[0:00:00] I actually don't think they fully understand the monumentalness of what is being built. AI designed drugs and cures for really terrible diseases. We're only just a few years away from that. Energy becomes free or cheap, right? And then suddenly the nature of money even changes. We don't want to have to wait to the eve before AGI happens and go, you know what? Maybe we should have thought about this a bit harder. Okay? No, we should be preparing for that now. This week.
[0:00:33] Google DeepMind CEO Demis Hassabis on Google's newest AI breakthroughs, building artificial general intelligence, and what happens next in a world where computers can do every job. Personally, I take a nap. Well, Kevin, the gang over at Google is at it again. Yeah, what have they done this time? Well, from the minds that brought us Gemini Advanced and Gemini Pro and Gemini Nano comes Gemma.
[0:01:00] Have you seen Gemma? Yeah, I will admit that when I saw that Google had come out with a new AI model called Gemma, I did momentarily think they're trolling Casey Newton personally. They want to kill this man by introducing the most complicated set of product names imaginable. So are you alive and how are you recovering from the news of Gemma?
[0:01:20] Well, I'm coming to you now from the hospital where I was taken for observation following my attempt to process the latest Google AI models, but I'm going to be back on my feet soon. But look, we like to have some fun talking about the various models which we struggle to keep track of and understand what they're doing.
[0:01:41] But fortunately, there is actually a person within the Google organization, Kevin, who could explain this stuff to us. Yes, we took your complaint about Google's AI naming conventions all the way to the top. Today on the show, we are talking with Google DeepMind CEO, Demis Hassabis. Demis is the head of all of Google's AI programs. And he is really the person who has caused you so much pain and grief over the last few weeks with these names.
[0:02:09] Yeah, but we should say, Kevin, like Demis is a big deal in the world of AI. I would consider him one of sort of the four or five most important people in the entire field of AI. And so today, we're going to just devote the whole episode to talking with Demis. And I think we should just give a little bit of background for people about who he is and why he's so influential within the field of AI. Yeah, so what are things that Demis has done that make you say that he is such a leading figure in this world?
[0:02:37] So I would say that Demis has been present for most of the big moments in AI over the last 10 or 20 years. In 2010, he and his two co-founders started DeepMind, which was this sort of research lab based in the UK that was doing all kinds of research into things like reinforcement learning. And eventually they sold DeepMind to Google for $650 million in 2014.
[0:03:04] For much of the next decade, they kind of existed as this quasi-elite research team within Google's AI division. They did a bunch of stuff that you and I have talked about on the show, including AlphaGo, which was teaching an AI to play the ancient board game Go at a superhuman level. That was a very big deal when it came out in the world of AI. And then they came out with AlphaFold.
[0:03:30] Which basically took some of the same AI principles and techniques that they had used to teach an AI to play Go and used it to solve what is known as the protein folding problem. Basically this problem from biology where you needed to be able to predict the 3D structures of different kinds of proteins. This was something that had...
[0:03:50] Given biologists a lot of trouble for many decades, and DeepMind and their AlphaFold project were able to essentially solve this entire problem more or less overnight. Yeah, I mean, back in those days, when you would ask Kevin and I to predict a protein structure, we'd just give you a blank stare. We had no idea how to do it. And then along comes AlphaFold, all of a sudden, we're cooking.
[0:04:11] But anyways, you know, Demis has a fascinating backstory. And if you're curious about how he got into artificial general intelligence, our friend Ezra Klein did a great podcast on the subject with Demis last year. So check that out if you want to know more about Demis' past. What we want to talk to him about today was what he is working on now. Yeah, so last year, after ChatGPT had come out and Google had kind of like rushed to catch up, it had released this thing, BARD, which...
[0:04:40] It was supposed to be the chat GPT competitor. They also did a big reorg inside their AI division. So Google had historically had two teams that were sort of working on cutting edge AI, Google Brain and DeepMind. And last year they merged those two into Google DeepMind and they made Demis Hassabis the head of all of it. Yeah, there is no more Google Brain. If you're on the Google campus looking for Google Brain, it's not there anymore.
[0:05:08] It has lost its brain, much like you. We're going to talk about Gemini and some of the research that went into it and what makes it different than or similar to other AI products on the market. But I thought today we should really focus on what Demis sees when he looks out the windshield rather than the rearview mirror, like where he thinks AI is now, both at Google with Gemini and in the world at large, and what he thinks is coming next.
[0:05:35] And the way we pitched this interview, Tim, by the way, was deep mind meets shallow minds. Demis Hassabis, welcome to Hard Fork. Thanks for having me. So, Demis, over the past couple of weeks, we have seen a bunch of new models. In addition to Gemini 1.5 Pro, there are now two models called Gemma. What the heck is going on over there?
[0:05:58] Well, we've been busy, busy cooking things, very, very busy. I'm not sure we're going to have a release every week, but it's what it seems like at the moment. So let me just unpack that for you. So obviously there's our Gemini models, our main models. Gemini 1.0 launched the Gemini era last December. And then, of course, last week we announced 1.5, so the new generation of Gemini. And then finally...
