Google DeepMind’s reported safety assessment finds no meaningful acceleration in the pace of its internal AI research and development, despite AI writing more than 75% of the lab’s production code. The finding offers a useful counterweight to warnings about recursive self-improvement, but it describes one lab’s internal measures, not the whole industry.
Google DeepMind Watch analysis
What happened
DigitalToday, citing Decrypt, says DeepMind’s Gemini 4 Argon safety report found no statistically significant change in the pace of AI performance progress since 1 April 2026. The report’s internal capability index rose by 0.1 points a month, according to DigitalToday. Gemini 4 Argon scored above the index’s earlier trend, which DeepMind attributed to more computing resources and wider organisational investment, rather than AI speeding up its own development.
AI involvement in coding is rising, but the account describes a long way to full autonomy. AI-generated pull requests at DeepMind rose from 7.5% in April to 11.9% in September. About 88% still required separate instructions, and the models independently solved around half of the lab’s internal research-engineering tasks.
Why it matters
The figures speak directly to a live argument: whether AI is already accelerating the research that produces more capable AI. On the account reported by DigitalToday, DeepMind’s own data does not show a meaningful end-to-end acceleration so far. The report also distinguishes routine, clearly defined work from open-ended scientific discovery, where automation was lower.
That is a narrower conclusion than “AI can’t improve AI”. The assessment is based on data from DeepMind’s lab, and DigitalToday says the report does not rule out risks from an unrestricted Gemini 4 Argon in areas such as cyberattacks or biological-weapons information.
Our read
This is valuable evidence to add to the debate, not a final whistle. The practical takeaway is that high AI use in code production does not, on its own, show that research is speeding up: the reported pull-request and task figures suggest substantial human direction remains. Keep the lab-specific limits attached to the headline; extrapolation is where a tidy statistic starts wearing a much larger hat.
What to watch
- Whether DeepMind publishes the underlying report and more detail on its capability index.
- Whether later measurements show changes in research pace or in how much work AI completes independently.
- How the findings compare with evidence from other AI labs. Activity teaser: DeepMind’s reported safety assessment finds no meaningful acceleration in its internal AI research pace, even though AI writes more than 75% of its production code. The figures suggest coding assistance and autonomous research are not the same thing, and the assessment is limited to one lab. Our full story examines what the reported measurements do, and do not, say about recursive self-improvement.
Discussion spark: What evidence would persuade you that AI is accelerating AI research: rising automation in routine coding, or clear gains in independently completed research work?
Sources and evidence
- Google DeepMind says no evidence AI self-improvement is accelerating, diverging from industry warnings – 디지털투데이 (16 October 2026, 04:45 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind