00:29:25) really quickly and um and and sort of (00:29:28) make really rapid progress and I think (00:29:30) what you're seeing over the last couple (00:29:32) years culminating in Gemini the Gemini (00:29:34) series which we're very happy with (00:29:35) Gemini three is as as you mentioned our (00:29:38) latest version um has sort of put us (00:29:41) back at you know near the top of you (00:29:43) know the top of the leaderboards where (00:29:44) we feel we belong and you feel like you (00:29:46) can stay there (00:29:47) >> I I I feel like we can stay there of (00:29:49) course yeah (00:29:50) >> amid all this competition there's (00:29:51) obviously a lot of talk about (00:29:54) >> bubbles in AI uh particularly around (00:29:57) valuations of certain companies (00:29:59) companies raing raising astronomical (00:30:01) sums of money the tech giant spending (00:30:03) hundreds of billions on infrastructure (00:30:06) uh and companies out there quite frankly (00:30:08) raising large sums of money with very (00:30:10) little product or or or even very little (00:30:12) profitability if any. And so where do (00:30:14) you think we are right now in terms of (00:30:17) this this kind of bubble discussion? Do (00:30:18) you think we're in a financial bubble (00:30:20) when it comes to AI industry? (00:30:21) >> I think it's not a binary thing this (00:30:23) bubble discussion. I don't I think um (00:30:25) some parts of the industry might be in a (00:30:27) bubble to me that's what it looks like (00:30:29) and and others probably not. you know, (00:30:31) fundamentally AI is going to be the most (00:30:33) transformative technology ever invented. (00:30:35) So that's there's that part that (00:30:37) underpins everything. So in the end, (00:30:39) it's a bit like the internet bubble in (00:30:41) the end. The internet was critical and (00:30:43) there were some generational companies (00:30:44) that were created in during that time, (00:30:47) right? Um so I think you know that's (00:30:49) sort of almost inevitable. There'll be (00:30:51) overexuberance once everyone realizes (00:30:53) how transformative a specific technology (00:30:55) is. uh and then there'll be probably a (00:30:58) reckoning and then the the things that (00:31:00) are real will survive and and flourish. (00:31:02) Um where it seems to me is you know (00:31:05) maybe like in the private markets where (00:31:07) there sort of seed rounds at tens of (00:31:09) billions of dollars where basically (00:31:11) there's just almost nothing there yet (00:31:12) and that seems a little bit (00:31:14) unsustainable over the long run. As far (00:31:16) as I'm concerned I don't really worry (00:31:17) about bubbles from my my point of view (00:31:19) is sort of leading Google deep mind. (00:31:21) I've got to make sure that what (00:31:22) whichever way it goes, whether um it (00:31:25) continues to go all rosy and exponential (00:31:27) like it is now or there's a bubble, you (00:31:29) know, there's some kind of bubble (00:31:30) bursting, that we're in the right (00:31:32) position to to to win either way and to (00:31:35) take advantage of that either way. And I (00:31:36) think we've got a good position given (00:31:38) Google's underlying business and how AI (00:31:40) fits with that. Um uh to to to benefit (00:31:43) uh whichever way it goes from here. some (00:31:45) I guess some of your biggest competitors (00:31:46) are the ones who have managed to raise (00:31:48) huge sums of money in the private (00:31:50) markets at this point. So do you feel (00:31:51) confident that even if there is some (00:31:53) sort of correction at some point that (00:31:55) you know you'll be able to weather it (00:31:56) out I guess? Yeah, I mean look, you (00:31:58) know, that's the whole point of uh (00:31:59) Google's balance sheet and and also all (00:32:01) the incredible products that um and (00:32:04) surfaces that that we have. You know, I (00:32:06) think it's you know, dozens of (00:32:08) multi-billion user products and and AI (00:32:11) kind of naturally fits into uh all of (00:32:13) those products, whether it's um you (00:32:15) know, email workspace or or you know, (00:32:18) new things like the Gemini app. (00:32:19) >> Yeah, you mentioned dynamics at play as (00:32:21) well. We talked competition. And the (00:32:23) other one is geopolitics which you (00:32:24) mentioned as well when huge discussions (00:32:26) around China of course in this kind of (00:32:28) competition battle between China and the (00:32:30) US. But you know there was a point where (00:32:32) people were discounting the