AlphaEvolve’s processing occurs in multiple steps and it uses Gemini to creatively solve problems.
Google DeepMind has introduced AlphaEvolve, a powerful AI agent that solves complex programming and mathematical problems. The technology is already being used by the company itself to design chips more efficiently and better manage datacenters.
Step-by-Step Thinking
AlphaEvolve processes in multiple steps. First, the model generates code using the language model Gemini 2.0 Flash. Then, an evaluator ranks the results, and AlphaEvolve selects and refines the best outcomes. When Gemini 2.0 Flash can no longer suggest improvements, AlphaEvolve switches to the more powerful Gemini 2.0 Pro.
The system has already been deployed for matrix multiplications, a crucial operation in AI. In one case, AlphaEvolve improved code for a chip that will soon be part of a new TPU processor. In another project, the AI agent made Gemini models 23 percent faster.
The tool also improved the internal platform Borg, which Google uses to control its datacenters. Thanks to AlphaEvolve, an average of 0.7 percent more computing power is now utilized.
Mathematical Problem Solver
According to DeepMind, AlphaEvolve is equally suitable for mathematics. The system was tested on more than 50 open problems in analysis, geometry, and number theory. In 75 percent of cases, Google claims it discovered highly advanced solutions.
AlphaEvolve will first be available to researchers through early access. This will likely be followed by a broader rollout to education and industry.
read also