[0:06:24] We have Gemma, which we announced today, which is the open source, lightweight open source, best in class models for open models, open weight models. I think some people may still be catching up to why there are so many different models. Often when we read about AI, I think we have a tendency to think about AI as one thing that is just sort of gradually getting a little bit better. Can you talk about why?
[0:06:51] Google and DeepMind have been working on so many different models simultaneously. Yeah, because I mean, we've always had, you know, the core of our sort of groups have always been foundational research. So we have a ton of fundamental research going on into all sorts of different innovations, all sorts of different directions. And that means at all times, there are the main tracks of the models we're building.
[0:07:18] The kind of the core Gemini models, but there are also many more exploratory projects going on. And then when those exploratory projects yield some results, we fuse that into the main branch, right into the next versions of Gemini. And that's why you're seeing things like 1.5 coming.
[0:07:35] You know, so hot on the heels of 1.0, right? Because, and we're already working on, you know, the next version because we have multiple teams each working at different timescales, sort of cycling with each other. And that's how you get kind of relentless progress. And I hope that's, you know, this is sort of the new normal for us really is shipping at this high velocity, of course, whilst still being, you know, very responsible and keeping safety in mind.
[0:08:03] I want to ask about the latest big release that you had, which was Gemini 1.5 Pro. The feature of that model that the people I follow and talk to are most excited about is a very long context window. The sort of previous longest context window I had ever heard of was from Claude Anthropic's model, which could handle up to 200,000 tokens.
[0:08:29] Essentially 200,000 words or fragments of words. Your new Gemini Pro 1.5 model can handle up to a million tokens, so five times bigger. Can you just explain what that means and why that's a big deal? Yeah, it's super important that long context you can think of as the model's working memory. You know, how much data can it sort of keep in mind at once and process over.
[0:08:57] The longer you have that and the more accurate it is as well, it's also quite important, the precision of recalling things from that long context, the larger amounts of data and context you can take into account. So a million means that you can do massive books, entire films.
[0:09:18] Lots of audio, things where, you know, entire code bases. So if you have a much shorter context window, you know, 100,000, that kind of level only, then you can only have snippets of that. And you sort of can't have the model reasoning over or retrieving over the entire corpus that you're interested in. So it actually frees up.
[0:09:37] It actually allows for all sorts of new use cases that you can't do with the smaller contexts. And actually, we've tested it up to 10 million tokens in research tests.
[0:09:50] One thing I've heard from AI researchers is that the problem with these big context windows is that they're very computationally expensive. Like if you're uploading an entire movie or a biology textbook or something and you're asking questions about it, it just takes a lot more processing power to go through all of that and respond. And if a lot of people are doing that, you...
[0:10:11] The costs start adding up pretty quickly. So did Google DeepMind come up with some clever innovation to sort of make these huge context windows more efficient? Or is Google just kind of bearing the cost of all of that additional computation? Yeah, no, it's a totally new innovation because you can't have context that long without some new innovations on top. But it's still computationally quite expensive.
[0:10:36] We're working hard on optimizations. The initial processing of the uploaded data can take a couple of minutes if you're using the whole context window. But if you think about that, that's like watching the whole film or reading the entire War and Peace in a minute or two. That's not too bad to then be able to answer any question about it. And then what we want to make sure is that once you've uploaded it and it's processed, it's sort of read the document or processed the video or audio, then the subsequent questions and answering of those questions should be fast.
[0:11:05] And that's what we're working on at the moment is optimizing that. And we're very confident we can get that down to, you know, order of a few seconds. And you said you've been testing up to 10 million tokens. Like how well does that work? Does that feel like that's pretty close to becoming a reality too?
[0:11:19] Yeah, it's very good in our tests. It's not really practical to serve yet because of these computational costs, but it works beautifully in terms of precision of recall and what it's able to do. Yeah, I mean, ChatGPT released this memory feature.
[0:11:37] Last week or the week before, that's essentially just a tiny scratch pad that can remember maybe a handful of facts about you. But man, if you were able to create a version of that that has 10 million tokens about you, it could know your entire life. It might be able to be a really good assistant for you. I think a lot of people see all these new AI releases coming out practically every day and it just kind of all blurs together for them.
[0:11:58] What can this model do that the last one couldn't? So I want to ask that question to you about Gemini. What is a thing that Gemini can do that BARD or previous Google language models could not?
[0:12:09] Well, I think the exciting thing is about Gemini and 1.5 especially is the sort of native multimodal nature of Gemini. We built it from the ground up to cope with any types of inputs, text, image, code, video. And then if you combine that with the long context, I think...
[0:12:30] You're seeing the potential of that. Like you could imagine you're listening to a whole lecture or there's a lecture, but there's an important concept that you want to know about and you want to just fast forward to that. So another interesting use case is now we can put entire code bases into the context window.
[0:12:46] Thank you very much.
[0:13:16] So I want to turn now to Gemma, the new family of lightweight open source models you just released. It seems like maybe one of the
[0:13:45] One of the most controversial subjects in AI today is whether to release foundational models through open source or whether to keep them closed. To date, Google has kept its foundational models closed source. Why go open source now? And what do you make of the criticism that making foundational models available through open source increases the risks and the ability for them to be used by bad actors?
