RESEARCH · RESEARCH · #2530
Anthropic's Opus 5.5 tops P‑Zero's TasteVal; Jack Clark moves RSI prediction to 2027
P‑Zero Research's new TasteVal benchmark, which measures models' ability to design and interpret experiments ('research taste'), finds rapid gains — Anthropic's Opus 5.5 reached a compute multiplier of 2.3 versus a 24‑expert human baseline and cost far less per run. Based on these results and a reported doubling of the metric roughly every three months since December 2025, Anthropic co‑founder Jack Clark has moved his forecast for recursive self‑improvement from 2028 to 2027, while the authors and other researchers note large uncertainties and narrow coverage of the benchmark.
KEY POINTS
- P‑Zero Research's new TasteVal benchmark, which measures models' ability to design and interpret experiments ('research taste'), finds rapid gains — Anthropic's Opus 5.5 reached a compute multiplier of 2.3 versus a 24‑expert human baseline and cost far less per run.
- Based on these results and a reported doubling of the metric roughly every three months since December 2025, Anthropic co‑founder Jack Clark has moved his forecast for recursive self‑improvement from 2028 to 2027, while the authors and other researchers note large uncertainties and narrow coverage of the benchmark.
- If research‑taste capabilities continue improving this quickly, models could plausibly automate more independent scientific work and enable faster recursive self‑improvement, changing timelines for high‑impact AI developments.
WHY IT MATTERS
If research‑taste capabilities continue improving this quickly, models could plausibly automate more independent scientific work and enable faster recursive self‑improvement, changing timelines for high‑impact AI developments.