Jack Clark, co-founder of Anthropic, wrote one sentence in the October 12 issue of his newsletter Import AI that almost disappears inside an edition about creativity, open weights and space: “I now think it’s more likely we’ll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028.”
RSI is recursive self-improvement, the point where AI systems start improving themselves. Clark pulled his own estimate forward by a year, and he says what changed his mind.
TasteVal measures what researchers call taste
The prompt is a benchmark called TasteVal, from P-Zero Research. It covers eight tasks drawn from frontier AI development: pretraining data curation, pretraining and fine-tuning language models, preference modeling, and robustness to adversarial prompts. What it measures isn’t knowledge but the experimental side of research taste — how well a model designs its own experiments and how well it reads the results.
The human baseline comes from 24 experts, at least two per task, who have recently worked at organizations including OpenAI, Google DeepMind, NVIDIA and Microsoft Research.
The numbers that caught Clark’s eye: experimental research taste in frontier models has doubled every 3.0 months since December 2025, with a confidence interval of 1.7 to 5.0 months. The best model is Claude Opus 5.5, with a compute multiplier of 2.30x against the human baseline at roughly one-thirtieth of the experts’ average cost per attempt. Before late 2025 the figure grew far more slowly; the slope only changed around the turn of the year.
Clark already named creativity as the key variable
In his analysis earlier this year, Clark argued that creativity was the variable deciding how fast AI systems could learn to improve themselves. At the time he wrote that AI systems hadn’t shown it in any transformative way, and that results in mathematics were at most suggestive. Since then he sees both: real progress on creative-looking math breakthroughs, and now signs that models are getting better at AI research itself.
A sentence in a newsletter is not a study
Clark isn’t making a company statement here, and that matters. Import AI is his own project with more than 144,000 subscribers, the sentence sits under a paragraph headed “Why this matters”, and it is marked as a personal revision. The benchmark behind it comes from P-Zero Research rather than Anthropic, and the top result belongs to Anthropic’s own model. If you want reasons to doubt it, there are two good ones: a single measurement series over nine months, and a confidence interval running from 1.7 to 5.0 months.
The direction is what’s worth watching. Clark’s worry isn’t that models write code, which they have done for a while. It’s that models are starting to pick the right experiments.