Hassabis：这场 AI 变革比工业革命更大
Hassabis on an AI Shift Bigger Than Industrial Age
So good to see you. Last time I saw you was in San Francisco, so I know you are all over the place. >> Um, >> I'm curious if this year feels different than the last time you were in Davos. Gemini 3 is out. We've heard that OpenAI called a go a code red internally. Do you feel like Google got its mojo back? >> Well, I'm not sure if that's for me to say, but I I feel like we had a very good year. So, it's been hard work, really hard work getting uh our technology and the models sort of to back to state-of-the-art.
I think we did that with Gemini 3 especially and Nanabun on our in imaging software. Um, and then I think we've also sort of adapted really to this new world of shipping very fast, kind of bringing a kind of startup energy to what we do. And >> do you think people underestimated Google or got something wrong? >> Yes, maybe. I'm not sure. Well, I mean, I think we always had the ingredients, you know, to to to be at the forefront of this.
Obviously, um, we got the long history in it. I think, you know, over the last decade, uh, Google and Deep Mind between us, we've invented most of the breakthroughs that the modern AI industry relies on now. Obviously, transformers, most famously, but alpha go, deep reinforcement learning, these things. Um, and we have these incredible product surfaces, billions of user services that are natural fits for AI actually from search to email to uh to to Chrome and um, but it's just getting all of that together uh, and uh, organized in the right way and I think we've done that in the last couple of years and there's still a lot more work to do but I think we're starting to see the fruits of that.
>> If you think you have an advantage, how big do you think your advantage is? How long does it last? Well, I think everything starts with research in my opinion and techn you know the models especially being uh you know state-of-the-art on all the different benchmarks and that's what we focused on first when we put Google and Deep Mind together. Um and I think with the Gemini series we're very happy with how that's going.
Um there's a lot more work to do there, but I think we're the only organization that has the full stack um from the TPUs and the and the hardware, the data centers, the cloud business, the frontier lab and all of these amazing products that can, you know, kind of natural fits for AI. So really structurally um from first principles, we should be doing very well. And I think um there's actually a lot more headroom to come from here.
>> I wonder what a day in the life of an a frontier model leading AI CEO is like. Like I, you know, I've read that you do most of your thinking between 1 and 4 a.m. >> Yes, that's true. >> Um, is it ever not a code red inside people? Like, do you ever feel comfortable? >> No, you never feel comfortable. I mean, we try, you know, code reds are for very special circumstances, but it's always I mean, for the last, I would say three, four years, it's been unbelievably intense.
um and you know 100 hour weeks uh 50 you know 50 weeks a year and that's the norm and I think that's what you have to do uh at the forefront of this unbelievably fastmoving technology it's ferociously competitive out there um maybe the most intense competition there's ever been in technology and the stakes are incredibly high um AGI and you know all of that that that can mean so commercially scientifically um and then if you add all the excitement of what we're doing and you know using it as my passion as you know is exploring scientific uh you know problems with AI accelerating scientific discovery itself um this is what I've always dreamed about and I've worked my whole life on AI towards this moment so it's sort of hard to sleep because there's so much work to do but also it's it's there's so much exciting things to to look into and to to push forward you >> I mean I know you're very focused on AI driving scientific progress discovering new materials Even we've seen now Gemini being integrated into humanoid robots.
>> Is the AlphaFold moment for the physical world and AI here? What is that and what does that look like? >> Yeah, I spent a lot of the last year actually looking very carefully into robotics. I do think we're on the cusp of uh a kind of breakthrough moment in physical intelligence. I still think we're about 18 months two years away from doing we need to do more research but um I think the foundation models like Gemini show the way forward.
I mean from the beginning we made Gemini multimodal so it could understand the physical world for multiple reasons. One was so we could build a universal assistant maybe that exists on your glasses or your phone that understands the world around you. But of course a second use of that would be for robotics. >> So what does the that moment for the physical world look like? I think it's it's having robots uh you know reliably do useful tasks um out in the world and I think there's a few things holding that back still.
Part of it is the algorithms are still not quite there. They they need a little bit more robustness. Um they have to work with uh less data uh than you get for the LLMs or the models that work on just digital realm. You can sort of create synthetic data. It's a lot harder to make that kind of data and there's still sort of some problems in the hardware that are not solved. specifically things like the arm and the hand.
