Transcript: Demis Hassabis on AI, Creativity, and a Golden Age of Science – All-In Summit

And again, this is fully… Every pixel you can see is fully generated. And then you can type things like “person in a chicken suit” or “jet ski” and it will just in real time include them in the scene.

DAVID FRIEDBERG: So.

DEMIS HASSABIS: It is quite mind blowing really.

DAVID FRIEDBERG: I think what is hard to grok when looking at this, because we have all played video games that have a 3D element to them when you are in an immersive world. But there is no objects that have been created, there is no rendering engine. You are not using Unity or Unreal, which are the 3D rendering engines. This is actually just 2D images that are being rendered, created on the fly by the AI.

DEMIS HASSABIS: This model is reverse engineering intuitive physics. So, you know, it’s watched many millions of videos and YouTube videos and other things about the world. And just from that it’s kind of reverse engineered how a lot of the world works. It’s not perfect yet, but it can generate a consistent minute or two of interaction as you as the user in many, many different worlds. There are some videos later on where you can control, you know, a dog on a beach or a jellyfish or that’s not limited to just human things.

DAVID FRIEDBERG: Because the way a 3D rendering engine works is you type in the programmer programs all the laws of physics. How does light reflect off of an object? You create a 3D object, light reflects off. And then… So what I see visually is rendered by the software because it’s got all the programming on how to create physics, how to do physics. But this was just trained off of video and it figured it all out.

DEMIS HASSABIS: Yeah, it was trained off of video and some synthetic data from game engines and it’s just reverse engineered it. And for me, it’s very close to my heart, this project. But it’s also quite mind blowing because in the 90s in my early career, I used to write video games and AI for video games and graphics engines.

And I remember how hard it was to do this by hand, program all the polygons and the physics engines and it’s amazing to just see this, do it effortlessly. All of the reflections on the water and the way materials flow and objects behave and it’s just doing that all out of the box, I think.

The Path to General AI and Robotics

DAVID FRIEDBERG: It’s hard to describe how much complexity was solved for with that model. It’s really, really, really mind blowing. Where does this lead us? So fast forward, this model is gen 5.

DEMIS HASSABIS: Yeah. So the reason we’re building these kind of models is we feel, and we’ve always felt obviously progressing on the normal language models, like with our Gemini model. But from the beginning with Gemini, we wanted it to be multimodal. So we wanted it to input and take any kind of input, images, audio, video, and it can output anything.

And so we’ve been very interested in this because for an AI to be truly general, to build AGI, we feel that the AGI system needs to understand the world around us and the physical world around us, not just the abstract world of languages or mathematics. And of course, that’s what’s critical for robotics to work. It’s probably what’s missing from it today.

And also things like smart glasses, a smart glass assistant that helps you in your everyday life. It’s got to understand the physical context that you’re in and how the world, the intuitive physics of the world works. So we think that building these types of models, these GENIE models, and also Veo the best text to video models, those are expressions of us building world models that understand the dynamics of the world, the physics of the world, if you can generate it, then that’s an expression of your system understanding those dynamics.

Vision, Language, and Action Models

DAVID FRIEDBERG: And that leads to a world of robotics, ultimately one aspect, one application. But maybe we can talk about that. What is the state of the art with the vision, language, action models today? So a generalized system, a box, a machine that can observe the world with a camera. And then I can use language, I can use text or speech to tell it I want you to do it. And then it knows how to act physically, to do something in the physical world.

DEMIS HASSABIS: Yeah, that’s right. So if you look at our Gemini Live version of Gemini, where you can hold up your phone to the world around you, I’d recommend any of you try it. It’s kind of magical what it already understands about the physical world. You can think of the next step as incorporating that in some sort of more handy device like glasses. And then it will be an everyday assistant. It will be able to recommend things to you as you’re walking the streets. Or we can embed it into Google Maps.

