Google Says AI Isn't Building Itself Yet—Even Though It Writes 75% of the Code
Summary
Google DeepMind’s safety assessment found no significant acceleration in machine-learning research and development from AI, despite AI writing more than 75% of the lab’s production code. Its internal capability index shows Gemini models improving at a steady rate; the latest model’s higher score is attributed to greater compute and investment, not AI-driven progress. Most code changes still require human steering or review, and AI agents are less reliable on larger tasks and open-ended research. The findings apply only to DeepMind and do not rule out risks from unrestricted models.
