Google engineer Haydn Belfield says self‑improving AI – the idea that AI can design better AI without human help – is not close to reality. DeepMind’s internal data shows AI now writes over 75 % of production code, yet about 88 % of those changes still require a human to check them before they are submitted.
Automation works for narrow tasks, not for research
The share of code changes fully delegated to AI rose from 7.5 % to 11.9 % as of the latest internal measurements. When tasks get more complex, AI success without human input drops sharply. In 1,389 runs of DeepMind’s Argon agents, humans remained essential for creative research, experimental design, diagnosing subtle training problems and long‑term planning. A small survey of twelve senior DeepMind researchers echoed this pattern, noting high automation for well‑defined, verifiable work and low automation for open‑ended scientific ideation.
Benchmarks show steady, not accelerating, progress
On DeepMind’s RE‑Bench – a set of 74 realistic machine‑learning engineering tasks – Argon solves about half of the problems, up from Gemini 3.7 Flash’s 27 % score. The General Capability Index (GCI), which tracks 123 models across 92 benchmarks, shows Gemini models improving at roughly 0.1 GCI points per month, with no statistical sign of speed‑up. Belfield says the modest bump for Gemini 4 Argon comes from more compute and organizational effort, not from AI‑driven acceleration.
Why it matters
For developers at Google and other companies, AI can handle routine coding but still requires human review, so jobs won’t disappear overnight. For the broader AI community, the lack of self‑accelerating research means predictions of a rapid “intelligence explosion” remain speculative, and safety concerns can be monitored over a longer horizon.