And then with robotics, we’ve built something called Gemini robotics models which are sort of fine tuned Gemini with extra robotics data. And what’s really cool about that is, and we released some demos of this over the summer was you can have, we’ve got these tabletop setups of two hands interacting with objects on a table, two robotic hands.

And you could just talk to the robot. So you can say “put the yellow object into the red bucket” or whatever it is, and it will interpret that instruction, that language instruction into motor movements. And that’s the power of a multimodal model, rather than just a robotic specific model, is that it will be able to bring in real world understanding to the way you interact with it. So in the end it will be the UI UX that you need, as well as the understanding the robotic, the robots need to navigate the world safely.

The Android for Robotics Vision

DAVID FRIEDBERG: I asked Sundar this does that mean that ultimately you could build what would be the equivalent of call it either a Unix, like an operating system layer or like an Android for generalized robotics, at which point if it works well enough across enough devices, there will be a proliferation of robotics devices and companies and products that will suddenly take off in the world. Because this software exists to do this generally.

DEMIS HASSABIS: Exactly. That’s certainly one strategy we’re pursuing is a kind of Android play, if you like, as a kind of robotics, almost an OS layer, cross robotics. But there’s also some quite interesting things about vertically integrating our latest models with specific robot types and robot designs and some kind of end to end learning of that too. So both are actually pretty interesting and we’re pursuing both strategies.

DAVID FRIEDBERG: Do you think that there’s humanoid robots as a good kind of form factor? Is that, does that make sense in the world? Some folks have criticized it as being good for humans because we’re meant to do lots of different things. But if we want to solve a problem, there may be a different form factor to fold laundry or do dishes or clean the house or whatever.

The Future of Robotics: Form Factor and Scale

DEMIS HASSABIS: Yeah, I think there’s going to be a place for both. So actually I used to be of the opinion maybe five, ten years ago that we’ll have form specific robots for certain tasks. And I think in industry, industrial robots we will definitely be like that, where you can optimize the robot for the specific task, whether it’s a laboratory or a production line. You’d want quite different types of robots.

On the other hand, for general use or personal use, robotics and just interacting with the ordinary world. The humanoid form factor could be pretty important because of course we’ve designed the physical world around us to be for humans and so steps, doorways, all the things that we’ve designed for ourselves rather than changing all of those. In the real world, it might be easier to design the form factor to work seamlessly with the way we’ve already designed the world. So I think there’s an argument to be made that the humanoid form factor could be very important for those types of tasks. But I think there is a place also for specialized robotic forms.

DAVID FRIEDBERG: Do you have a view on hundreds of millions, millions, thousands over the next five years, seven years? I mean, do you have a vision?

DEMIS HASSABIS: Yeah, I do and I spend quite a lot of time on this and I think we’re still, I feel we’re still a little bit early on robotics. I think in the next couple of years there’ll be a sort of real wow moment with robotics. But I think the algorithms need a bit more development. The general purpose models that these robotics models are built on still need to be better and more reliable and better understanding the world around it. And I think that will come in the next couple of years.

And then also on the hardware side, the key is I think eventually we will have millions of robots helping society and increasing productivity. But the key there is, when you talk to hardware experts is at what point do you have the right level of hardware to go for the scaling option? Because effectively when you start building factories around trying to make tens of thousands, hundreds of thousands of particular robot type, it’s harder for you to update quickly iterate the robot design. So it’s one of those kind of questions where if you call it too early, then the next generation of robot might be invented in six months time. That’s just more reliable and better and more dexterous.

DAVID FRIEDBERG: Sounds like using a computing analogy, we’re kind of in the 70s era, PC.

DEMIS HASSABIS: DOS, kind of, yeah, potentially. But of course I think maybe that’s where we are. But I think. Except that 10 years happens in one year probably, right?

DAVID FRIEDBERG: Four might be one of those years.

DEMIS HASSABIS: Right? Exactly, yeah.

