Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027
Summary
- A new benchmark called TasteVal measures AI capabilities in experiment design and result interpretation, capturing a quality called "research taste."
- Anthropic co-founder Jack Clark moved his prediction for recursive self-improvement forward to 2027 based on recent findings.
- Data shows that leading models like Anthropic's Opus 5.5 can match or exceed human expert scores on designated tasks using significantly less compute.
- The benchmarked research taste metric has reportedly doubled roughly every three months since December 2025 according to P-Zero Research.
- While TasteVal shows strong performance in experiment planning and evaluation, critics note it does not test long-term research agenda selection.
Why it matters
The emergence of AI systems demonstrating experiment design and "research taste" shifts the industry focus toward recursive self-improvement and autonomous AI-driven development. This could drastically accelerate AI progress while raising verification challenges.