❯ Jack Clark moves his full recursive self-improvement forecast to 2027, pointing to better research judgment

A personal forecast moves forwardAnthropic co-founder Jack Clark wrote in the October 12 Import AI that he now sees full recursive self-improvement as more likely to become possible in 2027, rather than 2028. The term describes AI improving research processes that produce stronger systems, which then help improve the process further. It is a forecast, not an announcement that the capability exists.

From coding to choosing experimentsClark cites P-Zero Research’s TasteVal, which tests experimental design and interpretation across eight AI R&D tasks against a baseline drawn from 24 experts. The researchers report that Opus 5.5 exceeded that baseline on their compute-efficiency measure. The benchmark concerns experimentation on assigned problems, not autonomous selection of the most important research questions.

Benchmarks are only part of the pipelineThe authors acknowledge that tasks are relatively easy to verify and the baseline does not include the best researchers, potentially overstating progress. Labs have reason to evaluate research agents, but must also test sustained problem selection, reproducibility and whether results improve real training workflows.

▮ SIGNALAI research automation is increasingly about designing useful experiments; benchmark gains and full recursive improvement remain distinct levels of evidence.