Google Deepmind engineer Haydn Belfield has revealed that self-improving artificial intelligence remains far off, contradicting fears of an imminent intelligence explosion. Although AI currently writes over 75 percent of Google's production code, about 88 percent of code changes still require human oversight and review before submission. Furthermore, the share of code changes fully delegated to AI rose only from 7.5 to 11.9 percent, showing that genuine autonomy is advancing at a measured pace.
Data from internal Argon agent runs and the RE-Bench benchmark demonstrate that AI struggles significantly as task complexity increases. While Argon successfully solves about half of the 74 realistic machine learning engineering tasks on RE-Bench—an improvement over Gemini 3.7 Flash at 27 percent—it remains far from the near-perfect performance required for full automation. Google's General Capability Index tracking 123 models shows a steady capability growth of about 0.1 points per month, with no measurable acceleration in AI-driven research.
Despite warnings from prominent researchers like Geoffrey Hinton and Yoshua Bengio regarding a potential rapid intelligence explosion, Google currently rates the risk of AI-driven research automation as low. Humans remain fundamentally essential for creative research, novel experiment design, long-term planning, and diagnosing training instabilities. This insight matters because it provides a grounded reality check on the actual speed of research automation amid widespread industry hype.