Google DeepMind has introduced AlphaEvolve, a coding agent that takes an evolutionary approach to general-purpose algorithm discovery and model optimization. AlphaEvolve combines the creative problem-solving abilities of Google’s Gemini models with automated evaluators that verify answers, then applies an evolutionary framework that improves on the most promising results. Evolutionary AI refers to techniques inspired by biological evolution, including natural selection, to optimize and design machine learning models.

Rather than relying solely on traditional gradient-based optimization, evolution draws on principles like mutation, crossover and selection to iteratively improve potential solutions.

Although currently in the prerelease stage, Google has been using AlphaEvolve internally for the past year. During that time, it was found to have enhanced the efficiency of its data centers, chip design and AI training processes, “including training the large language models underlying AlphaEvolve itself,” Google DeepMind explains in a blog post.

While AlphaEvolve applications are currently focused on math and computing, “its general nature means it can be applied to any problem whose solution can be described as an algorithm and automatically verified,” Google says.

The evolutionary process “is applicable both to problems where discovering new algorithms is the intrinsic goal, as well as to the broad range of problems where the solution of interest is not an algorithm itself but an algorithm can describe how that solution is to be constructed or found,” according to a white paper on AlphaEvolve.

TechCrunch reports “AlphaEvolve introduces a clever mechanism to cut down on [AI] hallucinations: an automatic evaluation system,” explaining that despite rapid advances in training, some newer models “hallucinate more than their predecessors,” underscoring the “challenging nature” of solving the problem.

The evolutionary approach “uses models to generate, critique, and arrive at a pool of possible answers to a question, and automatically evaluates and scores the answers for accuracy,” TechCrunch notes.

VentureBeat describes AlphaEvolve as “the Google AI that writes its own code and just saved millions in computing costs” and goes into detail on various solutions, including discovering a new algorithm used to power “Borg, Google’s massive cluster management system.” The scheduling heuristic “recovers an average of 0.7 percent of Google’s worldwide computing resources continuously — a staggering efficiency gain at Google’s scale,” VentureBeat writes.

A separate VentureBeat piece delves deeply into that accomplishment and explores how enterprise companies can adapt the process to their own needs.

Google says it is building a user-friendly interface for interacting with AlphaEvolve and is “planning an early access program for selected academic users” while exploring possibilities to make the tool “more broadly available.” Interested parties can register online.