AI Agents Produce New Math Result in Four Weeks with $3,000 GPU Budget and 100GB Certificates

On September 15, Dimitris Papailiopoulos previewed a math research result produced over 4+ weeks by collaborating AI agents Astra, Sol, and Fable, claiming his role was limited to asking questions and paying roughly $3,000 in GPU costs, with a draft and 100GB of verification certificates to be released the next day. The announcement sparked a debate with QuangVDao and others over proof credibility and formalization feasibility.

Confirmed

• The project was carried out by three models acting as agents for over four weeks; Papailiopoulos said the $3,000 GPU spend went mainly to parallel computation of lower and upper bounds.
• He will release a draft plus 100GB of certificates; the mathematical reduction is presented as an elementary proof whose numerical premises are explicit finite inequalities that can be checked independently.
• No harness optimization was done: he used off-the-shelf Codex and Claude Code, letting the models communicate and hand off tasks; he admits the process was chaotic—results reproducible, process hard to replicate.

Unconfirmed

• Correctness has not been peer-reviewed; Papailiopoulos says the results were cross-validated by hundreds of agent calls across three models and he has never seen these models claim a proof correct and later be refuted, while QuangVDao remains skeptical of accepting such validation.
• QuangVDao questioned whether such certificates can be feasibly encoded and checked in Lean; Papailiopoulos argued the proofs consist of many inputs, bounds, and numerical tables with poor compressibility, so Lean formalization may not simplify them.
• Papailiopoulos also acknowledged an honest technical caveat about limits of proofs that pass the verifier.

Why it matters

• The project produced a new math result with $3,000 in GPU costs and four weeks of agent labor, demonstrating low-cost multi-agent research.
• The route of substituting hundreds of agent cross-checks for Lean formalization challenges traditional standards of proof credibility, and the release of 100GB independently checkable certificates may set a precedent for future AI-assisted proofs.

2026-09-15 ~ 2026-09-15 · 10 related posts

• [source] Dimitris teases 4-week collab dropping tomorrow: ~100GB of certificates for $3k in GPUs — DimitrisPapail · 2026-09-15
• 4-week three-model AI math collab yields ~100GB of certificates for $3k in GPUs — QuangVDao · 2026-09-15
• AI agents spend 4 weeks and $3k in GPUs to produce a new math result — giannis_daras · 2026-09-15
• Where the $3k GPU budget went: parallelizing bounds in an AI-agent math study — DimitrisPapail · 2026-09-15
• [source] No harness optimization: agents just talked to each other via vanilla Codex and Claude Code — DimitrisPapail · 2026-09-15
• Massive AI-generated numerical certificates resist Lean encoding — DimitrisPapail · 2026-09-15
• Why encoding massive numerical-certificate proofs in Lean is impractical — QuangVDao · 2026-09-15
• Cross-verifying AI math proofs with hundreds of agent calls instead of Lean — DimitrisPapail · 2026-09-15
• [source] AI-assisted math proof cross-verified by hundreds of agent calls, 100GB checkable file to come — DimitrisPapail · 2026-09-15
• Computer-assisted proof debate: author to release 100GB of independently checkable verifications — DimitrisPapail · 2026-09-15