Google DeepMind: The Podcast 10 Years of AlphaGo: The Turning Point for AI | Thore Graepel & Pushmeet Kohli
211 snips
Mar 10, 2026 
Pushmeet Kohli, who leads DeepMind’s science work on problems like protein folding, and Thore Graepel, a key AlphaGo architect and accomplished Go player, reflect on AlphaGo’s turning points. They discuss why Go was the right challenge, the famous moves that shifted perceptions, how self-play led to AlphaZero, and applying search to big science problems like algorithm discovery.
AI Snips
Chapters
Transcript
Episode notes
Lee Sedol's Divine Move Exposed AlphaGo Fragility
- In game four Lee Sedol played move 78, an unusual wedge that confused AlphaGo and led to AlphaGo playing suboptimal moves and ultimately losing that game.
- Lee Sedol later called it the 'divine move' and was proud to have found a way to beat the machine once.
AlphaZero Learned Beyond Human Data
- AlphaZero trained from self-play with only game rules, rediscovering human patterns then discarding them when it found stronger refutations.
- Its play looked alien at first but often made sense many moves later as long-term planning emerged.
AlphaGo Opened Doors To Scientific Search Problems
- AlphaGo's success signaled that principled methods for navigating enormous combinatorial spaces generalize beyond games to scientific problems like protein folding.
- That belief motivated DeepMind to tackle problems such as AlphaFold soon after the Go victory.