Actually, when you look into robotics very carefully, you get a newfound appreciation, at least I did, for the human hand and how exquisite uh evolution has designed that. It's incredible. And it's very hard to match the reliability, the strength um and the dexterity that the human hand has. So, there's still quite a lot I in my opinion pieces to put together. Um but there's very exciting things. I mean, we just announced a new deep collaboration with the Boston Dynamics.
They've got some very exciting robots in Hyundai and actually applying it to automotive manufacturing and we'll see over the next year how that goes in sort of prototype form and then maybe in a year or two we'll have some really impressive demonstrations that we can scale up. >> A year ago deep sea seemed cataclysmic for the west. Um now a year later it's it's quiet. China seems to have been quieter. >> Yes.
>> Has your opinion on competition from China changed? >> Not really. Uh I mean I didn't think it was cataclysmic in the first place. I think it was a massive overreaction in the west. It was impressive and I think it shows that uh the Chinese are very capable that the leading companies um I think companies like by dance actually I would say are the most capable um and they're maybe only six months behind not one or two years behind the frontier.
So I think that's what deepseek showed. Some of the claims were overexaggerated about, you know, the the amount of compute they used and being so minimal and so on because they relied on some western models and also fine-tuning on the outputs of some of the leading western models. So it wasn't sort of denovo and um the other thing I think so far is not you know yet to be seen is can China actually uh the Chinese companies innovate beyond the frontier themselves.
So they're they you know they're gaining they're very good at kind of catching up uh to where the frontier is and increasingly capable at that but um I think they've yet to show they could innovate beyond the frontier. >> You helped define AGI. >> You have said we have a 50% chance of getting there by 2030. >> Is that still your timeline? >> It is. >> And is AGI still a useful target for you?
>> I think so. I think it's a very that is still my timeline. Um, and it's a little bit longer than some others that you some of my peers that you're here, but my bar is quite high. It's it's the it's the ability to, you know, a system that exhibits all the cognitive capabilities humans have. And I think we're still clearly quite far from that. Um, that means, you know, in things in like scientific creativity, not just solving a conjecture or solving a problem in science, but actually coming up with the hypothesis or the problem in the first place.
And as any scientist knows, uh, finding the right question is actually often way harder than finding the answer. So, um, I don't, you know, it's not clear that that these systems have that capability yet. In fact, they definitely don't right now. I think they will eventually, but it's not clear what's needed still. And there's things like continual learning, you know, online learning going beyond uh, what they've been sort of trained for and then they're static out in the world.
They need to learn on the fly. So there's quite a few in my view missing capabilities that are quite critical to uh what I would regard as an AGI system. >> Uh Google's a major investor in anthropic. Daario was here earlier today. >> Do you agree or disagree with his prediction that AI will wipe away 50% of entrylevel white collar jobs in 5 years? >> Um I think it's that's also my timelines and and my view on that would be a lot longer.
I mean, I think we're starting to see maybe the beginnings of that this year in terms of u maybe entry level jobs or internships, those types of things. Um, but I think there's uh we would have to solve a lot more of this consistency that AI doesn't have right now. I call it jagged intelligences. We're very good at certain things and it's very poor the current systems other things. And if you want to offload or delegate an entire task to say an agent um rather than having what we have today which are more like assisted programs, you're going to need a lot more consistency across the board.
It's no good for it to be good at 95% of that task. You need it to be good at the whole task um for you to be able to actually just sort of fire and forget on it. Um so I think there's there's there's still quite a lot more to be done before we'll see that kind of disruption. >> But that kind of disruption will happen. I think eventually sure I mean in the limit with AGI uh you know if you have systems like that I think that changes the whole economy actually it's way beyond the question of jobs I think potentially if we build it right we we're in a post scarcity world where we've solved some of the kind of fundamental root nodes of the world like energy sources um uh new clean renewable basically free energy sources if we solve fusion something like that uh with the help of AI new materials I think we'll be you know 5 10 years past AGI will be in a in a radically abundant world.
And so what does that mean with you know to the economy and and and how how society works? Actually >> before we get to a post scarcity world though if we get there there is so much anxiety about what happens in between you know I'm a mom. I know you have kids like >> what scares you most for them? What do you talk to them about? What do you tell them that's coming? I mean, we, you know, I've just heard so many, oh my gosh, for college graduates are going to have such a hard time.
