AlphaEvolve: DeepMind's AI Evolves Algorithms, Cuts Google Costs

AlphaEvolve: DeepMind's Gemini-based AI invents algorithms, optimizing Google's data centers & matrix multiplication, improving efficiency and hardware design.

• DeepMind's AlphaEvolve: Google DeepMind has created a new AI system called AlphaEvolve, based on Gemini LLMs, that uses an "evolutionary" approach to improve algorithms.

• Algorithm Invention: AlphaEvolve can invent new algorithms.

• Automatic Evaluation System: AlphaEvolve uses an automatic evaluation system. Researchers input a problem with possible solutions, and the AI generates multiple solutions, which are then analyzed. An evolutionary framework allows it to focus on and improve the best solution.

• General-Purpose AI: Unlike past AI systems trained on single domains, AlphaEvolve is a general-purpose AI that can aid research in any programming or algorithmic problem.

• Data Center Efficiency: AlphaEvolve suggested a change to the scheduling heuristics for Google's Borg cluster management system, saving Google 0.7 percent on its computing resources globally.

• Improved Matrix Multiplication: AlphaEvolve discovered a new algorithm that's more efficient than the one devised by mathematician Volker Strassen in 1969 for multiplying 4x4 complex-valued matrices. This could make generative AI more efficient.

• Tensor Chip Improvement: AlphaEvolve created a change to Google's next-generation Tensor processing hardware's Verilog hardware description language, dropping unnecessary bits to increase efficiency. This is still being verified.