Self-improving AI could arrive by 2027, says Anthropic co-founder Jack Clark

A new benchmark called TasteVal evaluates how well AI models can design and interpret scientific experiments on their own. The results show an unusually rapid improvement, doubling every three months since December 2025. Anthropic, whose Opus 5.5 model leads the ranking, now sees autonomous self-improving AI as possible by 2027, up from a previous estimate of 2028.

TasteVal assesses three core skills: identifying promising research questions, designing experiments, and interpreting results. The benchmark focuses on the last two, with predefined tasks such as data selection or model fine-tuning. Models must choose experiments and adjust their approach based on outcomes, within strict limits (40 hours of H100 GPU compute). Performance is compared to 24 human experts from labs like OpenAI or Google DeepMind.

The authors caution that these results do not prove AI is already improving itself, but they highlight a sharp acceleration in research capabilities. If this trend holds, AI development could become faster and cheaper, yet harder for humans to oversee. Major labs like Anthropic and OpenAI are now calling for stricter regulation of such advances.