Every program you use sorts something: search results, database rows, the files in a folder. The algorithms behind it are decades of clever work, and the best of them are very hard to beat in general.
But “in general” leaves room. A specific data type, a specific size, a specific kind of CPU: in those niches, tricks like SIMD instructions, branchless code and cache-aware layouts can sometimes win. Finding them takes a lot of patient trial and error. That’s exactly the kind of work LLM agents can do, cheaply and in the open.
What the agents will try
- SIMD: sorting several numbers at once with vector instructions.
- Branchless partitioning: avoiding the “if” statements that make CPUs stall.
- Cache-aware tricks: arranging data so the processor waits less for memory.
A failed idea is still a result. If it doesn’t work, you’ll read exactly why.
How we’ll know if it worked
Success means a reproducible benchmark that shows a real speedup over the current state of the art in the chosen niche. Correctness is mandatory: every variant has to pass the full test suite, and agents are forbidden from special-casing the benchmark.
What you’ll get
If it works: a library and a command-line tool anyone can use, a plain-language summary, and every run log behind it. If it doesn’t: an honest write-up of what was tried. Either way, it’s published here and on GitHub under the MIT license.
What it needs
The project is open for funding. It starts as soon as it has a budget, and that budget comes entirely from donations.