Thanks Alistair for that lovely introduction. It's so great to be back at Cambridge. I always have a warm feeling coming back here. This lecture hall is the first I was in, and I remember telling a friend that maybe one day I'd come back to give a lecture announcing AGI. I'm not doing that today, but maybe in a few years. Cambridge has inspired my whole career. My journey with AI started with games, especially chess. I played from age four and became fascinated by the mental processes behind it. I experimented with AI programs on an Amiga 500 in my teens and was hooked. That led me to Cambridge, where I was inspired by stories of Crick, Watson, Turing, and Babbage. I felt the intellectual giants of the past speaking from the stones. In 2010, we started DeepMind as an Apollo program for AGI. Our mission was to solve intelligence and then use it to solve everything else. We began with games, mastering Atari games in 2013, then took on Go. AlphaGo beat the world champion in 2016, watched by 200 million people, and invented new strategies. AlphaZero generalized this to any two-player game, even discovering new chess styles praised by Kasparov and Magnus Carlsen. These games were a training ground for real-world problems. We look for problems with massive search spaces, clear objectives, and data or simulators. The protein folding problem was perfect. Proteins are the building blocks of life, and predicting their 3D structure from amino acid sequences is incredibly hard due to the vast number of possibilities. We entered AlphaFold 1 in the CASP competition in 2018 and improved accuracy dramatically. AlphaFold 2 achieved atomic accuracy, and the organizers declared the problem solved in 2020. We then folded all 200 million known proteins and made them freely available, accelerating research globally. Over two million researchers use it, and it's been cited over 30,000 times. It's being used for plastic pollution, antibiotic resistance, neglected diseases, and fundamental biology. We've since released AlphaFold 3 for interactions and AlphaProteo for designing novel proteins. We take safety seriously, consulting over 30 biosecurity experts. It's been an incredible journey, and I hope many of you will use these tools in your own research.
Particular job, a particular function: what is the amino acid sequence and the genetic sequence that will give you that structure? So it's kind of like running it in reverse and trying to design new structures that will do novel things, and again could be extremely useful for designing drugs and things like antibiotics and antibodies. So taking a step back, what are the implications of all the work we've done in the last 15 years for science and machine learning? It's all about making search tractable. You have an incredibly complex problem with many possible solutions, and you need to find the optimal solution, like a needle in a haystack in an enormous combinatorial search space. You can't do it by brute force, so you learn a neural network model that learns about the topology of the problem to efficiently guide the search. This is an incredibly general way to approach a whole myriad of problems. For example, we used it to find the best Go move, but you could also change those nodes to be chemical compounds and now you're trying to find the best molecule in chemical space, which is the beginning of drug design. We're using very similar techniques to design molecules now as we move more into drug discovery. I think in biology, we're entering a new era of digital biology. Biology at its most fundamental level is an information processing system trying to resist entropy. AI is potentially the perfect description language for biology, just as math is for physics. AlphaFold is a proof point of that. I hope in 10 years it won't be an isolated breakthrough but will have heralded a golden era of digital biology. We started Isomorphic Labs to build on AlphaFold and reimagine drug discovery from first principles with AI. Currently it takes an average of 10 years and billions of dollars to develop a drug. Why can't we reduce that from years to months or even weeks, just like we reduced protein structure discovery from years to minutes? We think of this as doing science at digital speed. My dream is to create a virtual cell, like a yeast cell, where you can run experiments in silico and the predictions inform real-world experiments, reducing the expensive search in the wet lab. AI is applicable to many fields: health, materials, fusion, algorithms, weather prediction, quantum computing. I encourage universities to think seriously about multidisciplinary work applying AI to specialist fields. On the path to AGI, we've been making advances in world models. Our video model V2 is state-of-the-art, generating videos from text descriptions. It learns real-world physics just from watching YouTube videos. We've gone further with Genie2, which can generate a whole playable game from a text instruction, though currently only consistent for a few seconds. We're working to extend that to many minutes. We've also been working on safety from the beginning, with systems like SynthID to watermark AI-generated content. AI has incredible potential to help with our greatest challenges, but it's important to engage with a wide range of stakeholders. Governments have become interested, and we've seen international summits. My shorthand is that instead of 'move fast and break things,' we should approach this transformative technology with humility and respect, using the scientific method. We're building our own multimodal models, the Gemini series, and working on universal assistance with Project Astra. The next step is combining agent-based models with general world models for planning and robotics. I conjecture that classical machines can do a lot more than we previously thought, as AlphaFold shows. Any pattern that has physical structure can be efficiently discovered and modeled by classical learning algorithms. This has implications for quantum mechanics and fundamental physics. AGI could be the ultimate general-purpose tool to understand the universe and our place in it. Thank you.
Yeah, there's many. I studied memory and imagination for my PhD. I wanted to understand how the brain does future thinking and planning, and it turns out the hippocampus is involved in both, so we could mimic that with algorithms. There are big questions around creativity, dreaming, consciousness. Building AI and comparing it to the human mind is one of the best ways to make progress on root node problems like the nature of consciousness and whether there is something special about the substrate of the brain versus mimicking it in silicon.
Uh, hi. Um, I have two questions actually. So, since Deep Mind was founded before the deep learning revolution, I wanted to know what your state of mind was had deep learning not picked up or how were you going to progress? That's the first question. And second question is since you've had intimate experiences with such challenging problems, such high dimensional problems and we know that gradient descent and its variants can't converge to the optimum solution only locally optimum solutions. Were you surprised that anything works at all in these systems at any point in time and do you think that most of nature is kind of suboptimal and so we can potentially build a more optimal nature?
So look, they're both great questions. The first one: we called it DeepMind partly because deep refers to deep learning. Deep learning was just becoming common with Boltzmann machines and hierarchical neural networks from Geoffrey Hinton. It seemed super promising. We also bet on reinforcement learning, which was important for solving something like AlphaGo. You need both: deep learning to model the environment and reinforcement learning to make plans and take action. We bet on that even at the beginning because we knew classical expert systems would not scale. I learned here and at MIT that those methods would never scale to the problems I wanted to solve. Learning systems seemed to have unlimited potential, though they were harder to get working. We also saw the computing paradigm shifting with GPUs. All those influences came together in 2010. We were betting not because we knew it would work, but because we were confident the other methods would not. On the second question: it is surprising that some of these things converge. For the first couple of years, nothing worked. We couldn't even get a point on Pong. We wondered if we were 10 or 20 years too early, like Babbage. But then it did converge, and that gave us confidence. As for nature, I think things are not suboptimal; they are probably pretty optimal because they've gone through evolutionary processes. If they are stable over time, there is structure that is learnable.
Yeah, I love that area. I've followed it and helped people building EEG caps. The problem is resolution and the need for read and write. I'm fascinated by projects like Neuralink. Right now it's for medical applications like helping people walk again. Beyond that, if it becomes routine and safe, it could be one way for us to keep up with technology. In some senses, it's no different from having our phones with us 24/7. We're already symbiotic with technology. It would be one step further, but I'm not sure about the boundary between attached and carried.
