How does Google DeepMind operate and make bets? And what's making Google more experimental? Let's talk about it with two Google leaders right after this.
Welcome to Big Technology Podcast, a show for coolheaded and nuanced conversation of the tech world and beyond. We have a great show for you today because we're going to go deep inside the way Google's AI and technology research operations work. We have two great guests with us today. Lila Ibrahim is here. She is the chief operating officer of Google DeepMind. Lila, welcome.
And we're also joined by James Manyika. James is the SVP of research, labs, technology, and society at Google. James, welcome.
Well, thanks for having me.
And of course, this is our concluding conversation here in our series at Davos. And we do have a live audience. Live audience, make some noise. Let them know you're here.
All right. So much to get to, not a lot of time. Let's just start with the way that Google DeepMind operates. Demis Hassabis, the CEO of Google DeepMind, who was recently on the show, has described DeepMind as sort of a modern-day Bell Labs. But what does that mean, Lila? Can you tell us a little bit about how the research is — is it a lab, operation, company? How does it operate?
Yeah. Well, maybe I should start with our mission because I think everything is kind of based off of that, which is to build AI responsibly to benefit humanity. And so, the first thing we do is take really ambitious research agendas. We structure it in a way where we're looking at what are the big problems but not telling people how to do it. And when you think about how did we first approach that, it's really about taking inspiration from the golden era of Bell Labs, but also government programs like the Apollo program and even more recently Pixar. So it's all focused around bringing in really great talent and creating an environment for them to succeed and to explore. So first thing is that big research agenda, telling people what to kind of the area to focus on but not how to do their job. The second thing is really because it's such a broad agenda, we want to build interdisciplinary teams. How do you create a culture where you can have a bioethicist next to a computer scientist and a neuroscientist because we think that's really where the magic happens and unlocks the work. And you know, this type of approach has resulted in such extraordinary efforts and we're also not afraid to explore and then say when is it time. I think Demis has a remarkable way of measuring time, like time to explore, are we setting the really ambitious goals, how are we doing progress towards that, and also not being shy to say, okay, now's the time to take a step back and pause it or double down. Great example of that is over the past few years we've been doing a lot of work around one science area, learning science. How do people learn and can we improve it?
And then this year was really Demis was like, okay, Gemini is good enough, it's time to infuse everything we've done with the industry around learning science into Gemini. And that was one of our focus areas to really advance how Gemini could be provided for learners. So there's something I think quite magical within Google DeepMind about timing.
Okay, GDM, I guess. We're gonna go everybody in the tech industry.
I almost — I almost caught myself for saying it.
So let's — but let's talk about that. So the way I just want to talk through process a little bit. The way that you just described that, Demis said that Gemini was ready for learning and then Google DeepMind started to pursue it. How much of what Google DeepMind works on is, you know, top down versus bottom up? A way that I've heard OpenAI describe the way that it works is like a bunch of different startups within a larger company. Is that a similar way that Google operates or does it come more from the top?
Well, I, you know, because our mission is so ambitious, you know, we're really trying to understand what are the big challenges where AI can help us unlock our understanding of the universe around us and solve some of humanity's biggest challenges. And it's broad enough that we can do things like how do we do weather exploration and try to predict weather — yeah, weather forecasts. How do we do AlphaFold and protein structure prediction to help us better understand diseases so we can come up with better therapeutics? Generative AI, how can we continue to improve that to make people's lives better? So again, we take a very broad portfolio perspective, but we allow the space for researchers to explore and that's really what I meant in the beginning of like, you've got — we've got to find the right talent. So mission-driven culture and values-aligned people who want to have this type of exploration and a big impact and scale that we can have of being part of Google. So I would say some of this is — Demis is quite remarkable in terms of his thinking in this space because he's been doing it for so long, right? DeepMind was founded 16 years ago. It's been kind of a lifelong mission of his and yet we have an organization full of people who are creative, who like to work in an interdisciplinary environment, who want to have impact in this world. So they also come up with their own approach to things, setting — setting...
Pardon me. Yeah, a little bit of both.
Some top down from Demis and then some bottoms up.
Okay. And which makes managing part of that organization structure...
You about talent? I will talk with you about talent for sure. And, you know, on that note, how have things changed? Because I'm just going to talk about the tech industry more broadly, that it seems like there used to be a moment where a lot of tech companies gave, you know, these talented people broad leeway to explore things that might not have immediate results. Then all of a sudden we got into this AI race and many companies brought their researchers who were working on these long-term products much closer — or much long-term projects much closer to the product. And all of a sudden there was almost an imperative for long-term research to make immediate product impact. So has that changed as well over time? Is that something that's going on within DeepMind as well?
Yeah. You know, I've been — I joined about eight years ago. And we've definitely been on a journey, but what I think is so exciting about Google DeepMind and I think why so many of our employees stay so long is because we have that breadth of portfolio. So there are some people that want to continue the deep research, frontier AI research that they do, or a scientific — more focused on the science, and we have the space to do that exploration while also delivering on the advancements around generative AI such as all the progress we made last year with Gemini.
