What happened
On 14 May 2025 DeepMind described AlphaEvolve, an evolutionary loop over Gemini-generated code edits with automatic evaluators. The white paper and later arXiv preprint claimed a rank-48 tensor decomposition for 4×4 complex matrix multiplication, improving on Strassen’s rank-49 recursive scheme in that setting, plus scheduling and kernel changes used inside Google. Terence Tao publicly noted that the blog oversimplified the distinction between a tensor decomposition and an arbitrary 48-multiplication algorithm. Evidence remains a developer paper plus expert commentary, not a journal article. Update (2026-03-19): How we monitor internal coding agents for misalignment ↗. Monitoring model behavior in real internal workflows.