ability of (00:32:34) China and it and its companies to come (00:32:36) up with strong AI um models and and and (00:32:39) technologies. But actually we saw with (00:32:41) kind of what Deep Seek did um it kind of (00:32:44) brought a bit of shock to world but (00:32:45) actually more than that some of the big (00:32:47) tech companies like Alibaba coming up (00:32:49) with some very competitive open-source (00:32:51) models. So China's not out this game, (00:32:53) right? (00:32:53) >> Not at all. And actually, you know, I (00:32:55) think they are closer to the US front, (00:32:58) you know, US and West frontier models (00:32:59) than maybe we thought one or two years (00:33:01) ago. Um maybe they're only a matter of (00:33:04) months behind at this point. Um the (00:33:06) interesting thing is and they're very (00:33:07) there's some very capable teams of (00:33:09) course like the Deep Seeking and Alibaba (00:33:11) you mentioned. Um and uh the question is (00:33:14) is can they innovate um something new (00:33:17) beyond the frontier? So, I think they've (00:33:19) shown they can catch up, you know, and (00:33:21) and be very close to the frontier and (00:33:23) catch up very quickly. Uh, but can they (00:33:26) actually innovate something new like a (00:33:28) new transformers uh, you know, that gets (00:33:30) beyond the frontier? I don't think (00:33:31) that's been shown yet. (00:33:32) >> Is that going to be in your view (00:33:35) difficult because of restrictions on (00:33:36) access to technology like leading edge (00:33:38) chips for example? (00:33:39) >> No, I think it's more a mentality issue (00:33:42) you know. So I think it's something that (00:33:44) at least the leading labs the leading (00:33:45) frontier labs in the west have uh (00:33:48) nurtured I can say for ourselves you (00:33:49) know we you can think of deep mind as a (00:33:51) bit like a try to be a modernday bell (00:33:53) labs and encourage uh innovation and (00:33:57) exploratory innovation not just scaling (00:33:59) out what's what's known and and uh today (00:34:02) and of course that's already very (00:34:03) difficult because you need world-class (00:34:04) engineering already to be able to do (00:34:06) that um and and China definitely have (00:34:08) that the question is uh is the (00:34:11) scientific innovation part that's a lot (00:34:13) harder to you know to invent something (00:34:15) is about 100 times harder than it is to (00:34:18) to copy ed (00:34:21) is and I haven't seen evidence of that (00:34:23) yet but it's very difficult (00:34:28) so one of the most striking parts of (00:34:29) that part of the conversation for me (00:34:31) Steve was uh around China um I used to (00:34:34) live in China for just over 3 years (00:34:36) report out of China uh for CNBC covering (00:34:39) the tech sector there and there was this (00:34:41) growing view recently that actually (00:34:43) China is so far behind the US when it (00:34:46) comes to AI uh for for multiple reasons. (00:34:49) One of those is that oh it may not be (00:34:51) able to get its hands on the most (00:34:52) advanced chips so its industry could (00:34:53) fall behind. One view is that it's just (00:34:56) not innovating and it doesn't have the (00:34:57) capital the way US companies do. But (00:34:59) actually what was really interesting (00:35:01) from Demis is he said that he believes (00:35:03) Chinese AI models are are just months (00:35:05) behind uh where the U US is. So actually (00:35:08) not far behind and remember when uh last (00:35:10) year we had uh Deepseek really shock the (00:35:14) world and markets. Um it showed I think (00:35:17) China is in the game and since then (00:35:19) whilst Deepseek hasn't quite made the (00:35:21) waves it did uh when it first kind of (00:35:23) came out um Alibaba one of the world's (00:35:27) biggest or one of China's biggest tech (00:35:29) companies um has been a leader there. (00:35:31) gets developed some really interesting (00:35:33) models which if you look at the open- (00:35:35) source community such as uh on a site (00:35:37) called hugging face you see Alibaba's (00:35:39) models are amongst some of the most (00:35:41) popular experts who I've spoken to in (00:35:43) the space say they're amongst some of (00:35:45) the most advanced in the world so you (00:35:47) are seeing there and one of the things I (00:35:49) can tell you just from living and (00:35:50) working out there is Chinese companies (00:35:53) move fast they have the expertise and (00:35:55) they can innovate so you can't discount (00:35:58) them out of this kind of AI