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

AI systems can now design and interpret experiments more cheaply than human experts, accelerating AI development.

Anthropic co-founder Jack Clark has advanced his timeline for recursive self-improvement (RSI) in AI systems to as early as 2027, from a previous expectation of 2028. This shift is based on findings from a new benchmark called TasteVal, developed by P-Zero Research, which measures an AI's ability to design experiments and interpret their results. Clark noted that this capability has shown a steep improvement curve, roughly doubling every three months since December 2025. The best AI model tested on TasteVal outperformed a group of 24 human experts at a significantly lower cost. While TasteVal does not prove that AI systems are currently self-improving, it highlights a growing focus on AI's independent research potential beyond raw performance metrics. The benchmark specifically assesses an AI's capacity to select appropriate experiments and accurately interpret outcomes, with the top model demonstrating enhanced efficiency in this area.

The rapid doubling of AI's "research taste" every three months suggests a fundamental acceleration in AI development, moving beyond mere task completion to autonomous scientific inquiry. This has profound implications for the pace and cost of innovation, potentially making advanced AI research accessible to a wider range of entities, while also raising questions about the future role of human researchers in the discovery process. As AI systems become more adept at independent experimentation, the bottleneck in AI progress may shift from computational resources or coding to the conceptualization of novel research directions.

This advancement in AI's research capabilities, as evidenced by the TasteVal benchmark, signals a potential paradigm shift in how AI itself is developed. If AI can effectively design and interpret experiments, it opens the door to faster, more efficient AI model iteration and discovery. However, it also introduces challenges in verifying AI-driven research and maintaining human oversight, particularly as AI-generated findings may become increasingly complex and difficult for humans to independently validate.