Jack Clark (Anthropic) has brought his prediction forward by a year: full recursive self-improvement in AI is ‘more likely as early as 2027’

In *Import AI*, the co-founder of Anthropic draws on the TasteVal benchmark – which suggests that models’ search ‘taste’ doubles every three months – to revise his timeline.

In issue 476 of his Import AI newsletter, Jack Clark, co-founder of Anthropic, looks back at his own analysis of recursive self-improvement (RSI), i.e. the moment when AI systems could themselves advance AI research. He had identified creativity as a key variable and estimated, at the beginning of the year, that models had not yet demonstrated it in a transformative way. He now believes that a full ROI in 2027 is more likely, whereas he had previously been leaning towards 2028.

The "taste" of research measured

This change is based in particular on TasteVal, a benchmark from P-Zero Research that measures, on eight typical tasks of model development (pre-training data curation, fine-tuning, preference modeling, robustness to adverse prompts), the ability of a model to design experiments and interpret the results. According to its authors, this skill has doubled every 3 months since December 2025; the best model tested, Opus 5.5, would exceed the human benchmark, established with 24 experts who have worked for OpenAI, Google DeepMind, NVIDIA and Microsoft Research, for about one-thirtieth of their average cost per attempt.

Jack Clark adds to this the recent mathematical breakthroughs obtained with the help of models as indicators of progressing creativity. The same issue also mentions the visual reasoning benchmark Humanity's Sixth Sense, where the best model tops out at 53.6% compared to 93.1% for humans.