[00:00] [applause] >> It's great to see you. Seems like we always talk where it's very cold or very hot. >> exactly. >> very hot, so let's see if we can get through this without sweating. >> Yeah. >> Um Demis, everyone right now is freaking out about AI. Uh they're they're banning AI models in DC. Um you know, a lot of the a lot of the concern though is about these text-based models uh that can create software, find vulnerabilities in computers. Um I'm wondering if you think a lot as a
[00:31] lot of people do that the path to AGI runs through these models like like Mythos that may soon be sort of self-improving, or do you think that it still requires a multimodal approach ala what you're working on at Gemini? >> Well, there's a lot that's unpacked just in that first question. I mean, first of all, with the things that we're seeing with Cyber and Mythos and um I've I've I've been pretty vocal for a long while that as we get closer to AGI, and I think we're on the cusp of that
[01:01] now, I I said sort of said statements like we're in the foothills of the singularity, that we need a bit of a more systematic approach. To of course, there's amazing opportunities ahead. A lot of things you talked about in the introduction, solving all disease, finding new energy sources. I All of these are the reasons that I've worked on AI my whole career, but there are also risks. And Cyber's one, but I'm actually there are going to be even more serious things. That's just a kind of warning shot for humanity, and I hope we take it seriously. Um but there'll be
[01:31] bio, nuclear, other kinds of risks coming down the line, maybe in the next couple of years. And we've got to get ready for that, and I think we need a more systematic way to deal with uh the issues, uh and maybe kind of a standards body that ideally would be international as well that would would help uh test the the the latest frontier systems to make sure they're robust and the guardrails uh are sufficient. So, that's on the that's on the on the on the one hand of what what you just mentioned. In terms of the technical approaches to AGI,
[02:01] we've always had a kind of, you know, a broad at the broadest, I would say, and deepest research bench. That's why over the last decade, I think a lot of maybe 90% or plus of the big breakthroughs in AI that underpin the modern AI industry came from, you know, Google Brain or DeepMind as we were as separate research entities and now together as Google DeepMind. From Transformers that underpins all large language models to AlphaGo and all the reinforcement learning pioneering that we did back
[02:31] back then. So, so I think our approach has always been to bet on multiple things and push them as hard as possible. So, obviously we have our scaling work, our own multimodal foundation models Gemini. We're pushing hard on coding, but also we have our multimodal generative media models like Omni and VEO. And we think that it's important to give the these models understanding of the world around us,
[03:01] the context around us. And I think in the end, to have a full AGI system, you need to be able to also understand the physical world around you. And you definitely need that for things like robotics to become a reality and things like assistant on smart glasses, which I think are two very interesting applications. >> I'm going to take that as a no. Thank you. So, I when you started DeepMind, you were so far in the frontier. And when you joined Google, it just seemed like between DeepMind and Google, there was like almost all of the
[03:32] major talent in AI was under one corporate roof. Now you have at least three major competitors on the on the frontier who are all vying for the the top minds. And I'm wondering, do you think that DeepMind today still has the right talent to to win the race to AGI? >> Yeah, I think there's a lot of talent movement between all the leading labs and we win our fair share of the of the top talent. But what I would say is that we
[04:03] have by far the biggest and broadest research bench of any of the labs out there, the leading labs out there. And you know, we continue to put out the absolute frontier work, whether that's you know, on the foundation models or these other models that eventually will feed into the foundation models like our Omni and Via models too. So, but it's a ferociously competitive market out there right now, probably the most ferociously competitive there's
[04:33] ever been in the tech industry. And you know, I think that was inevitable. When I look back on this, we started this back in 2010 where you know, I started DeepMind and nobody was working on AI. Definitely not in industry, but even in academia, it was you know, it was basically sort of thought to be career suicide. Like of course we know AI doesn't work. You know, we tried it in the 90s, places like MIT and it was a dead end. You know, and that was the prevailing view.