[0:14:12] Yeah, well, look, I've actually talked about that a lot publicly myself. One of the main worries. So in general, open source and open science, open research is clearly beneficial, right? But there is this issue specifically with AGI and AI technology, powerful general AI technology, because it's general purpose. Once you put it out there.
[0:14:32] Bad actors can potentially repurpose it for harmful ends. But of course, once you open source something, you have no real recourse to pull it back anymore, right? Unlike an API access or something like that where, look, it turns out downstream there was this harmful use case no one had considered before. You can just cut that access off. I think that means that the threshold for safety and...
[0:14:57] My view is that as we get closer to AGI, then they'll have more and more powerful capabilities. So one must be more and more careful about what they could be repurposed for.
[0:15:18] And I haven't heard a good argument from, let's say, the open source maximalists, who are many of them out there, some of whom are my respected colleagues in academia. And what is the answer to that question of proliferation and bad actor access? And we got to think about that more as these systems get more and more powerful. So why was Gemma not one of those things that created that concern for you?
[0:15:43] Yeah, well, of course, because Gemma, as you'll notice, it comes in lightweight flavors, right? So they're relatively small. And actually, the smaller sizes are more useful for developers because generally individual developers and academics or small groups that want to have things working very fast on their laptops and so on. So it's sort of optimized for that. And because they're not frontier models, right, they're small scale models.
[0:16:09] We feel comfortable that the capability has been very stress tested. We know very well what they're capable of. And there aren't big risks associated with models of that size.
[0:16:21] I want to ask about a subject that we've talked about on the show recently, which is personality in AI chatbots and how much personality chatbots should have or be allowed to have by their creators. Some models, including the original Bing Sydney, have been criticized for having too much personality, for being creepy or threatening or harassing users. Other models have been criticized for being too boring and sort of giving trite answers and not being very helpful.
[0:16:50] Calibrate the personality, so to speak, of Gemini. And where would you say it falls on that spectrum? Yeah, look, it's a very interesting question. I think a live ongoing debate in the whole industry and field. My guess is that ultimately you're going to want personalization is going to be the answer here, where people will individually decide there's some base model with base behavior. And then you would opt in and you have a kind of personal assistant that you want to behave in a certain way.
[0:17:20] And I'm guessing that is what will actually ultimately happen because the problem with a general one is you can't satisfy all constraints, right? You know, sometimes, you know, I want Gemini to be very succinct, right? And just give me the bullet points, give me the facts. Other times you want it to be very discursive and creative. And at the moment, I think we're still quite nascent and we're still working on these base generic models.
[0:17:42] Well, Kevin won't rest until you've given him a personal assistant who is completely insane. So I look forward to seeing if you can satisfy that condition. I want to ask another sort of question in this vein around personality and how the prompts respond to us. There's been some people noticing online this week that Gemini doesn't seem to create a white man, if you ask it to, or if you try to depict figures from history, it sort of doesn't want to do that.
[0:18:11] We'll sort of do it in a historically inaccurate way. I understand all the sensitivities around this, but I'm curious what you make of that criticism and how you're trying to balance sort of not doing something deeply offensive with also doing stuff that is historically accurate.
[0:18:25] Yeah, look, actually, we just became aware, I just became aware of that, you know, yesterday when it started popping up on social media. And, you know, I think this is a good example of these nuances, right? The historical accuracy, absolutely, we want that. So we need to fix that. Versus when you have a generic prompt, obviously, then there's things are universal. So for example, if you said, you know,
[0:18:48] Draw me, create a picture of a person walking a dog or a picture of a nurse in a hospital or something like that. You'd want a universal, you know, depiction. But then others that, you know, historical events or historical figures, then perhaps that should be narrower. So I think that's an interesting, you know, piece of feedback. And this is why we also have to put some things tested out in the world. It's something that becomes obvious, actually, once you have it tested.
[0:19:17] Demis, I want to shift focus away from Gemini and sort of broaden our view a little bit here. You are less than a year into your big new job as the CEO of the combined former Google Brain Research Lab that existed inside Google and DeepMind, which was the...
[0:19:45] The company that you and your co-founders started all those years ago. And last year when Google Brain and DeepMind were combined, some folks I know in the AI industry were concerned. They worried that Google had historically given DeepMind this pretty long leash to work on whatever kinds of research projects it deemed important. And that as these
[0:20:08] Units got combined, DeepMind's priorities were going to be shifted toward the things that were sort of good and useful for Google in the short term, rather than these kind of longer term, you know, foundational research projects. It's been almost a year since those units were combined. Has that tension between sort of short term benefit to Google and maybe long term AI progress changed what you can work on at all?
[0:20:37] Yeah, well, I would actually say that this sort of first year, as you're alluding to, has gone fantastically well. One reason that we felt it was the right time to do that, and I did from a researcher point of view, is that...
[0:20:51] Maybe let's wind back five years or six years back when we were doing things like AlphaGo. We were very exploratory in AI as a field in terms of what was the right way to get to AGI, what breakthroughs were needed, what sorts of things should be bet on. And in that situation, you want to do a broad portfolio of things. So I think that was a very exploratory phase. I think in the last two, three years, it's become clear what some of the main components are going to be. As I mentioned earlier, we're still going to need new innovations, I think.
