AlphaEvolve

Google DeepMind system that designs novel algorithms. A live example of AI improving AI infrastructure — the concrete, deployed end of the Recursive Self-Improvement argument.

What it did (May 2025)

• Proposed a data-centre workload change saving ~0.7% of Google's worldwide compute — a recovered slice of capacity at hyperscale.
• Improved matrix multiplication, which in turn sped Gemini training ~1%.

The significance isn't the headline percentages but the loop: an AI system optimizing the very infrastructure that trains AI systems, with measurable gains fed back into the next training run.

Cross-links

• The thesis it instantiates · Recursive Self-Improvement
• Research thread it belongs to · Self-Evolving Agents (named in that page's lineage alongside Gödel Agent, Darwin-Gödel Machine, etc.)

Sources

• How AI Got Better at Building Itself (Economist) — the data-centre compute saving, the matrix-multiplication improvement, and the Gemini training speed-up.