Power up. From supercharging employee productivity and streamlining processes to inspiring innovation, Microsoft’s AI is designed to help you build the next big thing. No matter where you're starting, push what's possible and build your way with Azure's industry-leading AI. Check it out.
Amid growing discussion around how AI might be taught to “think” in novel ways, Google claims its new coding agent can apply creative problem-solving to tasks like chip design and data center improvement.
Earlier this month, Google DeepMind announced a system called AlphaEvolve that uses Gemini to create algorithms, then iteratively improves them with an automated evaluation system. The company said the agent has already improved data center efficiency, chip design, and even the training of its own AI models.
The announcement came ahead of a flurry of other news at its I/O developer conference, including advances in video, image, and coding models; a new AI Mode for search; and the replacement of Google Assistant with the more advanced Gemini.
But AlphaEvolve is more focused on back-end processes than consumer uses. The system “propose[s] computer programs [through LLMs] that implement algorithmic solutions” to given tasks. The automated evaluators then test the answers the program produces, and an evolutionary framework then makes improvements to optimize for a given outcome.
“This makes AlphaEvolve particularly helpful in a broad range of domains where progress can be clearly and systematically measured, like in math and computer science,” the research team wrote in the announcement.
Algorithms at work: Google said it’s already added certain algorithms discovered through AlphaEvolve to various aspects of its operations. For instance, one algorithm is helping its cluster management system, Borg, schedule processes to save an average of 0.7% worldwide on compute resources.
AlphaEvolve proposed tweaks to Google’s chip design process that were incorporated into an upcoming Google AI accelerator. Another AlphaEvolve-generated algorithm is saving 1% of the time it takes to train Gemini.
“Because developing generative AI models requires substantial computing resources, every efficiency gained translates to considerable savings,” the team said. “Beyond performance gains, AlphaEvolve significantly reduces the engineering time required for kernel optimization, from weeks of expert effort to days of automated experiments, allowing researchers to innovate faster.”
Google has also designed a user interface for AlphaEvolve that will be available to select academic researchers through an early access program.
AI discovery: The rollout comes as more tech companies have been looking for ways to move generative AI beyond regurgitation and amalgamation to producing genuinely original ideas. OpenAI has been pitching its latest reasoning models to scientists this way, and Microsoft just announced an agentic scientific platform called Microsoft Discovery.
While Google has aimed AlphaEvolve at math and computing problems thus far, the team believes the tool’s scope might broaden in the future.
“Its general nature means it can be applied to any problem whose solution can be described as an algorithm, and automatically verified,” the authors wrote. “We believe AlphaEvolve could be transformative across many more areas such as material science, drug discovery, sustainability and wider technological and business applications.”