AlphaEvolve breaks decade-old Ramsey theory records

DeepMind is shifting from using AI to solve math problems to using AI to *invent the algorithms* that solve them.

• –AlphaEvolve established new lower bounds for $R(3, 13)$, $R(3, 18)$, $R(4, 13)$, $R(4, 14)$, and $R(4, 15)$ using a single unified framework.
• –Unlike previous hand-crafted heuristics, the system uses Gemini to mutate and iterate on code snippets, effectively automating the role of the researcher.
• –This "algorithm evolution" approach proved robust enough to recover 28 state-of-the-art results where original methodologies were never even published.
• –The results suggest that any mathematical domain with a fast evaluator (like graph theory) is now ripe for automated discovery via code-generation agents.