DeepMind Finds No Evidence of Self‑Improving AI, Despite 75% Code Automation
Current overview (not historical)
Google DeepMind researcher Haydn Belfield reports that AI has not yet begun to accelerate its own development. While the system can automatically generate 75 % of production code, 88 % of code changes still require human review. In 1,389 runs of the Argon agent, success rates drop sharply as task complexity rises, and the Google General Capability Index shows no measurable increase in research speed. These findings contradict earlier claims that self‑improving AI was imminent. Belfield’s analysis suggests that, although automation is high, the bottleneck remains human oversight and the difficulty of scaling to more complex problems. The results imply that, for now, AI progress remains largely driven by human engineers rather than autonomous self‑improvement.
Written with AI from the reporting · updated 10/15/2026
Expand the detailed fact route and sources
Story route
A station-by-station route from the first public evidence onward; every station carries its own evidence and uncertainty.
- 1Unclassified fact stationOct 15, 06:24single sourcereport-time onlyrevised
Google Deepmind报告：自改进AI仍远未实现
Read original source ↗ The DecoderReplay this moment in the explorerEvidence & revisions · 1 source records
- Subject：
- Action：
- 发布关于自改进AI能力的分析报告
- Object：
- AI自动化研究能力
- Occurred (fact side)：
- not recorded
- Published (valid side)：
- Oct 15, 03:18
- First seen (known side)：
- Oct 15, 06:24
Fact stationSingle source · one-sided claimRevisedReport-time onlyCopy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.
- 2Unclassified fact stationOct 14, 00:00recorded Oct 16, 02:24single sourcerevised
DeepMind发现AI自研能力未显著加速
Read original source ↗ DecryptReplay this moment in the explorerEvidence & revisions · 1 source records
- Subject：
- Action：
- 发布Gemini 4 Argon安全报告并评估ML R&D自动化情况
- Object：
- AI模型自动化研究能力
- Occurred (fact side)：
- Oct 14, 00:00
- Published (valid side)：
- Oct 16, 01:36
- First seen (known side)：
- Oct 16, 02:24
Fact stationSingle source · one-sided claimRevisedCopy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.
Uncertainty marks flag where to check the sources; they are not proven causation.
Report timeline
Explore every side through dated source reports.
- DecryptRelated fact: single sourceSource report revisedGoogle DeepMind finds no significant acceleration in AI progress despite high AI code contribution
Report first recorded Oct 16, 02:24
In the Gemini 4 Argon Safety & Security Report, Google DeepMind found no statistically significant acceleration in its AI progress after April 1, 2026; the capability index has climbed a steady 0.1 points a month. Although over 75% of production code at the lab is AI-written, 88% of code changes still require human steering or writing. The lab concluded that there is currently no significant end-to-end acceleration of ML R&D, leading them to assess the risk from AI-driven ML R&D as currently low, contrary to concerns about recursive self-improvement.
Read original source ↗
- The DecoderRelated fact: single sourceFact occurrence time unverifiedSource report revisedGoogle says self-improving AI is distant despite significant code automation
Report first recorded Oct 15, 06:24
Google engineer Haydn Belfield states that self-improving AI is not yet measurable, even as AI now writes over 75 percent of production code. However, 88 percent of code changes still require human oversight, and the success rate of Argon agent runs sharply drops as task complexity increases. Google's General Capability Index shows Gemini model capabilities growing steadily at about 0.1 GCI points per month, without statistically significant acceleration, which the team attributes to compute and organizational investment, not AI-driven acceleration.
Read original source ↗
Event discussion
Reader opinions are hosted on GitHub Discussions; posting requires a GitHub account. Comments are not verified facts.