Google's DeepMind unit, part of Alphabet, has unveiled AlphaEvolve today, an artificial intelligence agent capable of addressing complex programming and mathematical challenges.
The company reports that AlphaEvolve has already been employed to enhance data center efficiency. Additionally, this AI agent shows promising potential in mathematical research and chip development.
AlphaEvolve operates through several stages. Upon receiving a programming task, the agent utilizes Google's lightweight Gemini 2.0 Flash language model to generate multiple code snippets. An automated evaluation system then ranks these snippets based on quality. Subsequently, AlphaEvolve selects the top code snippet and requests Gemini 2.0 Flash to refine it.
The agent optimizes AI-generated code over multiple iterations. When Gemini 2.0 Flash can no longer suggest improvements, AlphaEvolve switches to Gemini 2.0 Pro, a more potent model that trades some speed for enhanced output quality.
"The evolutionary process within AlphaEvolve leverages modern large language models' ability to respond to feedback, allowing discovery of candidates significantly different in syntax and functionality from the initial pool," DeepMind researchers elaborated in their research paper.
Google has implemented AlphaEvolve across various internal projects. Some focus on matrix multiplication, a mathematical operation AI models use to process data. A matrix is a collection of numbers organized in rows and columns akin to a spreadsheet.
Rather than drawing processor blueprints, chip designers write them using a programming syntax called Verilog. In one project, AlphaEvolve assisted Google engineers in improving the Verilog code for circuits executing matrix multiplication. The company has incorporated this circuit into upcoming products in its TPU series of AI processors.
In another internal initiative, AlphaEvolve developed methods enabling Google’s Gemini models to break down matrix multiplication into smaller, more manageable computations. According to the search giant, these enhancements increased the speed of one of Gemini's pivotal components by 23%.
AlphaEvolve also aided in boosting the efficiency of the company’s data centers. Google uses a software platform named Borg to manage its infrastructure resources. AlphaEvolve recommended improvements to this platform, recovering an average of 0.7% of Google’s global computing resources, as detailed by DeepMind researchers.
According to the search giant, the reasoning capabilities that enable AlphaEvolve to optimize data centers and chip designs make it useful for mathematical research. "To explore AlphaEvolve's breadth, we applied the system to over 50 open problems in mathematical analysis, geometry, combinatorics, and number theory," researchers wrote in a blog post accompanying the paper. "The system's flexibility allowed us to set up most experiments within hours. In about 75% of cases, to our knowledge, it rediscovered state-of-the-art solutions."
Google plans to provide academic circles with access to this AI agent through an early access program. Moreover, the company is exploring possibilities for expanding access to more users in the future.
"While AlphaEvolve is currently being applied in mathematics and computing, its general nature means it can be used for any problem whose solution can be described algorithmically and automatically verified," wrote DeepMind researchers. "We believe AlphaEvolve could drive transformative changes across many fields, including materials science, drug discovery, sustainability, and broader technological and commercial applications."
