Why AI cracking Navier-Stokes doesn't mean it can auto-discover better models
snikolov · x · 2026-11-03
The author argues that experimental AI research is a harder setup than symbolic proof-style discoveries:
- Proofs vs experiments: proof attempts are symbolic, fast to run, and failures point at what's missing; computational experiments take longer and bad outcomes are hard to attribute (bugs, wrong math, suboptimal training).
- Path 1: run cheap experiments and find modular improvements that scale and stack (e.g. iterating on optimizers on simple problems).
- Path 2: decompose verifiable subproblems, like finding equivalent but much faster computations.
- Path 3: let autoresearch systems invent open-ended, composable analysis tools—choosing the right conceptual lens is itself a research problem.
More from AGI Musings
- Contrarian take: verification is a special case, the fuzzy real world needs soft-verification — tensorqt · 2026-11-03
- Assisted vs. Replaced Thinking: Research Links Substitutive AI Use to Metacognitive Decline — DrKavner · 2026-11-03
- Carl Frey Calls for 'Learning Rights' Legislation to Decide Who Owns What AI Learns — carlbfrey · 2026-11-03
- Yacine predicts smaller models will start outperforming bigger ones, citing the bitter lesson — yacineMTB · 2026-11-03
- Frontier AI labs accused of 'extortion racket': selling cyber defense for risks they created — Graham_dePenros · 2026-11-03
- ICLR 2027 submission full of trivial claims sparks 'vibe science' quality warnings — IanArawjo · 2026-11-03