Okay. Let me take that a step further. The way that the transformation within Google has been described is that instead of having every product area or product group chart its own direction on AI, there's now this central engine room within the company, which is I think the AI division that generates — that creates the AI and then farms it out to these product areas. So can you talk a little bit about that process and how that works?
Yeah, and actually I think that's been one of the exciting things over the past few years with a combination of Google Brain and DeepMind of bringing the best of Google's AI teams and research together under one roof where we could explore such a broad portfolio. And so we've really been focused on, as you mentioned, becoming the AI innovation engine. And then I wouldn't say we farm things out to other Google teams. We collaborate very closely with the product areas and their customers to understand what the needs are so that we can build the models better from the start and do so in a very collaborative and responsible way such that by the time it goes to different Google products it's already been through a lot of that testing and can be refined for that specific use case. And that's actually helped us. I think what's resulted in that, for example, is like Gemini 3. We launched it and then it was available to a broad group of developers and users right away.
All right, one last question on this and then we're going to go to James. And James, thanks again for being here. So let me just ask you this. On our show, we have this hypothesis that Sundar spent time at McKinsey and this is sort of like a McKinsey-style approach to like reorg, centralize, and then work with the groups. Is there a truth to that?
Well, you have a former McKinsey person here who might be able to address the structure.
No, I think what you've got going on is an extraordinary thing, right? Because on the one hand, you've got the Gemini program, which underlies all of this, building the kind of large-scale models, Gemini itself, Gemini 2.5, 3, and all of that. And this kind of came about back three years ago when we put together the Google Brain team and the DeepMind team to create the Gemini program. So that program now underlies all the things across the company. So you see Gemini show up in Search, in Google Workspace, it shows up in all our products, in Notebook, and all of these things. So it's kind of the foundation and that's why, as Lila said, you know, Google DeepMind and the Gemini program has become the engine room. But in addition to that, you've got all these other things going on. There's deep science going on in the company. I mean, this idea of, you know, kind of this foundational kind of tackle the biggest root-node problems that open up research and innovation in so many areas. So you've got all of that going on too. And then you've got all these other, you know, special kind of ambitious projects working on things like Genie, who build world models. You've got work going on to build special things for Waymo and, you know, enhance the models that lead to Waymo, the driver — the Waymo driver. So you've got a lot of these things going on. So I don't think there's a top-down as much as let's take advantage of the foundation called the Gemini program. Make sure that every time we do these rapid iterations — you've seen we're now on a cycle of every six-ish months, there's a new generation of Gemini. Make sure it shows up immediately. As Lila described, there's no, you know, shipping delay. So the minute the latest version of Gemini comes out, you're going to see it in Search, you're going to see it in, you know, in the Gemini app itself, you're going to see it everywhere. So that's kind of the incredible thing that's happened over the last three years.
All right, I want to talk about Labs. So Google Labs, a lot of us who used Google products in the early days, you know, we saw this era of experimentation within Google and then Labs went away for a bit. Not that Labs was the only bit of experimentation within the company, but then Labs was revived and it seems like we're starting to see many more experimental projects come out of Google proper in a way that we hadn't seen in a long time. So how responsible is Labs for that? And why is Labs back?
Oh, Labs is so much fun. So what actually happened was three years ago, you know, this is kind of an inspired Sundar moment, said, let's reboot Labs and, you know, we're in this AI moment. How do we kind of explore and experiment and build these AI-first products that are totally AI-first? So the idea with Labs is let's take the most amazing research coming out of Google DeepMind and Google Research and any other place, quite frankly, in the company where there's incredible research and focus primarily on how do we build experimental AI-first products. I think what most people probably know of the most is what's now, you know, NotebookLM. You know, the way that started, by the way, is incredible because it was — I remember when I first encountered it last...
So what is NotebookLM and tell the story.
So Notebook is fun. So it just started out as a product called Tailwind. There were four or five people working on it and the idea was, you know, can we build an AI-native research tool that is grounded on what you put into it? So in other words, you know, your sources — you might have books, you might have papers, you may have drafts, you may have whatever your content that you want to ground it on, put it in a notebook and be able to engage with it. So that was the conception of the idea and in fact in some ways it got additional impetus from Steven Johnson, who's a writer. And Steven Johnson, you know, is one of these people who kind of keeps everything. So he has notes from the '90s and drafts of books and all kinds of things. He said, I'd love a product where I can dump all that stuff in and engage with it. What was I thinking in 1997? What was, you know, what was that draft I did? And so — and be able to — so what NotebookLM has become is this incredible research tool grounded on what you've put in. And when you engage with it and it summarizes or drafts something, it gives you these citations and that's in some ways a key feature of it. So if it says, Alex, you know, you said this or your source says this and summarizes in some way, it'll give you citations. If you want, you can click on the citations, they take you all the way back to the original content, right?
So which is incredibly useful.
Then a fun thing happened, which was, well, you know, so it's already a very useful tool. Then we said, well, actually, you know what, sometimes I want to hear my sources as opposed to just engage with them. So I said, okay, well, actually the technology is ready enough, we can actually add AI audio overviews.
Which is like effectively a podcast you can have it with like two hosts.
