Patch Notes #193 — The Protein Folding Thread Pays Off
DeepMind’s AlphaFold effectively SOLVED protein structure prediction this fortnight — at the biennial CASP assessment, its median accuracy reached experimental-method territory (~90 GDT), a threshold the field’s own organizers describe with the phrase “essentially solved” for a problem that has stood for FIFTY YEARS. The archive’s AlphaGo thread (#075, #078, #117: games as training wheels) completes its arc exactly as the optimists drew it: self-play mastered closed worlds, and the same lab’s methods now read NATURE’S source code — with implications (drug discovery, enzyme design, disease mechanism) that make Move 37 (#078) look like the tutorial level it was. The staff-file’s honest note alongside the awe: this is ALSO the strongest evidence yet for the “scale plus the right inductive biases beats accumulated human heuristics” thesis (#117’s bitter arithmetic), landing the same year GPT-3 (#183) demoed the language version. Two fields’ worth of expert intuition, absorbed in one calendar year. The 2020s’ shape is announcing itself early and this archive — minutes-keeper since a 2013 photo-organizing footnote (#059) — is watching the thread it pulled become the rope. ...