Google: Gemini models now help build their own successors

What's going on

Google is leaning more and more on its own AI to build its next generations of AI. That's according to ad-hoc-news.de (editorial team boerse-global.de) on October 24, 2026, citing statements from Google co-founder Sergey Brin and CEO Sundar Pichai. The report says the Gemini models are now helping create their own successors. Brin says Google is entering a cycle of gradual "recursive self-improvement". That means one AI helps build better AI, which in turn helps with the next step.

What the AI actually does

According to the report, the systems take on tasks such as searching for suitable model architectures (the basic structure of a model), optimizing training runs and systematically evaluating performance. In practice, the system proposes new "training recipes" and assesses so-called checkpoints, which are saved intermediate states of a training run. Human developers still make the final decisions.

Brin stressed that earlier Flash models (Google's fast, smaller model line) laid the technical groundwork for bringing the AI directly into the design of the next generation. The feedback loop is still at an early stage, he said, but it is already noticeably speeding up internal innovation cycles. Human scientists remain essential; the models are support, not a replacement.

Pichai's take

Pichai tempers expectations: Google is still far from a full end state of autonomous recursive self-improvement, and fully autonomous processes are years away. For now, research teams use specialized AI agents to assist with parts of the work. Over the next twelve months, he expects continued acceleration from feedback effects, and further progress in 2027.

On code: in April 2026, 75 percent of new code at Google came from AI, according to earlier statements. Pichai now called that figure meaningless, because virtually all code at the company is influenced by AI in some form. For Gemini 4 Argon, he said using AI cut the number of required code rollbacks (reverting faulty changes) by a third.

Demand and capacity limits

In its first 72 hours after release, Gemini 4 Argon processed almost 25 trillion tokens (the text chunks AI models work with). That's roughly three times what Gemini 3.0 handled in a comparable period. Because of capacity bottlenecks, Google had to temporarily throttle the compute it made available.

Changed product planning

According to the report, the faster cycles also changed planning. The originally planned Gemini 3.5 Pro was scrapped because, in Pichai's words, it would have had a very short functional half-life. Instead, in mid-2026 Google closed identified gaps in agent capabilities and coding in about three months through rapid Flash iterations from version 3.5 to 3.8. The results fed into the post-training (fine-tuning after the main training run) of Argon.

Industry context

The article points to an October 2026 report from Gladstone AI. It says most surveyed researchers at Anthropic and DeepMind expect effective recursive self-improvement to begin within 3 to 12 months. The source gives no details on the methodology.

Our take

AI helping to build AI is a central topic for industry watchers, because it affects both the pace of development and safety questions. However, the statements come from Google itself and haven't been independently verified. Apart from the rollback figure, concrete measurements (such as how much faster development has actually become) are missing. There's also a notable tension between Brin's wording ("recursive self-improvement") and Pichai's toning down to "assisting agents". The source is also a finance portal with ads for EU AI Act guides mixed in. The information is secondhand, and we don't have the original statements.