Jack Clark Moves Self-Improving AI Forecast to 2027 After TasteVal Benchmark

Anthropic co-founder Jack Clark now believes recursive self-improvement, where AI systems develop stronger successors, could become possible as early as 2027. He previously expected 2028. The shift follows results from P-Zero Research’s TasteVal benchmark, which measures how well AI systems design experiments and interpret the results.

The key point for readers: this is not evidence that AI systems are already improving themselves. It is a signal that AI models may be getting better at parts of the research process that matter for future AI development.

P-Zero Research defines “research taste” as the ability to identify worthwhile research questions, design experiments, and interpret results. TasteVal tests only the last two: experiment design and result interpretation. The benchmark gives models fixed research goals, so it does not test whether they can choose the most promising long-term research agenda.

The tasks cover AI model development areas including training data selection, pretraining, fine-tuning, training on human preferences, and defenses against adversarial prompts. A fixed coding agent runs experiments on a single Nvidia H100 GPU to reduce the influence of coding skill on the score.

According to Clark, the measured capability has roughly doubled every three months since December 2025. The top model tested, Anthropic’s Opus 5.5, reached a compute multiplier of 2.3, meaning it matched expert-level scores with less than half the compute by P-Zero Research’s definition. The reported statistical range was wide, from 1.15 to 4.37.

The human baseline came from 24 experts who had recently worked at organizations including OpenAI, Google DeepMind, Nvidia, and Microsoft Research. But the article notes the benchmark may overstate how quickly research taste is improving, and the comparison group did not include top-tier researchers.

TasteVal lands in a broader debate over whether AI systems can take on the research decisions needed to accelerate their own development. The benchmark pushes against claims that models lack the ability to plan and evaluate experiments, but it does not settle the harder question of long-term scientific direction.

If similar trends hold across other tests, AI development could become faster and cheaper. The tradeoff is oversight: as models plan and evaluate more of their own experiments, parts of AI research may become harder for humans to verify.