Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027

Why it matters

P-Zero Research released TasteVal, a benchmark measuring whether AI systems can design experiments and interpret results on their own, a capability the lab calls research taste. The top-performing model, built by Anthropic, matched the best results of a comparison group of human experts from leading AI organizations while using far less compute. Across recent model generations, this compute efficiency has roughly doubled every quarter, though the authors warn the tasks may favor what labs optimize for easily and that the expert group lacked top-tier researchers. Anthropic co-founder Jack Clark, who had previously been cautious about AI creativity, responded by moving his forecast for recursive self-improvement earlier by a year. The benchmark matters because it targets the exact capability skeptics like Herbie Bradley had identified as a barrier, while leaving open the harder question of whether models can choose promising research directions. A separate Fudan University study found AI agents still depend heavily on humans for final research decisions.