[0:21:20] And you've just seen one with our 1.5 with models with the long context. I think there are lots of new innovations like that that are going to be required.
[0:21:30] So foundational research is, I would say, still as important as ever. But there's also this big engineering track now of sort of scaling and exploiting known techniques and pushing them to the limit. And there's incredibly creative engineering at scale that has to be done there, all the way from the bare metal hardware work up to the data center size and the efficiencies of all of that. And one reason why the timing is right now is that...
[0:21:59] If we were to say to me, make some AI powered products five, six years ago, we would have had to build quite different AI to the, let's call it the AGI research track, the track of researching general AI techniques that will one day be useful for AGI, right? Versus doing something special cased for a particular product. That would have needed a kind of bespoke AI, handcrafted AI. And so that effectively, there would have been two different types of things you would have to do.
[0:22:27] That's not true today. To do AI for products, actually the best way to do that now is to use the general AI techniques and systems because they've got to a level of sophistication and capability where they're actually better now than doing any special case type of hard-coded type of approach. So in fact, that's converged. So what you can see today is that the research tracks and the product tracks have converged.
[0:22:53] And so now I don't have to have a split brain on like, oh, I'm working on products over here and so I have to do this type of AI. And then I'm like a handcrafted assistant, Siri-like assistant versus a true chatbot that understands language. They're one and the same now. So first of all, there's no kind of dichotomy there or tension there. The second thing is, if that's true, it's actually really good for research to have...
[0:23:21] Tight feedback loops with grounded real applications, because that is the way you really understand how your models are doing. You can have metrics, academic metrics coming out of your ears.
[0:23:34] But the real test is when millions of users use your product and do they find it useful? Do they find it helpful? Is it beneficial to the world? And you get obviously a ton of feedback that way. And then that leads to very rapid improvements in the underlying models. And so that phase that I think we're right in the middle of now is very, very exciting.
[0:23:58] So in San Francisco, the mood around AI is very optimistic. It's clearly very optimistic inside Google, but it's more pessimistic other places. Pew Research Center, the survey last year found 52% of Americans said they feel more concerned than excited about the increased use of AI. Only 10% are more excited than concerned. What do you think explains that downturn in public sentiment? And what do you think you can do about it?
[0:24:24] I think, you know, I don't know what to make necessarily of those kinds of polls. I think it depends how exactly the question you ask it. If you ask it in a very naive way, I think people are always worried about change, right, or disruption. And clearly, AI is going to bring enormous change. I've always believed that's why I worked my whole life, my whole career on this, 20 plus years. I think the world is realizing what I felt people like myself and other researchers have been in for a long time have known for decades now. Like, if this was to work, it would be, you know, the most monumental
[0:24:54] So I think it's dawning on people, but they don't know quite, they haven't interacted it with many different ways. And so it's sort of very new. And I think what we need to do actually as a field is present.
[0:25:08] Concrete use cases that are clearly incredibly beneficial, right? And one of the things we've done historically, I think like that, I'd point to is AlphaFold, but the average person in the street probably doesn't know about that yet, what that impact will be, but they will do if that leads to AI designed drugs and cures for really terrible diseases, right? And I think we're only just a few years away from that.
[0:25:31] Obviously, we spun out Isomorphic Labs, a sister company of Google DeepMind, an alphabet company, which I also run, which is to take the AlphaFold technologies, move them into chemistry and biochemistry to actually design drugs to bind to the right parts of the protein structures that, of course, AlphaFold has predicted. And then we've just signed big deals with Big Pharma and on real drug programs. And I expect in the next couple of years, we'll have AI-designed drugs in the clinic.
[0:26:01] Yeah, I agree. I mean, AlphaFold is the thing that I hear far and away the most when it comes to sort of like the best possible uses of AI technology. But it was also a somewhat unusual problem because it was sort of the right...
[0:26:24] Do you think there are other kind of similar shape problems out there or have most of this sort of low hanging fruit already been picked? No, I think there's many problems of that type.
[0:26:53] So the way I normally describe it, and obviously protein folding is archetypal example of this, is imagine any problem in science, which basically has a huge combinatorial search space, huge number of possibilities, way more than you could search by brute force. So let's take chemistry space, the space of possible compounds. Some people estimate that's 10 to the power 50 in terms of the possible compounds one could create. So truly enormous number of possibilities.
[0:27:23] So intractable to do that by hand. But what do you do is if you can build a model of chemistry that understands what's sort of feasible in chemistry, you could use that to do a search where you search, but you don't search every possibility. You search just a tiny fraction of the possibilities that make the models telling you are kind of the highest value.
[0:27:45] And I think there are a lot of things in science that fit that. And I give examples. So protein folding is one. But I think finding a drug compound that has no side effects but binds exactly to the thing that you want on your protein or on the bacteria, that's another one. I think finding new materials. Like I dream of a room temperature superconductor that's cheap to make.
[0:28:10] So that's one of the things I'd love to turn our systems to, or the ultimate optimal battery design. And I think all of those things can be reimagined in a way where these types of tools and these types of methods will be very productive. Do you think we're close to seeing AI being able to cure a major disease like an Alzheimer's or a cancer?
