Google DeepMind’s Demis Hassabis with Axios’ Mike Allen (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) Thank you very much. Big finish. I'm (00:00:02) Mike Allen, co-founder of Axio on bealf (00:00:05) of my co-founders, Roy Schwarz, Jim (00:00:08) Vanhey. Thank you to all of you who for (00:00:10) coming up nine years now have been fans (00:00:12) of Axios and thank you for turning out (00:00:15) here in San Francisco in this historic (00:00:17) bank, this very cool uh setting for this (00:00:20) Axios AI plus SF summit. Uh welcome to (00:00:23) all of you around the world. for our big (00:00:26) finish. Deis Hassabis, PhD co-founder (00:00:30) and CEO of Google Deep Mind. He's a (00:00:34) neuroscientist and entrepreneur and AI (00:00:37) pioneer. Demis was a chess prodigy at (00:00:41) five, a Nobel laureate at 48. He's a (00:00:45) Britishborn genius. He's been kned. (00:00:47) Deisabas, welcome to Axios. (00:00:52) [applause] (00:00:57) Thank you so much. Thanks for having me. (00:00:59) >> We've been looking forward to this. We (00:01:00) appreciate (00:01:01) >> to be here. (00:01:02) >> It was just over 400 days ago that you (00:01:05) found out you were a Nobel laureate. And (00:01:08) you said in that moment, you said this (00:01:10) is surreal. (00:01:12) >> This is the big one. (00:01:13) >> Yeah. (00:01:14) >> What has changed since then about your (00:01:16) life and work? What has it made (00:01:17) possible? (00:01:19) >> Um, well, look, it's still pretty (00:01:20) surreal actually. still hasn't fully (00:01:22) sunk in, but uh it has made quite a big (00:01:25) difference. The the thing it makes a (00:01:26) difference to is when you speak to (00:01:28) people not in your field, including you (00:01:31) know big government people, things like (00:01:32) that who maybe don't know that much (00:01:34) about AI. If you, you know, you have the (00:01:36) Nobel Prize, it's a sort of shortcut to (00:01:39) almost anyone to to know that you're, (00:01:41) you know, you're expert in your field. (00:01:42) So, it's it's going to be useful, I (00:01:44) think, in the future. and [laughter] (00:01:47) and you had endless resources at your (00:01:51) disposal. Are there new resources that (00:01:53) you have or that you think you can tap (00:01:55) now? (00:01:56) >> Not really. I mean, you're right. We (00:01:57) we're lucky at at Google at Deep Mind. (00:01:59) We have we have a lot of resources. (00:02:01) They're not endless. We always need more (00:02:03) more compute. Uh no matter how much (00:02:05) compute you have, but um but we have, (00:02:08) you know, a lot of great things, which (00:02:09) is why we're able to do such a broad (00:02:11) portfolio of things. But it's mostly (00:02:13) again this uh this platform it gives you (00:02:16) to basically speak out about things that (00:02:18) you care about. And I haven't done a lot (00:02:20) of that yet, but I think it will be (00:02:22) important. Maybe we're going to talk (00:02:23) about AI safety and other things. I (00:02:25) think uh the Nobel and the platform that (00:02:28) gives you uh could be useful for that. (00:02:30) >> And what's on the short list of in (00:02:33) addition to AI safety that you think (00:02:34) you'll be talking more about using your (00:02:37) platform? (00:02:37) >> Yeah, well it's not just about uh safety (00:02:40) in the long term. AGI safety obviously I (00:02:42) think we think a lot about that but it's (00:02:44) also about responsible use of AI today. (00:02:47) Uh what are the kinds of things we (00:02:49) should be using AI to to improve and to (00:02:52) to power up and to accelerate and maybe (00:02:55) you know what sorts of things we should (00:02:56) be careful about um uh uh in the even in (00:02:59) the near term. So I think that's one (00:03:01) thing I think also just getting society (00:03:03) ready for what's to come. you know, AGI (00:03:06) probably the most transformative moment (00:03:08) in human history is on the horizon and (00:03:11) um we need to get prepared as a society (00:03:15) um and as a species and I think of (00:03:17) course governments and other important (00:03:19) people uh uh other important leaders are (00:03:22) going to be critical in that and I think (00:03:24) having having something like the Nobel (00:03:26) platform opens pretty much any door. One (00:03:29) of the things that distinguishes you is (00:03:31) you're deep in the science and yet you (00:03:35) also are on the front line of this fight (00:03:38) and this race among companies, (00:03:40) hyperscalers, superpowers and you sort (00:03:44) of in the mold of Steve Jobs, you also (00:03:46) have a product mind. You want to create (00:03:50) delightful things for people, but you (00:03:52) always say you're a scientist first. (00:03:55) >> Yeah, science. I'm a scientist first. (00:03:57) The reason