Blindly porting LLM ideas into protein models slammed: more compute makes algorithms matter more
anshulkundaje · x · 2026-11-08
samsinai endorses MillionInt's argument that "the more compute you pour into training, the more important the algorithm you're using," adding that in biology, careful thinking about data matters far more than for natural language.
He criticizes the blind transfer of LLM ideas into protein language models (PLM/GLMs) — always defended with the "bitter lesson" when scrutinized — which is why the only real step-change came from structure-based models that carefully scaled algorithms, data, and compute together. The original post argues that misreading the Bitter Lesson plus a brief window when scaling transformers worked has reward-hacked a generation of scientists, with whole dimensions of scaling left unexplored.
More from AGI Musings
- METR Progress per R&D Dollar Fell 63% Even as Yearly Gains Doubled — sebkrier · 2026-11-08
- As intelligence gets Jevons-paradoxed, 'dark data' becomes the first jackpot — hardimanjames · 2026-11-08
- Rutgers mathematician: AI acknowledgements are a fad — judge all writing as AI-generated — AlexKontorovich · 2026-11-08
- tszzl: Comparing AI regulation to nuclear control is nonsense — models replicate infinitely — tszzl · 2026-11-08
- BitTorrent's Bram Cohen to mathematicians: your turn to debug AI proofs — hargup13 · 2026-11-08
- "Less worried about AI passing the Turing test than humans failing it" — TinfoilTricorn · 2026-11-08