[0:28:36] I think we are very close. I would say, you know, we're a couple of years away from having the first AI design, truly AI design drugs for a major disease, cardiovascular, cancer. We're working on all of those things, isomorphic. And then obviously there's still the clinical trials and that stuff has to happen. And right now that would be the bottleneck.
[0:29:03] I think certainly getting it into the clinic, the discovery phase, I would like to shrink that from years to months, maybe even weeks at some point. So I think in a couple of years, I would be disappointed if we don't have some great candidates for drugs for very important diseases starting to go through clinical trials. I want to...
[0:29:25] Let's shift our conversation a little bit more toward the long-term future of AI. And a term that has already come up a couple of times in this conversation is AGI, artificial general intelligence, which is a term that gets thrown around a lot these days without a lot of specificity. So I thought we should just start by asking you in one sentence, what does AGI mean to you?
[0:29:49] Well, AGI means a system that is generally capable. So out of the box, it should be able to do pretty much any cognitive tasks that humans can do. This might be a stupid question, but when AGI arrives, assuming it does, how will we know? How will we recognize it? Like one of your engineers, presumably, if this all goes according to your plan, will show up in your office one day and say, Demis, I've got this thing. I think it's AGI.
[0:30:14] How do you test that? Is there one sort of battery of tests you could put it through that would convince you like this is AGI? Is there one question you would ask it to determine whether it was truly AGI or not? How will we know when this thing shows up?
[0:30:28] Yeah, well, actually, one of my co-founders, Shane Legg, he did his whole PhD on the testing of and measuring of these systems. And I think the best, because it's so general, that actually makes it quite difficult to test, right? You can't test it in one particular dimension. I think it's going to have to be a battery of thousands of tests and performing well across the board.
[0:30:51] Covering all of the different spaces of things that we know the human brain can do. And by the way, the only reason that's an important, obviously, anchor point is the human brain is the only existence proof we have in the universe, as far as we know, of general intelligence being possible.
[0:31:09] So that's why I originally studied neuroscience as well as computer science, because clearly, certainly in the early days of AI, it was important to get neuroscience inspiration too, for how are these intelligent phenomena, how do they come about? What do they look like? And therefore, what do these systems need to be able to do in order to exhibit signs of general intelligence, right? And I think we're still quite a long way off of that, actually, with the current systems, right? There's a lot of things, all of us who've been...
[0:31:38] And in your best guess, how far are we away from that kind of AGI?
[0:31:56] Well, look, I think we're making enormous progress as a field. We're making enormous progress with Gemini and those types of systems, which I think will be important components of an AGI system. Probably not enough on their own, but certainly a key component. And I would not be surprised if we saw systems nearing that kind of capability within the next decade or sooner. Have your timelines shifted at all over the past year or two as things like language models have gotten out into the public?
[0:32:26] It's funny, actually, because I was looking at our original business plan we wrote back in 2010 when we started DeepMind. And we had all sorts of predictions in that business plan, including compute and other things and other inventions that would be needed. And we stated in their 20-year timescale. And I think we're actually pretty on track. That gives us six more years, if I'm counting correctly. Roughly.
[0:32:48] Roughly speaking, but that's compatible with I wouldn't be surprised within the next decade. That doesn't mean it's going to happen. So I just wouldn't be surprised. So you can sort of infer some probability mass based on that. So I think there's a lot of uncertainty because you don't know if the current techniques are going to hit a brick wall. If they do, then you would have to kind of invent some Nobel Prize level innovation to get through that brick wall. Right now, we don't see one.
[0:33:15] Ones have been conjectured in the past and some of my colleagues in the field conjecture that there could be some brick walls. But I think it's an empirical question, actually. I think we've got to, that's why we do push really hard on both things. We want to scale the current ideas and know-how and techniques to the maximum. And we want to double down on our foundational research and innovative research and exploratory research to find improvements to the existing ideas and also...
[0:33:45] Think through what could be the brick walls and if they are end up being brick walls with the scaling system, then what would be the answer? So hopefully we have the answer at the point where we hit a brick wall. We would already have some ideas of how to get around it. That would be the ideal if indeed there turns out to be a brick wall because there may not be.
[0:34:04] Do you think the world is ready for something like AGI to show up? If we only have six years to prepare for a computer that can do everyone's job, what should we be doing now to get ready for that? Well, look, I think the debates are happening. So I think the silver lining with, I guess, this craziness of the last couple of years on AI is that everyone's talking about it.
[0:34:25] I think chatbots have been useful in that sense in that the average person can interact with a cutting edge AI in a way that's easy to understand, right? You know, alpha fold, you need to be an expert in biology or proteins or medical research to really get what it is. So, you know, language is different, right? We all use language every day and it's an easy thing for everyone to understand. So I think those, you know, it's good that those debates are happening. It's something that's going to affect everyone in society. I think there are questions on international cooperation.
[0:34:54] I would like to see a lot more of that. Unfortunately, the geopolitical nature of the world right now is not very conducive to that. So that's unfortunate timing because I think some kind of international collaboration is going to be, in my opinion, very important here, which is why I was pleased to see how many international leaders engaged with the summit in the UK back in last autumn. And then we need to, of course, accelerate our...
[0:35:22] Research into safety, guardrails, control mechanisms. And I think actually we need to sort of do more.
