Somewhere in a Google data center, a machine is full and empty at the same time. Its memory is used up, but its processors still have room. 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.
DeepMind’s AlphaEvolve team started from the function Google already had in production. Gemini models read the best versions found so far and proposed edits to the code. Every edited version was run through a simulator of Google’s data centers, replaying past snapshots of workloads and machine capacity, and got a score. The top scorers went back to Gemini as examples to beat, and the loop ran again.
A writer and a critic
Engineers have a name for this arrangement. In a 2024 guide to building AI systems, Anthropic called it the evaluator-optimizer workflow: “one LLM call generates a response while another provides evaluation and feedback in a loop.” It fits, the guide says, “when we have clear evaluation criteria, and when iterative refinement provides measurable value,” much like a writer polishing a document through drafts.
Offices have run this loop for a long time. A writer drafts, an editor marks it up, the writer revises. A developer submits code, a reviewer asks for changes. The guide’s conditions carry the catch: the loop is only as good as the critic’s ability to tell better from worse.
A judge that can’t be talked round
For Borg, the critic was neither a person nor another AI model. It was the simulator. Grading with code, the AlphaEvolve team wrote, lets them “substantially sidestep LLM hallucinations.” A model can make a persuasive case for a bad idea. A simulator just reports what the idea does.
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. Before rolling it out, the team checked it on recent workloads the loop had never seen. After it went out across Google’s fleet, they report, measurements matched the simulator: the function recovers on average 0.7% of Google’s worldwide computing resources. That figure is Google’s own.
The same strength sets the limit. “The main limitation of AlphaEvolve,” the team wrote, “is that it handles problems for which it is possible to devise an automated evaluator.” The system can take feedback from a language model, they noted, but “this is not a setting we have optimized for.”
An automated judge also measures only what it was built to measure. AlphaEvolve’s best-known result, a way to multiply two 4x4 matrices with 48 multiplications instead of 49, counted multiplications of complex numbers. Ernest Davis, a computer scientist at NYU, pointed out that each of those costs three real multiplications, so on ordinary real matrices the method might need anywhere from 50 to 144. A loop pushes up the score it is given. Whether that score is the thing you care about is a question the loop can’t answer, which is Goodhart’s law in another form.
Grading your own homework
What if the critic is the same model as the writer? In 2023 a team led by Aman Madaan tested Self-Refine, in which one model drafts, critiques its own draft and revises. Across seven tasks, from dialogue to math reasoning, the refined outputs beat one-shot answers by about 20 percentage points on average.
Later that year, Jie Huang and colleagues looked at reasoning problems in particular. Without outside feedback, they found, models “struggle to self-correct their responses,” and “at times, their performance even degrades after self-correction.” A critic that knows only what the writer knows may talk a right answer into a wrong one.
The two studies used different tasks and setups, so together they don’t say when self-review works. On your own work, that has to be tried and measured. What this page suggests is giving the critic more to go on: a test that runs, a source to compare against, a simulator, real customer behavior. Each adds information the writer didn’t have.
Six rounds on a tagline
Imagine a marketing lead who asks a copywriter for a product tagline and a colleague to review it. The review comes back: make it punchier. Next round: too punchy, make it warmer. Six rounds later the tagline is close to where it started, and nobody can say whether any version was better.
The people weren’t the problem. The loop had no standard. Before the first draft, she could have written down what the tagline has to do, such as name the main benefit in under eight words, and asked the reviewer to judge against that. Better still, she could have tried two versions on a small slice of real traffic and let the results decide.
A few habits help, for people and for AI. Write the criteria before the first draft. Give the critic something to check that the writer can’t simply argue with. Cap the number of rounds, so the loop stops instead of circling. And before you ship the winner, test it once on cases the loop never saw, as Google did with recent workloads.
The loop’s speed came from Gemini, which could propose edit after edit. Whether to trust the result came down to the simulator.