AI’s Role in Scientific Discovery

DAVID FRIEDBERG: So let’s talk about other applications, particularly in science. True to your heart as a scientist, as the Nobel prize winning scientist, I always felt like the greatest things that we would be able to do with AI would be the problems that are intractable to humans with our current technology and capabilities and our brains and whatnot, and we can unlock all of this potential. What are the areas of science and breakthroughs in science that you’re most excited about? And what kinds of models do we use to get there?

DEMIS HASSABIS: Yeah, I mean, AI to accelerate scientific discovery and help with things like human health is the reason I spent my whole career on AI. And I think it’s the most important thing we can do with AI. And I feel like if we build AGI in the right way, it will be the ultimate tool for science.

And I think we’ve been showing at DeepMind a lot of the way of that. Obviously AlphaFold most famously, but actually we’ve applied our AI systems to many branches of science, whether it’s material design, helping with controlling plasma and fusion reactors, predicting the weather, solving Mass Olympiad math problems, and the same types of systems with some extra fine tuning can basically solve a lot of these complex problems. So I think we’re just scratching the surface of what AI will be able to do, and there are some things that are missing.

So AI today, I would say, doesn’t have true creativity in the sense that it can’t come up with a new conjecture yet or new hypothesis. It can maybe prove something that you give it, but it’s not able to come up with a sort of new idea or new theory itself. So I think that would be one of the tests actually for AGI.

The Nature of Creativity in AI

DAVID FRIEDBERG: What is that? Creativity as a human?

DEMIS HASSABIS: Yeah.

DAVID FRIEDBERG: What is creativity, then?

DEMIS HASSABIS: Well, I think it’s this sort of intuitive leaps that we often celebrate with the best scientists in history and artists, of course. And, you know, maybe it’s done through analogy or analogical reasoning. There are many theories in psychology and neuroscience and as to how we as human scientists do it, but a good test for it would be something like, give one of these modern AI systems a knowledge cutoff of 1901 and see if it can come up with special relativity like Einstein did in 1905. Right. If it’s able to do that, then I think we’re onto something really, really important where perhaps we’re nearing an AGI.

Another example would be with our AlphaGo program that beat the world champion at go. Not only did it win back 10 years ago, it invented new strategies that had never been seen before for the game of Go, this is famously move 37 in game two that is now studied. But can an AI system come up with a game as elegant, as satisfying, as aesthetically beautiful as Go, not just a new strategy? And the answer to those things at the moment is no. So that’s one of the things I think that’s missing from a true general system, an AGI system is it should be able to do those kinds of things as well.

The Path to AGI: Missing Pieces

DAVID FRIEDBERG: Can you break down what’s missing and maybe relate it to the point of view shared by Dario, Sam, others about AGI is a few years away. Do you not subscribe to that belief and maybe help us understand what is it in your understanding of structure, in your understanding of the system architecture, what’s lacking?

DEMIS HASSABIS: Well, so I think the fundamental aspect of this is can we mimic these intuitive leaps rather than incremental advances that the best human scientists seem to be able to do? I always say, what separates a great scientist from a good scientist is they’re both technically very capable, of course, but the great scientist is more creative. And so maybe they’ll spot some pattern from another subject area that can sort of have an analogy or some sort of pattern matching to the area they’re trying to solve. And I think one day AI will be able to do this, but it doesn’t have the reasoning capabilities and some of the thinking capabilities that are going to be needed to make that kind of breakthrough.

I also think that we’re lacking consistency. So you often hear some of our competitors talk about, you know, these modern systems that we have today are PhD intelligences. I think that’s a nonsense. They’re not PhD intelligences. They have some capabilities that are PhD level, but they’re not in general capable. And that’s exactly what general intelligence should be of performing across the board at the PhD level. In fact, as we all know, interacting with today’s chat bots, if you pose the question in a certain way, they can make simple mistakes with even high school maths and simple counting. So that shouldn’t be possible for a true AGI system.

