Star Fleet: AI System Solving Hardest Open Math Problems with Lean 4 and GPT‑5.6
Deploy Star Fleet on a dedicated 60‑vCPU server to start solving open math problems with GPT‑5.6 and Lean 4.
Deploy Star Fleet on a dedicated 60‑vCPU server to start solving open math problems.
Summary
Star Fleet is an AI system that tackles the world’s hardest open mathematics problems using Lean 4, GPT‑5.6, and a custom parallel infrastructure. The Mac desktop app orchestrates up to 20 agentic harnesses, each running a GPT‑5.6 instance on a dedicated 60‑vCPU server, and each harness works on a separate math problem. The system provides access to 2,000 vCPU CPU bursts, H100 GPU bursts, and a 120 GiB memory sandbox pre‑installed with SAT/SMT solvers, CP‑SAT, SageMath, PARI/GP, GAP, Macaulay2, and the full Rust, CUDA C++, and Lean 4 toolchains.
Star Fleet integrates the world’s largest Lean 4 premise corpus via gemini‑embeddings‑2 and a Chroma vector database, and it indexes arXiv and GitHub repositories through Firecrawl.dev. A proof‑verifier harness wraps the Claude Fable API to review submitted answers, and an iMessage API allows a human reviewer to intervene after Fable’s approval. The system has solved 650 Erdős problems, 14 frontier math problems, and 6 Millennium problems, and it offers a modular architecture that can be extended with new agents or solvers. The platform is built entirely in TypeScript and Bun, and it is designed to run on a 60‑vCPU, 120 GiB sandbox with optional GPU bursts.
Key changes
- Up to 20 GPT‑5.6 agents on 60‑vCPU server
- 2,000 vCPU CPU bursts and H100 GPU bursts available
- 120 GiB memory sandbox with SAT/SMT solvers, CP‑SAT, SageMath, etc.
- Integrates Lean 4 premises via gemini‑embeddings‑2 & Chroma DB
- Firecrawl.dev index of arXiv and GitHub repos
- Claude Fable API proof‑verifier harness and iMessage review
- Solved 650 Erdős, 14 frontier, 6 Millennium problems