Google DeepMind's Self-Improving AI Ambitions May Not Be as Revolutionary as Claimed

Google DeepMind is working on self-improving AI models, but recent developments show that the concept might not be as far-reaching as initially thought.

The company's leaders have mentioned recursive self-improvement in various interviews and public comments, but a closer look at their actual progress reveals a more nuanced picture.

Gemini 4, Google DeepMind's ambitious pre-training effort, has been running on the Ironwood TPU fleet since July 21, 2026. However, despite its 'ambitious' nature, the project doesn't appear to be rewriting its own core weights unsupervised.

The distinction is crucial in AI safety research, where recursive self-improvement refers to a system that gets better at improving itself without human intervention. Google DeepMind's AlphaEvolve product, which automates algorithm discovery and optimization, is the closest thing to self-improving currently available from the company.

AlphaEvolve uses an ensemble of two Gemini models to iteratively refine its performance on specific tasks, but it still relies on human-defined goals and constraints. The system's ability to improve itself is limited to the task at hand, rather than operating as a standalone agent.