Google DeepMind sees no AI-driven speedup in AI research, engineer says

Google engineer Haydn Belfield says self-improving AI remains far off. According to Belfield, Google DeepMind has seen no measurable acceleration of AI research driven by AI itself. The Decoder, which relays the findings, stresses that this does not mean automation is absent: AI now writes more than 75 percent of production code. But about 88 percent of code changes still need human oversight or review, and every change is reviewed by a human before submission, according to Belfield. The share of code changes fully delegated to AI rose from 7.5 to 11.9 percent; the source does not say over what period.

The more complex and time-consuming a task gets, the less likely AI is to handle it on its own. Across 1,389 internal Argon agent runs, the success rate without human input dropped sharply as complexity increased. Humans remain essential for creative research, novel experiment design, diagnosing subtle training instabilities and long-term planning, Belfield said. A survey of twelve senior DeepMind researchers confirmed the pattern, with automation running highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation.

On DeepMind's in-house benchmark RE-Bench, which includes 74 realistic ML engineering tasks, Argon solves about half of them. That is a big jump from Gemini 3.7 Flash at 27 percent, but The Decoder notes it is still far from the near-perfect performance full automation would require.

To track progress systematically, the team built a General Capability Index (GCI) covering 123 models and 92 benchmarks. Gemini model capabilities are growing at a steady pace of about 0.1 GCI points per month, with no significant change in speed. The team attributes Gemini 4 Argon sitting slightly above the trend line to more compute and organizational investment rather than AI-driven acceleration. It currently rates the risk from AI-driven research automation as low, but is building measurement infrastructure to track signs of acceleration going forward.

Not everyone agrees. More than 20 AI researchers, including Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, recently warned of a possible "intelligence explosion" in a research paper. They argued that AI systems could automate all AI research and development within a few years, compressing what would normally be years of progress into just months.

Key facts

Google engineer Haydn Belfield says Google DeepMind has seen no measurable acceleration of AI research driven by AI itself.
AI now writes more than 75 percent of production code, but about 88 percent of code changes still need human oversight or review; fully delegated changes rose from 7.5 to 11.9 percent.
Across 1,389 internal Argon agent runs, success without human input dropped sharply as task complexity rose; on the 74-task RE-Bench, Argon solves about half, against 27 percent for Gemini 3.7 Flash.
DeepMind's General Capability Index (123 models, 92 benchmarks) shows Gemini capabilities growing at about 0.1 GCI points per month with no significant change in speed.
More than 20 researchers, including Hinton, Bengio and Pachocki, recently warned in a paper of a possible "intelligence explosion".

Why it matters

The debate over whether AI will soon speed up its own development usually runs on argument. Here it gets internal numbers from one of the labs best placed to see it. Belfield's account says the capability trend for Gemini models is steady at about 0.1 GCI points per month, and that the one model sitting slightly above the line is explained by more compute and organizational investment, not by AI doing the research. It also sits against a recent paper by more than 20 researchers who argued AI could automate all AI research and development within a few years.

Who it affects

Mainly people who follow AI timelines and safety: researchers, policymakers and anyone weighing claims of an imminent "intelligence explosion". It also touches engineers and research teams, since the findings describe where human work still matters at DeepMind: creative research, novel experiment design, diagnosing subtle training instabilities and long-term planning.

How to use it

There is no product or tool here. The useful part is the set of reference points: what share of code AI writes (more than 75 percent), how much still gets human review (about 88 percent of changes, and every change before submission), and how agent success falls as tasks get longer and more complex. Anyone judging claims about research automation can compare them against these figures. The team says it is building measurement infrastructure to track signs of acceleration going forward.

How solid is it

The figures are internal Google DeepMind data as reported by Belfield and relayed secondhand by The Decoder; the original publication is not linked in the article. The RE-Bench, Argon-run and GCI results come from the team's own benchmark and index, and the explanation for Gemini 4 Argon sitting above the trend line is the team's own attribution. The survey covered twelve senior DeepMind researchers.

Risks and caveats

The source does not say where or when Belfield published or presented the findings, and gives no date. The period over which fully delegated code changes rose from 7.5 to 11.9 percent is not stated, and the source does not say the 88 percent and 75 percent figures are measured over the same period or code base. It does not define the General Capability Index or its scale beyond the 0.1 points per month rate, and gives no timescale for when self-improving AI might arrive; it only says the prospect remains far off. It also does not name the paper by the 20-plus researchers or give a Google response to it. The team rates the risk as low today, which is a current assessment, not a forecast.

“The more complex and time-consuming a task gets, the less likely AI is to handle it on its own.”

— The Decoder, summarising Haydn Belfield's findings