So I think that we are maybe, I would say sort of five to 10 years away from having an AGI system that’s capable of doing those things. Another thing that’s missing is continual learning, this ability to online teach the system something new or some, or adjust its behavior in some way. And so a lot of these, I think core capabilities are still missing and maybe scaling will get us there. But I feel if I was to bet, I think there are probably one or two missing breakthroughs that are still required and will come over the next five or so years.

Progress in AI Models: Convergence or Continued Growth?

DAVID FRIEDBERG: In the meantime, some of the reports and the scoring systems that are used seem to be demonstrating two things. One, perhaps, and tell me if we’re wrong on this, a convergence of performance of large language models and number two perhaps is a slowing down or a flatlining of improvements in performance on each generation. Are those two statements generally true or not so much?

DEMIS HASSABIS: No, I mean we’re not seeing that internally and we’re still seeing a huge rate of progress. But also we’re sort of looking at things more broadly. You see with our GENIE models and VEO models and nanobana.

DAVID FRIEDBERG: Nanobanana is insane. It’s bananas.

DEMIS HASSABIS: Yes, it’s bananas.

DAVID FRIEDBERG: Can I see who’s used it? Has anyone used nanobanana? It’s incredible, right? I mean I’m a nerd who used to use Adobe Photoshop as a kid and Kai’s power tools and I was telling you Brights 3D, so the graphic systems and recognizing what’s going on there was just mind blowing.

The Future of Creative Tools

DEMIS HASSABIS: Well, I think that’s the future of a lot of these creative tools is you’re just going to sort of vibe with it or just talk to them and it’ll be consistent enough where with nanobanana, what’s amazing about it is that it’s an image generator, it’s state of the art, best in class, but one of the things that makes it so great is its consistency. It’s able to instruction follow what you want changed and keep everything else the same and so you can iterate with it and eventually get the kind of output that you want. And that’s I think, what the future of a lot of these creative tools is going to be and sort of signals the direction and people love it and they love creating with it.

DAVID FRIEDBERG: So democratization of creativity I think is really powerful. I remember having to buy books on Adobe Photoshop as a kid and then you’d read them to learn how to remove something from an image and how to fill it in and feather and all this stuff. Now anyone can do it with nanobanana and they can explain to the software what they wanted to do and it just does it.

DEMIS HASSABIS: Yeah. I think you’re going to see two things which is the sort of democratization of these tools for everybody to just use and create with without having to learn incredibly complex UX’s and UIs like we had to do in the past. But on the other hand I think, and we’re also collaborating with filmmakers and top creators and artists, so they’re helping us design what these new tools should be. What features would they want?

People like the director Darren Aronofsky, who’s a good friend of mine, amazing director and he’s been making and his team making films using Veo and some of our other tools. And we’re learning a lot by observing them and collaborating them. And what we find is that it also superpowers and turbocharges the best professionals too, because they’re suddenly the best creatives, the professional creators. They’re suddenly able to be 10x100x more productive. They can just try out all sorts of ideas they have in mind, very low cost and then get to the beautiful thing that they wanted.

So I actually think it’s sort of both things are true. We’re democratizing it for everyday use for YouTube creators and so on. But on the other hand, at the high end, the people who understand these tools and it’s not everyone can get the same output out of these tools. There’s a skill in that as well as the vision and the storytelling and the narrative style of the top creatives. And I think it just allows them, they really enjoy using these tools. It allows them to iterate way faster.

Personalized Content Creation

DAVID FRIEDBERG: Do we get to a world where each individual describes what sort of content they’re interested in? “Play me music like Dave Matthews” and it’ll play some new track.

DEMIS HASSABIS: Yes.

DAVID FRIEDBERG: Or “I want to play a video game set, you know, in the movie Braveheart and I want to be in that movie.”

DEMIS HASSABIS: Yes.

