Verifiable vs non-verifiable is the wrong lens; slow-to-verify is the real bottleneck
rosstaylor90 · x · 2026-10-09
Ross Taylor argues the verifiable/non-verifiable distinction misunderstands model jaggedness—fast-to-verify vs slow-to-verify is the right frame. Many "non-verifiable" skills like research judgement are verifiable under long-term objectives, but are hard to hillclimb because such objectives train slowly with sequential compute or are very expensive with parallel compute. Other slow-to-verify cases reflect real-world friction—scientific instruments, new hardware and facilities—driving new physical labs, though the overarching difficulty of optimizing long-term objectives remains under-appreciated.
More from AGI Musings
- Ex-UK AISI comms officer goes independent and questions why AI companies don't act on their own warnings — AaronBergman18 · 2026-10-09
- Why hard workers hate AI: fear of a 'Brahmin caste' of Idea Guys — menhguin · 2026-10-09
- Prediction: frontier models as real-time robot policies at 10k-100k tok/s within 1-2 years — ATTlKA · 2026-10-09
- AI won't take all jobs soon: replacement needs scarce, costly human specialists — boxenluder · 2026-10-09
- Mathematician: Terence Tao's strategy assumes AI never explains math better — burny_tech · 2026-10-09
- After 3 days at an AI-Talmud workshop: how do you eval AI where experts disagree? — EhudReiter · 2026-10-09