Anthropic co-founder Jack Clark has updated his forecast for recursive self-improvement (RSI) in AI models, moving his expected arrival date forward to 2027 from 2028. The shift follows results from TasteVal, a benchmark developed by P-Zero Research that evaluates an AI’s ability to independently design scientific experiments and interpret results.

According to research data, Anthropic’s Opus 5.5 achieved a compute multiplier of 2.3 on TasteVal, reaching human expert-level experimental output while utilizing less than half the compute budget. P-Zero Research noted that experimental evaluation capabilities across leading models have roughly doubled every three months since late 2025.

While TasteVal measures experimental decision-making rather than fully autonomous AI research, the results suggest that automated experiment selection is rapidly reducing compute bottlenecks in frontier model development.

Why it matters

  • Accelerates expected timelines for fully automated AI research and development cycles.

  • Provides a concrete framework for measuring AI research efficiency and compute optimization.

  • Signals rising strategic importance of agentic experimental design over raw scaling.

Source: the-decoder.com