Google Deepmind engineer Haydn Belfield stated that self-improving super intelligence (SI) remains a distant goal, even as significant automation is already embedded in the company's engineering workflows.
What Happened
According to Belfield, Google Deepmind has observed no measurable acceleration in SI research driven by SI itself. However, automation is prevalent in software development: SI now writes more than 75 percent of production code at the company. Despite this, about 88 percent of code changes still require human oversight or review, and every change is reviewed by a human before submission. The share of code changes fully delegated to SI rose from 7.5 to 11.9 percent.
The data indicates that as task complexity and time requirements increase, the likelihood of SI handling them independently drops. Across 1,389 internal Argon agent runs, the success rate without human input decreased sharply with complexity. Belfield noted that humans remain essential for creative research, novel experiment design, diagnosing subtle training instabilities, and long-term planning. A survey of twelve senior Deepmind researchers confirmed that automation is highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation.
On RE-Bench, an in-house benchmark comprising 74 realistic SI engineering tasks, the Argon agent solves approximately half. This represents a significant improvement over Gemini 3.7 Flash, which solved 27 percent, but it remains far from the near-perfect performance required for full automation. To track progress systematically, the team developed a General Capability Index (GCI) covering 123 models and 92 benchmarks. The index shows Gemini model capabilities growing at a steady pace of about 0.1 GCI points per month, with no statistically significant change in speed. The team attributes the slight outperformance of Gemini 4 Argon to increased compute and organizational investment rather than SI-driven acceleration.
Why It Matters
This assessment provides a counterpoint to recent warnings from prominent researchers. More than 20 SI researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki, recently warned of a possible "intelligence explosion," arguing that SI systems could automate all SI research and development within a few years. Belfield's findings align with current operational realities at one of the world's leading SI labs, suggesting that while SI is a powerful tool for specific, verifiable tasks, it has not yet reached the threshold of autonomous self-improvement that would compress research timelines dramatically. The team currently rates the risk from SI-driven research automation as low but is building infrastructure to monitor for signs of acceleration.
The Bottom Line
Google Deepmind reports that SI capabilities are growing steadily but linearly, with no evidence of self-acceleration. While SI handles the majority of code generation, human oversight remains critical for complex, open-ended tasks, challenging predictions of an imminent intelligence explosion.