Anthropic co-founder Jack Clark believes that self-improving AI could arrive next year in 2027, a year earlier than he had expected.
"I now think it's more likely we'll see full RSI (recursive self-improvement) become possible in 2027, whereas before I thought it was more likely in 2028," the Claude maker's co-founder wrote in his personal newsletter, 'Import AI'.
Clark's view is partly based on a new benchmark test (TasteVal, from P-Zero Research) that found AI's "experimental research taste" has been doubling roughly every three months since December 2025. The central argument is that if AI can do AI research well, it could start making itself better with less human help.
That's what recursive self-improvement or RSI basically is - where an AI gets good enough at AI research to build a better AI, which builds a better one still, and it goes on like that in a self-improving loop.
Clark's timeline shift is also based on recent AI maths breakthroughs that look creative.
In that scenario, AI progress would no longer depend only on how many human researchers there are.
Coming back to "research taste" - it is basically good scientific judgment, the ability to choose which problems matter, designing experiments and reading the results correctly. In a way it's like bridging the gap between a lab assistant and a good scientist. The test, however, measures only the experiment-and-results part, not the choice of problem.
What The Test Found
The test gave AI eight tasks drawn from real AI development, such as preparing training data and fine-tuning models. The results were compared against 24 human experts, with at least two per task, from leading firms including OpenAI, Google DeepMind, NVIDIA and Microsoft Research.
The best model, Anthropic's Opus 5.5, beat the human baseline at about 1/30 of the experts' average cost per attempt.
Clark had argued in an earlier essay titled 'AI Systems Are About To Start Building Themselves' that creativity was the key variable for how fast AI could improve itself. At the beginning of this year he said AI hadn't yet shown creativity in a major way. However, he now points to creative-seeming maths results plus tests like this one in his newsletter as major reasons for his belief in the RSI timeline.
