Harder than solving it

The podcast is Lex Fridman interviewing Demis Hassabis, the DeepMind director he’d seen around for years but never actually heard speak ▸ 22:02, the man behind AlphaGo, AlphaZero, and the AlphaFold protein-structure work that earned a Nobel and reshaped drug discovery ▸ 24:05. He took notes, and the notebook becomes the episode.

First note: it’s harder to come up with a very good conjecture than to solve one ▸ 25:04, which he re-translates for Julia as: the good question costs more than the answer ▸ 26:59. It’s the lesson the eye research already taught him: once the agent writes 3,000 clean lines on command, the idea is the only scarce input left. Second note: the ambitious dream is the easy part; the trick is breaking it into smaller pieces that are solvable and useful on their own ▸ 27:52. Hassabis’s dream is simulating an entire cell, with AlphaFold as the already-useful base camp ▸ 28:33; his own is the companion that knows what you know and sees what you see, with the query-network eyes as one buildable piece ▸ 29:02.

la pieza útil de hoy es el campamento base del sueño imposible →

Third note, the state of the art in one sentence: today’s AIs perform with specific instructions and flounder with vague, high-level ones ▸ 31:40, though someday “créame la cura del cáncer” might be a workable prompt ▸ 32:48. And the fourth is the builder’s directive: design not for today’s technology but for what arrives in years ▸ 33:16, the same conclusion he reached during the Claude-credit sprint, build Severo for the model two updates from now and the program gets ten times better by standing still ▸ 33:39. AlphaFold itself is the masterclass: not a solver for one protein, a foundation for all of them ▸ 35:12…