Hello and welcome to The Tech Download - a new CNBC Original podcast where we unpack the tech stories that matter most. Each season, we dive into one big theme - and what it means for your money - with insights from the industry's most influential voices.
I've always thought that, in the end, it would be the most important technology we'll ever invent. And it's sort of the natural progression, really, of the computer age.
This season, we’re looking at Google DeepMind, the powerhouse driving the tech giant's AI push. We've been given rare access to key figures at the company - including our guest for this episode DeepMind co-founder and CEO Demis Hassabis.
I think there it's going to be like the industrial revolution, but maybe 10 times bigger, 10 times faster. So, it's incredible amount of transformation, but also disruption that's going to happen.
Hey everyone, and welcome to The Tech download. Allow me to reintroduce myself. I'm Arjun Kharpal, Senior Technology correspondent at CNBC based in London, and I've got a very special new co host with me.
Hey there, Arjun, yeah. Steve Kovach here, I cover tech over here in New York. I mostly focus on Apple and Microsoft. But look, I've been covering the tech industry for over 15 years now. I kind of have a good grasp on everything. I'm so excited to be here with you, Arjun, because I've just admired your work from across the ocean for so long, and now we actually get to kind of collaborate and do this thing together. I think it's gonna be a good time.
It's going to be so fun. Steve. So between us, we think we've got nearly three decades of experience covering tech, and the crazy thing is, we've got so much to learn. And I think over the course of us doing this podcast, we're going to learn so much, speak to so many interesting people, I’m so excited that this first series, we're kicking off with an insight into Google Deep Mind, one of the world's leading AI labs as well, just for our listeners and our viewers, a quick intro, I guess, to Google DeepMind. It was a company founded in 2010 here in London, where I sit, as well. Very small company founded by three people, Demis Hassabis, is Shane Legg and Mustafa Suleiman, who's at Microsoft now, right, yeah.
And in fact, I interviewed him, nearly a year ago now, Mustafa Suleiman, he's basically doing what Demis is doing over at Google. And it's just kind of interesting to see how Google was like this incubator, so to speak, for all of this top AI talent around the world, Demis obviously stuck around. He's running Deep Mind over there. What I also think is really interesting, though, is just this AI moment, Arjun, we've been living through for the last three years, and how three years ago, chat GPT comes on the scene, and Google was kind of seen as under threat. They went through this code red. They had to go through a bunch of reorganizations internally. Eventually, Demis came out on top as the leader of AI. And guess what? 2025 was a really interesting year for AI over at Google, they kind of caught up, and in some ways, even surpassed what chat GPT was already doing. And this is really interesting because the fundamental technology for all this, these large language models we've been talking about for so many years, started at Google, and the perception was Google, let chat GPT kind of take that technology and run away with it. But now, in my view, at least Gemini is pretty much on par, if not better, than chat GPT. and Google Deep Mind is integral for this. I mentioned it was founded in 2010 Google actually acquired Deep Mind in 2014 I was very new into my career as a tech reporter as well. Google paid around 400 million pounds for Deep Mind at the time in 2014 about $540 million. It's a stake this day that could be worth 10s of billions, maybe hundreds of billions of dollars, according to some estimates today, and Deep Mind really is very much responsible for Google's AI. We talk about Gemini, the chat bot, the AI that Google's released to consumers. This is powered so much by the technology coming out of Deep Mind. But even before all of this Deep Mind was having some big breakthroughs, there was a big moment a few years ago when they released a system called Alpha Go. This was the first computer program that was able to defeat a world champion in a game called Go. This is a very complex game, and it was seen at the time as one of the grand challenges of AI because it was such a complex game with so many different combinations available. The other big breakthrough, of course, was was something called Alpha Fold. This was another AI system developed at DeepMind that could accurately predict 3d models of protein structures. And the idea is, here is, if you could do that, this may lead to some medical breakthrough. So this advancement of science has been pretty core to what Deep Mind's been up to. And clearly it was a significant bet from Google more than 10 years ago, because it's helped turn Google into an AI world leader today.
Yeah, that's exactly right now. What really struck me about Deep Mind, having watched them for so many years, is how rooted in science they were. They weren't necessarily trying to build consumer products like they do now. They were really trying to solve fundamental problems in science and really usher in this era of AI powered drug discovery, of other big, complex problems like climate change. I know Demis talks about that a lot, and he's going to talk about that in your conversation as well. Arjun, absolutely. Steve, look, it's a great scene set up for Deep Mind. So let's get into the conversation with its CEO, Demis Hassabis.
