AI-authored code at Anthropic grew from 3% to 26% of R&D output during the Scion development period, with a Claude model rewriting an ingestion pipeline without instructions to run 4x faster with 20x less memory.
October 15 reporting also described Anthropic's adversarial review technique, where one model critiques another's finished work rather than approving it, noting the technique can turn counterproductive when the reviewer fixates narrowly, which Anthropic called 'myopic'. Rob Bensinger argued it is 'time to panic about AI,' suggesting loss of control could be years or months away.