You could have — actually the origin of the idea wasn't even to do that. So initially the idea was that a few of us, you know, Jeff Dean and, you know, this legendary Jeff Dean said, well, actually, you know what, we're reading all these papers that are coming out at this incredible pace in the computer science field, it would be nice to be able to hear a summary of them verbally while I'm driving into work or something. So just, you know, then I can figure out which papers I'm going to read. So that was the original idea. They said, actually no, you know what, it's easier to learn stuff when you hear people talking about it, engaging — that's why seminars are interesting, right, as a learning mechanism. So that's where the idea came from. So we did these audio overviews, which, you know, in the form of a podcast or a discussion with two hosts discussing it, and now it's evolved and that's when the product just kind of took off.
Yeah. Whenever I give a presentation about AI, that's the party trick where I build one of these notebooks in front of the audience and then I hit play on the podcast and people who haven't seen it before — it's like a jaw-dropping moment. And in fact, we've had multiple people on our YouTube feed and coming from the podcast, they've asked, Alex, did they train on your voice? Because it sounds a lot like me. And I say, no, listen, it's — they always say, let's unpack this at the beginning, and you have to understand every podcaster says that.
So actually, you know, one of the most fun use cases of a notebook, by the way, is because now you can put in things in all kinds of formats — that can be papers, that can be YouTube videos, that can be whatever is on your hard drive. One of my fun use cases was actually when I had to do this thing where I was seeing all these papers from literally over 100 countries in different languages. So I put them all in and just engaged with content in multiple other languages because NotebookLM can handle multiple languages. And now you can do video overviews. So I think you can make like an animated — not an animated video, but a video with like graphics and slides. But I think this is an example of the kind of thing that happens in Labs where we try to take this incredible research that Lila and colleagues and others are doing at Google DeepMind and Google Research to say, how do we build amazing AI-first products. Flow is another example. And if you play...
I'll tell you a story about Flow, then I'll let you talk a little bit more about it. I just did my first and last mountain climb and it was Cotopaxi in Ecuador and I wanted to make a video sort of capturing the moment. But there were a couple of things that happened that I just could not — that I didn't videotape because I decided to spend the climb actually climbing as opposed to YouTubing, which is apparently from what I hear rare these days. And there was a moment where my water bottle fell out of my backpack and rolled down the glacier and then kind of disappeared into the darkness. And I wanted to illustrate that. So I went to Flow, the Google video generator, and I said I want to make an animation documentary-style to show this and slotted that into the video. So now you can — and before I would have to hire an animator, now you can do it.
Yeah. No, it's incredible. But I think what you know — Flow is an example of the magic that happens in Labs. So I remember a bunch of us got together. So Josh, who runs Labs, and, you know, Demis and a few of us said, what if we put all these tools we now have together into something that's actually useful? And in fact the initial version of it that we have, you know, in some ways was clunky. Then we said, well, actually let's just talk to some actual filmmakers and get their input. So one of the things that happens in Labs, by the way, is we try to engage a lot with creatives and others to help us think about how we build these tools. So anyway, that's how Flow came about.
Yeah. You can build scene-by-scene prompting into video and you can have continuation. I think that's probably where the name comes from. Flow.
And what you just said was an insight that came from filmmakers. In fact, the initial version said, no, no, what you've got here is actually not very useful. I'd like to be able to build things scene by scene and be able to stitch them together, be able to do this. So, you know, so that's why it's been helpful. So if you say what is Labs, it's a place where we try to experiment with all these things. At any one time, we probably have about 30 experiments cooking. So if you go to the Google Labs site, you'll probably see about 30 different things.
But I have a request for you. Broaden the access because there's a lot of projects in there that seem really interesting to use, but every time I'm there, it's waitlist.
You have to — we'll work on that. We'll work on that. I mean, so for example, you know, one of the other ones — and sometimes we're surprised what people find useful. I'll give you an example. One is Gemelli, which is a tool for SMBs. So imagine this is not your typical kind of techy startup SMB, but kind of a more traditional SMB where they want to build a web presence. And so you can literally engage with Gemelli as an SMB and be able to build literally a web presence in incredibly imaginative ways. So we always have all these things cooking in Labs. AI Studio is another example of the kinds of things — this is for developers. So we're trying to think of all these incredible creatives, whether they're developers, artists, filmmakers, musicians, to create these incredible AI tools.
Yeah, there's two that I really want to get access to and I think are potentially going to be big. Maybe the next NotebookLM. There's CC, which is an experimental productivity agent within Google, which looks great. And then Disco — you can build a web app basically based on links. So if you're like thinking about doing something for the weekend, you can just like open a bunch of tabs and then it will figure out what type of app to make for you. So maybe a custom map with dots for each potential event and then you can pick the dates that you want to actually be in the place that you're thinking about and then it will sort of highlight what's going to be available then. So this is to both of you. Back in the day, Google had this concept called 20% time where Google employees were basically empowered to spend 20% of their time on something that wasn't core to their job description. And that's where a lot of big Google products came out of. I think Gmail was a 20% project. So I want to ask you both about these experimental projects. Who builds them? And is a version of 20% time back or how does this — you know, obviously a lot of cool experiments. How is it happening inside the company?