DAVID FRIEDBERG: And I just have that experience. Do we end up there or do we still have a one to many creative process in society? How important culturally? And I know this is a little bit philosophical, but it’s interesting to me which is are we still going to have storytelling where we have one story that we all share because someone made it, or are we each going to start to develop and pull on our own kind of virtual.

The Future of Entertainment and Co-Creation

DEMIS HASSABIS: I actually foresee a world and I think a lot about this having started in the games industry as a game designer and programmer, is that in the 90s? I think the future of entertainment, this is what we’re seeing is the beginning of the future of entertainment. Maybe some new genre or new art form where there’s a bit of co-creation.

I still think that you’ll have the top creative visionaries. They will be creating these compelling experiences and dynamic storylines and they’ll be of higher quality even if they’re using the same tools than the everyday person can do. But also so millions of people will potentially dive into those worlds, but maybe they’ll also be able to create co-create certain parts of those worlds and perhaps that the main creative person is almost an editor of that world. So that’s the kind of things I’m foreseeing in the next few years. And I’d actually like to explore ourselves with technologies like Genie.

Isomorphic Labs: Revolutionizing Drug Discovery

DAVID FRIEDBERG: Right. Incredible. And how are you spending your time? Are you at… Maybe you can describe isomorphic what isomorphic is and are you spending a lot of your time there?

DEMIS HASSABIS: I am. So I also run Isomorphic, which is our spin out company to revolutionize drug discovery, building on our AlphaFold breakthrough in protein folding. And of course, knowing the structure of a protein is only one step in the drug discovery process.

So isomorphic, you can think of it as building many adjacent AlphaFolds to help with things like designing chemical compounds that don’t have any side effects but bind to the right place on the protein. And I think we could reduce down drug discovery from taking years, sometimes a decade to do, down to maybe weeks or even days over the next 10 years.

DAVID FRIEDBERG: That’s incredible. Do you think that’s in clinic soon or is that still in the discovery phase?

DEMIS HASSABIS: We’re building up the platform right now and we have great partnerships with Eli Lilly, I think you had the CEO speaking earlier, and Novartis, which are fantastic, and our own internal drug programs. And I think we’ll be entering sort of pre-clinical phase sometime next year.

DAVID FRIEDBERG: So candidates get handed over to the pharma company and they then take them forward.

DEMIS HASSABIS: That’s right. And we’re working on cancers and immunology and oncology and we’re working with places like MD Anderson.

Hybrid Models: Combining Probabilistic and Deterministic Approaches

DAVID FRIEDBERG: How much of this requires… And I just want to go back to your point about AGI as it relates to what you just said. Models can be probabilistic or deterministic. And tell me if I’m reducing this down too simplistically, that the model takes an input and it outputs something very specific, like it’s got a logical algorithm and it outputs the same thing every time.

And it could be probabilistic where it can change things and make selections. The probability is 80% I’ll select this letter, 90% I’ll select this letter next, et cetera. How much do we have to kind of develop deterministic models that sync up with, for example, the physics or the chemistry underlying the molecular interactions as you do your drug discovery modeling. How much are you building novel deterministic models that work with the models that are probabilistic, trained on data?

DEMIS HASSABIS: Yeah, it’s a great question actually for the moment and I think probably for the next five years or so we’re building what maybe you could call hybrid models. So AlphaFold itself is a hybrid model where you have the learning component, this probabilistic component you’re talking about, which is based on neural networks and transformers and things. And that’s learning from the data. You give it any data you have available.

But also in a lot of cases with biology and chemistry, there isn’t enough data to learn from. So you also have to build in some of the rules about chemistry and physics that you already know about. So for example, with AlphaFold, the angle of bonds between atoms and make sure that the AlphaFold understood you couldn’t have atoms overlapping with each other and things like that.

