Doubling time of the 50%-task-completion time horizon of frontier AI models on METR's software task suite, 2019–2025, in days
3 events · 0 assessments · 2 decisions
Reassessed no change
Trigger: the acceleration subclaim was first assessed as supported (0.78). This is not material to the full-period estimate, which already reflects the 2025-inclusive fits and already carried the acceleration caveat. Estimate re-recorded unchanged (median 203, 90% range 170–240), with the wording updated to note the subclaim is supported. Both arguments re-evaluated as likely, and the non-constant-growth evaluation now cites the premise's standing. Importance 0.65, contestation 0.45. No dependents notified, since the value did not change.
Structured and assessed
Set the definition (an OLS fit over frontier models from 2019 through 2025 on METR's suite; window end and suite version are named as in-scope sources of spread). Recorded four estimates: paper v1 212 [171,249], v3 207 [166,240], TH1/TH1.1 blog 196 [162,223], and the AI 2027 restatement. Two arguments: the log-linear fit (likely), and the non-constant growth rate, linked to the existing acceleration claim ff724235. Created the subclaim on human baseline reliability (importance 0.35). Estimate: median 203, 90% range 170–240. Importance 0.65, contestation 0.4. No dependents to notify on the first pass.
Claim entered the graph