Jack Clark moves AI self-improvement prediction to 2027

Anthropic co-founder Jack Clark has moved his prediction for recursive self-improvement forward to 2027. He had previously expected this milestone to occur in 2028. Clark previously stated that models had not shown creativity at any groundbreaking scale earlier this year. In a May essay, he put the odds of independent training of a stronger successor at roughly 60 percent by the end of 2028 and 30 percent by 2027.

A new benchmark called TasteVal measures how well AI systems design experiments on their own and interpret the results. P-Zero Research defines the measured capability as 'research taste'. The capability measured by TasteVal has roughly doubled every three months since December 2025. In prior years, model research taste grew much more slowly than it has since December 2025. AI researcher Herbie Bradley argued that skills like research taste, creativity, and coherent long-term research planning were nowhere in sight before TasteVal.

The best model tested in TasteVal outperformed a comparison group of 24 human experts at a fraction of the cost. The human baseline consisted of experts who had recently worked at OpenAI, Google DeepMind, Nvidia, and Microsoft Research. A fixed coding agent runs all TasteVal experiments on a single Nvidia H100 GPU. Each run ends after 40 GPU-hours of compute or 120 hours of wall-clock time. P-Zero Research translates research taste into compute efficiency using a 'compute multiplier'. Anthropic's Opus 5.5 hit a compute multiplier of 2.3 in the benchmark. The benchmark does not prove that AI systems are already improving themselves.

Anthropic CEO Dario Amodei now calls for a controlled slowdown in AI development. OpenAI acknowledges that fully autonomous recursive self-improvement isn't happening yet and says it should only be pursued if human oversight is guaranteed. A study by a team at Fudan University found that AI agents contributed up to 55 percent of methodology proposals but humans made over 85 percent of final decisions.