I say that is that's the (00:03:58) that's the sort of default approach I (00:04:00) take to everything. So um and what I (00:04:03) mean by that is the scientific method (00:04:05) really that way of thinking. Um I really (00:04:08) love the I mean I think it's the most (00:04:10) the scientific method is is the most (00:04:12) important maybe idea humanity's ever (00:04:14) had. Um you know created the (00:04:16) enlightenment and then modern science. (00:04:18) So basically, modern civilization (00:04:20) depends on on on this on this idea of (00:04:22) scientific method and experimentation (00:04:25) and then updating your hypothesis and so (00:04:27) on. And I think it's an incredibly (00:04:29) powerful method, but I think it can be (00:04:31) applied to more than just science. I (00:04:32) think it can be applied to everyday (00:04:34) living and indeed business. Um, and (00:04:36) that's what I've tried to do is sort of (00:04:38) take that uh uh to its limit. And I (00:04:42) think that's what gives us um you know (00:04:44) advantage in some ways as a as a (00:04:46) research organization as an engineering (00:04:48) organization. Yes, we're in the middle (00:04:50) of this ferocious probably the most (00:04:52) ferocious competitive battle maybe tech (00:04:54) has ever seen. Um and uh but one of the (00:04:58) things that I think gives us an edge is (00:04:59) is the rigor and precision we bring to (00:05:02) our work. um because um we have a (00:05:05) scientific method sort of at the heart (00:05:06) of it and we blend world-class research (00:05:09) with world-class engineering with (00:05:10) world-class infrastructure and I think (00:05:12) you need all three of those things to be (00:05:14) at the frontier of something like AI and (00:05:17) I think you know we we're we're sort of (00:05:18) pretty unique in having uh worldclass (00:05:20) capabilities in in all those areas. Um (00:05:24) yeah so in Axio fashion we're going to (00:05:27) divide our conversation between zoom out (00:05:29) and zoom in. So, zoom out uh getting (00:05:32) your priceless uh mind on the state of (00:05:35) AI. So, we're going to talk about the (00:05:37) blunt state of AI. And what I'm going to (00:05:39) ask from you is given the known knowns (00:05:42) today, be blunt, (00:05:44) >> clinical, no hype, no soft selling. Can (00:05:47) we do that? (00:05:47) >> I'll do my best. (00:05:48) >> All right. Um (00:05:51) what does the next 12 months of progress (00:05:54) look like? What do you believe that if (00:05:56) if we sit here a year from today and I (00:05:58) would love to uh what will have changed (00:05:59) in the world? (00:06:01) >> Um I think the things that that we're (00:06:02) we're pressing hard on are um uh the (00:06:06) convergence of modalities. So you Gemini (00:06:10) which is our main foundation model has (00:06:11) always been multimodal from the (00:06:12) beginning. It takes images, video, uh (00:06:15) text, audio and then can produce now (00:06:18) increasingly produce those uh uh types (00:06:20) of outputs as well. Um, and I think (00:06:22) we're getting some really interesting uh (00:06:25) cross-pollination by being multimodal. (00:06:28) One the best example of that is our (00:06:29) latest image model NO Banana Pro which (00:06:32) um I think shows some astonishing sort (00:06:34) of understanding of visuals and it can (00:06:37) kind of you know create infographics (00:06:38) that are really accurate and so on. So I (00:06:40) think over the next year you're going to (00:06:42) see that uh uh progress a lot and I (00:06:44) think for example in video when that (00:06:46) converges with the language models (00:06:48) you're going to be see some very (00:06:49) interesting combinations of capabilities (00:06:51) there. I think the other things we're (00:06:53) going to see over the next year and I'm (00:06:54) personally working on is world models. (00:06:56) So uh we have this um uh uh system (00:06:58) called Genie Genie 3 which is like an (00:07:01) interactive video model you can think (00:07:03) about. So you can sort of generate a (00:07:04) video but then you can start walking (00:07:06) around it like you're in a game or (00:07:07) simulation and it stays coherent for a (00:07:09) minute. I think that's very exciting. Um (00:07:12) and then uh you know maybe the other (00:07:15) thing is a agent based systems. So we I (00:07:17) think the field's been talking a lot (00:07:18) about agents but then they're not (00:07:20) reliable yet enough to do full tasks. (00:07:22) But I think over the next (00:07:23) >> we've heard a lot about that today here (00:07:24) on the Axia stage. What would you say a (00:07:26) year from now? How will agents have (00:07:28) progressed? What's an example of how it (00:07:31) will work in everyday life a year from (00:07:33) now? (00:07:34) >> Well, look, I