Google is using Gemini models alongside its researchers to help develop their successors, according to co-founder Sergey Brin. In a Google for Developers discussion, Brin said the company’s internal models are already ahead of the versions available publicly.
The approach points to recursive self-improvement: a model assists with building a more capable system, which can then contribute to work on the generation that follows. Google is already applying AI agents to monitor training runs, flag incorrect configurations, identify duplicate data and suggest fixes for coding errors.
DeepMind’s Gemini 4 Argon safety report offers a more measured view of how much of that work can currently be handed to AI without direct intervention. The report says code changes are still reviewed by people before they are integrated.
- 11.9% of proposed code changes were fully delegated to AI in September, up from 7.5% in April.
- About 88% still required human involvement, either through repeated instructions or code written by programmers.
- Argon completed roughly half the tasks in an internal AI research-and-development evaluation, below the level needed for full automation.
That gap matters because generating or repairing code is not the same as directing research. Human researchers remain responsible for selecting experiments and diagnosing why a training run is producing poor results, according to the report.
Google’s analysis, based on 92 benchmarks, did not find a statistically significant acceleration in overall model progress. Although Argon outperformed the previous trend, the report says greater investment and additional compute are the more likely explanations. The key open question is whether AI-assisted research can eventually produce a measurable improvement beyond those conventional inputs.