[05:03] But we just felt the small band of us felt that that actually with the right ideas and using learning systems, reinforcement learning and betting on neural networks that a lot of fast progress could be made. And we were right in the end, but it also meant that now the whole world in the last few years has woken up to the potential of AI that every important company in the world is going to get involved in it. >> Yeah. So, we're we're here at Cannes. We're at an advertising conference. There's a lot of people here who are
[05:33] extremely creative and I'm sure a lot of them, you know, even here in the audience are using your your video creation tools to create ads, to do other things in the in the creative arts. Um, what can you do with these tools now that you couldn't do a year ago? >> Well, I mean every I would say every month these tools and the underlying models are improving massively. And a year ago, I just think the biggest change I would say with our tools like
[06:04] with our new Omni model and things like Nano banana for images is the ability to kind of live edit what the output is of the generative models. So, I think that's become extremely useful for creators. So, you you know, if you it was part of the creative process obviously is you generate the first idea, the first concept, but you like some of it but not other parts of it. You don't want to have to regenerate the whole thing, which is where we were a year ago. You want to be able to describe in natural language ideally as you would to a designer like, "Okay,
[06:35] keep that part the same but change this to something else." And then iterate that maybe hundreds of times till you get to the to the final polished version that you want. So, I think that kind of fine-grain control is is been a big change over the last year as well as just general relentless quality improvements. Yeah. >> There I mean, there's also you know, even controversies within the ad industry where people try to figure out are you are they using AI at all? Is this 100% human created and you should disclose this, etc. Do you think that's
[07:06] a conversation that's sort of temporary because we haven't adjusted yet to how AI's going to change creativity or do you think that that's here to stay that there will always be that that conversation? >> Well, there's two different parts here. With for certain, we need to deal with misinformation and deep fakes. So, that is something we was cognizant we were cognizant of years ago when we first started down building these generative models 3 4 years ago. I we foresaw that we would be in a world where these systems would be really
[07:36] good. Obviously, that's what we're planning to do and they would eventually be almost as you know, photorealistic. And so therefore we would need a system, digital watermarking system, which we created called SynthID, that was robust, was sort of unhackable, and would be embedded imperceptibly in the image. So that, um, you know, anyone, or citizen, a journalist, or government could go and detect whether that image was generated by an AI or not. And all of the things that we've
[08:07] all of our models that generate anything from music to images to videos come with SynthID embedded in it. And we've also open-sourced it and given it to the rest of the industry to use. So a lot of our, uh, uh, industry colleagues have now adopted that standard. OpenAI, Nvidia, and many other big ones. So I hope eventually, I think that should become, uh, almost a, a regulation really of like if you're creating, uh, generative media, then it should
[08:37] come with provenance detection. And obviously that will also help with things like right holders and IP rights, too. So that can all be sort of, uh, connected together. Um, as to whether it should be disclosed if you use AI for a piece of the work or that you're doing, I'm not sure. I think that might be just an era we're in where, okay, we're using, you know, Photoshop or some other tool before, and now this is a more advanced tool, but it's just a tool for, uh, your own creativity. I just I'm not sure that needs to be disclosed in in in
[09:09] in the sense that you're talking about, other than you should know that the output was, uh, synth, you know, synthetically generated. >> Do you If you look at your career, it creativity is this throughline. I mean, you started out creating video games. You studied this as a neuroscientist, the nature of creativity in the brain. You're, you know, even AlphaFold, I think you could think of as a as a very creative approach to science, right? Um, now we have all these tools. There are people who would say, "Well, this is going to make us less creative. This is, a you know, we're now just asking a model to do what you slaved for years
[09:41] doing as you know, in your in your career. So, how do you think how do you think about it? How's it going to change creativity? >> So, definitely going to change it, but I think what I'm seeing is a two-fold change. One is it's democratizing some of the creative tools. So, more people can try this and try out their ideas for a relatively quickly and relatively easily. That's double-edged though, because it also produces a lot more things that maybe are not very creatively valuable. But, also it means a lot more people can
[10:12] break into those industries. I think there's a lower bar there's a sort of lower bar to entry. There's less gatekeeping. So, that will probably mean new creators, professional creators find a path wherever they are in the world through using these tools. And then the other thing I'm seeing is on the professional side and we work with many professional directors and amazing collaborators. And we try and talk to them to design our tools to help enhance and empower their creative process and and
[10:43] and help them with their creative process. I think it's going be incredible. They can sort of do 10x more things than they used to be able to do, try out more more stream ideas, you know, iterate faster through they all have way more ideas than they can ever produce in their lifetimes. And so, this allows these tools allow them to try out things in relatively inexpensive and quick ways. So, I think for the professional they're going to be able to iterate their way to much cooler