we we have this concept of (00:07:35) a universal assistant that we want (00:07:37) Gemini eventually to become. Uh I think (00:07:40) this is also you're going to see from us (00:07:41) over the next year. This will be on on (00:07:42) on more devices as well. By universal, (00:07:46) we mean it's not just on your computer (00:07:47) or your laptop or your or your phone, (00:07:49) but maybe comes around with you on (00:07:51) glasses or other devices. And um I think (00:07:55) it needs you know we want to create (00:07:56) something that is useful to you in your (00:07:59) everyday life that you consult many (00:08:01) times a day. it becomes a part of the (00:08:03) fabric of your life and it just improves (00:08:05) your productivity but also your personal (00:08:07) life you know recommendations for books (00:08:08) and films and other or activities that (00:08:11) you'd like and but yeah so but agents at (00:08:14) the moment they can't comp you can't (00:08:15) delegate to them uh a whole task and be (00:08:19) sure they're going to complete that (00:08:20) entire task uh uh completely reliably (00:08:23) >> but a year from now you think they will (00:08:25) >> I think a year from now we'll start (00:08:26) having agents that uh are close uh to (00:08:30) doing that and bullcase, barecase, what (00:08:34) is the best case for what AI can do for (00:08:38) the world and what do you fear most? (00:08:41) >> Well, look, the the the the the best (00:08:43) case scenario that that I've always (00:08:44) dreamed about and why I've worked my (00:08:46) whole life on on on AI and you know (00:08:48) getting closer to this moment we've been (00:08:50) working towards for decades now, many of (00:08:52) us is um uh a kind of I somes call it (00:08:56) radical abundance. So this idea we (00:08:58) solved a lot of the biggest issues (00:09:00) confronting uh society and humanity (00:09:03) today. So whether that's free uh uh (00:09:06) renewable clean energy, maybe we sold (00:09:08) fusion or better battery optimal (00:09:10) batteries and and solar uh materials, (00:09:13) semiconductors, you know, material (00:09:15) science. We've solved a lot of diseases. (00:09:18) So then we're in a situation where, you (00:09:20) know, we're in this new era, post (00:09:21) scarcity era, and we're potentially, you (00:09:25) know, humanity's is is flourishing and (00:09:27) traveling to the stars and spreading (00:09:29) consciousness to the to the galaxy. (00:09:31) >> And what do you fear most? (00:09:34) >> Well, even that utopian kind of view has (00:09:38) some questions around it about what will (00:09:40) be um our purpose as humans if there are (00:09:43) these technologies and that are out (00:09:44) there that are solving all these (00:09:46) problems. all be left to solve. You (00:09:47) know, I worry about that as a scientist (00:09:49) and you know, the scientific method (00:09:50) even. So, there's that, but there's also (00:09:52) obviously the the well-known uh down (00:09:55) challenges and risks with AI of well, (00:09:57) twofold. One is bad actors um uh using (00:10:01) AI for harmful ends um or the AI itself (00:10:04) as it gets closer to AGI and becomes (00:10:06) more gentic um it goes off the rails in (00:10:09) some way that harms humanity. (00:10:11) >> So, you mentioned going off the rails. (00:10:14) Um, how worried are you about these (00:10:17) catastrophic outcomes? Your level of (00:10:19) concern? I'm just going to rattle them (00:10:21) off. One, pathogens created by an evil (00:10:24) actor using AI. (00:10:25) >> Mhm. (00:10:27) >> I think that's definitely one of the one (00:10:29) of the bad use case scenarios that we (00:10:30) have to guard against for sure. (00:10:32) >> Energy or water cyber terror using AI by (00:10:36) a foreign actor. (00:10:37) >> Yeah, that that's probably almost (00:10:39) already happening now, I would say. (00:10:41) Maybe not with very sophisticated AI (00:10:43) yet, but I think that's the most obvious (00:10:45) vulnerable vector. Um, and which is why (00:10:48) we focus quite a lot and we are focusing (00:10:50) quite a lot as Google and as DeepMind on (00:10:52) on AI for cyber security. So, so to (00:10:54) power up the defensive side of that (00:10:56) equation, (00:10:57) >> AI operating outside human control on (00:11:00) its own. (00:11:02) Well, this goes back to the agentic (00:11:03) stuff where I think as that becomes more (00:11:06) sophisticated and it's clear why the (00:11:08) industry will build those things because (00:11:09) they'll be more useful as things like (00:11:11) assistance. Um, so they're definitely (00:11:13) going to happen, but the more aentic and (00:11:16) autonomous they are, the more room there (00:11:18) is for these things to uh deviate from (00:11:21) what you maybe had intended when you (00:11:23) gave