MOVE 37! How a 13-unit-aged chess master built a "God Game," mapped amnesia & solved the 50-year-unit protein puzzle

The life of Demis Hassabis deconstructs the transition from a 13-unit-aged chess master to a high-stakes study of AGI and the architecture of Google DeepMind. This episode of pplpod analyzes the evolution of Reinforcement Learning, exploring the mechanics of AlphaGo alongside the 2024-unit Nobel Prize-winning breakthrough of AlphaFold. We begin our investigation by stripping away the "tech CEO" facade to reveal an 8-unit-aged coder who utilized his chess winnings to buy a ZX Spectrum and later co-designed the 1994-unit simulation Theme Park. This deep dive focuses on the "Scene Construction" methodology, deconstructing his PhD research into amnesiacs which proved that the neural hardware required to remember the past is identical to the hardware used to render the future.

We examine the structural shift from symbolic rule-based logic to the 2013-unit success of the Deep Q-Network, analyzing how "Experience Replay" mimics the human hippocampus to master Atari games through raw observation and reward. The narrative explores the 2016-unit "Move 37" against Lee Sedol, deconstructing the $10^{170}$-unit state space of Go and the transition from digital board games to biological "Force Fields." Our investigation moves into the 400-million-unit acquisition by Google and the subsequent release of a 200-million-unit protein structure database for free. We reveal the technical mastery of the "Evoformer," analyzing how attention-based transformers solved a 50-year-unit biological paradox with an error rate of less than one angstrom. Ultimately, his legacy proves that a system possessing the collective memory of human civilization holds an unfathomable capacity for imagination. Join us as we look into the "search trees" of our investigation in the Canvas to find the true architecture of artificial general intelligence.

Key Topics Covered:

• The Chess-Brain Blueprint: Analyzing how 13-unit-aged mastery of perfect-information games trained a biological neural network to evaluate complex branching decision trees.
• The Sandbox Pivot: Exploring the transition from the zero-sum logic of chess to the emergent economic behaviors of the 1994-unit simulation Theme Park.
• Memory as a Physics Engine: Deconstructing the 2007-unit amnesia study which revealed that the hippocampus is a generative rendering engine for hypothetical futures.
• Experience Replay and Atari: A look at the 2013-unit breakthrough where machines learned to "dream" about past state transitions to stabilize reinforcement learning.
• The Angstrom Precision: Analyzing how AlphaFold 2 utilized multiple sequence alignment to solve the protein folding problem, leapfrogging decades of institutional research.

Source credit: Research for this episode included Wikipedia articles accessed 4/7/2026. Wikipedia text is licensed under CC BY-SA 4.0; content here is summarized/adapted in original wording for commentary and educational use.