>> Well, I don't know about that. Look, I think it's it's it's it's going to be an age of disruption just like the industrial revolution was. Maybe 10x of that, which is kind of unbelievable to think about, and 10 times faster. So, I usually describe it, it's going to be 10 times bigger and 10 times faster than the industrial revolution. 100x, so 100x of it. Um, no, I say this to everyone, but I think um that comes with it huge opportunities and and I also am just a big believer in human ingenuity.
We're extremely adaptable because our minds are so general. You know, the human mind is very general. We've adapted. Look at the modern world around us. We, you know, our huntergatherer minds have managed to build modern civilization. So, I think we'll adapt again. I think it's a little bit unprecedented because of the speed of it. Usually, it takes one generation or two generations for a transformation like this to happen.
and the magnitude of the the transformative power of this technology. But I I think um the kids today, you know, I'd be encouraging them to get incredibly proficient with these new tools and and native with them. And they're almost equivalent of giving uh them superpowers, you know, in the creative arts that you could probably do the job of, you know, what would have taken 10 people uh in in one. And I think that means um you know if you're entrepreneurial, if you're creative with game design, films, projects, you can probably get a lot more done and break into those industries a lot more easily than you could in in the past as you know a new upcoming talent.
>> Some folks have advocated for a pause to give regulation time to catch up to give society time to sort of adjust to some of these changes. In a perfect world, if you knew that >> every other company would would pause, if every country would pause, would you advocate for that? >> I think so. I mean, I've been on record saying what I'd like to see happen. This was always my dream of the the kind of the road map at least I had when I started out deep mind 15 years ago and started working on AI, you know, 25 years ago now, was that as we got close to this moment, this threshold moment of AGI arriving.
Um, uh, we would maybe collaborate, uh, you know, in a scientific way. I I sometimes talk about setting up an international CERN equivalent for AI where all the best minds in the world would collaborate together uh uh and and and do the final steps in a very rigorous scientific way involving all of society, maybe philosophers and social scientists and economists as well as technologists um to kind of figure out what we want from this uh technology and how to utilize it in a in a in a way that benefits all of humanity.
And I think that's what's at stake. Um unfortunately it kind of needs international collaboration though because even if one company or even one nation or what even the west decided to do that um it has no use unless the whole world agrees at least on some kind of minimum standards and you know international cooperation's a little bit tricky at the moment. So um that's going to have to change if we want to to to to have that kind of um rigorous scientific approach to the final steps to AGI.
So if AGI comes in, let's say 2030, we don't have the regulations set up yet. Are we destined for something difficult? >> Well, then we're, you know, I I doesn't I think I'm still optimistic that enough of the leading players will uh uh uh uh kind of communicate together and hopefully collaborate at least on things like safety and security protocols. There's a lot of that already. We work, you know, quite closely with Anthropic, for example, on those things.
And um that would be needed then you know maybe kind of more uh uh peer-based cooperation if we can't get that international thing to work but I would you know that would be a lot more pressure >> Sam to cooperate with you. >> Um potentially I'm you know I think I'm on pretty good terms with pretty much all the other leaders of all the leading labs. I I I mean I think if the stakes are high enough and I think a lot of it is uh understanding what's at stake um and what the risks are and uh and I think that will become clearer to everyone in the next 2 three years.
>> So let's talk about the technology and the the next curve. Uh Yan Lun has said he doesn't think transformers and LLMs >> alone will get us to AGI. >> Do you agree or disagree? And and you know if they're dead ends then what are we doing? >> Yes. Uh, no. I I disagree that they're dead ends, but uh uh obviously I and I think that's clearly wrong. I mean, they're so uh amazingly useful already, but I think uh the way I say it's an empirical question.
I think it's a scientific question whether um they're going to be enough on their own. I think it's a 50/50 that you know just scaling up existing methods with some tweaks will be enough. It might be um and you have to do that and I think that's useful work because at a minimum the way I look at it is these LLMs will be a component a massively important component of the final AGI system. The only question in my mind is is it is it the only component right and I could imagine there are one or two breakthroughs maybe a small handful you know less than five that are still needed from here right yes and so you know and those might be things like world models that's something that Yan talked about we're working on that in fact we have the best world model currently which is genie our genie system and I work on that directly and I think it's very important but also things like continual learning and uh uh having consistent systems that are you don't have these jagged edges that they're good at and not good at.