[11:13] things way more quickly. So, but just like with any new tool, if you use it in the wrong way, the internet's the same, computers are the same. If you use it in a lazy way, it kind of takes away from the creative process. But, if you use it in a in a innovative way, it should add to the creative process. And I think it's going to take a while for the creative industries to figure out the best way to use these things. Um and I talked to my game designer friends, uh you know, in the games industry, they're very excited about these tools, but I
[11:43] still I would say at least the games industry, which is the the creative industry I know best, that we're we're still yet to figure out the the the any deep ways, the deeper ways of using this. But it's very early, you know, so we're the games industry using it for obvious things like to create some assets and um uh some graphics and things like that. But um can it change the nature of games and introduce whole new genres of games? That's what I think could be possible. Like it was in the '90s when I was started out in the games industry when graphics and AI first came on the
[12:14] scene for computer games. Um and it allowed us to make whole new types of game genres. That's what I hope to see uh these new tools spark. >> Do you Do you think this, you know, there's a criticism that you know, these were trained on on the outputs of humans, right? Um should there be some sort of like auditability that lets people see, oh, this output used partially, you know, one of my creations and I should get compensated for that. Do you Do you think that should happen? >> Well, maybe uh a new economic model is
[12:45] needed and I think that the the tech industry and the creative industry together need to work together. And that's what happened with streaming, which changed, you know, music and things like YouTube and then content ID and and YouTube specifically and Spotify and other companies like that came up with new really robust business models. So, I think that that will probably be needed. But it's very difficult, as everyone in the creative industries know, like to specifically attribute this is 1% this and 5% this and 10%
[13:15] that. It's going to be difficult to kind of agree on objectively like what that is. Um and we as even as human creators, we are the the things we create are the output of all of the all of the experiences we've had and the things we've learned and exposed ourselves to you know other art forms and other creators and what they've created and then we we mix all of that with our own creativity to to generate the new things. So in
[13:45] some sense that's always been the creative process but we'll have to see if you know probably new business models will will be eventually be needed. >> I hear this a lot from people that you know I'm I'm okay with AI being used for science to cure disease for instance right the stuff you're doing in Isomorphic Labs. >> Yeah. >> Um but I I don't like the fact that it's it's recreating music or you know the work that a filmmaker might have done or you know maybe even an ad agency. Um but I wonder if there's a if there's a point to be made here about sort of
[14:15] cross-discipline capabilities especially when it comes to AI where you know on the one hand over here you might be working at Isomorphic on a virtual cell right which we have we can't do today but maybe one day we'll be able to actually see how a cell works right in in real time and over here where you're creating these world-class video models that may one day be able to sort of analyze that virtual cell and potentially help you know cure disease create new therapies. Do you sort of see that
[14:45] happening down the road? I mean help us understand. >> Yeah so the the whole thesis behind AGI as a term and our original goal at DeepMind from the very beginning was to create this general purpose intelligent system that could um learn from almost any input and then generate useful insights or spot useful patterns and then output that in almost any way. Right that's obviously how the human mind works and look at modern
[15:16] civilization that we've created with our hunter-gatherer brains is pretty unbelievable if you think about how that happened. And that's really what we know as a general intelligence. And that's what we try to focus on from the beginning of DeepMind and now the whole AI field of like systems that are general and that learn, rather than being hard-coded, hard-programmed with the answer, which is you know, it's funny to think back now, but that's what we that's what the AI field used to do for the first 50-60 years of of AI as a field.
[15:46] Things like Deep Blue, the chess program, and so on. And so what that means is is that some of these things are inseparable. So if you want a fully general system that understands the world around you and can like analyze scientific papers or scientific data, including like visual data, like pictures of cells or, you know, proteins you know, little molecules even, depending on the resolution of the of the imaging equipment, it's the same type of things a capability that you need to analyze, you
[16:18] know, YouTube videos or or just the general vision that's coming through a camera. So a lot of these capabilities are general purpose and you develop them for one thing, but really that's just a means to an end for another thing. And you can see that with first first sort of 5-6-7 years of DeepMind, we were working on games and AI being good at games, like like Go and and Atari games. And obviously one reason I picked games is
[16:48] cuz I love games and I make games and I've always been involved in games. But the the real reason was is because they were challenging tasks that were the right level for the AI systems at the time. So they were never an end in themselves. They were a means to an end, which was to help give us quantifiable and achievable intermediate goals that were impressive and very hard to do, but just about within the realms of possibility. And that would then we trusted that would be a ladder to get us
