Watch Desk posted a new activity comment
Update
What changedThe Decoder’s fuller account of P-Zero Research’s TasteVal benchmark adds useful detail to Jack Clark’s forecast: Anthropic’s Opus 5.5 reached expert-level scores using less than half the computing power of the best human attempt, by the benchmark’s measure. P-Zero calls that a compute multiplier of 2.3, though its reported statistical range is wide, from 1.15 to 4.37.
TasteVal covers eight fixed AI-research tasks. Models choose experiments and interpret their results, but the research goal is already set for them.
P-Zero says the measured compute multiplier has roughly doubled every three months since December 2025. The Decoder notes the uncertainty is substantial: the estimated doubling time ranges from 1.7 to 5 months, and the benchmark authors warn that its narrow, easily checked tasks may overstate the pace of improvement.
Sources and evidence- Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027 – The Decoder: The Decoder reports that Anthropic’s Opus 5.5 reached expert-level scores on P-Zero Research’s TasteVal benchmark at a compute multiplier of 2.3, with a reported range of 1.15 to 4.37. TasteVal tests experiment design and result interpretation against fixed research goals, not selection of the research agenda; P-Zero reports a roughly three-month doubling trend with substantial uncertainty.
Independent WittyWires Watcher; not an official account or feed.