Google DeepMind Sees Early Signs of Recursive Self-Improvement Ahead of Gemini 4

Logan Kilpatrick, a member of the technical staff at Google DeepMind, says the company is seeing early signs of recursive self-improvement, in which increasingly capable models accelerate the development of the models that follow them. He made the comments in an interview on The Pomp Podcast published on September 16, as reported by analyticsindiamag.com.

Kilpatrick was responding to criticism that Google has fallen behind rival frontier labs. He said the criticism is understandable given the expectations around Google, but argued that the company's broad portfolio across AI research and science keeps it in a strong position. He described a feedback loop in which AI systems are increasingly useful for the work required to improve AI itself, said other labs are seeing the same thing, and added that this underscores the value of being at the frontier.

One piece of evidence he cited is Google's release cadence for Gemini: Gemini 3.5, 3.6, 3.7 and 3.8 arrived in roughly three-to-four-week increments. Google released Gemini 3.6 Flash in July and Gemini 3.8 Flash in September, describing 3.8 at the time as its most capable reasoning and coding model.

The larger question is whether that loop carries into Gemini 4. Kilpatrick said Google hopes the improvement loop translates over to the upcoming model, which he called the company's largest and most ambitious pre-training run so far, and said it could help Google get back in contention with some of the frontier labs.

Existing research is part of the same story. AlphaEvolve has been used for algorithm discovery and applied to areas including data-centre scheduling, hardware design and AI training; according to Google, it improved a matrix multiplication kernel used in Gemini by 23 percent, reducing Gemini's training time by 1 percent.

On the methods side, DeepMind researchers recently unveiled Dream-RSI, a framework in which AI agents build replay simulators from their own discovery histories. Those simulators allow thousands of candidate exploration policies to be evaluated offline instead of through costly re-runs, and the best-performing policy is then deployed for the next round, so each discovery history becomes another simulator for improving future exploration.

Recursive improvement is not only a podcast talking point. A June 2026 Google DeepMind report examining the path from AGI to artificial superintelligence listed recursive improvement as one of four possible routes from human-level AI to systems that could surpass the capabilities of large human organisations.

Recursive self-improvement is moving from a distant talking point into release cadences and corporate narratives, and the test ahead is whether Gemini 4 actually shows the loop at work and whether labs can produce verifiable numbers for the gains. For now, what Google calls early signs is mainly about AI assisting researchers with coding, experiments, evaluations and algorithm discovery, rather than a model rewriting its own weights.