Overview

• AlphaEvolve, DeepMind's latest AI system, combines Gemini language models with an iterative evaluation loop to tackle complex algorithmic challenges.
• The system rediscovered known solutions 75% of the time and improved them in 20% of cases during benchmark tests on mathematical problems.
• AlphaEvolve optimized Google’s infrastructure by recovering 0.7% of compute resources and reducing Gemini model training time by 1%.
• The AI produced provably novel algorithms, including an improvement on the Strassen algorithm for matrix calculations, a 56-year-old standard.
• While promising, AlphaEvolve is limited to machine-gradeable, numerical problems and is set to launch an early-access program for academics.