Colin Snyder shipped Star Fleet Math — a Mac desktop app that runs up to 20 parallel math agents ("starships"), each on its own 60-vCPU server with GPT-5.6 via Codex, Lean 4 verification, Claude Fable review, and Ton 618 — a local premise graph where verified lemmas compound.
By July 15, 2026, the dashboard showed 27 Erdős problems as Solution Proposed, 623 untested, and 0 Frontier Math or Millennium claims closed. A Hacker News thread (92 points, title: Solving 20 Erdős Problems with 20 Codex Accounts Running in Parallel) split praise for the harness engineering from skepticism about readability, cost, and what "solved" means.
explainx.ai maps architecture, verification chain, and how this fits next to Leanstral, Erdős #728, and the formal-math agent wave.
TL;DR
| Item | Detail |
|---|---|
| Builder | Colin Snyder · advised by Mike Kim · starfleetmath.com |
| Stack | TypeScript + Bun Mac orchestrator · 20× GPT-5.6 Codex · Lean 4 |
| Status (Jul 15) | 27 Erdős Solution Proposed · 623 untested · 650 Erdős tracked |
| Verify | Lean 4 → Fable harness → optional iMessage human review |
| Memory | Ton 618 — verified premises in dependency graph |
| Funding | Self-funded weekend project — not employer-linked (Colin on HN) |
| HN | Harness praised · writeups vs Lean artifact quality debated |
Architecture — one starship per problem
Each starship is a full agentic harness, not a single chat session:
| Layer | Capability |
|---|---|
| Reasoning | GPT-5.6 via Codex — one account per parallel problem |
| Compute | 60 vCPU, 120 GiB sandbox · bursts to 2,000 vCPUs (x86 search) · H100 GPU bursts |
| Formal | Lean 4 toolchain · CaDiCaL, kissat, Z3, CP-SAT · SageMath, PARI/GP, GAP, Macaulay2 |
| Retrieval | Largest Lean 4 premise corpus Colin cites · gemini-embeddings-2 + Chroma English search |
| Literature | Firecrawl index of arXiv + GitHub |
| Review | Claude Fable proof-verifier harness |
| Human gate | iMessage API to Colin after Fable approval |
| Compounding | Ton 618 — every verified theorem/lemma woven into a dependency graph |
explainx.ai read: This is meta-harness design — the Mac app is a fleet commander; heavy lifting lives in remote sandboxes and API models. That matches how GPT-5.6 Sol coding agents use Codex for execution while local UI coordinates sessions — here scaled to 20 math problems at once.
The 27 "Solution Proposed" — what that label means
starfleetmath.com groups problems into Solution Proposed, Untested, and solved counts for Frontier Math / Millennium tiers.
Solution Proposed (27) includes Erdős #123, #129, #130, #254, #267, #320, #321, #336, #394, #415, #450, #489, #521, #522, #538, #584, #638, #662, #709, #769, #793, #796, #959, #1183, #1186, #1188, #1189 — each with a Lean-linked writeup on the site.
Colin's team avoided problems with informal partial answers already online — a deliberate filter against literature regurgitation.
| Claim level | What it implies |
|---|---|
| Solution Proposed | Lean + Fable pipeline accepted a candidate |
| Erdős wiki 🟢 | Community / Terence Tao tracker treats as resolved |
| Peer review | Independent mathematicians validate novelty and correctness |
Aristotle + GPT-5.2 Pro reached wiki full solution on #728 (Jan 2026) with a published arXiv writeup. Star Fleet's batch is volume + parallelism — acceptance requires the same external refereeing, not only formal syntax.
Hacker News — harness vs human proof
vessenes (HN, ~3h): praised Lean corpus, vCPU sharding, embedding search — but said Fable proof summaries read like opaque technical walls vs the Cycle Double Cover leak's more human-readable narrative. Lean 4 bytecode remains the compelling artifact; distillation into teachable technique is still hard.
Other threads:
| Topic | HN take |
|---|---|
| Cost | 20 × 60-vCPU dedicated servers — Colin: self-funded; not Xinobi AI employer project |
| GPT-5.6 local? | API/Codex, not self-hosted weights |
| Open source? | Asked; no public repo yet |
| Millennium overnight? | Joked — Colin's scope is Erdős catalog, not Clay prizes |
| Fun vs jobs | Split: conjecture + referee stay human vs proof search automates |
Colin replied he will work on more human-readable writeups — acknowledging the artifact / exposition gap formal AI keeps hitting.
Comparison — Star Fleet vs other Erdős AI stacks
| System | Shape | Notable outcome |
|---|---|---|
| Star Fleet | 20× parallel Codex + Lean + Ton 618 | 27 proposals Jul 2026 |
| GPT-Erdos | Catalog + Lean attempts + reviewers | Taxonomy of new vs literature hits |
| Aristotle / Harmonic | Autoformalization pipeline | #728 wiki full solution |
| Leanstral 1.5 | Open MoE prover model | PutnamBench, miniF2F saturation |
| OpenAI geometry | General reasoning model | Planar unit distance milestone |
Star Fleet optimizes orchestration and memory compounding — not a single frontier model drop. Whether Ton 618 beats raw model scale depends on how many proposals survive human + wiki scrutiny.
Practical takeaways for builders
- Parallelize problems, not tokens within one proof — Colin's 20 Codex accounts map cleanly to 20 Erdős IDs; intra-proof search uses vCPU/GPU bursts inside each starship.
- Verifier stack matters — Lean 4 ground truth + Fable harness + human iMessage is a three-layer trust ladder; skip any layer and "Solution Proposed" deflates.
- Memory graph > chat history — Ton 618 naming aside, dependency-linked lemmas are how formal systems compound — same design pressure as agent memory graphs in coding agents.
- Budget reality — Self-funded 20-server parallelism is not a hobby GPU; compare to running SOTA locally economics before copying the fleet model.
- Read the Lean, not the summary — HN's critique applies to any Fable-written exposition; treat summaries as pointers, proofs as evidence.
Summary
Star Fleet Math (July 2026) is Colin Snyder's Mac fleet orchestrator: 20 GPT-5.6 Codex starships, Lean 4 + SAT/SMT/CAS sandboxes, embedding search over a large premise corpus, Fable verification, and Ton 618 compounding memory — yielding 27 Erdős Solution Proposed entries on starfleetmath.com. Hacker News celebrated the harness while questioning proof writeups, cost, and resolution status. Solution Proposed ≠ community accepted; compare to #728 and Leanstral for how formal artifacts graduate to recognized math.
Related on explainx.ai
- Leanstral 1.5 — Proof Abundance for All
- OpenAI planar unit distance Erdős problem
- Will AI replace mathematicians? IEEE Big Mathematics
- GPT-5.6 Sol + Claude Code Claudex setup
- OpenAI GeneBench Pro — GPT-5.6 Sol computational biology
- Obsidian vault agent memory graph
- How to build enterprise AI benchmarks — Nadella
Sources: starfleetmath.com · erdosproblems.com · Hacker News thread, Jul 15, 2026 · Erdős #728 arXiv writeup · Colin Snyder HN replies
Status counts from starfleetmath.com July 15, 2026. Formal acceptance requires independent mathematical review beyond this article.