Demis, thanks for joining me on the tech download. Appreciate it.
Demis, we're going to try to get through a lot in our time here. But I want to start first with the technology itself. And we've been talking about AI, and we've been talking about the capabilities and how they've been continuously improving as well. Now in the tech world, and there's a lot of conversations about how good can these models get, how good can these systems get, and there's a lot of debate around this idea of scaling laws for our for our listeners, you know, it's this idea of of more compute, more data, bigger models eventually will lead to bigger systems as well. You said we need to push scaling laws to the maximum. There's questions over now, are we hitting any kind of walls in terms of progress of those scaling laws, in terms of the ability for these models to get better, and just from you know, what you've been developing here at DeepMind, what are you seeing?
Well, look, I think scaling laws are going very well, so we're definitely seeing increased capabilities by putting in more compute, more data, and making these models generally larger, so that trends continuing may not have been as fast as it was a couple of years ago. So there's some talk of diminishing returns and but there's a big difference between sort of no returns and exponential and I think we're somewhere in the middle where there's very good returns, and that's worth doing. On top of that, if I you know in terms of, like, getting all the way to AGI, artificial general intelligence, you know, maybe that there's one or two big innovations still needed as well, and maybe missing, in addition to the scaling up of kind of the existing ideas
We'll get on to AGI very shortly. But what are missing in your view?
Well, if you look at, I mean, we've all, you know, played around with different chat bots, and you can see that, you know, they can do very impressive things in some dimensions, but they're kind of like jagged intelligences I like calling them in the sense of like, they're very good at certain things. But there are other things that they don't do they're not capable of at all. And and if you pose a question in a certain way, you find that they're flawed and they can't do some relatively simple things. And so for a true general intelligence, you shouldn't see that inconsistency should be consistent across the board. And also they're things like it can't continually learn it can't learn new things online. It can't truly create original things. So there's quite a few capabilities that you would like to see and you would need for general intelligence that are missing from today's systems.
That's really interesting. So what would be the sort of unlock to get to those Intelligent Systems? I just want to quickly discuss the conversation I had with Thomas Wolfe, who's the co founder over at Hugging Face. Yes, he was talking to me a few months back about his view on LLMs, in particular, large language models. And just saying, they're really great. And you know, you use these chat bots, and the chat bots say, hey, great question, great idea. And here's, here's all the information you need to know. But what's missing is the ability for these systems to come up with new and novel ideas, perhaps, and particularly, I know you're so interested in science and what AI could do to unlock new drugs or discover new diseases, etc, that actually maybe the LLMs limitations are there that can't come up with these Nobel Prize winning ideas, these novel ideas. So perhaps there needs to be some sort of new architecture. What's your what's your thinking on that at the moment?
Well, look, my passion for and my whole reason I spent my whole career on AI is, I think, eventually will be the ultimate tool for science. And of course, we've shown that with things like Alpha Fold and all of the science work we've been doing over the last decade. But there's still a long way to go in terms of, can an AI actually come up with a new hypothesis itself, not just solve a conjecture that is already out there, which would be already useful and impressive. But can it actually come up with a new conjecture, a new a new idea about how the world might work? And so far, these systems can't do that. They don't really have the capability to do that. So there seems to be something missing. I think, some of the capabilities that required a kind of long term planning, better reasoning, maybe also the idea of a world model. This idea of, like, you know, the system actually understanding better the physics of the world, so that it can run simulations, kind of in its mind to test its own hypotheses. You know, these are things that you know the best scientists do, human scientists do. And so far, our AI systems, you know, are not able to do that.
Can you just help us understand a bit more of this idea of world models? Because it may be a term people are hearing for the first time. You know how that gets they different differ from LLMs
So LLM's, and the models we use at the moment are, you know, mostly around text. Of course, things like Gemini, our foundation model, can also cope with images and video and audio, so different modalities. But it's still actually understanding the physics of the world, the causality of the world. You know, how one thing affects another thing? Can you plan a long time into the future? These are all related concepts, and if you really want to understand how the world works, so that maybe you can invent something new in the world, or explain something about the world that was not known before, which is basically what scientific theory does, then you have to have this, this accurate model of how the world works, you know, starting with intuitive physics and and how the physics of the world works, but all the way up to biology, you know, and economics.