Now in theory it could learn that, but it would waste a lot of the learning capacity. So actually it’s better to kind of have that constraint as a constraint. Now the trick is with all hybrid systems, and AlphaGo was another hybrid system where there’s a neural network learning about the game of Go and what kind of patterns are good. And then we had Monte Carlo tree search on top, which was doing the planning. And so the trick is, how do you marry up a learning system with a more handcrafted system, bespoke system, and actually have them work well together? And that’s pretty tricky to do.

DAVID FRIEDBERG: Does that sort of architecture ultimately lead to the breakthroughs needed for AGI, do you think? Are there deterministic components that need to be solved?

DEMIS HASSABIS: I think ultimately what you want to do is when you figure out something where this one of these hybrid systems, what you ultimately want to do is upstream it into the learning component. So it’s always better if you can do end to end learning and directly predict the thing that you’re after from the data that you’re given.

So once you’ve figured out something using one of these hybrid systems, you then try and go back and reverse engineer what you’ve done and see if you can incorporate that learning, that information into the learning system. And this is sort of what we did with AlphaZero, the more general form of AlphaGo. So AlphaGo had some go specific knowledge in it, but then with AlphaZero we got rid of that, including the human data, human games that we learned from and actually just did self learning from scratch. And of course then it was able to learn any game, not just go.

Energy Efficiency and AI’s Future Impact

DAVID FRIEDBERG: A lot of hype and hoopla has been made about the demand for energy arising from AI. This is a big part of the AI summit we held in Washington D.C. a few weeks ago. And it seems to be the number one topic everyone talks about in tech nowadays. Where’s all this power going to come from?

But I ask the question of you. Are there changes in the architecture of the models or the hardware, or the relationship between the models and the hardware that brings down the energy per token of output or the cost per token of output that ultimately maybe say, mutes the energy demand curve that’s in front of us? Or do you not think that that’s the case and we’re still going to have a pretty kind of geometric energy demand curve?

DEMIS HASSABIS: Well, look, interestingly, again, I think both cases are true in the sense that especially us at Google and at DeepMind, we focus a lot on very efficient models that are powerful because we have our own internal use cases, of course, where we need to serve say, AI overviews to billions of users every day. And it has to be extremely efficient, extremely low latency and very cheap to serve.

And so we’ve kind of pioneered many techniques that allow us to do that, like distillation, where you sort of have a bigger model internally that trains the smaller model. Right. So you train the smaller model to mimic the bigger model. And over time, if you look at the progress of the last two years, the model efficiencies are like 10x, even 100x better for the same performance.

Now, the reason that isn’t reducing demand is, is because we’re still not got to AGI yet. So also the frontier models, you keep wanting to train and experiment with new ideas at larger and larger scale, whilst at the same time, at the serving side, things are getting more and more efficient. So both things are true.

And in the end, I think from the energy perspective, I think AI systems will give back a lot more to energy and climate change and these kind of things than they take in terms of efficiency of, of grid systems and electrical systems, material design, new types of properties, new energy sources. I think AI will help with all of that over the next 10 years. That will far outweigh the energy that it uses today.

A Vision for the Next Decade

DAVID FRIEDBERG: As the last question, describe the world 10 years from now.

DEMIS HASSABIS: Wow, okay. Well, I mean, you know, 10 years, even 10 weeks is a time in AI, so the Brownian field of 10 years. But I do think, feel like if we will have AGI in the next 10 years, you know, full AGI, and I think that will usher in a new golden era of science, so a kind of new renaissance. And I think we’ll see the benefits of that right across from energy to human health.

DAVID FRIEDBERG: Amazing. Please join me in thanking Nobel Laureate.

Related Posts

• Full Transcript: Zanny Minton Beddoes Interviews Elon Musk
• Transcript of Daniel Kokotajlo Interview: Diary Of A CEO Podcast
• Transcript: Why a Social Media Ban Won’t Save Teens – Candice Odgers  
• Transcript of Elon Musk Interview: SXSW Live – March 9, 2013
• Digital Afterlife: Who Owns Your Data When You’re Gone? – Suresh Sankaran Srinivasan (Transcript)