race but (00:36:00) also take Demis' point that he said (00:36:02) whilst the Chinese companies are sort of (00:36:04) catching up and and and are very much in (00:36:06) this race, one thing they haven't proven (00:36:08) is their abilities to kind of make these (00:36:10) big breakthroughs. So, you know, I (00:36:12) thought that was a really interesting (00:36:13) and nuance for you. I guess the other (00:36:15) part here, Steve, is something you (00:36:16) picked up on is Dis's comments on (00:36:18) bubbles and AI bubbles. (00:36:20) >> Yeah. And that and by the way, just (00:36:22) talking, let's go back to what he said (00:36:23) first about the months thing. uh (00:36:26) Deepseek a year ago. It wasn't just (00:36:27) about the fact that China can do it and (00:36:29) make a really good large language model (00:36:31) or a chatbot. It was also the idea that (00:36:33) they did it without the most powerful (00:36:35) Nvidia chips that kind of rattled the (00:36:37) markets as well. And that's what we're (00:36:38) seeing here in the United States now, (00:36:40) Arjuna, is trying to limit China's (00:36:43) ability to get those NVIDIA chips. (00:36:45) There's all this talk about maybe (00:36:46) they'll get those H200 chips, which (00:36:48) aren't the best chips, but they're (00:36:50) better probably than what China has (00:36:51) access to. And then you get into the (00:36:53) whole smuggling thing. But to Dennis's (00:36:55) point, you know, if they really are (00:36:57) months behind without full access to (00:36:59) these chips, you know, that kind of (00:37:02) questions Nvidia's prominence and (00:37:04) dominance uh in the chip space as well. (00:37:06) But yes, what what you said about the (00:37:08) bubble is also super uh interesting too (00:37:11) because you asked him about that. Are we (00:37:12) in a bubble? What do you think? All this (00:37:14) sort of things. And he basically said (00:37:16) we're Google, we're rich. It doesn't (00:37:18) matter. We have the money. We have the (00:37:20) free cash flow to spend this. Our (00:37:21) balance sheet is our superpower. If for (00:37:23) some reason we need to rein back the (00:37:25) spending, we can do it and we'll be (00:37:27) fine. But guess who can't do that? (00:37:29) That's OpenAI and Enthropic. The other (00:37:31) two leaders, XAI, we can throw them in (00:37:34) here, too. Their whole thing is they (00:37:36) have to raise money indefinitely in (00:37:39) order to get to the point where they can (00:37:40) finally show some revenue and and (00:37:42) revenue growth to uh sustain themselves (00:37:45) without continuous fundraising. If (00:37:47) things start to dry up, OpenAI and (00:37:50) Enthropic are at extreme risk. Google, (00:37:53) Microsoft, Meta, they have the cash flow (00:37:55) to move on to another project. Meta's (00:37:57) already done it with the metaverse. (00:37:59) These companies can pivot very easily (00:38:01) because they have these big high margin (00:38:03) businesses already. (00:38:08) De um a lot of people I guess forget how (00:38:11) much of Google's AI capabilities come (00:38:13) out from DeepMind and and yourself and (00:38:15) your teams. Um how do you work with (00:38:17) Google? There's a lot of fascination (00:38:18) around that. (00:38:20) call you up one day and say, "Hey, Deis, (00:38:22) we need this thing or we have this idea (00:38:24) for Gemini or for some other AI product. (00:38:27) Um, can you build it?" How was that (00:38:29) relationship? (00:38:29) >> Yeah. So, the last three years we've (00:38:30) combined everything together as into (00:38:32) Google Deep Mind. This this one entity (00:38:34) that that that all the AI research at (00:38:36) Google goes on in and it's a kind of (00:38:38) combination of uh Google research, (00:38:40) Google Brain and and and DeepMind. And I (00:38:42) run that group and it's it's like the (00:38:44) engine room of Google. You should think (00:38:45) of it like that. So uh all the AI (00:38:48) technologies is is done by this group by (00:38:50) our group and then it's diffused across (00:38:52) you know all of these incredible (00:38:54) products uh right across Google and the (00:38:57) last couple of years we've been building (00:38:58) that backbone so not just the models but (00:39:01) also almost rearchitecting the entire (00:39:03) infrastructure of Google so that it can (00:39:05) you know these things can ship (00:39:06) incredibly quickly these models it's (00:39:08) almost sim shipped to all the main (00:39:10) surfaces so you know when we release a (00:39:12) new Gemini model