StoryResearch
AlphaEvolve agent discovers improved algorithms, including 4x4 complex matrix multiplication
Google DeepMind's AlphaEvolve, a Gemini-powered coding agent, discovered new algorithms for complex tasks. For 4x4 complex-valued matrix multiplication, it found an algorithm requiring 48 scalar multiplications, improving on Strassen's 1969 method. It also discovered a scheduling heuristic that recovers 0.7% of Google's worldwide compute and accelerated Gemini's training by 1%.
- Discovered matrix multiplication algorithm for 4x4 complex matrices using 48 scalar multiplications, beating prior work
- Found scheduling heuristic that recovers average 0.7% of Google's global compute capacity without new hardware
- Accelerated Gemini's critical kernel by 23%; achieved up to 32.5% speedup for FlashAttention
- On 50+ open math problems, rediscovered state-of-the-art solutions in 75% of cases and improved 20% of known solutions