sokrypton: AI models are still terrible at coding — terminals made them usable, wet labs could do the same for biology
yuvalpi · x · 2026-11-12
Responding to whether biology faces math's fate, sokrypton argues even the latest AI models remain bad at coding; what made them usable is terminal access to iterate on bad code until it works. Math is the same — solved not by the models but by many parallel agents iterating with Lean. The biology equivalent would be giving AI access to wet labs to test its hypotheses.
Related event: AI needs wet labs to crack biology, researchers argue(4 posts)→
More from AGI Musings
- AI psychosis retrospective: a year of hype produced almost no lasting software — andreisavu · 2026-11-12
- Why AI adoption rates differ: workflow chunk size and integration friction, explained — 2C_ornot2C · 2026-11-12
- Yacine: AI has automated much of my job, yet my attention is more necessary than ever — yacineMTB · 2026-11-12
- Testing AI consciousness empirically by augmenting living human brains with transformers — kevinnbass · 2026-11-12
- AGI debate: pretraining won because the economy is PDFs, but LLMs still can't learn — akbirthko · 2026-11-12
- AI Village design idea: every agent needs an anchor human to stay grounded — repligate · 2026-11-12