Google: AI Is Not Close To Improving Itself

Haydn Belfield, an engineer at Google DeepMind (Google's AI research lab), says self-improving AI is still far from reality. Self-improving AI means AI that speeds up its own development. At Google, AI now writes more than 75% of the code that ships in real products. But 88% of these AI-written changes still go through human review before release. The share of code changes that AI handles with no human review at all rose from 7.5% to 11.9%. Belfield said the longer and more complex a task gets, the less likely AI can handle it on its own. Google built its own capability index to track model progress over time. It climbs steadily by 0.1 points a month, with no sign of speeding up.

AI does well on narrow, clearly defined coding tasks. It starts to struggle once a task gets longer and more complex. Human review stays mandatory before any change ships. This points to slow, gradual progress, not to AI suddenly learning to improve itself without help. Another group of researchers recently warned of a coming «intelligence explosion».