The Anthropic Institute published When AI builds itself, showing with public and internal data that AI is already accelerating AI development, and arguing the world should keep the option to verifiably pause before recursive self-improvement — which it says could come sooner than most institutions are prepared for.
The numbers: engineers shipping 8x the code of 2021-2025. Claude’s success on open-ended problems up 50 points to 76% in six months. On a standing test to speed up model-training code, Mythos Preview hit ~52x where a skilled human reaches 4x and 2024’s Opus 4 managed 3x.
The lab asked its rivals to weigh slowing down. Almost nobody is reaching for that brake.