Altman asks how much smarter than humans AI should be allowed to get as task length doubles every 105 days

新智元 · wechat · 2026-11-05

In a Raj Shamani interview, OpenAI CEO Sam Altman posed the question of how much smarter AI should be allowed to get than humans.

METR's task-length metric: 50 minutes in early 2025, over 16 hours by April 2026; for post-2024 frontier models it doubles roughly every 105 days (previously every 7 months)
Human side flat or declining: Norway's 730k conscription IQ records show 7 points lost per generation post-1975 (reverse Flynn effect)
RSI risks: Altman flags both loss of control and power concentration, saying OpenAI monopolizing the technology would also be bad
Layered AI-on-AI monitoring: OpenAI's Oct 29 writeup describes activation probes, agent-conversation auditors, forensic investigators with sandboxed code execution, and system-log sentinels, plus a 30-minute no-response forced shutdown
The verification gap: AI produced a Navier–Stokes result in 88 hours; human mathematicians have audited it for nearly two months. Unverified results feeding the next training loop risk model collapse (Nature, Oxbridge teams)

More from AGI Musings