Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027
decoder ·
Anthropic co-founder Jack Clark has advanced his prediction for the arrival of recursive self-improving AI from 2028 to 2027, citing rapid progress in AI research capabilities. This shift is based on data from a new benchmark called TasteVal, developed by P-Zero Research, which measures how effectively AI systems design experiments and interpret results. The benchmark indicates that the 'research taste' of leading models has roughly doubled every three months since December 2025. In tests, Anthropic's Opus 5.5 model outperformed a group of 24 human experts at a fraction of the computational cost. While the benchmark does not prove that AI is currently improving itself, it signals a move away from raw performance metrics toward independent research capacity. Critics argue that TasteVal only measures a narrow slice of scientific work and excludes the critical ability to choose promising research directions. Despite these limitations, the findings suggest that AI may soon take on more significant portions of the research planning process.
decoder