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

Google DeepMind is seeing early signs of recursive self-improvement in AI, with increasingly capable models accelerating the development of subsequent models, according to Logan Kilpatrick, a member of the technical staff at Google DeepMind.

“I think we are like laser-focused right now at the frontier. We’re seeing all these early signs of recursive self-improvement,” Kilpatrick said in an interview on The Pomp Podcast published September 16.

Kilpatrick was responding to criticism that Google has fallen behind rival AI 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 gives it a strong position.

More importantly, he said Google is seeing a feedback loop in which AI systems are increasingly useful for the work required to improve AI itself.

“I think the other labs are seeing this as well,” Kilpatrick said, adding that the development is “underscoring the value of being at the frontier.”

Recursive self-improvement generally refers to AI systems contributing to the development of better AI systems, which can then contribute to further improvements. The process does not necessarily mean that a model autonomously rewrites its own weights. It can also involve AI systems assisting researchers with coding, experiments, evaluations, algorithm discovery and other parts of the model-development process.

Kilpatrick pointed to Google’s recent Gemini release cadence as an example.

“I think you've seen this with the Gemini 3.5, 3.6, 3.7, 3.8, lined up in literally three-to-four-week increments,” he said. “We’re seeing very reasonable progress.”

Google released Gemini 3.6 Flash in July and Gemini 3.8 Flash in September, with the company describing 3.8 as its most capable reasoning and coding model at the time.

Kilpatrick said Google hopes the improvement loop will translate into its upcoming Gemini 4 model, which he described as the company's “largest, most ambitious pre-training run so far”.

“So hopefully we'll see that sort of like translate over in the same way to Gemini 4,” he said. “And it'll be our largest, most ambitious pre-training run so far.”

He added that the model could help Google “get us back in contention with some of the frontier labs”.

Google’s existing research also points towards a broader strategy of using AI to accelerate scientific and technical work. AlphaEvolve has been used for algorithm discovery, while DeepMind is applying AI to areas including genomics, weather forecasting and scientific research.

Google said AlphaEvolve has been used to improve data-centre scheduling, hardware design and AI training. It also improved a matrix multiplication kernel used in Gemini by 23%, reducing Gemini's training time by 1%.

Google DeepMind researchers recently also unveiled Dream-RSI, a framework where AI agents build replay simulators from their discovery histories to test thousands of exploration policies cheaply.

The framework uses recorded exploration trees to test alternative exploration strategies offline, allowing thousands of candidate policies to be evaluated without rerunning costly experiments. The best-performing policy is then deployed for the next round, creating a loop in which each discovery history becomes another simulator for improving future exploration.

Kilpatrick’s comments are significant because recursive improvement has increasingly become part of the discussion around what comes after today's frontier models. 

A June 2026 Google DeepMind report examining the path from AGI to artificial superintelligence identified recursive improvement as one of four possible routes from human-level AI to systems that could surpass the capabilities of large human organisations.