That's a general system shouldn't have that. So, I think um you know, better reasoning, more long-term planning. There's quite a few uh capabilities that are still missing. And it's an open question whether a new uh architecture or new breakthrough is needed or more of the same. And we're just from my point of view, from Google Deep Mind's point of view, we're pushing as hard as possible on both those things. Both inventing new things and scaling up existing things.
So, uh, slightly different but related. Uh, Ilas Suscover said, "The era of scaling and making bigger models make improvements is is nearly over. Is that something you agree with?" >> No, I don't agree with I think his exact quote was this. We're back to the age of research, but but and you know, I love Ilia and we we're very good friends and we we agree on a lot of things, but my view is we never left the age of research.
At least from the point of view of Deep Mind, we've always invested. We've in my view, we've always had the the deepest and broadest bench. Um actually Google and deep mind together uh we if you look at over the last decade we've invented about 90% of the breakthroughs that you know the modern industry relies on of course transformers most famously uh but also you know deep reinforcement learning alpha go these kinds of reinforcement learning techniques we pioneered all of that so if some new breakthroughs are required in the future I would um back us to be the ones just like in the past to be the ones to make those breakthroughs uh you know in the future Uh last agree or disagree.
Elon says we have entered the singularity. >> No, I don't I think that's very premature. Um you know I I think the singularity is is is another word for you know a full AGI arriving and I think I explained earlier why I think that we're still uh you know nowhere near that. Um I think we will get there. Um and you know 5 years even 5 years is not a long amount of time if you think about what that is but I think there's still a lot of work to be done before we have anything that looks like the singularity.
So talk to us a little bit about the culture inside Google right now, you know, to, you know, win this race but do it right. You know, the leadership, how involved are Larry and Sergey >> right now? How often do you talk to them and what are their priorities? >> Yeah, they're very involved and and you know, Larry more on the strategic side, you know, see him at board meetings and other times when I visit the valley.
Sergey is more hands-on on the, you know, he's involved in the coding of, you know, on the Gemini team specifically more in the in the in the algorithmic uh details. Uh, and it's great having them both energized around where we are. And who wouldn't be, you know, at this moment, absolutely incredible moment for computer science. So, just from a scientific point of view, which both of them are, it's an incredibly exciting uh moment in in history, human history really.
And so, of course, everyone wants to be uh hands-on and heavily involved in that. So that's great and uh and and and for us just as as a as an entity, you know, I'm trying to combine the best of many worlds. So, you know, startup energy of shipping things fast and taking risks and doing things like that, which I think you're seeing the the benefits of, you know, the big company uh resources is amazing, useful, but also protecting the space for long-term research still and exploratory research, not just uh researching what we'll deliver in, you know, three months in a product.
I think that would be a mistake, too. So, I'm trying to balance all of those different factors together. And, you know, I think in the last year, uh, things have been going well, and I think we can still do better, and I think we still will do better this year. Um, but I I'm very happy with our trajectory. I think it's the steepest trajectory of improvement and progress of anyone in the industry. >> You are a Nobel laureate and I know how obsessed you are with, you know, AI powering scientific research.
If AI itself, let's say, makes a Nobel worthy discovery, who should get the prize? The AI or the human? >> Um, I think still the human, I would say, because I feel like I mean it depends what you mean by completely on its own, right? For so for now, these are still tools and I view them as, you know, like very uh maybe the ultimate scientific tool, but sort of better versions of telescopes and microscopes. We've always built tools so that we can investigate the natural world better.
uh we're tool making animals basically that's what distinguishes humans from from the other animals and that's our superpower and I include computers in that of course and AI being the ultimate expression of that so in some ways I think of AI and I've always thought of it as the ultimate tool to do science and um and I think for the foreseeable future that's going to be a collaboration with top scientists uh putting in the creative ideas and maybe the hypotheses with these amazing tools that enhance you know data processing and pattern matching and the exploration part of science.
>> You you obviously could have sold Deep Mind to anyone. >> Um, and I think all of these companies are asking a lot of us to trust >> Yes. >> you, especially if regulation doesn't keep up with technology, which history shows it probably won't. Um why why should we trust you and why do you think Google which I would implicitly think you believe is is the mo the place that we should believe in >> the most when it comes to something that feels so risky.
