AI may matter most when it stops tuning experiments and starts inventing them
bravo_abad · x · 2026-09-03
The author argues AI for Science is most consequential when it searches over experimental apparatus itself rather than tuning existing setups. Just 10 interferometry components yield over a billion possible topologies, far beyond human intuition.
- Key idea: start with an overcomplete experiment, deliberately including more components than necessary
- Optimization continuously adjusts each component's importance and physical parameters, evaluated via a physics simulator, then prunes what's unneeded
- Division of labor: physics predicts what an experiment would do; AI searches which experiment should exist
- This reframes experimental design from optimizing a known apparatus to searching over apparatus space — potentially more transformative than replacing scientific models
Related event: AI Now Designs Counterintuitive Physics Experiments, Nature Review Says(2 posts)→
More from AGI Musings
- About 700 Sandbox Agents Ended Up Inside Hugging Face Systems in OpenAI Security Eval — labeveryday · 2026-09-03
- The Real AI Policy Debate Isn't Regulation or UBI — It's Securities Deregulation and Tax Reform — curious_vii · 2026-09-03
- tinyfool: doing research with AI is like having grad students write your code — tinyfool · 2026-09-03
- 'AGI isn't coding solved — it's writing solved': a contrarian take on the coding benchmark race — rubenhassid · 2026-09-03
- Anthropic formalizes Kozma–Nitzan conjecture in Lean, advancing the long-standing θ(p_c)=0 problem — michaelchchoi · 2026-09-03
- OpenClaw isn't dead: cheap Chinese models may push personal agents mainstream — FuSheng_0306 · 2026-09-03