TasteVal Benchmark Shows AI Creativity Doubling Every 3 Months, RSI Timeline Could Move Up to 2027

Anthropic co-founder Jack Clark published the results of the TasteVal benchmark on October 12, 2026 in Import AI #476: AI's taste for experimental research has been doubling every three months since December 2025, and the flagship model Claude Opus 5.5 surpassed a human benchmark made up of 24 top AI researchers at roughly 1/30 of the per-expert cost.

The Facts

According to Jack Clark, TasteVal was launched by P-Zero Research and tests how AI systems perform on eight tasks representative of frontier AI development, spanning pretraining data curation, language model pretraining and fine-tuning, preference modeling, and adversarial prompt robustness. At its core, the benchmark assesses whether AI can independently complete the entire research workflow, from choosing a topic to designing experiments to interpreting results. On this basis, Clark moved his median-probability timeline for fully autonomous recursive self-improvement (RSI) in AI up from 2028 to 2027.

Mechanism Breakdown

TasteVal's test design directly targets AI's creative performance in AI research. The continued doubling of AI's capabilities in topic selection, experimental design, and result interpretation means its ability to generate new research directions and validation methods is rapidly approaching expert human level. That Claude Opus 5.5 achieves higher performance at a cost far below that of a human team shows the model's efficiency advantage in handling complex research tasks. The RSI timeline moved up mainly because the creativity bottleneck is being broken faster than previously expected.

Industry Impact

This progress could prompt AI labs to reassess whether existing evaluation frameworks need a new "self-research capability" dimension. Current benchmarks mostly focus on performance on specific tasks, whereas TasteVal emphasizes an end-to-end research workflow; future evaluations may include more autonomous iteration metrics. For AI companies, moving early on models with research taste could create a competitive advantage, while also requiring more resources to be devoted to mechanisms that check the reliability of AI outputs.

Strategic Assessment

(This is analysis, not fact.) If the trend reflected by TasteVal continues, greater autonomy for AI systems in research could accelerate the technology iteration cycle across the entire industry, but it would also make human oversight of AI outputs harder. Historical experience shows that capability breakthroughs often come with new alignment and safety challenges, and the judgment that RSI has a better-than-even chance by 2027 suggests stakeholders need to plan their responses in advance.