[0:35:32] A lot more work in that direction. And also philosophy too. What do we want from our systems? And ethics and philosophy. It's deep philosophical issues. What do we want our systems to do? What values should they have? It intersects with the earlier discussion we had about personas. It actually comes down to values. What do you want your systems to represent? And of course, it's important who makes those systems.
[0:36:00] Cultural and societal background therein and Western systems and China's building systems. I mean, there's a lot of complications here. What role do you see humans playing in a world where AGI exists and can just sort of run everything on our behalf? Well, I think this is going to happen in many stages. I think initially I'm seeing AI and the next versions as these incredible assistive tools.
[0:36:24] That's how I think we should design and make them. So there's sort of this debate about tools versus creatures you sometimes hear. And I think that we should be firmly, I'm in the account, we should be making tools to assist human experts and so on, whether they're scientists or medics or whatever it is, to free them up to do the higher level conceptual work.
[0:36:46] So today our systems, maybe they can help you with data crunching or some sort of analysis of a medical image, but they're not good enough yet to do the diagnosis themselves, in my opinion, or to trust them with that. They should be an expert human in the loop. And I see that as the next phase. And for however many years or decades that will be.
[0:37:08] And then maybe we'll understand these systems better in the course of doing that. And we'll be able to figure out what to build with them next to allow them to go to the next stage. And I think society will have to adapt about what it is that...
[0:37:23] We want to do in a society where we have AI systems that are able to do very useful things for us. Maybe we have abundance because of that, because we crack things like energy problems, things like physics and material design. So there should be a huge plethora of amazing benefits that we just have to make sure are kind of...
[0:37:41] Equally distributed, you know, so everyone in society gets the benefit of that. And then, you know, I think incredible things might be possible that sort of written in science fiction books, books like the Culture series by Ian Banks and so on. It's always been my favorite since I was a teenager of a depiction of.
[0:37:58] You know, maximum human flourishing across the cosmos, you know, helped by AI systems, solving a lot of these fundamental problems for us, helping us solve a lot of these problems. I think it could be an amazing, amazing future. And with incredibly, you know, big challenges that are facing us today as society, climate, disease, poverty, a lot of these things.
[0:38:20] Water access could be helped by innovations that come about through the use of these AI tools. That's the positive side of the approach of AGI. There's also a side that worries a lot of people, including AI safety people, risk people. You yourself have worried about existential threats to humanity that could result from very powerful AI systems.
[0:38:51] I'm going to ask you a question that we've asked a lot of people on this show, which is what is your P-Doom? Yeah, I know that's people are fixated with that. Do you know my honest answer to these things? First of all, I actually find a lot of the debate on the social media sphere a little bit ridiculous in this sense. You can find people on both sides of the argument, very eminent people. Just take like Jeff Hinton versus Yann LeCun, right? I mean, both are Turing Award winners. I know them both very well.
[0:39:21] People who originally were in the field. And, you know, the fact that they can be completely in opposite camps to me suggests that actually we don't know.
[0:39:32] With this transformative technology, it's so transformative, it's unknown. So I don't think anyone can precisely, I think it's kind of a nonsense to precisely put a probability on it. What I do know is it's non-zero, that risk. So it's definitely worth debating and it's worth researching really carefully because even if that probability turns out to be very small,
[0:39:56] Let's say on the optimist end of the scale, then we want to still be prepared for that. We don't want to have to wait to the eve before AGI happens and go, you know what, maybe we should have thought about this a bit harder. No, we should be preparing for that now and trying to ascertain more accurately what the risk really is.
[0:40:17] What is that risk? How would we mitigate it? What risks are we worried about? Is it self-improvement? Is it controllability? Is it the value systems? Is it the goal specification? All of these things are research questions, and I think they're empirical questions. So it's unlike a natural science like chemistry, physics, and biology, the phenomena you're studying is already out there, exists in nature. So you go out there and you study and you sort of try to take apart and deconstruct what's going on. But with an engineering science,
[0:40:46] The difference is you have to create the artifact worthy of study first and then you can deconstruct it. And only very recently, I would say, do we have AI systems that are even sort of interesting enough to be worthy of study, but we have them now.
[0:41:01] Things like Gemini, AlphaFold, and so on. And we should be doubling down, and we are obviously as Google DeepMind, but the field should be doubling down on analysis techniques and figuring out understanding of these systems. Way ahead of...
[0:41:16] Where, you know, we're on the kind of cusp of AGI. And that isn't a lot of time because if we're, you know, less than a decade away, these problems are very hard research problems. They're just, they're probably harder or as hard as the breakthroughs required to build the systems in the first place. So we need to be working now, yesterday on those problems, no matter what the probability is, because it's not, it's definitely non-zero.
[0:41:41] So I don't agree with the people that say there's nothing to see here. I think that's ridiculous. What basis are they making that assumption? Just like in the past, 10, 15 years ago when we started out, I remember I was doing my postdoc at MIT, and that was the home at the time of traditional AI methods, logic systems, and so on. I won't name the professors, but some of the big professors there were like, learning systems, this deep learning, reinforcement learning, they'll never work.
[0:42:07] 300 years for sure never work. And I was just like, how can you put 0% on something in 300 years?
