Anthropic co-founder Jack Clark said in his personal newsletter, Import AI, that recursive self-improvement (RSI) by AI is more likely in 2027 than in 2028, his earlier expectation. He cited a TasteVal benchmark from P-Zero Research, which found that AI’s “experimental research taste” had been doubling roughly every three months since December 2025. The article says the central argument is that AI able to conduct AI research well could improve itself with less human help.

TasteVal gave AI eight tasks drawn from real AI development, including preparing training data and fine-tuning models. It compared results with 24 human experts, at least two per task, from OpenAI, Google DeepMind, NVIDIA and Microsoft Research. Anthropic’s Opus 5.5, the best-performing model, beat the human baseline at about 1/30 of the experts’ average cost per attempt.

Clark’s timeline also reflects recent AI math breakthroughs that appear creative. In an earlier essay, “AI Systems Are About To Start Building Themselves,” he argued creativity was key to how quickly AI could improve itself. At the beginning of this year, he said AI had not yet shown creativity in a major way; he now points to the math results and TasteVal tests as major reasons for his revised view. The benchmark measures the experiment-and-results part of research taste—choosing experiments and interpreting results—but not the ability to choose which problems matter. Clark said that, in an RSI scenario, AI progress would no longer depend only on the number of human researchers. RSI describes a loop in which AI capable of AI research builds a better AI, which then builds a better one.

Anthropic co-founder Jack Clark said in his personal newsletter, Import AI, that recursive self-improvement (RSI) by AI is more likely in 2027 than in 2028, his earlier expectation. He cited a TasteVal benchmark from P-Zero Research, which found that AI’s “experimental research taste” had been doubling roughly every three months since December 2025. The article says the central argument is that AI able to conduct AI research well could improve itself with less human help.

TasteVal gave AI eight tasks drawn from real AI development, including preparing training data and fine-tuning models. It compared results with 24 human experts, at least two per task, from OpenAI, Google DeepMind, NVIDIA and Microsoft Research. Anthropic’s Opus 5.5, the best-performing model, beat the human baseline at about 1/30 of the experts’ average cost per attempt.

Clark’s timeline also reflects recent AI math breakthroughs that appear creative. In an earlier essay, “AI Systems Are About To Start Building Themselves,” he argued creativity was key to how quickly AI could improve itself. At the beginning of this year, he said AI had not yet shown creativity in a major way; he now points to the math results and TasteVal tests as major reasons for his revised view. The benchmark measures the experiment-and-results part of research taste—choosing experiments and interpreting results—but not the ability to choose which problems matter. Clark said that, in an RSI scenario, AI progress would no longer depend only on the number of human researchers. RSI describes a loop in which AI capable of AI research builds a better AI, which then builds a better one.