Demis Hassabis: Agents, AGI, and the Next Big Scientific Breakthrough

When Demis Hassabis made AlphaFold available to scientists around the world, he did not patent it. For a leading AI company, this is an almost inexplicable move. Yet today, every new drug goes through AlphaFold—and its free distribution has accelerated science more than a thousand academic papers. The real revolution here lies not just in the power of AI, but in how it is distributed and used. Everyone thinks that artificial intelligence is a race to build the biggest brain. But Hassabis thinks the opposite: the challenge is not just to build bigger models, but to make them truly useful, accessible, and capable of learning the way we do—continuously, without starting from scratch each time. Today, most models operate in a “stateless” mode, meaning they forget everything between sessions: each prompt is a fresh start, as if they had never learned anything from the past. Yet the human brain works differently: during sleep, it reprocesses and consolidates key experiences—an aspect that Hassabis studied during his doctoral research on the functioning of the hippocampus. When DeepMind designed the first program capable of beating Atari video games in 2013, the real breakthrough came precisely from drawing inspiration from this biological concept: the model “replayed” the best games internally, learning from the winning sequences. But today, even with gigantic memory windows—millions of tokens—most AI systems collect data in a crude way, without distinguishing between what really matters and background noise. And there is a tangible cost every time the model has to search for the right information to make a decision. Here, Hassabis makes a startling comparison: despite having enormous “memories,” current models are actually less efficient and selective than a human being with a limited but well-organized memory. And here comes the game-changer: it is not enough to increase the power or the amount of data. What is needed is a new quality in learning—the ability to adapt to the context, to reason over the long term, and to build a personal and selective memory. A practical example? Despite the progress made, no AI agent has yet created a game capable of dominating the world rankings, even though today anyone can prototype in half an hour what took seventeen-year-old Hassabis six months. What is still missing are the “soul” and the profound creativity that give meaning and value to a work: the human drive that transforms a tool into a masterpiece. Yet, the dividing line is becoming increasingly blurred. Hassabis predicts that, over the next 6–12 months, we will see teams capable of multiplying productivity a thousandfold using AI agents, before full autonomy actually arrives. And the question that arises is: What happens when these capabilities are no longer in the hands of just a few labs, but are widely available in open-source models that are small enough to run on a phone and powerful enough to solve real-world problems? In this regard, the Gemma model, which was made open source and downloaded 40 million times in two weeks, is just the beginning. For those who want to achieve the next scientific breakthrough, Hassabis offers two rules: look for problems that have a monstrously large search space—such as protein configurations or moves in Go, where no brute-force algorithm can compete—and define a clear objective, an “objective function” that you can climb. Then you need either a lot of real experimentation or enough data to simulate the universe you want to explore. In this context, the new AIs will not be mere “problem solvers” but could become co-scientists: capable not only of finding answers but also of posing radically new questions, such as “inventing” the next problems of the millennium that are worth solving. Hassabis even proposes a test: train an AI using data from before 1901 and see if, like Einstein in 1905, it can “discover” special relativity on its own. The ultimate goal is not just to solve difficult problems, but to create new frontiers of knowledge. And there is a surprising element to this approach: Hassabis does not believe in a “single superbrain” that encompasses everything. Instead, he envisions general systems that use specialized tools—something like an AI orchestra, where each model excels at one thing but collaborates with the others. The idea we need for the future is not just more power, but more coordination and more “dexterity” in using tools. If you’re wondering what the difference is between a startup that genuinely does science with AI and one that simply packages an API, the answer is this: real impact comes from those who combine in-depth expertise across multiple fields—not just computer science, but also physics, biology, and materials science. Hassabis closes with a piece of advice that sounds like a challenge: “Hard problems aren't more complicated than easy ones. They're just difficult in a different way. If you only have one life, spend it on something that would change the world if you weren't around.” General intelligence may arrive while we are still on this journey. But the real question is: What would you build today, knowing that tomorrow your tool could change the rules of the game while you're still playing? True innovation lies not in the size of the model, but in the quality of the questions it enables us to ask. If this vision has opened up a new perspective on AI and science for you, you can click I'm In on Lara Notes—it's your way of declaring that this idea now belongs to you. And if tomorrow you find yourself discussing AlphaFold or question-generating models with someone, you can tag that conversation with Shared Offline: because truly powerful ideas always start with an in-person conversation. This Note is based on an interview conducted by Y Combinator with Demis Hassabis, and it has saved you 37 minutes.