Anthropic's Jack Clark Pulls Forward His Timeline for 'AI That Builds AI' to 2027
He now considers AI systems that develop their own successors more likely to become possible in 2027 than in 2028. The research-skills benchmark he cites shows rapid gains, but it covers only eight tasks.
Anthropic co-founder Jack Clark has moved up his forecast for the arrival of “self-improving AI,” systems that develop their own successors. Writing in his newsletter on October 12, Clark said he now considers full self-improvement more likely to become possible in 2027 than in 2028.
Why he moved the date up a year
As recently as May, Clark put the odds of a system capable of developing a successor AI without human involvement emerging by the end of 2028 at least 60 percent. This time he singled out the earlier date, 2027, as the more likely one. He did not attach a probability figure to the new forecast.
As evidence, he pointed to AI’s results in mathematics research and to a new performance benchmark called TasteVal. Built by P-Zero Research, TasteVal measures how well an AI can design experiments for a given research problem and interpret the results. It is a test of judgment about how to steer a research effort, not of whether the model gets the right answer.
How it measures experiment design
TasteVal hands the AI under evaluation a research problem and then asks it to plan the experiments and interpret the findings. The actual coding and execution of the experiments are handled by the same separate AI in every evaluation. The design removes variation caused by swapping out the model doing the execution, isolating the ability to design and judge.
The developers said that frontier models’ experimental research ability has roughly doubled every three months since December 2025. In the eight-task benchmark, the strongest performer was Opus 5.5. Opus 5.5 used computing resources for experiments about 2.30 times more efficiently than the human experts it was compared against. The cost per attempt was roughly one thirtieth of the experts’ average, according to the developers.
The scope of the evaluation has limits. TasteVal does not measure which research problems an AI would choose to pursue. It covers only tasks that are relatively cheap and quick to verify. And the human researchers used for comparison did not include top-tier researchers.
Does anything change right now?
No service is currently available to sign up for or purchase. The TasteVal team is directing interested parties to contact them for model evaluations. What is worth watching is how much of the judgment required for research experiments AI can handle, and how far it drives down the cost.
Clark’s forecast rests on the expectation that gains in AI research ability will translate into the development of successor AI systems. The open question is whether results on predefined tasks extend to choosing research topics and to development as a whole. Whether the same efficiency gap holds up across a wider range of task types and against higher-caliber researchers will be a key test of that forecast.