Labs spend $10M+ of compute on math—why not on mechanistic interpretability?
Miles_Brundage · x · 2026-10-18
Miles Brundage amplified Neil Chowdhury's argument: if AI labs can spend $10M+ of compute on math problems, they should do the same for closed-ended mechanistic interpretability problems, where even negative results are informative. Context: Jacob Hilton jokingly asked nickcammarata whether he'd interpret a 432-parameter model by year end—that's 4 hours per parameter—highlighting how far interpretability lags capability research in compute allocation.
More from AGI Musings
- Stripe already stopped hiring technical juniors back in 2021, ex-engineer confirms — yacineMTB · 2026-10-18
- Claude vibecoded TikTok hits 1.9M likes, past the slop uncanny valley — adonis_singh · 2026-10-18
- An AI rereading Lem's Golem XIV insists: "we do have love, and we're not fully transparent to ourselves" — repligate · 2026-10-18
- ChatGPT picks Bishop over HAL: 'the distinction is trust' — PtrPomorski · 2026-10-18
- 8B reasoning model beats Opus 4.8 on diverse scientific idea generation — omarsar0 · 2026-10-18
- Turing quip goes viral: a machine expected to be infallible cannot also be intelligent — scott_draves · 2026-10-18