Google DeepMind has announced a new artificial intelligence (AI) agent called AlphaEvolve. It helps solve complex problems in math and computing.
AlphaEvolve uses large language models (LLMs). It combines creativity from Gemini models with an evolutionary framework. This framework improves ideas over time. AlphaEvolve also uses automated evaluators to check if solutions work well. It can summarize documents, generate code, or brainstorm ideas, but now it targets bigger challenges.
AlphaEvolve improves Google’s data centers, chip design, and AI training, and even helps train the LLMs that power AlphaEvolve itself. It designs faster algorithms for matrix multiplication, a key math operation in computing. It also finds new solutions to open math problems, showing its wide potential. In 2023, researchers showed LLMs can write code to solve scientific problems. AlphaEvolve takes this further by evolving entire codebases, creating complex algorithms.
Improving Technology and Math
AlphaEvolve uses two Gemini models: Gemini Flash explores many ideas, while Gemini Pro gives deeper insights. Together, they suggest programs that solve problems. AlphaEvolve tests these programs with metrics, which are objective measures of quality and accuracy. This works well in math and computer science, where progress can be measured clearly. Over the past year, AlphaEvolve’s algorithms improved Google’s data centers, hardware, and software, making the digital system more efficient and sustainable.
In data centers, AlphaEvolve found a simple rule to manage tasks better, saving 0.7% of Google’s computing resources. It also rewrote code for a chip circuit, making matrix multiplication faster in Google’s Tensor Processing Unit (TPU), an AI accelerator. In AI training, AlphaEvolve sped up matrix operations by 23%, cutting Gemini’s training time by 1%. It also improved GPU instructions, boosting speed by 32.5% for AI models. In math, AlphaEvolve found a new way to multiply 4x4 complex matrices using 48 steps, beating a 1969 record. It also tackled 50 math problems, improving solutions in 20% of cases, like the kissing number problem in geometry.
AlphaEvolve keeps growing as LLMs improve at coding. Researchers plan to share it with academic users through an Early Access Program. Its ability to evolve algorithms can help in material science, drug discovery, and sustainability, transforming many fields.
