Google DeepMind Explores Self-Improving AI with Gemini 4 Pre-training and AlphaEvolve

Google DeepMind is engaged in an ambitious pre-training effort for its next-generation AI model, Gemini 4, with leaders openly discussing recursive self-improvement as a near-term goal. While a fully autonomous self-improving model has not yet been achieved, a live product called AlphaEvolve is already automating aspects of this work, indicating significant advancements in AI capabilities.