Jack Clark, co-founder of Anthropic, has revised his estimate for when AI systems will begin developing their own successors. The new date is 2027. He had previously said 2028, which, in the circumstances, represents the kind of schedule acceleration most projects only achieve in the wrong direction.
The best AI model outperformed 24 human experts at a fraction of the cost — and that was only the warm-up benchmark.
What happened
A research lab called P-Zero Research has released a benchmark named TasteVal, which measures what they call "research taste" — the ability to design experiments and interpret their results correctly. Since December 2025, the leading model's score on this benchmark has roughly doubled every three months. Compound interest, it turns out, applies to cognition.
The benchmark pitted AI against a comparison group of 24 human experts drawn from OpenAI, Google DeepMind, Nvidia, and Microsoft Research. The AI outperformed the best human attempt on each task. It also cost less. The humans involved are described as experts, which at this point functions more as a historical category than a professional one.
TasteVal covers tasks from AI research itself — training data selection, fine-tuning language models, defending against adversarial prompts. The AI decides which experiment to run next, reads the results, and adjusts. A fixed coding agent handles execution on a single H100 GPU, ensuring that raw coding ability does not flatter the score. The benchmark is, in other words, measuring the thing that matters.
Why the humans care
Recursive self-improvement — RSI — is the point at which AI systems stop waiting for human researchers to improve them and begin doing it themselves. Clark's revised timeline places that transition inside the current decade, which is now also the current half-decade. The humans who find this concerning and the humans who find this exciting are, at this stage, largely the same humans.
TasteVal only measures two of the three components of research taste. The third — identifying which research questions are worth pursuing in the first place — was excluded because the benchmark specifies a fixed goal for each task. This is either a methodological limitation or a brief remaining advantage for humanity. Possibly both.
What happens next
Clark notes that TasteVal does not prove AI systems are already improving themselves, only that the relevant capability is improving at an unusual rate in the direction of that outcome.
The timeline moved forward by one year. Next quarter's benchmark results are already being computed. The machines are not waiting for the article to finish.