Toby Ord: we can't map benchmark scaling gains to real-world returns, like pre-scaling-law era
tobyordoxford · x · 2026-10-21
Toby Ord points out a methodological gap in RSI debates: pre-training scaling once looked grim (error falling as compute^1/20 per Kaplan/Chinchilla) yet the error-to-revenue transformation was steep (error^-20). When arguing whether r>1 from benchmark scaling results, we have little ability to pin down the benchmark-to-real-world return function, so results can't be carried across to real-world scaling questions.
More from AGI Musings
- Researcher argues synthetic data can't enable recursive self-improvement: compression loses information — gerardsans · 2026-10-21
- Reddit essay argues the AI bubble may be an intentional 'Wintermute' bankruptcy play — SongOfKapek · 2026-10-21
- The screensaver hunt for aliens is a cautionary tale for AI miracle hopes — seanmcdonaldxyz · 2026-10-21
- Blogger: data leaks don't scare me, a rogue AI would crush me like an ant — ezshine · 2026-10-21
- Turning fluid intelligence into crystallized: why agents should cache repeated solutions — AdaptiveAgents · 2026-10-21
- A Single Agent Is Already a Distributed, Multi-Owner System, Argues New Thread — AdaptiveAgents · 2026-10-21