Google DeepMind reportedly bets on recursive self-improvement for Gemini 4

Multiple reports indicate Google DeepMind has made significant advances in recursive self-improvement (RSI), where AI systems autonomously refine and optimize their own algorithms — a capability widely regarded as a critical, and dangerous, step toward artificial general intelligence. The research is described as a top priority for DeepMind's leadership and is speculated to be integrated into the upcoming Gemini 4 model, enhancing its adaptability and letting it optimize performance without external intervention.

The strategic subtext, per a 36kr report, is that DeepMind's flagship is reportedly running roughly two months behind competitors, and RSI is being framed as a high-risk, high-reward bet to dramatically accelerate iteration speed and close that gap. The results remain unverified and the framing is speculative, but the direction is significant precisely because RSI is the exact mechanism that this week's departing Anthropic researcher cited when warning about 'racing straight to self-improving superintelligence.'

Competitively, the timing is pointed: DeepMind is pursuing autonomous self-improvement in the same week OpenAI's Astra reached a 'Critical' cyber threshold and the industry's leaders publicly endorsed a slowdown. If Gemini 4 does ship with RSI-derived capabilities, it would reframe the frontier race around self-optimizing systems rather than raw scale.

What to watch: whether DeepMind publishes any concrete RSI results or benchmarks, how it squares an RSI push with Google's participation in the 'pace the frontier' conversation, and whether Gemini 4 arrives with verifiable evidence of self-improvement rather than marketing framing.