the initial instruction or the (00:11:25) initial goal. So this is a very active (00:11:27) area of research which is to how to make (00:11:29) sure that systems that maybe are capable (00:11:32) of continual learning or online learning (00:11:34) stay uh within the guard rails that that (00:11:37) you set. I mean, I think the good news (00:11:39) is um because AI is become such so big (00:11:43) commercially and for enterprises, if you (00:11:46) think about renting or selling one of (00:11:48) your agents as a model provider, leading (00:11:50) model provider to another big business, (00:11:53) those businesses will want guarantees (00:11:56) around the agents behavior, what it does (00:11:58) with their data, what it does with their (00:12:00) the customers. And if those things go (00:12:03) wrong, they're not going to be (00:12:04) existential in any way, but you'll lose (00:12:06) the business for sure. So because why (00:12:08) would that business enterprise go with (00:12:10) that provider? They would choose a (00:12:11) different provider that was more (00:12:13) responsible and had better guarantees. (00:12:15) So I think what's great about that is um (00:12:17) that that will it will sort of (00:12:18) capitalism will reward sort of naturally (00:12:22) uh ideally more responsible actors (00:12:24) >> but it's possible that the AI could jump (00:12:28) the moat, jump the guard rail (00:12:30) >> potentially if done wrong. I mean it's (00:12:31) there was always a possibility. We're we (00:12:34) nobody really knows what the um that's (00:12:36) one of the big unknowns. I think it's (00:12:38) non zero that potential. Uh so it's (00:12:41) worth very seriously considering and (00:12:43) mitigating against but um you know I (00:12:45) hear people talk you know give very (00:12:47) precise percentages about what the (00:12:49) chances of these poom (00:12:50) >> a p doom which I think is kind of (00:12:52) nonsense because no one knows what it (00:12:54) is. What I know is it's (00:12:56) >> so you don't you don't quantify it but (00:12:57) you say it's (00:12:58) >> it's non zero. So clearly if your PDM is (00:13:02) non zero then you you you know you must (00:13:04) put significant resources and and and (00:13:07) attention on that. (00:13:08) >> Where is the US winning the AI race (00:13:11) against China and where are we losing? (00:13:13) >> Um I I I think that we're still in the (00:13:16) in the US and in the west um in the lead (00:13:20) uh if you look at the at the latest (00:13:22) benchmarks and um the latest systems but (00:13:25) they're not you know China is not far (00:13:26) behind. If you look at the latest (00:13:28) DeepSseek or the latest smallers, (00:13:29) they're very good and they there are (00:13:31) some very capable teams there. So maybe (00:13:33) we're, you know, the lead is only a (00:13:35) matter of months as opposed to years at (00:13:37) this point. (00:13:38) >> Because when you put chips aside, AI, (00:13:40) China probably is winning. (00:13:42) >> Um, no, I think chips is one thing, but (00:13:45) I think algorithmically, innovation (00:13:47) wise, I think the West still has the (00:13:49) edge. So I don't think any of the (00:13:51) Chinese models or or companies have (00:13:55) shown they can innovate on (00:13:57) algorithmically something new that um (00:14:00) beyond the state-of-the-art they they (00:14:03) they've been very good at um uh fast (00:14:06) sort of following the the current uh (00:14:09) state-of-the-art. (00:14:10) >> Our last zoom out question and you're (00:14:11) going to like this one. What's the most (00:14:13) astonishing thing about AI that you (00:14:16) think gets shockingly little attention? (00:14:19) The most astonishing thing about AI that (00:14:20) gets shocking little little attention. (00:14:22) >> Wow. Yeah. I think if I think of the (00:14:24) things we're working on and already have (00:14:25) working, it's the um multimodal (00:14:29) understanding these models have. Like if (00:14:31) you (00:14:31) >> and multimodal video, (00:14:33) >> yes, video uh image and and I mean (00:14:36) audio, but I'm thinking specifically (00:14:38) video actually. So if you if you give (00:14:40) Gemini a YouTube video to process, you (00:14:43) can ask it all sorts of incredible (00:14:45) things about the video that it's just (00:14:47) sort of mind-blowing to me that it can (00:14:49) understand sort of conceptually in a lot (00:14:51) of cases like not always but in many (00:14:53) really impressive cases what's (00:14:55) happening. Can understand (00:14:56) >> example of a question. Um well I've (00:14:58) asked questions on on like um you know (00:15:00) one of I mean look this was just (00:15:02) something I tested Gemini on the other (00:15:03) day was was um I love the film Fight (00:15:06) Club and uh there's some scene