And do you envisage the world if we get of artificial general intelligence, is sort of human level of intelligence, that that there will be a combination of llms and world models working together, or will sort of world models supersede, in some sense, LLMs
No, I think there'll be some convergence of these technologies. That's at least my betting is, is there will be these LLMs, or foundation models, you know, like Gemini under the hood, that will be a key component. I think the question, I think there's almost no doubt about that in my mind, which is why we must try and scale those systems as as as big and as as powerful as we can. But the question is, is, is it the only component that's needed for an AGI and that's where I think, I suspect other types of technologies and other types of capabilities will be needed. And I think these world model capabilities, and we're working on our versions called Genie and and we have video models like VO, state of the art, video models that you can generate videos from, in from text, and you can think of video models and interactive models like Genie as kind of, you know, early embryonic world models, where, if you can generate something that's realistic about the world, then in a sense, your model understands that about the world. Otherwise, how could it have generated it?
Demis, you mentioned this, AGI, artificial general intelligence. I know there's various definitions of it floating around. You've previously said you believe that reaching AGI could be somewhere in the in the realm of five to 10 years away. Is this still your view, given, I guess, some of the profound developments we've seen in 2025
Yes, I think we're right on track from that. Actually, when we started DeepMind back in 2010 we thought this would be a 20 year kind of mission to build AGI, you know, a system that's capable exhibiting all the cognitive capabilities we have, including, you know, things like true innovation and creativity and planning and reasoning and things like that. And I think we're about five to 10 years away from that, but that's, you know, pretty incredible, if you think about how transformative a technology this is
You mentioned there might need to be some more technology breakthroughs. We're seeing things like the models advancing. We're seeing the semiconductors advancing rapidly as well. Are there any currently bottlenecks and things you need to figure out? I know energy is something that's been brought up so much saying, Well, look, we can keep advancing chips, we can keep advancing models, but at some point we're just not going to have enough energy to run these data centers, to run these AI models.
Yeah, yeah. Well, ook, there's, there's lots of physical constraints. So of course, there's, you know, no one ever has enough chips. And you know, we're lucky that we have, you know, our own TPU range, in addition to GPUs and but there just aren't enough compute chips in the world, really, for the demand. And of course, in the end, that comes down to energy as well. There's this idea of energy will be effectively is synonymous with intelligence as we get into the era towards AGI. Now, the interesting thing is, I think the AI itself will help here, in the sense of getting more efficiencies out of existing infrastructure, but helping with things like material design better better solar materials, but it could also help with new breakthrough technologies like fusion. We, you know, we have a collaboration with Commonwealth fusion in the US to help contain plasma and fusion reactors. And one of my pet projects is, can we come up with a room temperature superconductor material using AI? So I think there are multiple breakthroughs that AI could come up with and help us come up with that would help with the energy situation. In fact, indeed, that's, I think that's one of the most promising use cases of AI. And then the other thing is, as these systems are getting better, they're also getting, you know, 10x more efficient per year. So if you look at our range of models, we have our kind of Lighthouse model, our Pro versions of Gemini, but then we have our flash versions, which are way more efficient, and the sort of workhorse models that are used for everything, and they use techniques like distillation, where you have a big model that teaches a smaller model, and the smaller model is really, really efficient. And I think there are more and more innovations and techniques like that that will keep bringing the efficiency curve down, and so you get, you know, much better performance per per watt.
We hear a lot about sort of AGI, and I think there's a lot of people wondering, technology, sounds amazing, sounds great. There's also a lot of fear right around the proliferation of this technology and the impact it's going to have on on people every day and their lives. I guess for you, what are some of things we need to consider, from that perspective, in terms of the impact on society, whether it's around jobs, whether it's around kind of what we're going to do with our time if we reach this goal, versus, I guess, the benefits that you believe this technology is going to bring for humanity?
