Google DeepMind Audit Finds No AI Self-Acceleration, Despite AI Writing 75% of Code

The finding that AI writes most of DeepMind's code yet produces no measurable acceleration in AI progress itself is the more interesting signal: automation of implementation is not the same as automation of discovery, and the bottleneck appears to sit in research judgement rather than typing. This matters for anyone extrapolating from coding-assistant adoption to recursive self-improvement timelines, including labs, policymakers and the wider AI sector. The open question is whether that gap holds as tooling improves, or whether the audit simply captured a stage before the effect would show.