The Next Intelligence
Breakthrough

I'm so happy we live in this time and space right now. My background is in chemistry — I did medicinal chemistry. In drug discovery, a lot of people don't think about where our drugs actually come from. You start with target identification, then ligand-based design, then structure-based design. You're lucky if you have a protein structure — and at the time there weren't many. Some people's entire PhD was about crystallizing out a single protein structure. Where I'm excited now is with the release of models like AlphaFold. When it came out, it was literally a miracle, because suddenly you didn't have to do thousands of experiments and get lucky to crystallize an enzyme. And now with generative drug design, it's even more exciting, because you can be creative in ways you never thought about. The whole pipeline of drug discovery is so much faster than it's ever been.

Something I think about a lot as an anesthesiologist: when we put a patient under anesthesia, there's a middle state as they move from minimal to moderate to deep sedation to general anesthesia, and their bodies reveal themselves. For people with a history of chronic anxiety, depression, or PTSD, you put EEGs on their head — they're deep from an anesthetic perspective, so thought isn't happening, but you move certain parts of their body and their heart rate and blood pressure go through the roof. If I could do any experiment, I'd want to record everyone's subconscious — we do hundreds of thousands of surgeries a day. If you actively quantified that and did machine learning on it, could you map out someone's subconscious? Coupled with their conscious thought — which honestly might be their diary to ChatGPT — that would be a true picture of who you really are.

When I think about multimodal in biology, I think biology is made up of many different languages. You have grammar around DNA — DNA is our genetic code, with all the instructions for making an organism, and it follows rules. It turns out to be one of the most complex languages ever, and it's just one of them, because there's also proteins and RNA. They all stem from DNA but are specialized, and they all connect at the same time. That's been the challenge: people think we have to specialize in each domain — a model for proteins, a model for DNA, a model for RNA. But these things are all part of the same system. When disease happens, or when a drug is in your body, it's inherently a multimodal system, with a cascade of effects. The ability to model that is what's been holding back the field. That's why it takes $2.5 billion to make a drug over ten years.

This year we plan to preview progress on our next generation of models. We started off with genomics as a team. One thing we learned with the Evo projects is that DNA was not enough — it's a key building block, but it's just one of the languages. We want a single system to understand not just DNA, but also proteins, RNA, epigenomics, metabolomics — all of these omics. The idea is to fuse all of these modalities into a single system, and we believe this unlocks applications in every domain in bio, pharma, diagnostics, and synthetic biology. We're designing whole genomes as a company now. That's also one of the reasons we work on biosecurity — making sure the technology is safe, and actively defending against potential threats that biological AI enables.

The multimodal bet Google made — way before I joined DeepMind — was a very bold bet: that we could have all those modalities of data, and that handling all of it would enrich the model. It was incredibly hard. But if you look at the reasoning process, we think in terms of images, or in terms of dynamics. There was a strong conviction that to do something that would really be AGI, you need to work in all those modalities. Being multimodal adds a large burden to the model's complexity — the Gemini model needs to generate images, do speech-to-speech, understand legal, write code, design good user interfaces. So the fact that we see something excellent emerging out of that — there's a bit of a relief. Like, yes, this is working.

You look at Gemini writing code: it writes code, runs the application, and looks at the resulting user interface and interaction — which means posing a judgment on visual things that are very human-specific, and it works. If all you do is understand text, it's difficult to get a sense of what a good human experience is from a UI standpoint. I think this was a bold bet taken by the people who started DeepMind, and the correct one. And you see it in specialized things like biology — how much does knowledge of biology help in other domains? The story of science is induction: you see things, look at anomalies, but often it's transferring from one domain to another — analogous thinking, where you take something from biology and apply it to economics.