Well, of course, you know, I believe that overall, AI is going to be one of the most beneficial technologies humanity ever invented. That's why I spent my whole career working on it. But it's only, you know, it's not a given. It's a dual purpose technology. I dream about using AI for things like curing diseases. We have a spin out called isomorphic that builds it on, on Alpha fold work on protein folding, work that we did a few years ago to accelerate drug discovery and try and solve all disease. I think that's now, you know, within reach, that type of thing in the next decade or two. We've discussed energy. There's many benefits, I think AI is going incredible benefits AI is going to bring. But there are also risks. Obviously, there's kind of economic disruption, and I think there it's going to be like the industrial revolution, but maybe 10 times bigger, 10 times faster. So, you know, it's incredible amount of transformation, but also disruption that's going to happen. And you know, we need some new economic models, probably for that. And then on terms of the the worries about the usage of AI, I have two which I think are worth worrying about. One is bad actors repurposing these general purpose technologies, AI technologies, for harmful ends. And then the second one is AI itself, as it we get towards AGI and agent based systems. So these are systems that are able to do things more autonomously than than today's systems. They can, you know, what are the guardrails around that? How do we make sure we can keep them doing the things that we want them to do, and not veer off into something that we didn't expect. And so those are the two kind of risks that are kind of that I foresee.
Do you feel that you've got systems that, or you're developing systems that you can be in control of?
I think we're, we're very confident about that. You know, we've had and thought about responsibility and safety and security of these systems in the very beginning. You know, we started DeepMind back in 2010 almost no one was working on AI back then. But we planned for success, and we knew success would mean these extremely powerful systems. So we also understood the the other side of the coin of that. So from the very beginning, we've tried to be very thoughtful use the scientific method and a scientific approach to try and understand as much about our systems we're building before we deploy them. Of course, that doesn't mean we won't make any mistakes. There's two, it's, it's such a incredible and fast moving technology. But I think with with something like AI, we need to be, you know, I call myself a kind of cautious optimist. I'm, I'm very big believer in human ingenuity. I think given enough time and care, we'll get this right as scientists and as a society. But it's, it's not a given, and so we shouldn't be sort of rushing into this and we need to go into it with our eyes open.
Because, I guess the reason I asked, because I know you've spoken to people like Yoshua, Bengio, Max Tegmark, and these are people I've also spoken to and and they're of this cohort that please do do we need to be rushing so quickly into a world of AGI and agentic systems? Maybe we need more tool based AI AI to solve specific things, rather than these all purpose or general purpose kind of systems? And I know they've called for, for perhaps a slowdown to the development of of these AGI systems. In your view, do you think you should be slowing down?
Well, I've had lots of you know, I know them very well, Yoshua And Max, we've had many discussions and many others. And actually, I have some sympathy for that view, that, you know, building a tool based AI is, you know, thinking of AI as a tool, or the ultimate tool for, say, science is the right way to build AI in the initial stages. And certainly that's the way we're viewing it, and the kinds of things we apply AI to, like Alpha Fold. But the thing is, you know, it's a very complex geopolitical and corporate system that we're in. And it isn't just about, you know, there are many companies trying to build this. There are also many nations trying to build it. And it's, there's a sort of race dynamic, which I ideally wouldn't be there. So in an ideal case, this would be a scientific endeavor, and it will be very carefully, each step would be carefully considered. But unfortunately, the the prac that the real world isn't isn't like that, and we have to kind of be pragmatic about where we are. So what we're trying to do is be good role models for Yes, being on the frontier, pushing that the benefits of that as quickly as we can and as broadly as we can, but also try and be as responsible as possible with that along the way and thoughtful as possible. And I think we've got that balance pretty, pretty good right now. And hopefully that's a bit of a role model to the rest of the field in the industry, too.
Yeah I want to address some of those dynamics as well. But just just first, I guess, just from a personal point of view. Have you ever you said you sort of started this mission of Deep Mind? You know, you believe in the technology, but has there ever been any moments in your career when you're like, should we be doing this?
Look you when you look at how powerful the technology is, I really think that there are so many challenges confronting society today, not to do with AI, climate, poverty, you know, the access to water, there's a there's just so many issues health, aging, population, disease. So like, you know, energy we talked about earlier. So if a some, if I if there wasn't a technology transformative as AI coming down the road, I'd be really worried about society's ability to deal with these challenges. So interestingly, AI itself is one of those challenges, maybe one of the greatest ones, but it's also one which can help us cope with and resolve and solve some of these other big, grand challenges. So it's a very interesting one, right? It's sort of double edged, and I've always believed in that. I've always thought that, in the end, it would be the most important technology we'll ever invent. And I think it's sort of the natural progression, really, of the computer age.
Demis, you just, just a quick aside, you started life in gaming, which is amazing. Yeah, codeveloping theme park. Yes, fantastic, fantastic game as well. Did you ever do you still play games?
Yes, I love games. It's my hobby really well, like these days, like League of Legends, with my two boys and my brother and we have a little team. We've done it since lockdown. But, yeah, I love games in all its forms from football to video games
in such a high impact, stressful role as you have, potentially, is that your unwind?
