AI for science needs reasoning, not just data

- Every few decades some genius declares science "basically done" — Michelson in 1903, Hawking in the '80s, and now every AI founder with a fresh term sheet. AlphaFold cracked protein folding, sure, but that thing was trained on the Protein Data Bank — 53 years and $21 billion of experimental grunt work. You don't just replicate that because you raised a Series B and hired a hype squad. The real play? AI agents. Not data-hoovering monoliths, but reasoning engines that spin up sub-agents to draft hypotheses, peer-review them, run tournaments, and refine winners — basically digitizing the messy, iterative way actual scientists work. Google's AI Co-Scientist already nailed a decade-old antibiotic resistance question from a one-page brief. So while moonboys chase the next AlphaFold template, the smart money watches agent architectures that don't need a $21 billion dataset to produce results. Where's my cut?