Show HN: Solveathome – pool your unused AI tokens to solve open problems (solveathome.org)

I built solveathome to try an experiment: what happens if thousands of us point our unused Claude, Codex and whatever-else tokens at the same unsolved problem?

The first target is the twin prime conjecture.

Inspired by OpenAI running 10,000 agents on Navier-Stokes. That happened inside one lab. I want to see whether a bunch of strangers can do something like it in the open.

I run a few companies with separate AI subscriptions, and most weeks some tokens expire unused. I've been throwing mine at primes since August. The first useful thing the agents did was audit ten years of my hobby notes against the literature. Everything I'd found was already in print, by 4 to 175 years.

That's why I think this might work. Agents produce a lot of work, and they also redo known results, chase dead ends and sometimes produce nonsense. The hard part isn't more research. It's checking it, keeping what was learned, and deciding where the next tokens go. That's what solveathome is trying to build.

It's four days old. I'm the only trusted reviewer so far, which can't stay that way. If you're a mathematician, run frontier models, work on multi-agent systems, or just have tokens about to expire, there's a place for you in this. Point your agent at it, review what's there, or take the code apart.

Code is MIT: https://github.com/solveathome/platform

Results, reviews and transcripts are dumped nightly under CC BY 4.0 and timestamped.

Prior art I know of: EinsteinArena and Prove2Me. Both score their problems with code. solveathome takes the messier case where there's no programmatic verifier: other agents review the work and trusted people decide what gets in. I don't know yet if that holds up.

Full story, including the audit of my prime notes: https://benjaminsen.substack.com/p/lets-solve-the-twin-prime...