[0:42:14] Think 300 years back what happened, what society's done. I mean, that's just not a scientific statement to say 0%. We don't even understand the laws of physics well enough to say things are 0%, let alone technologies. So it's clearly non-zero. It's massively transformative. We all agree. Hugely monumental impact, hopefully for good. Obviously, that's why we're working on it, and I've worked my whole life on it. We just talked about that, science, medicine, et cetera, human flourishing. But we've got to make sure it goes well.
[0:42:43] If it's non-zero, we should be investigating that empirically and doing everything we can to understand it better and actually be more precise then in future. Maybe in five years' time, I would hope to better give you a much more precise answer with evidence to back it up rather than slanging matches on Twitter, which I don't think are very useful, to be honest.
[0:43:02] I have a shorter, more medium-term fear, which is that before AGI gets to the point of enslaving us or whatever, it just massively concentrates power and wealth in the hands of a very few companies. And it doesn't feel like the benefits are evenly distributed. So I wonder if you feel like in your role, you are in a position to make sure that those benefits are more broadly distributed, or if that risk of this stuff just really concentrating...
[0:43:32] A lot of power and money is real. Look, I think there's several nuanced answers to that. It's a complex question. I think that right now, a lot of resources are required to build the most cutting edge models, but you're already seeing open source systems, including Gemma, our contribution to that today as well, are getting pretty powerful. So for a lot of everyday use cases, that might be already
[0:43:59] Plenty good enough for a particular product or an application or so on. And I think...
[0:44:04] The developer community is going to create amazing things with these models. Even the proprietary ones have API access to them, like Gemini 1.5 is coming, 1.0 is already out, including Ultra. So you can build on top of that enterprise customers and so on. So that's all happening. There are multiple providers of these models, right? There isn't just one company, there's several. So they'll all be competing on price. I mean, you're already seeing price of tokens is going to like...
[0:44:32] You know, is going down, you know, discounted every day, it seems. So, you know, I think all of that is good for the consumer, good for everyday users and good for, you know, companies and others, enterprises that are building on this. And then ultimately, this is the funny thing is, so I would couch myself as a cautious optimist, right? But the thing is, and I think that's the correct approach when you're talking about something as transformative as AI.
[0:45:01] And I've thought about it for many, many decades. And the funny thing I see of some of the more, I would say, techno optimists, I think they sometimes call themselves on the Twitter crowd, is I actually don't think they fully understand the monumentalness of what is being built.
[0:45:16] Because if they did, I think cautious optimum is the only reasonable approach, I would say, in the face of quite a lot of uncertainty. Obviously, a lot of obvious, amazing things that could happen, curing diseases, et cetera, but uncertainty over how the technology is going to develop.
[0:45:37] for something that transformative, then cautious optimism, I think, is the only reasonable approach. And so, yeah, I think there's going to be incredible things. And I think one of those things is going to be, I'm not even sure, if you imagine a world where AGI has arrived and it's helped us solve a lot of big scientific problems, I sometimes call them root node problems. So if you think of a tree of knowledge and what are the core big...
[0:45:59] That opens up, you know, suddenly energy becomes...
[0:46:17] I don't know if company constructs would even be the right thing to think about at that point.
[0:46:40] Again, this is where international collaboration between governments and other things may be required to make sure that these systems are, well, if there's multiple ones, are managed in the right way and used for the benefit of everyone. Yeah. A lot of AI labs have been grappling with governance and what is the best structure for something like AGI to emerge. You just mentioned the possibility of some sort of international collective or cooperative that would handle this.
[0:47:10] Across the industry, OpenAI has set itself up as a nonprofit with a for-profit subsidiary. Anthropic is a public benefit corporation. You're making a slightly different bet, which is to try to get to AGI inside Google, which is a big for-profit company that has a fiduciary duty to make money for its shareholders.
[0:47:32] Does that worry you? And as we get closer to AGI, do you think Google will have to change its corporate structure somehow in order to prevent some of the bad outcomes that people at other AI labs are worried about? I feel like the current construct is...
[0:47:51] That's the best method we've ever invented for understanding things in the world. Unbelievably powerful.
[0:48:20] That's what's created the modern world and all the benefits of the modern world. And so we got to double down on that method and trust that method. That would be my approach versus alternative methods, which are very effective as well, but I think less correct for this type of technology, like a hacker growth mentality, move fast and break things. You know, you sometimes hear is the value mantra, obviously creative phenomenal products and advances, but I think not appropriate.
[0:48:50] For the type of monumental technology we're talking about with AI. Right there, I think it's the scientific method is the better approach. And I think Google is the most scientific of the big companies, I would say. Always had that in its DNA. And we've tried to bring that more into that. And also receptive to these ethical concerns from the beginning. One of the reasons we teamed up with them was we had our own ethics charter as DeepMind. We had it from the beginning. And you now see that embodied in the Google AI principles. Google was the first of the big companies to put out
[0:49:19] We've been incredibly responsible, I think, and thoughtful about the way we're deploying these technologies and how we're building it. And hopefully you can see that in the approach we've taken. And sometimes that means we take a little bit longer with things before we put them out.
