Accelerating Scientific Discovery with AI (Nobel Lecture)

Type: lecture Slug: nobel-prize-lecture-accelerating-scientific-discovery-with-ai—hassabis Sources: nobel-prize-lecture-accelerating-scientific-discovery-with-ai—hassabis Last updated: 2026-05-13

---

Summary

Hassabis (2024) delivered his Nobel Prize Lecture in Chemistry on 8 December 2024, titled “Accelerating Scientific Discovery with AI.” The lecture traced the intellectual arc from his neuroscience background through DeepMind’s game-playing achievements to AlphaFold2’s solution of the protein folding problem, and articulated the “learnable nature conjecture” — the idea that many laws of nature may be expressible as learnable patterns in data, enabling AI to accelerate discovery across all sciences.

Core content

Personal intellectual trajectory:

• “Thinking about thinking” — the neuroscience motivation for studying intelligence (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §Opening).
• From neuroscience PhD to games as a testbed for AI capabilities (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §Background).

Game-playing as a stepping stone:

• Go, chess, and Atari as progressively harder challenges demonstrating that deep RL can discover novel strategies (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §Games).
• The lessons from game-playing (self-play, learned representations, scaling) that transferred to scientific applications.

AlphaFold2:

• The protein folding problem and CASP14 breakthrough (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §AlphaFold).
• The decision to release predictions for the entire human proteome (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §Impact).

The learnable nature conjecture:

• The central thesis: many physical, chemical, and biological laws may be discoverable by pattern recognition in data, not just by human theorizing (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §Vision).
• Implications for the future of science — AI as a tool for hypothesis generation, simulation, and discovery.

Connections

• Theme: AI-for-science, learnable-nature-conjecture
• Project: (none — personal lecture)
• Collaborators: (none — solo lecture)
• Era: post-alphafold
• Venue: venue—Nobel-Prize
• Synthesizes: paper—highly-accurate-protein-structure-prediction-with---hassabis, paper—a-general-reinforcement-learning-algorithm-that-masters-chess-shogi-and-go, paper—mastering-atari-go-chess-and-shogi-by-planning-with-a-learned-model
• Notable quote: The “learnable nature conjecture” — the proposal that many natural laws are expressible as learnable patterns (lecture—nobel-prize-lecture-accelerating-scientific-discovery-with-ai §Vision)

Honest Gaps

• The extracted text is only ~8K characters — this appears to be lecture slides or a very condensed abstract, not a full transcript. A complete wiki page cannot be written from this extraction alone.
• No co-authors (solo lecture) — metadata is correct on this point.
• The lecture was delivered orally; the written extraction may miss nuances, visual demonstrations, and Q&A.
• The “learnable nature conjecture” is presented as a vision rather than a rigorous claim — its epistemological status is unclear.
• Without a full transcript, the specific arguments and examples used cannot be verified.