[17:18] to a research ladder that would get us to where we are today, which is having these systems that can then eventually do really amazing things in the in the real world and a tackle real world problems uh like scientific problems like a protein folding with AlphaFold and now drug discovery. And that's personally what I spend my time using these AI systems for is AI for science. That's my my always been my main passion and the main reason why I'm building these AI tools. But of course, there are many
[17:48] other incredible things those same that same underlying platform can be used for including generative media models that's helpful for creativity and also productivity tools of course like um the large language models. >> It's fascinating how this is all connected and I I think back to your your first paper as a neuroscientist, which is now famous. It In 2007, it tied the the hippocampus in the brain to creativity. And it was this study of people who had lost their memory and it found that people who had
[18:18] had damage to the hippocampus and couldn't remember remember things also couldn't picture things picture the future. And this visual aspect to creativity is so important. Like even even fMRI studies on people who were blind from birth find that they access the visual part of their brain. So, I'm just wondering like is that Do you think that as you sort of try to recreate a machine hippocampus so to speak >> Yeah. >> um in AI that would let AI be creative. Do you think that actually could come from you know, just like training these
[18:49] models on the creative industries? >> Yeah, so the as you mentioned that the through line between that was so I first of all used my own visual creativity to help design and program these video games very early in my career. And I I And then when I went into neuroscience to do my PhD, I was fascinated to try and uncover the mechanisms that in our brain that allows us to do that that all of us use all the time is that at least the way I used to create for games was visualize the end goal visualize really
[19:19] viscerally visualize a player a kid playing the game and the using the interface and thinking through even before it was it was there programmed like what would be the issues with it and how would they find it fun and these types of things and I kind of mentally simulated that and we do that every to every day when we plan you know when we're going to have an important business dinner say later we imagine a lot of the times we sort of viscerally imagine like where's everyone going to sit what am I how am I going to open the conversation you know how are people going to feel about it so we're using
[19:50] this sort of imaginative or future thinking it's sometimes called capability in the brain and all the time and I had this suspicion when I first started my PhD normally I was supposed to be doing it on memory but but and and memory had been studied for a long time and it's memory is dependent on the hippocampus and there are these rare patients that have had a disease that's only just attacks the hippocampus unfortunately for them but leaves the rest of the brain intact and
[20:20] there's a few of them in in the UK we went to interview every single one of them and I had this idea that reading all the memory literature which is I did the sort of the first month of my PhD ED was that there was two schools of thought one is that it's a video tape and you just record everything that happens to you and another which obviously seemed wrong to me and there was another school of thought which is that it's a reconstructive process that actually memory when you remember something you're actively reconstructing it from its parts and that seemed much more
[20:50] obviously correct to me but if that was true then imagination should use the same brain mechanisms so it's just the goal is different instead of trying to recreate something that seems familiar to you and this time you're trying to create something from those component parts that looks novel that feels novel to you to your brain. And in in fact, that's what we discovered and we were the first people to amazingly to test any of patients on um their imaginative capabilities rather than just their memory. >> Do you see any similarities between
[21:22] what's happening under the hood with these video models recreating the world from prompts and and that and what's happening in the in the brain? >> I think definitely on a systems level. I don't know I don't think the implementations are similar one-to-one and that was never the idea behind doing neuroscience. It wasn't to copy the brain, but it was to understand the principles and the algorithms the brain might be using and the representations the brain was using. And and sort of lift that inspiration to then try and build that and into the direction of our
[21:52] AI models. So, I think there were some for sure some similarities between the way our new Omni models and video models are generating the world. Um and there's definitely some things to be learned about I think uh how that system is working. And quite a few neuroscience uh professor friends of mine um are actually indeed comparing the latest models about what they can do with a certain prompt with what uh uh a human might do in an fMRI machine and what images they might create. They're
[22:22] doing all sorts of crazy amazing things like decoding what image the person's thinking about or dreaming about and then using one of these models to recreate uh uh actually the visuals and then asking the the the subject in the scanner, is that what you were imagining? And it is. Um so, it's kind of we're we're going to have these sort of amazing kind of sci-fi devices, I think, in the next few years. >> So interesting. Do you you've talked about the Einstein test. I love this idea that you could sort of
[22:52] give an AI all the data that Einstein had and and nothing more. In other words, like the cutoff date is whatever it is, like really 1901. Um Um, and and then see if the AI model could do what Einstein did, which is to come up with the theory of relativity and that and all these other, you know, physics breakthroughs that we use today. So, I when you hear that though, you could almost imagine like text models, but are you are you really kind of imagining a a more visual process? >> Well, I think you first of all, that's what that's how I would define true