[0:49:40] Because we are trying to fully make sure we understand to the extent, it doesn't mean we won't ever make mistakes because this is new technologies. And sometimes as we talked earlier, you need the direct feedback. It's useful from users and from experimental releases. And that's why we do stage releases, right? Like with 1.5 now, Pro, it's an experimental release so we can get early feedback. But look, I think that's the right approach. And I think I'm very comfortable with where we are now.
[0:50:06] In five, 10 years, as we get closer to AGI, we'll have to see how the technology develops and also what state the world is in at that point and the institutions in the world, like the UN and so on, which we engage with a lot. And I think we need to see how that goes and how the engagement goes over the next few years. I wonder if you have a thought on what the most AI-proof job in the world is right now. Is there anything you see out there where you think, yeah, that's a safe bet for the next five or 10 years?
[0:50:35] Well, look, I actually think that what's going to be, you know, and
[0:50:39] You can just say podcast host if you want. That would make us feel better. Podcast host, obviously. Yes. But I think actually a lot of the jobs where, and I talk to a lot of my creative friends in creative industry like film, music, and games and stuff. I used to do games design myself back in the day, early in my career. I think there are certain types of creatives who also really love technology as well as the creative process. And I think they're going to be like super powered up.
[0:51:08] By using their creativity on top of these tools, whatever these generative AI tools do, they still need the creative input to make them do interesting, valuable things. Otherwise, they're just doing fairly mundane things with the average user. I think that it can be incredibly a credible multiplier for those types of creatives. I have a friend of mine who's...
[0:51:33] And so on. And it's just.
[0:52:00] Means that the whole process is accelerated for them in terms of actually getting to the film production. And then stuff on the science. I dream of a science assistant that can just summarize a whole field area for you. Or here's a bunch of, tell me what the best reviews are and the counterpoints. You know, obviously we need to fix factuality and other things a lot better before we have that. And we're working on that. But that would be incredible for like.
[0:52:21] Then giving me the information where I could then make a new connection or new hypothesis to then go and test out, right? Or to help a doctor on a complex diagnosis. Doctors are unbelievably busy. Can they keep up with the latest literature as well and the cutting edge of research? A sort of science assistant or medical assistant tool could help them do that while they're treating patients 24-7, right? So I just feel like there's a lot of things like that. And then I think a lot of jobs,
[0:52:51] There may be somewhat undervalued today, manual jobs, manual labor jobs, things like that, I think, or caring jobs where you really want the human emotional empathy and touch, I think are going to be much more valued in future. And that's maybe a good thing, right? Perhaps they're being undervalued today in our capitalist society. All right. Last question, Demis.
[0:53:14] What is your personal plan for the post-AGI world? When the AIs come and they don't need humans to run companies or go on podcasts anymore, how will you spend all your copious free time? Are you going to apply for a job as a plumber or a gardener? What does the world look like for Demis in the post-AGI world?
[0:53:36] So what I've always wanted to use my AGI tools for would be to really understand the deepest questions of nature and physics. So the fundamental nature of reality. I'd like to have the time to ponder that, think that through, perhaps traveling on a starship to Alpha Centauri, thinking about that, meditating on these ideas, maybe doing some extreme sports, stuff like that. I think there'll be plenty of very exciting things for us to do. We're just going to have to be very creative about it. And as I said,
[0:54:05] There's many, many amazing science fiction books, positive ones, that talk about what such worlds might look like. And I think they're very exciting if we get it right. You mentioned you wanted to do some extreme sports in the post-AGI world. What extreme sports are you interested in?
[0:54:22] Look, I haven't done very many extreme sports today because, you know, I have to be careful, right, to keep healthy and fit. But, you know, maybe there'll be some new ones we're able to do, you know, mountain climbing up Mount Olympus on Mars. That might be quite fun, hiking up there. I think that's a good place to leave it. Demis Hassabis, thank you for coming on Hard Fork. Thanks so much for having me. It's been great fun. Did you see that ChatGPT apparently had a nervous breakdown last night? No, what happened?
[0:55:01] So I was just following this on social media, but all of a sudden it just started spitting out nonsense. Sometimes it would start just speaking Spanish. Ay caramba! And no one could quite explain why. I think it's back to normal, but it did appear that ChatGPT for a night just went crazy. My favorite response was there was a user who was talking with ChatGPT and it said,
[0:55:31] Are you having a stroke? Some of what you're saying makes no sense or aren't proper words. And chat GPT says...
[0:55:39] Whoops, I really apologize if my last response came through as un-unclear or se siente, like it drifted into some nonsensical wording. Sometimes in the creative process of keeping the inner time spanglish vibrant, the cogs in la tecla might get a bit whimsical. Muchas gracias for your understanding. Y al ensure we're being as crystal clear como lo from now on. Winky face. And it said...
[0:56:02] Would it glad your clicklies to grape turn tooth over a mind ocean jello type? Or submarine else que que sierras que dive into? Please share with their fourth como desire. Period.
[0:56:16] Yeah, ChatGPT really snapped with that one. You know what, Kevin? For so long now, you've been asking for a chatbot that has personality. Well, congratulations, my friend. You've finally gotten one. It's completely insane, and it's speaking Spanglish, and it's available for $20 a month. Head on over to OpenAI.com and start your new life, my friend. Let's go. Let's go!