Somewhere in a Google data center, a machine is full and empty at the same time. Its memory is used up. Its processors still have room, but no new job can land there, because every job needs some of both. Google’s researchers call that idle capacity stranded.
Which machines end up like this depends partly on a few lines of code. When a job arrives, Google’s cluster manager, Borg, lists the machines able to run it and asks a scoring function to rank them. The job goes to the top scorer, among other objectives.
DeepMind’s AlphaEvolve team started from the function Google already had in production. By Google’s own account, in a May 2025 blog post and white paper, the system runs as a loop. Gemini models read the best programs found so far and propose edits to the code, the fast Flash model for volume and the stronger Pro model now and then. Every edited program is run and scored automatically. The top scorers go back into the next prompt as examples to beat.
For Borg, the scorer was neither a person nor another AI model. It was a simulator of Google’s data centers, replaying historical snapshots of workloads and machine capacity. Grading with code, the team wrote, lets them “substantially sidestep LLM hallucinations.” (The system can ask a model for feedback, but they called that a setting they had not optimized for.)
The winner fit in a few lines: it divides a job’s CPU and memory needs by what each machine has free and folds the two ratios into one score. After a check on recent workloads the loop had never seen, Google rolled it out across the whole fleet, choosing it over a deep reinforcement learning approach partly because engineers can read and debug it.
Google says measurements after deployment matched the simulator. The function, in production for over a year by May 2025, recovers on average 0.7% of Google’s worldwide compute. That is Google’s own figure. Ernest Davis, a computer scientist at NYU, wrote that not enough information had been given to evaluate it in depth, though the gains “do seem impressive.”
The same loop then went after a record that, in one setting, had stood for 56 years. Two rounds of Strassen’s 1969 method multiply two 4x4 matrices with 49 multiplications. After scoring 16,000 candidates from Gemini, AlphaEvolve found a way with 48. Those 48 are multiplications of complex numbers.
One complex multiplication costs three real ones, Davis noted, so on real matrices the method might need anywhere from 50 to 144 real multiplications. He doubted the result had much practical value, and wrote that the launch and its press coverage had been “accompanied by hype, misleading statements, and outright untruths.” Sanjeev Arora of Princeton told WIRED the advances were relatively small, while adding that search applies in many settings.
On June 16, 2025, Jean-Guillaume Dumas, Clément Pernet and Alexandre Sedoglavic posted a 48-multiplication method with only rational coefficients, removing, in their words, “the requirement for complex-number arithmetic.” In November, a paper with Terence Tao and Javier Gómez-Serrano among its authors ran AlphaEvolve on 67 math problems; it matched the best known answers in most and improved several.
The white paper names the catch itself: “The main limitation of AlphaEvolve is that it handles problems for which it is possible to devise an automated evaluator.”