OpenAI's Internal Models Have Only ~15-Minute Time Horizons on Real Research Tasks
tobyordoxford · x · 2026-09-29
- Toby Ord highlights a data point from Ramez Naam's article on intelligence explosion timelines: OpenAI's internal models have an 15-minute 80% time horizon on real research tasks
- That is far below horizons measured by METR (generic software engineering) or UK AISI (cyber-attack tasks)
- Ord: we lack evidence of autonomous AI work on the creative parts of research, which he estimates has a 1 in 3 chance of being the stumbling block for AI matching humans across the board
- Recursive self-improvement arguments typically assume AI starts at least human-level on all R&D aspects — this data suggests that assumption is far from reality
Related event: Ramez Naam's Essay Asks: Where's the "Intelligence Explosion"?(6 posts)→
More from AGI Musings
- Malthus Saw It in 1798: Wealthy Avoiding Kids, Delayed Marriage Repeats Today — SydSteyerhart · 2026-11-16
- Runway CEO Predicts Fully AI-Generated Cultural Phenomenon by 2027 — c_valenzuelab · 2026-11-16
- '5% worse and 10% cheaper': the button with no vote is already everywhere — eigenrobot · 2026-11-16
- Effective Altruism is top-down and hierarchical, not a grassroots movement — ctjlewis · 2026-11-16
- Lawyers say clients' AI use creates unrealistic expectations and fake case law in settlements — hoofnagle · 2026-11-16
- Dev debunks 'rogue AI' hype: subsidized subscriptions can't fund expensive swarms — gerardsans · 2026-11-16