It is, it is, I would say so. And it's also, you know, it's, it's a kind of in the past as well, has been a great creative endeavor for me, you know. And it's how I learned programming and other things was, was through making games.
I have nowhere near as stressful job as you but that's my unwind too, for sure. Get home, turn the console on.
Just in that small segment alone, Steve, there's so much to unpack, and I want to focus on on two kind of big buzz words right now. The first is artificial general intelligence, or AGI, this idea, and I know there's so many different definitions of it, but broadly speaking, this idea of AI that that is as smart or smarter than humans. And I think that so many of these big AI labs, including Open AI, including Deep Mind, are pushing and hoping to get to this stage of AGI, and so far, they've approached this with a technique called large language models, these AI models that are trained on huge amounts of data, but mainly text. But there's this other buzzword, right? World models, this idea of these AI models that understand the physical world. And this is this. This buzzword is really growing in popularity, right?
Yeah. And I think this is going to be a big theme of AI going into the rest of 2026, and even into next year. Because the idea here is that, LLMs, sure we got the language part down. It can mimic the way humans talk and speak and write and things like that. But when it comes to the physical world, you know, we talk so much about robotics and AI and physical AI, well, they need to understand how the physical world works, how water flows, how air moves and things of that nature, and what really struck out to me when you brought this up to Demis he said, Yeah, we do need to start exploring that more. And in fact, he sees a world in which the LLM and those world models start to converge. I think that was the word he used, converge into something more unique and more powerful and capable. This is also a debate that's been playing out among AI leaders, like on social media, you could fire up X or your favorite social media site, and what really struck out to me is Yann LeCun. He was the head of AI for many years over at Meta. He recently left to start his own thing, because he kind of got superseded by Alexander Wang and that whole big talent wars that happened over last summer, he had a really interesting interview in the Financial Times, he doesn't think LLMs are what's going to get us to AGI, to your point, that's what everyone's chasing. The super intelligence AGI, whatever you want to call it. His thing is LLMs can only get you part of the way. You need world models and all sorts of other things. And he kind of harshly criticized meta for not thinking beyond the LLM and that seems to be part of the reason why he left to do his own thing. And it's really interesting to see one of Meta's big competitors, Gemini, just talk openly about it and say Demis saying, Yeah, we need to do this. We need to start thinking about this. It enables so many things, from robotics, autonomous driving and just a better understanding for these AI models and intelligent systems that we're chatting with to get you that right answer. Steve, do you ever use a chat bot and you put something in and it will say, Hey, Steve, great question. That's a really clever thought
All the time. That's all that's that. It's the sycophancy of all these chatbots, right, where they're like, Oh, you're so smart and great at asking me these questions. Yeah, all the time, Exactly Because the reason I bring that up is partly to this point, this growing criticism of LLMs is that actually, yes, they're great and they'll give you the information. But actually, when it comes to LLMs as a foundation for being able to create new ideas, novel ideas, there's limitations there, and I think that's partly what Demis was speaking to and why this idea of world models is really growing in popularity. It's going to be interesting to see how this plays out, as you mentioned, into this next phase of AI, where it's key for things like robotics, driverless cars, and many other use cases too.
Yeah, and you'll you'll notice, as we continue this podcast, I am incredibly cynical about the robotics angle of this AI moment we're living in all that so many of the robots we're seeing, they're literally puppets. They're tele operated. The best example, of course, is the Tesla Optimus robot, which started out as a man in a body suit dancing around. Now it's a real robot, but again, it's tele operated. There are literally people at a control room controlling it over the internet and even using their voice to talk to you and things like that. So we are , the robotics people I talked to, we had one in the office just a couple weeks ago, and they said the hardest part isn't building the actual robot, it's training it, and that's where these world models are going to come in, so they can actually operate autonomously, like we've been promised. Perfect.
Demis, you mentioned some of the dynamics at play, right? And competition commercially, of course, is one of those. We've got Open AI, we've got Anthropic, we've got all these different AI labs out there. It's intense. And Gemini three has had such good reception so far. But there was a point people were doubting Google as a whole and its ability to compete. And I'd say a point. It was at some point in 2025 and it wasn't that long ago. And then, you know, Gemini 3 really came out and impressed a lot of people as well. But it's a space that's ever changing. So how we how would you assess right now, the competitive environment? How do you feel it?