in it I (00:15:09) think where Brad Pitt or or or maybe (00:15:11) it's Ed Norton I can't remember takes (00:15:12) off his ring uh uh before having a fight (00:15:16) and the sort of um I asked you know (00:15:19) Gemini like what's the significance of (00:15:21) of of of that of that [snorts] action (00:15:23) and you know he came up with a very (00:15:25) interesting sort of philosophical point (00:15:26) about leaving behind uh everyday life (00:15:28) and and just sort of symbolically (00:15:30) showing that um was you know very (00:15:33) interesting kind of meta insight that (00:15:35) that you know these systems have now and (00:15:37) I think if you use it the other thing (00:15:39) that's sort of not appreciated is like (00:15:40) we have this thing called Gemini Live (00:15:41) where you can point your phone at (00:15:43) something and say you're a mechanic uh (00:15:46) uh it can actually just help you with (00:15:48) whatever you know task you have in front (00:15:50) of you ideally that should be glasses (00:15:51) because you want to have your hands free (00:15:53) really for that um but I think people (00:15:55) don't realize how um how powerful that (00:15:58) multimodality capability is yet. (00:16:00) >> All right, you've given us the perfect (00:16:02) bridge in transition to zooming in. Uh (00:16:04) congratulations on Gemini 3 last month. (00:16:07) Uh your gamechanging uh model, you say (00:16:10) it reasons with unprecedented depth and (00:16:13) nuance. Tell us what's unique about the (00:16:15) nuance part of Gemini 3. Yeah, I think (00:16:18) it's just um uh uh we're really pleased (00:16:21) with the the the the almost the (00:16:23) personality of it, the style of it as (00:16:25) well as its capability. I I I I like the (00:16:27) way um that it answers succinctly. It (00:16:30) pushes back a little bit if you're (00:16:32) doesn't just agree with whatever you're (00:16:33) saying. It pushes back gently on some (00:16:35) ideas that if they're not if they don't (00:16:37) make sense. And I think people are (00:16:39) appreciating uh it seems you know sort (00:16:41) of I feel like it's a you can feel it's (00:16:43) a bit of a step change in its kind of (00:16:45) intelligence and therefore usefulness. (00:16:48) >> And what's something that Gemini has (00:16:50) answered or produced where you said I (00:16:52) didn't know it could do that or I didn't (00:16:54) know it would do that. (00:16:56) >> Well actually this is the the amazing (00:16:58) thing of when you why we love what what (00:17:00) we're doing so much is that the this era (00:17:03) we're now in with research connected to (00:17:05) product. The great thing about that is (00:17:07) that you get millions and potentially at (00:17:10) Google billions of users immediately (00:17:12) take advantage of the new technology you (00:17:14) put out there. And uh we're continually (00:17:17) surprised by the cool things that people (00:17:19) figure out very quickly um to use these (00:17:22) models for. Um and a lot of those things (00:17:24) sort of, you know, tend to go viral. But (00:17:26) the thing I I most enjoyed with Gemini 3 (00:17:28) was oneshotting uh games. So back to my (00:17:31) very first career of making AI for (00:17:32) games, I think we're very close now with (00:17:34) these models. maybe the next version (00:17:36) models where you could start really (00:17:37) creating perhaps commercial grade games (00:17:40) uh you know vibe coding them uh with you (00:17:43) know in a few hours which used to take (00:17:45) years (00:17:46) >> and that shows nuance. What does that (00:17:48) show about the model? Well, I think it's (00:17:50) just incredible uh uh depth and and and (00:17:54) capability of these models to understand (00:17:57) very high level instructions and and (00:18:00) produce you know very detailed outputs (00:18:02) and the other things that uh uh Gemini 3 (00:18:05) particularly is good at is front-end (00:18:07) work and developing you know websites (00:18:09) and it's it's pretty good aesthetically (00:18:11) and creatively as well as um (00:18:14) technically. Something we've written a (00:18:16) fair amount about at Axios is that even (00:18:18) the authors, creators of these models (00:18:20) don't totally understand them. What's (00:18:23) something about Gemini 3? Yeah. (00:18:24) >> That you feel like you don't totally (00:18:26) get? (00:18:27) >> Well, actually I feel like with all (00:18:29) these models um and and maybe all of the (00:18:32) the audience are feeling this too is (00:18:33) that it there's such a fast pace of of (00:18:36) of innovation and improvement. Um we're (00:18:39) spending almost all of our time building (00:18:42) these things. We have we don't even have (00:18:44) I I have to have this feeling every time (00:18:45) we release a new version that I