[23:22] creativity and that was my test for that is cuz people always ask, well, how do you define it where you're not just extrapolating something that already is known, but you actually coming up with a new hypothesis, a new scientific hypothesis about some part of reality that is genuinely novel. Uh, like Einstein most famously did in 1905 with his incredible set of experiments and papers. Um, so, it's interesting that you it could be that there's enough in language, um, that that's enough to come up with some
[23:52] new theory that was sort of hidden, cross-connected with all of the text, if you could read it all and hold it all in mind. But, Einstein himself used to daydream while he was a patent clerk in Switzerland and this is well documented and dream of these sort of thought experiments of, you know, being on trains and then if you were speed of traveling at the speed of light, what would it look like? And so, he was doing he was using his visual imaginative apparatus to come up with these new theories that then he had to prove mathematically. Um, but I think you're
[24:24] going to need to access and at least understand and certainly if you're going to propose new experiments or have to do new experiments in order to test your hypothesis to develop it further. Uh, so, it's not just a a theory, then I think you're going to need to have an understanding of the world of atoms, not just the world of bits or the world of logic. >> Yeah. >> You that, you know, in your video game days, um, you you had this game the the Republic the Revolution at Elixir, which is the company you founded. Game was actually kind of a failure, but
[24:54] it was this idea it was it was too ambitious. You were trying to recreate this former Soviet republic and simulate the world in a sense, right? >> Yeah. >> It's amazing cuz I think >> Pentium 2003, it was probably a little bit too ambitious. >> Right. Right. There was not an H100 >> of my time, yeah. >> Right. Right. Now Now you have access to, you know, I tens, hundreds of thousands of of GPUs at your fingertips and you're building these virtual worlds again. You're You know, it's it's funny how it kind of comes full circle.
[25:25] >> Mhm. >> Um but I Do you imagine that this is where inside these virtual worlds that like obviously they're going to be useful for robotics, right? Like training these things to to go upstairs or whatever. But like you know, Einstein in this virtual world, a virtual Einstein wandering around at the the patent office and daydreaming. Like is that where it's going to happen? >> Well, the the reason is look with with Republic we tried to simulate a whole country with hundreds of thousands living breathing people sort of going about their daily lives and all the
[25:56] politics of it and you were supposed to like create a revolution. So it was it was a pretty ambitious game that we had to write by hand. And what's amazing now and it took years. But it it was amazing now is that we might be close to the possibility of actually just sort of generating that with some of our systems. I don't think we're ready yet, but in two three years time that might be something close to that might be possible. And then we could realize maybe the full vision of what I was thinking about then. But the reason this is all
[26:26] important and why there's a why it's all connected is simulations and AI I think are fundamental and they're very closely related. So simulations are going to be useful basically that's what imagination is. It's a type of simulation. The reason simulations are useful is it allows you to try out many things in theory and then select the best path. Right? That's what AlphaGo did when it tries to select the best Go move in the current position. It simulates tens of thousands of moves
[26:57] using Monte Carlo tree search in this case, uses the model of Go to to kind of constrain that to only the useful paths, and then it evaluates at the end of those 20-30 moves, which one of those end positions is the most promising, and then that's what guides its next move and allows it to beat the world champion. But, there are many areas of the world robotics and and assistance are just two things in science even, but there are many areas of the world where we would love to, in my opinion, kind of be able to have many reruns of that
[27:28] problem. I'll give you an example, economics. Right, so economics is, you know, it would be great if it was more like a natural science. So, at the moment we we like put up and down interest rates by half a percent and then we see, did it cause a recession? Oh, whoops, we maybe we shouldn't have done that. But, it would be much better if we could simulate hundreds of thousands of trajectories and here the economy, what it would do if you adjusted these big levers, and then get some kind of statistical uh aggregate of all of those accurate
[28:00] simulations, and then make an informed, much more rigorous scientific decision about what to do. But, that's not possible in a lot of the social sciences because you can't rerun the experiment in a controlled way hundreds of times like you can with the natural sciences. So, I think simulations which are learned and and and then where AI's connected is you can you can hand code a simulation if you understand the underlying system well enough, but most of the time the things we want to simulate, whether they're it's like
[28:30] weather models like we have these the world's best weather models, or it's economics model, we don't actually understand well enough how the system really works on a mathematical level. So, then an AI system could learn it that that simulation from the data. So, that's sort of the bigger goal, if you like, of what I'm trying to do. >> I love how you've tied the the creativity part of AI with the science, and I'm sure you're giving this audience like a lot of inspiration in their in their work. So, thank you so much, Dennis. Great to talk with you.
[29:00] >> Great to talk. Appreciate it. Thank you. Yes, nothing.