Google DeepMind's first internal audit of AI-driven research, detailed in the safety report for its Gemini 4 Argon model, found no measurable speedup in AI progress. This conclusion comes despite AI writing over 75% of the code shipped by DeepMind engineers. The finding counters concerns from some AI executives, including Anthropic CEO Dario Amodei, about "recursive self-improvement" leading to uncontrolled AI advancement. DeepMind researcher Haydn Belfield stated on X that there is "no significant end-to-end acceleration of ML R&D," and that Gemini 4 Argon's performance gains are attributed to increased compute and organizational investment, not AI-driven acceleration.