METR task-horizon debate: logistic rollover by 2027 or hyperbolic singularity?
teortaxesTex · x · 2026-11-09
A debate over the METR task-horizon trend: Joel Sercel argues that while p80 horizons (skilled-human software/ML tasks an agent completes with 80% success) still double every few months, evaluation ceilings, scarce long-horizon demo data, and real-world messiness point to a rollover — his logistic fit puts the inflection around now to mid-2027, then 65 hours by 2028, 550 hours by 2030, plateauing near 1,000–1,400 hours. teortaxesTex mocks zero-evidence predictions of imminent trend collapse, arguing a hyperbolic fit with a finite-time singularity is the better prior. The attached chart notes solar installations historically beat five-year forecasts by 3x+, a jab at underestimating exponential trends.
More from AGI Musings
- AGI leader's 2010s plan — build a lead, then unilaterally slow down for safety — never worked — S_OhEigeartaigh · 2026-11-09
- Economist: AI's academic productivity gains and moral hazards are both gigantic — paulnovosad · 2026-11-09
- Austrian economics paper: AI excels at optimization but can't replace entrepreneurial judgment — ArtificialOther · 2026-11-09
- Survey Maps 119 Personal-Agent Systems Across Six Levels of Loop Closure — liuziwei7 · 2026-11-09
- In the AI era, moats built over years are replicated in days, argues popular account — DeryaTR_ · 2026-11-09
- InstaDeep CEO maps frontier progress onto AI2027: unreleased models halfway to Agents-2 — HankYeomans · 2026-11-09