July 14, 2026
Proof Wars: Revenge of the Nerds
Solving 20 Erdős Problems with 20 Codex Accounts Running in Parallel
AI math armada claims 20 puzzle wins as commenters ask who’s paying and who’s losing the fun
TLDR: A developer says his 20-account AI system solved 20 hard math problems in parallel using massive computing power. Commenters were less busy celebrating than asking who paid for it, whether the code will be shared, and if AI is turning math from art into automation.
A lone builder says he launched a mini fleet of 20 AI “starships” to attack famously hard unsolved math questions at the same time — and the comments instantly turned into a mix of awe, side-eye, and existential panic. The setup sounds almost comically overpowered: desktop control center, giant server farms, bursts of insane computing power, and bots checking each other’s work before pinging the human. To some readers, it felt like watching a sci-fi movie where the calculator got a military budget.
But the real action was in the reactions. One camp was plainly fascinated, peppering the post with practical questions like what exactly is the “harness,” where did the proof library come from, and is any of this open source? Another camp went straight for the money question: who is funding this? That became the thread’s biggest eyebrow-raiser, because even non-experts could tell this was not a cheap hobby.
Then came the culture-war twist. One commenter dropped the most dramatic line in the thread by asking if this was “sucking the fun out of math” and joking that maybe mathematicians should get to keep their jobs. That gave the whole discussion a deliciously awkward vibe: is this a thrilling breakthrough, an extravagant tech flex, or the beginning of machines muscling into one of humanity’s most elite intellectual playgrounds? Either way, the community sounded equal parts impressed, suspicious, and wildly curious.
Key Points
- •Star Fleet Math is presented as a Mac desktop AI system that uses up to 20 parallel agentic harnesses, each running GPT-5.6 on a dedicated 60-vCPU server to work on separate open math problems in Lean 4.
- •Each starship is described as having access to burst CPU and H100 GPU compute, a searchable Lean 4 theorem corpus, indexed arXiv and GitHub data, proof-review tooling, and a long-term memory system called Ton 618.
- •The software stack is described as being built from scratch in TypeScript and Bun, with sandboxes preloaded with SAT/SMT solvers, CP-SAT, computer algebra systems, and Rust, CUDA C++, and Lean 4 toolchains.
- •The page lists 27 proposed solutions and says the team tried to avoid problems that already had informal or partial answers available online.
- •For Erdős Problem #123, the article states a proposed result for pairwise-coprime integers a, b, c > 1, provides a Lean theorem statement, and says the main challenge was satisfying the non-divisibility condition and overcoming a finite-seed gap in prior induction-based methods.