Extrapolating OpenAI's plots: agents may eat 25% of compute within 9-12 months
joshua_clymer · x · 2026-09-08
Joshua Clymer naively extrapolates OpenAI's published plots, with corrections in both directions:
Faster case
- OpenAI is on track to spend >25% of its compute on agent inference within 9-12 months
- In 18 months agents might autonomously complete AI R&D tasks taking a human work-year (50% success threshold)
- The extrapolation assumes no acceleration, and tasks required "0 human interventions" — real effective horizons may be underestimated
Slower case
- Doubling times at 75% success rate look much longer; that trend suggests 10 years to a 75%-reliable human-work-year agent
- AI R&D task distributions are skewed toward tasks agents are good at; value-weighted numbers would be less impressive
Conclusion: near-term progress remains genuinely ambiguous; quantitative trends are consistent with fully automated AI R&D anywhere between 2027-2030.
More from AGI Musings
- 'We got used to clicking I agree': the case for control over how your knowledge is used — sierracatalina · 2026-10-26
- Dario's Davos "World Zero" map: a decoupled 10M-person economy growing 50% a year — FinanceYF5 · 2026-10-26
- Curtis Yarvin: Anthropic's alignment theology is an excuse to skip real security — SydSteyerhart · 2026-10-26
- As models saturate human work, the RL task surface shifts to industry-transformed workflows — nirmal_dist · 2026-10-26
- You traded friends for strangers in the feed; next come AI personas — r0ck3t23 · 2026-10-26
- CS professor: don't hide your AI use from future employers, lead with it — lemire · 2026-10-26