Google DeepMind unveils AlphaEvolve, an AI agent that can generate new algorithms

Google DeepMind has announced a new AI agent that can solve complex coding and math problems, called AlphaEvolve. In a blog posted about the release, the company said that AlphaEvolve can improve the efficiency of data centers, chip design and AI training processes, including “training the large language models underlying AlphaEvolve itself.” 

The algorithm generated by AlphaEvolve was deployed at Google’s Borg cluster management system for their data centers which led to a 0.7% recovery in their fleet-wide compute resources on average, CEO Sundar Pichai noted. 

“AlphaEvolve discovered new efficient algorithm for matrix multiplication, a fundamental problem in Computer Science, improving Strassen’s method that has stood for 50 years. Applied to 50 open Math problems, it matched the best known answers 75% of the time and improved on 20%,” Pushmeet Kohli, head of AI for science at DeepMind added. 

The agent is able to generate code using the lightweight Gemini 2.0 Flash language model after which a self-evaluating method to rank the code in terms of quality. Then, AlphaEvolve picks out the best pieces of code and improves them over multiple rounds. 

Once it reaches a point where no more suggestions are made, the agent starts using Gemini 2.0 Pro. 

This reduces the risk of hallucinations which are common even in the most advanced large language models. 

DeepMind said that AlphaEvolve also made changes to Google’s Tensor processing hardware by removing unnecessary bits from the chip’s Verilog hardware description.