it's there the next day (00:39:14) or the same day in in search and uh and (00:39:17) that's been going really well. I think I (00:39:18) would say we've really got into our (00:39:19) groove uh with the 2.5 Gemini models and (00:39:23) and for the last sort of year uh that's (00:39:25) been coming really uh a smooth process (00:39:28) now and I think you'll see that more uh (00:39:30) over the next next 12 months. Um and so (00:39:33) you know we think of ourselves as the (00:39:34) and describe ourselves sort of as the (00:39:36) engine room for that and you know Sundar (00:39:38) and I pretty much talk every day about (00:39:40) strategic things and where should the (00:39:42) technology go and what does uh uh the (00:39:44) wider Google need um and then you know (00:39:46) we adjust the road maps and the plans uh (00:39:49) you know on a daily basis whilst keeping (00:39:51) in mind the long-term goals of uh you (00:39:54) know getting to AGI first fast and (00:39:55) safely. So we should we should expect (00:39:57) more of the ability to come up with with (00:40:00) new things, new AI tools and that be (00:40:02) shipped across the Google portfolio etc. (00:40:04) because of that kind of change you've (00:40:06) made in that relationship. (00:40:06) >> That's right. So it's an incredibly uh (00:40:08) tight sort of uh iteration loop and and (00:40:11) and and you know we're all on the same (00:40:13) tech stack and so on. A lot of what (00:40:15) you're building is going into Google (00:40:16) products, but I know kind of covering (00:40:18) companies like Samsung, you help (00:40:20) companies like Samsung to build out some (00:40:21) of the AI tools um within, you know, (00:40:23) their smartphones for example and that (00:40:25) kind of thing as well. Well, look, we (00:40:27) work with a lot of partners as you as (00:40:29) you mentioned um you know, we're very (00:40:31) proud of the fact that our technology (00:40:32) selected by those partners um because (00:40:35) they see how capable it is and and (00:40:37) actually you know, it comes to Samsung (00:40:38) and other devices. Um, I think there's (00:40:41) really interesting way I'm very (00:40:42) interested in the idea of uh uh edge (00:40:45) compute and and faster versions of these (00:40:47) models working on these edge devices be (00:40:49) those phones, but also new devices like (00:40:51) glasses that we're working on um and you (00:40:54) know partners like WBY Parker and the (00:40:56) idea of smart glasses and I think um (00:40:58) Google's worked on smart glasses for a (00:41:00) long time as you know but I think the (00:41:02) day you know finally we have the killer (00:41:05) app I would say for it which is this (00:41:06) idea of a universal assistant and um and (00:41:10) uh and and sort of helping you in your (00:41:12) everyday life. And I think all the all (00:41:14) the all the big uh device players are (00:41:16) going to be interested in that type of (00:41:17) technology. (00:41:18) >> Demis, we've only got a few minutes (00:41:19) left, but I do want to ask a little bit (00:41:21) about I was a brand new tech reporter (00:41:23) when Google bought DeepMind 2014. I (00:41:26) think it was a 400 million pound deal (00:41:27) back then. Um so many people didn't know (00:41:30) what you what you did. Um and why is (00:41:32) Google buying this British company? (00:41:34) What's going what's going on here? Um, (00:41:36) do you ever look back to that and and (00:41:38) think, "Oh, maybe we should have stayed (00:41:39) independent at all, or are you happy (00:41:41) with how things have turned out?" (00:41:42) >> Well, look, I we I knew it's funny. So, (00:41:45) so the the head of search at the time, (00:41:47) Alan Eustace, he he was sort of in (00:41:48) charge with Larry Larry was sponsoring (00:41:50) the Larry Page was sponsoring the the (00:41:52) deal uh as he was CEO at the time, but (00:41:54) Alan Eustace was delegated, the head of (00:41:56) search to kind of close the deal. And I (00:41:58) did tell Alan that this would be the (00:41:59) most important acquisition Google ever (00:42:01) made, which is which is quite something (00:42:03) given they've, you know, there's YouTube (00:42:04) and and and (00:42:06) uh Adwords and other things that they (00:42:08) they previously acquired. But I kind of (00:42:10) knew how important this was going to be. (00:42:12) Uh and also how good a fit it was with (00:42:14) Google's um uh mission, which is (00:42:16) organize the world's information. And AI (00:42:18) is a very natural fit to that uh and (00:42:21) organizing