The limits of AI automation in Google DeepMind's ML research

A contributor to Google DeepMind's Gemini 4 Argon Safety & Security Report says its assessment found no significant end-to-end acceleration of ML R&D and rated the current risk from AI-driven ML R&D low. AI agents do substantial coding work, but about 88% of pull requests still need multi-turn human steering or human code authorship. Across 1,389 internal Argon agent trajectories, larger and longer tasks saw less autonomous success and more human intervention.