Google enters recursive self-improvement phase with Gemini models

Google co-founder Sergey Brin has announced that the company is entering the early stages of “recursive self-improvement” (RSI), a phase where AI models assist in developing their own successors.

During a discussion regarding the new Gemini 4 Argon model, Brin noted that researchers and AI models are now working “hand in hand” to prepare subsequent versions. This process aims to significantly accelerate the pace of AI advancement. Brin indicated that Google has already joined the “RSI treadmill,” with internal teams testing even more advanced models, such as a version known as Carbon.

According to Google’s Argon security report, certain aspects of the AI research and development cycle are already automated, including much of the production code and bug fixes, though human oversight remains a requirement before deployment. Industry competitors Anthropic and OpenAI have also indicated they are approaching similar RSI capabilities, with researchers predicting RSI could become common practice within the next three to twelve months.