haven't (00:18:47) even explored a tenth had time to even (00:18:49) explore a tenth of probably what the (00:18:51) existing systems can do because of (00:18:53) course we're on to immediately you know (00:18:55) we're referencing back to the ferocious (00:18:56) race and competition we're in we're (00:18:58) immediately focusing on the next (00:19:00) innovation uh and obviously making sure (00:19:02) it's safe and reliable and all those (00:19:04) things. So again, our users end up uh uh (00:19:08) taking them much further than often uh (00:19:10) we we've tried internally. (00:19:12) >> And one more question on Gemini 3, a (00:19:15) little back story and you had a number (00:19:19) of irons in the fire, but LLM's the (00:19:22) textbased uh large language models. uh (00:19:26) you didn't necessarily go all in on that (00:19:29) as the holy grail. Something that Walter (00:19:31) Isacson, the great author and thinker (00:19:33) and your friend said to me is that when (00:19:35) you saw the power of the LLM, you did a (00:19:39) pivot, a pureette, as Walter said it, (00:19:42) and were able to leapfrog to great (00:19:45) success. And Walter's point was that (00:19:47) most business people would have been (00:19:49) stubborn, might have doubled, triple (00:19:51) down on their other bets. How did you (00:19:54) make this decision to go allin on your (00:19:57) LLM? (00:19:57) >> Well, I think this is again the the the (00:19:59) beauty of and the strength of the (00:20:00) scientific method. If you're a true (00:20:02) scientist, you can't get too dogmatic (00:20:05) about some idea you have. You you need (00:20:07) to go with where the empirical evidence (00:20:09) is taking you. So, first of all, this is (00:20:11) this is Walter is probably referring (00:20:13) back to the 2017 2018 era. So, there we (00:20:16) had a lot of irons in the fire. As we (00:20:18) said, we had our own very capable (00:20:20) language models. They were called (00:20:22) Chinchilla and then Sparrow and we had (00:20:23) these various different code names for (00:20:25) them. Um they weren't publicly released (00:20:26) but they were internal. In fact, some of (00:20:28) the scaling laws were originally figured (00:20:30) out by our team. They're called the (00:20:31) Chinchilla scaling laws. Um but we also (00:20:34) had other types of programs alpha zero (00:20:36) things that were building on Alpha Go (00:20:37) pure RL systems and we also had some (00:20:39) cognitive science more neuroscience (00:20:41) inspired architectures as well. And at (00:20:43) the time all we weren't sure my job is (00:20:46) to make sure we build AGI uh first fast (00:20:49) and safely, right? That's always been (00:20:51) our our solve intelligence, our mission (00:20:53) at DeepMind. And and so I'm kind of (00:20:56) agnostic actually to the to the approach (00:20:58) that's taken. I'm pretty pragmatic on (00:21:00) that. That's maybe my engineering side (00:21:02) of me is I have some theories as as a (00:21:04) good scientist would, but I'm I'm I'm at (00:21:06) the end of the day, it's got to (00:21:07) pragmatically work. And so when we (00:21:09) started seeing the beginnings of scaling (00:21:11) working, then we increasingly put more (00:21:14) and more resources onto that branch of (00:21:16) the of the of the research tree. (00:21:18) >> Something that's refreshing about your (00:21:20) approach is with artificial general (00:21:21) intelligence, human capable uh AI. You (00:21:25) don't shy away from it. Some other (00:21:26) people say, "Well, we won't know or (00:21:28) we're already there or it doesn't (00:21:30) matter." You say that it does matter and (00:21:32) we will know. And you say it's not far (00:21:35) off. (00:21:35) >> Yeah, we're definitely not there now. (00:21:38) So, and and I and (00:21:39) >> actually quite close is how you say. (00:21:41) >> Yes, quite close. I think we're like (00:21:42) five to 10 years away if you were to ask (00:21:44) me. I'm sorry. I think Say that again. (00:21:45) >> Five to 10 years away. I think my bar (00:21:48) though is quite high. So, this is the (00:21:50) the we define AGI as you know the a (00:21:52) system that that exhibits all the (00:21:54) cognitive capabilities we have and that (00:21:56) includes uh inventive and creative (00:21:58) capabilities. I think there are missing (00:22:01) there's as all of you have used the (00:22:02) current LLMs there are they're they're (00:22:04) amazing in some ways. They're really (00:22:06) impressive in some senses in some (00:22:07) they've got incredible almost PhD levels (00:22:09) uh key skills in some areas IMO gold (00:22:12) medals and so on but in other areas (00:22:14) they're very flawed still and so they're (00:22:16) these sort of jagged intelligences so (00:22:18) the you would expect across the board (00:22:20) consistency