Google engineer says self-improving AI remains far from reality
Google engineer Haydn Belfield argues that self-improving AI is not yet achievable, despite automation advances. At Google DeepMind, AI handles over 75% of production code but still requires human oversight for 88% of changes. Only 11.9% of code changes are fully delegated to AI, up from 7.5%, with success rates dropping sharply for complex tasks across 1,389 internal Argon agent runs.
Key points
- Google DeepMind’s AI writes 75% of production code but needs human review for 88% of changes
- Only 11.9% of code changes are fully automated by AI, up from 7.5% previously
- DeepMind’s GCI shows no measurable acceleration in AI research from self-improvement
Belfield’s findings contrast with warnings from over 20 AI researchers, including Geoffrey Hinton and OpenAI’s Jakub Pachocki, who argue AI could automate all AI research within years. DeepMind’s General Capability Index (GCI) shows steady but unspectacular growth—0.1 points per month—with no evidence of AI-driven acceleration. The team attributes recent gains in Argon’s performance to increased compute and organization, not self-improvement.
Model page: Gemini 4 Argon →
Google says self-improving AI isn't anywhere close
The Decoder · 14 October 2026
Loading the full article…
This text was published by The Decoder and written by Matthias Bastian. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Stanford economists argue AI should focus on human collaboration over automation · 1 src
- Anthropic warns small AI models may self-replicate on IoT devices by end of 2026 · 22 src
- Meta publishes ten health studies generated with its Muse AI platform · 1 src
- MIT study finds AI solved 35% of theory problems by mid-2026 · 1 src
- Researchers release Mr.Chatterbox-2.7B-Chat model trained on 19th-century British texts · 1 src
Comments
via GitHub Discussions