TasteVal — research-taste doubling accelerates to 3.0 months; Opus 5.5 passes the human baseline
| Date | 2026-10-12 |
|---|---|
| Category | Benchmarks |
| Judgement | Supports |
| Gauge weight | 2 / 3 |
| Source | P-Zero Research TasteVal benchmark (arXiv 2610.06824), via Import AI #476 by Jack Clark, 2026-10-12 |
Summary
Opus 5.5 exceeded a baseline of 24 human experts at a compute multiplier of 2.30x and one-thirtieth of the cost. Research-taste doubling accelerated to every 3.0 months after December 2025, from 14 months before, and a naive extrapolation pulled the ASI median from mid-2030 to late 2028.
A/B Impact Analysis
As a measured acceleration, this supports Scenario A. The rebuttal in the record is equally clear: TasteVal measures hill-climbing inside a defined space, not paradigm invention, and straight-line extrapolation of a short acceleration window can repeat past extrapolation errors. The evidence therefore did not move the median; it only shifted probability mass slightly toward 2029 within the range. It is logged as support with an explicit no-overweighting condition.