from a true AGI and they're (00:22:22) missing other capabilities like (00:22:24) continual learning online learning (00:22:26) long-term planning and reasoning they (00:22:28) can't do any of these things currently I (00:22:30) think they will be able to but maybe one (00:22:32) or two more breakthroughs are going to (00:22:33) be required (00:22:34) >> and a question from the great Ena (00:22:35) Frereded who we've uh seen today and (00:22:37) whose uh coverage from day zero of Axios (00:22:40) has helped make Axios what it is. Uh she (00:22:44) says you're obviously (00:22:46) um uh you've said that AI might be one (00:22:51) advance two advances away from AGI. (00:22:54) >> Yes. (00:22:54) >> Will we get there just by improving LLM (00:22:58) and generative AI or do you think that (00:23:00) there might be a different approach (00:23:01) that's needed to hit a GI in your 5 to (00:23:04) 10 years? I think I think again this is (00:23:05) an empirical question but what I do know (00:23:08) this is this would be my best guess is (00:23:10) um the scaling of the current systems (00:23:13) you we must push that to the maximum (00:23:16) because at the minimum it will be a key (00:23:18) component of the final AGI system it (00:23:20) could be the entirety of the AGI system (00:23:23) there's a chance that just scaling will (00:23:25) get you there but my guess is if I was (00:23:27) to guess from where I my vantage point (00:23:28) now is that one or two more big (00:23:30) breakthroughs when I mean there's (00:23:32) innovation going on all the time by the (00:23:33) way even including in scaling um (00:23:36) existing techniques but I'm talking like (00:23:38) a transformer level or alpho level type (00:23:40) of breakthrough. I think we might I (00:23:42) suspect when we look back in once AGI is (00:23:45) done that one or two of those things (00:23:47) were still required in addition to (00:23:49) scaling. (00:23:49) >> We're about to get the hook. So a super (00:23:51) rapid round. Another question from uh (00:23:53) Ena. you obviously are a big believer in (00:23:56) AI, but if you look at what's being (00:23:58) spent, that doesn't mean that there (00:24:01) might not be a big enough bubble to (00:24:02) rattle the economy. How worried are you (00:24:04) about that? (00:24:05) >> Um, I think we there I think it's not a (00:24:08) binary. I think some parts of the AI (00:24:10) industry are probably in a bubble like, (00:24:12) you know, I don't know, like the seed (00:24:13) rounds of, you know, you know, $50 (00:24:16) billion seed rounds and things like that (00:24:18) seems a little bit unsustainable. But um (00:24:20) on the other hand, of course, I more (00:24:22) than anyone believes that AI is the most (00:24:24) transformative uh technology ever. So I (00:24:27) think in the fullness of time, this is (00:24:29) all going to be uh more than justified. (00:24:31) And my job as head of Google Deep Mine (00:24:34) and and the engine room of Google is to (00:24:36) make sure we win either way. If if the (00:24:39) bubble the so-called bubble bursts or if (00:24:42) things continue to be good like they are (00:24:43) now, we're in a strong position. (00:24:44) >> The AI recruiting wars, what's the end (00:24:47) state of this competition for talent? (00:24:49) Well, look, it's gone pretty crazy (00:24:50) recently. Things like what Meta have (00:24:52) been doing and, you know, everyone's got (00:24:54) to do what what makes sense for them. (00:24:56) Uh, what we found for us is that we want (00:24:59) people who are missiondriven. We have, I (00:25:01) think, the best mission. We have the (00:25:03) full stack. So, I think if you want to (00:25:05) do the most impactful work and have the (00:25:06) most positive impact on the world, then (00:25:08) I think there's nowhere better uh than (00:25:10) than at Google DeepMind. And in the end, (00:25:13) I think the best scientists, the best (00:25:15) researchers, the best engineers, they (00:25:16) want to work on the most cutting edge (00:25:18) stuff. So if you're the sort of top of (00:25:20) the leaderboards with the best systems, (00:25:22) uh that's that's sort of a self fueling. (00:25:25) This is a question from James (00:25:26) Vanderhigh, an entrepreneurial young (00:25:28) mind at High Point University in North (00:25:30) Carolina. He says, "There's a lot of (00:25:32) conversation about AI gaining a mind of (00:25:35) its own. Is there a scenario where AI (00:25:38) could act in its selfinterest?" (00:25:41) Well, that's a great question and and (00:25:42) it's related to the some of the the the (00:25:44) the more sort of catastrophic outcomes (00:25:47) is if that went wrong, that would be one (00:25:49) of the issues that with agentbased (00:25:51) systems or very autonomous systems if (00:25:53) somehow they developed a self-interest (00:25:55) that was some in some sense sense (00:25:58) conflicting with what the designers or (00:26:00) even perhaps humanity wanted it to do. (00:26:03) >> And finishing with a fun thing, you're (00:26:06) still a gamer. What does gaming teach us (00:26:09) about the world and what does gaming (00:26:11) teach us about where these machines are (00:26:13) headed? (00:26:14) >> Well, look, I think uh my chess (00:26:16) background and and my training in that (00:26:18) and then other games subsequently has (00:26:20) been critical to how I do my work and (00:26:22) both in business and in science. Uh I (00:26:25) think the thing I love about games and (00:26:26) there's many things I've loved about (00:26:28) them, but I love the creativity of (00:26:30) making them. But I also just playing (00:26:32) them I think is the best way to train (00:26:33) your mind because the best games whether (00:26:35) that's chess or go or whatever or poker (00:26:38) they're microcosms of something in the (00:26:40) real world right but in general you (00:26:42) don't get in the real world to have (00:26:44) several practice goes at making the (00:26:47) decision correctly in that moment. Maybe (00:26:49) in the real life you only get a dozen of (00:26:51) those critical moments, but you can (00:26:52) practice your decision-m capabilities as (00:26:55) much as you want uh w within the the the (00:26:58) the almost the simulation really of the (00:27:00) world with games. Um and as long as you (00:27:02) take the games very seriously, so you (00:27:04) put you put a lot of thought into your (00:27:05) decision-m, it really does train your (00:27:08) your decision-m and planning (00:27:09) capabilities in my opinion. Now, you've (00:27:11) pointed out that our squishy brains uh (00:27:13) evolved uh to be hunter gatherers and (00:27:17) yet we're facing a disruption that as (00:27:18) you put it to the Guardian will be 10 (00:27:20) times bigger than the industrial (00:27:22) revolution and maybe 10 times faster. (00:27:25) Are we facing a situation where most (00:27:27) humans can't keep up and maybe no human (00:27:31) including you can keep up? (00:27:33) >> Well, the good news is and I think my (00:27:35) point on the hunt gather was look how (00:27:37) adaptive our brains have been. We we (00:27:40) evolved to be hunter gatherers and yet (00:27:42) here we are sitting in our modern (00:27:44) cities, modern civilization with all the (00:27:46) technology around us and um you know the (00:27:50) same human brain pretty much has been (00:27:52) able to adapt to that. So I'm a really (00:27:55) uh big believer in uh human ingenuity (00:27:58) and um and I think we're infinitely (00:28:00) adaptable. We are the only existence (00:28:02) proof our brains are the only existence (00:28:04) proof of general intelligence perhaps in (00:28:06) the known you know universe that we know (00:28:08) of so far. So we are general (00:28:10) intelligences ourselves and so we should (00:28:12) be able to infinitely adapt. There is a (00:28:14) question about when AGI post AGI what (00:28:17) kinds of technologies can we create (00:28:19) brain computer interfaces other things (00:28:21) that some of us may choose to to use in (00:28:24) addition to our existing technologies (00:28:26) and that could be one way for us to keep (00:28:28) up. (00:28:28) >> And as we say goodbye you're a lifelong (00:28:30) Liverpool fan. You've helped them with (00:28:32) their analytics. How will AI affect and (00:28:37) inform the World Cup here in North (00:28:39) America? (00:28:40) >> Well, we've had a lot of we've had a lot (00:28:42) of teams approach us for for help, too. (00:28:44) And um and I have to be try and be equal (00:28:46) with that, but it's hard having a (00:28:48) lifelong spot of Liverpool. But I'm (00:28:49) looking forward to trying to make it out (00:28:51) here maybe at least for the World Cup (00:28:52) final. (00:28:52) >> But but let's be serious. What what what (00:28:55) will it change between now and then? (00:28:57) It's a it's a lifetime in AI between now (00:29:00) and then, right? (00:29:01) >> Yeah. Well, what in AI or AI for sport (00:29:03) or just in (00:29:04) >> Yes. Yeah. Well, I mean, look, sport has (00:29:05) immense amount of data and it's all (00:29:07) about uh extreme elite performance. So, (00:29:10) it's actually a natural bed fellow for (00:29:12) for AI to to come in and and help (00:29:14) optimize that process even further. (00:29:16) >> And without giving away a trade secret, (00:29:18) what will it be able to do for a World (00:29:20) Cup team? (00:29:21) >> Uh maybe score more headers from from (00:29:23) corners, you know, if you place the (00:29:25) that's one of the things I think our (00:29:26) system found out like precise (00:29:28) positioning of the players. Deus, thanks (00:29:31) for making a