and understanding (00:42:22) information. I mean, what better tool (00:42:24) than AI for that? So, I kind of knew (00:42:26) that would be a natural fit. And we sort (00:42:27) of knew that this, you know, maybe it's (00:42:29) now worth, I don't know, 100x thousandx (00:42:32) of, you know, what of what we sold it (00:42:34) for. But the thing is, I wanted to get (00:42:36) back to the science at the time and and (00:42:38) and push forward the research, which was (00:42:40) still very nent back in 2014. And and (00:42:43) you know, fair play to Google is they (00:42:44) were one of the few companies in the (00:42:45) world, I think, that could recognize uh (00:42:47) and specifically Larry at the time, how (00:42:49) important this technology was going to (00:42:50) be, what it could become, and what we (00:42:52) see it for it today. And I don't think (00:42:54) we could have done the the great work we (00:42:56) did with Alph Go and Alpha Fold and all (00:42:58) the science we've done um and uh if we (00:43:00) hadn't had their backing and the amount (00:43:02) of compute that they could bring uh uh (00:43:04) to to play. So I don't have any regrets (00:43:06) at all. (00:43:07) >> So tech CEOs, AI CEOs, new rock stars of (00:43:10) the world. I've seen Jensen Hang here in (00:43:12) Europe and the CEO of Nvidia, you know, (00:43:14) being followed around by everyone as (00:43:16) well. Um Jensen I think said recently (00:43:19) that that you and him talk he had great (00:43:21) things to say about Nano Banana the new (00:43:23) image generation 2 as well. What what (00:43:25) what do you guys discuss? (00:43:26) >> Oh we disc I mean Jensen's great you (00:43:28) know he's incredible pioneer also (00:43:30) somebody you know I admire him for (00:43:32) sticking to his vision for 20 30 years (00:43:35) now. In fact I first started using GPUs (00:43:37) in the '9s on for gaming of course for (00:43:40) for for writing graphics engines and (00:43:42) physics uh engines. So it's funny that (00:43:44) it's come full circle to me that that (00:43:46) you know my my early gaming days even (00:43:48) the hardware that was pushed then is now (00:43:50) useful for AI ironically. Um but yeah we (00:43:52) talk about he's very interested in (00:43:54) science and AI for science and actually (00:43:56) you know alpha fold was trained on GPUs. (00:43:58) So we and he loves Alpha Fold and the (00:44:00) work that we're doing you know uh in (00:44:02) drug discovery. So we mostly talk about (00:44:05) um AI for science. I I know a lot of the (00:44:08) data centers are built in Nvidia (00:44:10) systems, but I know Google also has its (00:44:12) its tensor processing units, TPU chips. (00:44:14) Is there any kind of competitive (00:44:16) friendliness there? (00:44:17) >> Yeah. Well, look, we we're lucky we have (00:44:18) our own we love our TPUs. We we (00:44:20) generally use them internally for (00:44:22) training our um our best models and (00:44:25) actually we found there's a big demand (00:44:27) for that from the elite AI teams uh who (00:44:31) are trying to build large models or (00:44:33) serve uh very large AI models. uh (00:44:36) they're specifically built for that. So (00:44:38) TPUs are sort of they're a little bit (00:44:39) more special case than GPUs. You can (00:44:41) think of GPUs as being more general. So (00:44:43) you know maybe we would use a GPU when (00:44:45) we're trying to um explore some new (00:44:47) architecture uh like Alpha Fold was or (00:44:49) some new application. Um but then once (00:44:52) we're when we're trying to sort of um (00:44:54) scale to the maximum things we know then (00:44:57) um you know custom silicon can be a lot (00:45:00) more efficient. Um so we're lucky we (00:45:02) have we have both. We get to use both (00:45:04) here at at Google and Deep Mind. (00:45:06) >> Great. Damus, just looking to the (00:45:07) future, you're obviously so focused on (00:45:09) science and the potential for AI to (00:45:12) create new drug breakthroughs, do (00:45:15) discover new diseases, lots of potential (00:45:17) things there. Um, you've also got (00:45:19) isomeorphic labs, of course, as well. (00:45:21) Where are we on this path to your your (00:45:24) vision of of AI unlocking all of these (00:45:27) these kind of breakthroughs in the world (00:45:29) of science? Well, look, I I I love I (00:45:31) always point to AlphaFold as probably (00:45:33) the best example so far of AI applied to (00:45:36) science. You know, I'm very proud of (00:45:38) that project and you know, we solved a (00:45:39) 50-year grand challenge