Max Jaderberg and Tuomas Sandholm on AlphaFold, Games and Scientific Discovery

Lauren Crichton introduces Max Jaderberg of Isomorphic Labs and Tuomas Sandholm, a researcher in computational game theory. She outlines protein structure prediction, imperfect-information games and Sandholm’s work on kidney exchange before Joel Hellermark joins them.

Hellermark asks how Jaderberg moved from StarCraft to medicine. Jaderberg says deep learning, generative models and reinforcement learning form a shared toolbox: games test the methods, while biology and chemistry offer another domain.

At Isomorphic Labs, he says, models represent proteins, DNA and RNA, while drug design seeks molecules that can bind to and modify disease-related proteins. The intended output changes from game actions to candidate molecules.

Jaderberg recalls AlphaFold beginning as a hackathon experiment before becoming a multiyear research program. He points to the CASP blind protein-structure prediction competition as evidence of the jump in performance with AlphaFold 2.

AlphaFold 3, announced shortly before this 2024 conversation, extends structure prediction to interactions involving proteins and other molecules such as DNA, RNA and small molecules. Jaderberg says this helps researchers study molecular machinery rather than isolated proteins.

He recalls experts expecting breakthroughs in games and protein folding to take much longer, even as results were emerging internally. He stresses that drug design still requires many further advances; AlphaFold 3 is not the end of the process.

Jaderberg expects broader foundational models for biology and chemistry to complement earlier models fitted to small datasets. He describes structure prediction as a task from which useful capabilities for molecular interaction and design may emerge.

Why not simply use a language model? He says much scientific information is not prose: laboratory measurements, sensor outputs and simulations require suitable data representations and architectures.

Specialized models can incorporate relevant physics, chemistry and biology while reusing some core deep-learning components. Jaderberg hopes future systems may eventually represent much more scientific data in a common form.

Hellermark turns to Sandholm’s poker research. Sandholm says imperfect-information games matter because many real interactions involve other agents and hidden information, unlike chess with its visible board.

Poker also had a long history as a game-theory benchmark. Sandholm’s team pursued superhuman performance in heads-up no-limit Texas hold’em, building on a benchmark the AI community had already chosen.

He says Libratus learned through self-play without reading human poker textbooks or game records. It rediscovered familiar tactics but also differed from common human advice.

One example is limping: calling the blind instead of raising or folding. Sandholm says their systems used the move in a minority of cases even though traditional advice often discouraged it.

For real-world applications, Sandholm begins with two-player zero-sum settings because game-theoretic guarantees are especially clear there. He says an appropriate strategy can protect against exploitation even by surprising opponents.

He argues that strong game play needs specialized search and reasoning, highlighting imperfect-information subgame solving. In his view, a general generative model alone cannot match a purpose-built game system.

Jaderberg discusses possible ways to give language models more computation: search can explore and score alternative continuations, while diffusion-style generation might refine an entire sequence. At the time, he notes, next-token prediction remains the approach scaling most successfully.

Can systems trained on existing data create new science? Jaderberg points to AlphaGo’s famous move 37 against Lee Sedol as a surprising move whose strategic value became clearer later in the game.

He says generalization and verification are central, though he does not claim a complete explanation of how novelty emerges. Sandholm distinguishes learning from historical data, optimization and simulation-based self-play as different routes to AI solutions.

Sandholm values self-play where data is scarce, secret or misleading. He argues that systems need not inherit all the limitations of human examples when they can experiment within a useful simulator.

Asked why language models have not yet disrupted more of the economy, Sandholm expects broader disruption but argues that expert tasks will need specialized AI that can generate knowledge beyond existing human material.

Jaderberg says scientific models can make mistakes, just as navigation systems do. He compares laboratory checks to looking at road signs: an inexpensive experiment or human check can verify a proposal more easily than discovering it from scratch.

On whether to wait for AGI, Sandholm argues that useful systems will need specialized components such as strategic reasoning engines. He is skeptical of AGI; Jaderberg expects it eventually but does not claim a reliable date.

Jaderberg argues that urgent scientific and medical problems justify focused work now rather than waiting for general intelligence. His optimistic medical timeline is a prediction, not evidence that cures have already been achieved.

Looking ten years ahead, Sandholm expects a broader understanding of which AI methods suit which applications. Jaderberg imagines progress may feel incremental locally while producing dramatic differences over a decade.

Sandholm closes with a question about science itself. Some computed solutions and theoretical descriptions may be too complex for people to understand directly, he argues, challenging publication and the use of knowledge when no short human-readable account is available.