>> Yes. I think you you need to judge these companies by their actions and also look into you know the motivations I would say of the leaders involved in those you know those endeavors and for me and for us and that's one reason I picked Google as the right home for deep mind is several reasons the the main one being that the founders of Google and the way Google was set up by them was as a scientific company you know many people forget Google itself was a PhD project right it was Larry and Sergey's PhD project so I felt an immediate affinity with with them um and and Larry led the acquisition but also the board who they collected on the board you know you have John Hennessy is the chair who's a chewing award winner himself and Francis Arnold another Nobel Prize winner you know these are unusual people to have on a corporate board so the whole environment is very scientific um and science science and researchled and engineering as a culture and it's deeply ingrained in the culture and that means doing science at the highest level means being really rigorous being thoughtful and applying the scientific method everywhere you can and I think that's not just to the technology but it's also to the way you operate as an organization.
So I feel like um you know we we we're very uh we try to be very thoughtful and responsible and have as much foresight as possible over the technologies we put out in the world. Doesn't mean we'll get everything right because it's so complex and so new and nent and so transformative this technology but we hope to coast correct as quickly as possible if there you know something does go wrong. Um and then the final thing I would say is just the um I was just attracted by the the types of things Google tries to do in the world.
You know, organizing the world's information I think is a very noble uh uh goal which is obviously Google's uh mission statement. And I think it fitted very well with that with DeepMind's mission statement of you know solving intelligence and using it to to solve everything else. And and that was our and I think those two mission statements are natural fit. AI and organizing the world's information naturally go together.
And um I think those are the types of products, the types of products that Google's well known for from maps to to to to Gmail to obviously search. I think they're genuinely useful products in the world. And I think AI is an easy fit how that would work with those products to enhance them for everybody to use in their everyday lives. And I think that's a great thing for the world. So I'm, you know, I'm happy to be contributing to that.
>> Okay. So post scarcity world, we're there. people no longer have jobs. What do you personally plan to do with your time once you have achieved all your technical goals? The research is just automating itself, >> right? Um well, I would love to use it for what I will do post the singularity is to use it for exploring the the the limits of physics. I think that was my my favorite subject at school was the big questions.
You know, what what is the fabric of reality? What's the nature of reality? What about the nature of consciousness? um the answer to the Fermy paradox, all of these things. What is time? What is gravity? I'm amazed more people, you know, we just go around our daily lives, you know, not really thinking about these massive questions that for me are always kind of almost screaming at me for like like what is the answer to these things, these deep mysteries and I would like to use AI uh and to explore all of those things.
Maybe traveling to the stars as well with the help of uh you know and new energy sources and materials and other things that's unlocked by AI. Will we all have meaning and purpose if we don't have work? >> Well, I think that's the to be honest with you, that's the thing I worry more about than the economics. I think the economics is a almost a political question of like when we get all of these extra benefits and productivity, can we make sure that it's shared uh uh for the benefit of everyone and I think obviously that's what I believe in.
But then the bigger question than that is what about purpose and meaning that a lot of us get from our jobs and scientific endeavors. Um how will we find that in the new world? Um, and I think, you know, we all need some new great philosophers in my opinion to to help with that, uh, and thinking that through. Um, maybe we'll, you know, be get much more sophisticated with our art and, um, and and exploration that we do and things like, you know, extreme sports.
There's many things we do today that aren't just for economic gain and perhaps we'll have very esoteric versions of those things in the future. >> All right. So, everyone in the room is wondering what they should be doing. Like, what should I do about AI? What do I do? sitting here in Davos in 10 years, >> what is the biggest mistake do you think people in this room will have made about AI? >> Well, look, I I think there's two things I would say.
One is uh for the younger generation and our kids and so on is the only thing we're certain of is there's going to be huge amount of change. So, I think in terms of learning skills, get ready to kind of learning to learn is the most important thing. How can quickly can you adapt to new situations, absorb new information using the tools that we have for the CEOs and and business folks in the room? I think the most important thing to do now is uh there are many providers of leading models and leading uh uh service providers and there'll be more for these AI models.
You know, pick the partners that you uh feel are approaching it in the right way. And so, you know, kind of partner with those that are making the changes and and approaching this technology in the way that you would like to see in the world.