Aemon Set a World Record on an NP-Hard Problem with Less Than $10 of Compute

The Macro: Most Engineering Problems Have Solutions Nobody Has Found Yet

There is a class of problems in engineering and science where the solution space is so vast that even expert teams explore only a tiny fraction of it. Algorithm optimization, materials design, circuit layout, scheduling, and logistics all fall into this category. A skilled engineer might try 10 or 20 approaches over weeks of work. The optimal solution might be the 847th variation that nobody had time to test.

The traditional way to handle this is to rely on domain expertise and intuition. An experienced engineer knows which approaches are worth trying based on years of experience. This works well enough, but it leaves enormous optimization potential on the table. The best human solution is almost never the best possible solution.

Evolutionary and genetic algorithms have attacked this problem for decades, but they typically operate on narrow problem formulations and require significant setup by domain experts. The recent wave of AI coding agents like Devin and Cursor can write and test code, but they are designed for software development, not research-grade optimization.

Aemon, backed by Y Combinator, is positioned differently. It is an autonomous research engineer that reads relevant literature, generates hundreds of solution variants, evaluates them against user-defined metrics, and iteratively improves across generations. The claim: optimal solutions beyond what human experts can find.

The Micro: From Literature Review to World Record

Ray Xu and Richard Zhou are the kind of founders who make you do a double take. Ray is a UIUC CS dropout who published at top AI conferences (ICLR and EMNLP) before age 20. Richard is a University of Waterloo CS dropout and international medalist in mathematics and robotics. Both left school because they found something more interesting to build.

The proof point that gets attention is the circle packing result. Aemon set a new world record on an NP-hard math optimization problem, beating a result set by a major AI research lab, using less than $10 of compute. That is a remarkable demonstration of the platform’s ability to search solution spaces efficiently.

The workflow is methodical. Aemon reads the codebase and relevant research, maps the solution space, generates and evaluates hundreds of variants, scores and recombines solutions across generations, and allows human experts to adjust constraints and priorities mid-run. The output is a finalized, justified solution ready for production use.

The evolutionary approach is well-suited to optimization problems where you can define a clear evaluation function but the search space is too large for exhaustive exploration. Engineering optimization, hyperparameter tuning, algorithm design, and combinatorial problems all fit this pattern.

Competitors include FunSearch (from a major research lab) for mathematical optimization and various AutoML platforms for hyperparameter search. But Aemon’s scope is broader, handling arbitrary optimization problems rather than specific domains.

The Verdict

Aemon is making a bold claim: AI that can out-engineer human experts. The circle packing result provides one concrete proof point.

At 30 days: are paying customers using Aemon for real engineering optimization problems, and are the results consistently better than human-designed solutions?

At 60 days: how well does the platform generalize across problem domains? A system that works for mathematical optimization but struggles with engineering design would be too narrow.

At 90 days: are customers building Aemon into their standard development workflow, or using it as a one-off optimization tool?

I think Aemon is building something genuinely interesting. The evolutionary search approach applied to engineering problems with modern AI capabilities is a powerful combination. The world record result is not a gimmick. It demonstrates that machine-speed exploration of solution spaces can find things that human expertise misses. The question is whether this generalizes to enough problem types to build a large business.