in science of (00:45:41) protein folding, how the structure of 3D (00:45:43) structure of proteins and over 3 million (00:45:46) researchers around the world are using (00:45:47) it in their critical work. So, I can't (00:45:50) imagine a more transformative sort of (00:45:52) technology. Um, and what I would love is (00:45:55) to see have be able to point to a dozen (00:45:57) alpha folds and, you know, each of them (00:45:59) revolutionizing their area of science or (00:46:02) mathematics. And I think we're well on (00:46:04) the way to that. And we're working on (00:46:05) half a dozen projects like that in (00:46:07) material science, in physics, in in (00:46:09) maths, in weather prediction. Um, and (00:46:12) uh, and I think that the next 10 years, (00:46:14) if if AI goes well and progresses well (00:46:16) and we use it in the right way, it could (00:46:18) usher in a new golden age of um, (00:46:20) scientific discovery. What do you think (00:46:23) are going to be the big things in AI in (00:46:24) 2026? Any big breakthroughs, any big (00:46:26) progresses that you think will happen? (00:46:28) >> Agentic systems, systems are able to do (00:46:30) things more autonomously are going to (00:46:31) start becoming reliable enough to be (00:46:33) useful. Um, I think we're going to see (00:46:36) some really interesting things in (00:46:37) robotics in the next 12 to 18 months. (00:46:39) We're working really hard on some very (00:46:40) ambitious projects with Gemini robotics. (00:46:42) And then finally, maybe um, you know, AI (00:46:46) assistants on devices. I think we're (00:46:49) going to start seeing them really useful (00:46:50) in the in the real world. Uh and then (00:46:52) maybe the thing I'm most excited about (00:46:54) is advancing world models further, (00:46:56) making them more efficient so they can (00:46:57) actually be used maybe for planning in (00:46:59) our general models. (00:47:00) >> Great, Damis. I'm going to take that (00:47:02) last answer as a sort of teaser trailer (00:47:04) for the next time you and I get to catch (00:47:06) up hopefully at some point this year. (00:47:08) Thank you so much for joining me, Dis. (00:47:09) >> Thank you. Thanks for having me. (00:47:14) So Steve, just in that final part of the (00:47:16) conversation, I thought what was (00:47:17) interesting is the relationship between (00:47:20) kind of the deep mind entity and the (00:47:22) broader Google business. And there was a (00:47:24) part where Deis was saying he speaks to (00:47:26) Sundar Pichi, the CEO of Google or (00:47:29) Alphabet every day. Um, and and how sort (00:47:33) of more integrated they've become. And I (00:47:36) think if I'm thinking about that in this (00:47:37) AI race, what that signals to me is that (00:47:40) Google has clearly figured out how to (00:47:43) become speedy at getting AI products to (00:47:46) market. But also, you got to think about (00:47:48) all these Google products, right? (00:47:49) Whether it's Chrome, whether it's uh (00:47:51) Gmail, whatever it might be, they are (00:47:54) wanting whatever Google AI is being (00:47:56) developed to spread all across of all (00:47:58) across those products. that gives them (00:47:59) an absolutely mammoth user base to kind (00:48:03) of almost instantly tap into with some (00:48:05) of these products. And I I've always (00:48:06) I've said this for for a while now. I (00:48:08) think one of Google's biggest strengths (00:48:10) really is that when you think about the (00:48:12) Android operating system and and you (00:48:14) know how large it is, 70% odd market (00:48:17) share globally. You know, that is a huge (00:48:19) amount of people and devices where (00:48:22) Google AI could be effectively installed (00:48:24) on and used quickly. So they're in a (00:48:27) good position in terms of going to (00:48:29) market, I think. And and and clearly um (00:48:31) this relationship between DeepMind and (00:48:33) the broader Google business is going to (00:48:35) be integral for Google to sustain any (00:48:38) success over the over the longer run (00:48:39) here. (00:48:40) >> Yeah. And on the Android front alone, I (00:48:42) mean, Samsung, the biggest manufacturer (00:48:44) of Android phones, they're already (00:48:45) putting Gemini is their main chatbot. (00:48:47) Gemini is their main AI. I'm I was a (00:48:49) little surprised Samsung didn't try to (00:48:51) build their own which like they have in (00:48:52) the past but no they've completely gone (00:48:54) all in on Gemini. They're partnering (00:48:57) with uh Google on those uh the new mixed (00:49:00) reality headset that they have. There (00:49:02) are some upcoming glasses that they're (00:49:04) working on in partnership also uh with (00:49:06) companies like Warby Parker to design (00:49:08) them. Uh so yeah, Samsung has like (00:49:11) really adopted this and that is a huge (00:49:13) platform for Gemini. Just just that just (00:49:16) the Samsung uh angle of it. Just that (00:49:19) huge market share they already have is (00:49:21) is great. And then let's talk about (00:49:22) Apple. Gemini is actually going to be (00:49:24) the engine that powers this new version (00:49:27) of Siri we're expecting in just a couple (00:49:29) months time. He did talk about his (00:49:32) excitement to see Gemini kind of spread (00:49:34) on more devices. So, I think it's a (00:49:37) really smart move by Apple to kind of (00:49:39) realize it can't build this on its own (00:49:41) and honestly do what Samsung is doing (00:49:43) and say, "Okay, let's just integrate (00:49:45) this proven technology. We already have (00:49:47) a great relationship with Google." And (00:49:48) this is honestly a different kind of (00:49:50) Google that I've been see that I've seen (00:49:52) for so many years where you had so many (00:49:54) different groups kind of working on the (00:49:56) same thing. I mean before this big reorg (00:49:59) and and Demis got all that control over (00:50:01) all of AI, there were multiple groups (00:50:04) within Google working on artificial (00:50:05) intelligence uh kind of bumping against (00:50:08) each other and Sudar Pachai was really (00:50:10) smart saying we got to this is a huge (00:50:12) moment. We got to reorganize everything. (00:50:14) He folded everything under Demis Hbasus (00:50:17) and put it into uh DeepMind and that's (00:50:20) where we are now and it's it's really (00:50:22) paid off in 2025 uh in a big way with (00:50:25) Gemini 3. (00:50:26) Yeah. And and that consumer space really (00:50:29) is uh getting more and more intense when (00:50:31) it comes to the the AI side of things (00:50:33) particularly as you know you mentioned (00:50:35) before when you were talking about some (00:50:36) some of the talk about bubbles. These (00:50:38) competitors like Open AI you know Google (00:50:41) has um big uh balance sheet strong cash (00:50:44) flow and it has a huge user base of (00:50:46) users and it continues to innovate. And (00:50:47) I think this really does given the kind (00:50:49) of that reorg and and this kind of speed (00:50:52) you're seeing now from Google, I think (00:50:54) this is adding going to add a lot of (00:50:56) competitive pressure onto OpenAI (00:50:58) particularly on the consumer side uh in (00:51:01) 2026. So it's all up for grabs. (00:51:04) >> Yeah. And we're going to see a lot of (00:51:06) different stuff I I anticipate from (00:51:08) OpenAI this year. They're going to throw (00:51:10) all the spaghetti at the wall they can (00:51:11) to see what sticks because they've put (00:51:14) enormous pressure on themselves to (00:51:16) generate enormous amounts of revenue in (00:51:18) order to fulfill all of these promises (00:51:20) they made about you know capital (00:51:22) expenditure build out of these big data (00:51:24) centers with Oracle and all these sorts (00:51:26) of things like that. Uh it cannot happen (00:51:28) all these committed spending they have (00:51:30) unless they productize it better and (00:51:32) more effectively. But like to your point (00:51:34) we're seeing this with Meta by the way. (00:51:36) Meta has a huge opportunity to leverage (00:51:39) its user base and it hasn't figured out (00:51:41) how to do that in the way Google has. Uh (00:51:43) so right now Google feels like they're (00:51:45) kind of on top of things. Well, look, (00:51:47) part two of this miniseries on DeepMind (00:51:49) is going to be out next week and we're (00:51:51) speaking to Laya Ibrahim who is the COO (00:51:54) over at DeepMind. So catch that. And if (00:51:57) you got any comments uh or or thoughts (00:51:59) about this episode, please reach out to (00:52:01) us. Uh you can reach us pretty much (00:52:02) everywhere. I think uh you're you're on (00:52:05) uh multiple media. You're a blue sky (00:52:07) guy. (00:52:08) >> Yeah, we're all over the place. (00:52:09) >> No Instagram. I quit Instagram 7 and 1/2 (00:52:12) years ago and I do not regret it. (00:52:13) >> Wow, that's amazing. (00:52:15) >> No more doom scrolling. Love it. (00:52:16) >> No more doom scrolling for this guy. (00:52:19) >> Oh, thank you all for listening and (00:52:21) watching. We'll catch you next time.