I'd like to welcome up an amazing return guest here for us, and we always have great conversations — the COO of Google DeepMind, Lila Ibrahim. Thank you. You know, this moment, like all these other moments we've had in AI, it feels like, you know, you hear so much in terms of, you know, super AGI, super intelligence is around the corner. You know, at the same time people are like, it's overhyped. And you're in such an interesting place because you're really looking at the big things AI can do around science, drug discovery, all these things. Maybe a good place to start us off is to talk about where are we, what is actually happening today with the technology that your company really helped pioneer?
Well, thank you, Ana, for having me here and starting off with such a simple question like that.
Here, let's see. So, you know, I think we're still in the very early days of artificial intelligence in many ways, even though it's all around us. I see some folks on their phones — if you have an Android phone and you have a longer battery life, you can thank AI for that. Your movie recommendations, song recommendations, that's all driven by AI. We just don't call it AI. And so I think in many ways there's parts of AI that are really well integrated.
I have heard people say that, you know, AI is always — when something is new, like once it's actually in use, people don't call it AI.
Right, exactly. And that's why we're still at generative AI, because we're still learning how to make it really useful. And so I think what's been interesting the past few years is generative AI has really captured the imagination of what we might be able to do with a next-generation technology tool. And what we've seen over the past year was AI became more multimodal. That meant you can put in text, you can put in voice, images, you can get out video. So this aspect of having many ways of interacting with the technology, much like we do with other human beings, has really changed how we think about generative AI. But generative AI — and there's still a lot to do, by the way, on generative AI. Anyone type in anything in generative AI and get something not factual, right? Yeah, I see a lot of heads nodding. We need to make it more factual. We need to think about how do we bring in things like memory or reasoning in a multimodal way as well. So there's still a lot of work to do in order to make it more of an assistive technology. It's a lot of value today, but there's still work to do, and there's work around the fundamental research of how to improve AI as a field, as a science in itself. The area I'm particularly excited about is using AI for scientific discovery. Much like, think about it as like the next generation of a microscope or telescope — like, how can we understand the universe around us? And if we better understand the universe, can we unlock possible solutions to things like healthcare and diseases, to climate, to food security and crops. So I think there's a lot of really interesting things right now. We've already demonstrated that we can use AI for scientific discovery.
And where are we with that? Because I hear that all the time. I hear it's going to help us discover new drugs, find new materials, that sort of thing. And then I hear people say it hasn't actually discovered anything new yet. Like, it's, you know — not which isn't a dig in the sense of, you know, it's programmed with existing human knowledge. It's great at finding things that we already know, but it's hard to think of today an LLM — you know, is it true discovery if it's finding a pattern that we haven't seen? My sense is that's really what we're talking about, right?
Yeah, and maybe what I'll do is I'll give an example of AlphaFold, which is an advanced AI system that we developed about four years ago. And it predicts the 3D structure of a protein. Now, why is that important? Because diseases like, let's see, Parkinson's or even malaria — these are protein-based diseases. So we can also use proteins to help deal with industrial waste. So there was a 50-year grand challenge in the field of science of, can you predict a 3D structure of a protein? And if you can, you can understand it as a basic building block of life and how to address some of these issues. Every year — it would take, actually not every year, it would take a PhD student about four to five years to understand a single protein. We published a database with all 200 million plus known proteins known to humankind, where it's as simple as like a search or a map search. You type something in, it gives you the predicted 3D structure, and made that freely available to science worldwide. What's significant about that now is we've basically taken a billion years of PhD time and shrunk it into a database that's accessible. And in fact, we have 2.7 million researchers — who knew there were that many — looking for protein folding in 190 countries, using this tool now to advance things like understanding diseases, which we can now work on therapeutics, including some of the neglected diseases in the Global South that were underfunded. There's also work happening on like, why are some crops more resistant to disease than others? And then work, as I mentioned, in dealing with industrial waste. The University of Portsmouth in the UK is actually working on plastic-eating enzymes and accelerating their work in that space. So we've actually been quite surprised by what this has unlocked and being used in areas that we could never have imagined. In fact, we're also excited because last year it was awarded with a Nobel Prize — the first time in history that a Nobel Prize was awarded to an application of AI. And I think this is entering a new era of discovery now. So your question of like, how far are we — I think we're, you know, we're still somewhat in the early days, but we are showing progress in this. And it's not just protein folding. There's work happening with like 15-day weather forecasting, which you can now predict the trail of a hurricane. There's work happening in materials science discovery now in the hands of researchers worldwide, saying which ones can help us actually advance battery technology for electric vehicles, work with supercomputers. So I think we're kind of in this new era, new era of discovery.
And a lot of the talk this year is agents — you know, really allowing these tools not just to give information back to humans, but taking action on their own. How do you feel in terms of, are we really ready to give — as you pointed out, it's still sometimes, you know, giving us hallucinations. I still feel like if I use Gemini, if I use ChatGPT, any of the tools, I'm going to get back a lot of useful information. But if I really am going to rely on it, I got to go do a quick Google search and make sure it's not making something up. It feels to me like an agent isn't going to do that.
It could, like, I mean, that's the thing — an agent could go back, do a Google search, say, make sure.
But how ready is the technology, the underlying technology, to be given these new powers?
Yeah, I think there's two parts to that: how ready is the technology, and how ready are we as humans. So, and this last year we did a technology demonstration around Project Astra, which is a research platform which introduces agentic features. And this basically means, as Ana mentioned, being able to take actions, but you as the person in still has control and approves any action that's taken. And that might be something like you're planning a trip and then you want to add it to your calendar, or you want to purchase tickets, activities like that.
