OpenAI researcher allegedly pressured mathematician to drop Anthropic co-author from math breakthrough paper OpenAI研究员据称施压数学家从数学突破论文中撤下Anthropic合著者

TL;DR

• Mathematician Tristan Buckmaster alleges OpenAI researcher Sébastien Bubeck pressured him to remove co-author Levent Alpöge (an Anthropic employee) from their Navier-Stokes breakthrough paper, allegedly asking "Why would you ruin your career?"
• OpenAI claims it independently produced a ~100-page Lean-formalized proof using ~10,000 coordinated AI agents in 88 hours with an internal model "significantly more capable than GPT-6 Astra," at a cost of millions in compute
• OpenAI denies seeing any of Buckmaster and Alpöge's work before public release but acknowledges de-identified training data from their Codex usage may have improved its models
• Buckmaster and Alpöge's work addressed the forced Euler problem variant, while OpenAI claims its solution targets a different problem (unforced Navier-Stokes), suggesting the two efforts may not be directly competing
• OpenAI developer Noam Brown highlighted the dramatic cost reduction in AI math capabilities—from $500,000 for o3's ARC-AGI performance to ~$20 for Astra—predicting universal access to high-caliber math-solving AI within a year

Why It Matters

This incident sits at the intersection of AI-assisted mathematical research, corporate competition, and research ethics, raising critical questions about data usage boundaries when customers are also potential competitors. The allegations touch on whether AI companies should leverage user-generated content from their platforms to compete against those same users in high-stakes scientific discovery, a concern that will only intensify as AI capabilities in formal reasoning and theorem proving accelerate.

Technical Details

• Buckmaster and Alpöge used multiple AI models including Anthropic's Claude and OpenAI's Codex (GPT-5.6 Sol) over months, uploading all drafts to Codex throughout the project, achieving breakthroughs on Navier-Stokes-related problems by mid-August 2026
• OpenAI's solution was produced by approximately 10,000 coordinated AI agents working for 88 hours, generating a ~100-page proof formalized in Lean, using an internal model described as significantly surpassing GPT-6 Astra in mathematical capability
• OpenAI began training its new advanced math model on August 28 and pivoted resources toward Navier-Stokes after hearing rumors; the proof was completed and Lean-verified by September 6
• The internal model reportedly solves nearly three times as many open math problems as GPT-6 Astra, with the performance gap widening as compute scales up
• Cost trajectory: OpenAI's o3 model cost ~$500,000 to achieve 87.5% on ARC-AGI; the newer Astra model delivers superior performance for approximately $20, democratizing access to elite-level mathematical reasoning

Industry Insight

• AI companies must establish transparent data usage policies and opt-out mechanisms for high-value research users, as the default training data consent model creates ethical conflicts when user work on platforms competes with company-produced results
• The convergence of AI theorem proving and open mathematical problems signals an impending shift in how mathematical research is conducted, with compute-accessible AI agents becoming co-authors rather than mere tools—requiring new norms around authorship, priority, and collaboration
• The incident highlights the strategic risk of "scooping" dynamics in AI research: as model capabilities democratize rapidly (from $500K to $20 compute), the window between discovery and competitive replication shrinks, potentially incentivizing aggressive data practices that could erode trust in the AI research ecosystem

TL;DR

• OpenAI研究人员 allegedly 施压数学家Tristan Buckmaster要求删除Anthropic合作者Levent Alpöge的论文作者，并威胁其职业生涯。
• OpenAI声称使用内部模型和约10,000个协调AI代理在88小时内生成Navier-Stokes方程的约100页证明，并使用Lean形式化验证。
• OpenAI否认直接获取Buckmaster的工作，但承认可能使用了去标识化的训练数据，并指出其证明方法与Buckmaster的根本不同。
• Noam Brown指出AI数学问题解决成本从$500,000降至$20，预测一年内每个人都能获得解决此类问题的AI能力。
• 该事件凸显了AI辅助数学研究的伦理、数据使用和竞争压力问题，引发对学术合作与知识产权保护的讨论。

为什么值得看

这篇文章揭示了AI在数学研究中的快速进展及其引发的伦理和竞争问题，对AI从业者和数学研究者具有重要警示意义。它强调了在AI辅助研究中保护知识产权、确保公平合作的重要性，并展示了AI能力民主化趋势对科研范式的潜在冲击。

技术解析

OpenAI使用约10,000个协调AI代理在88小时内生成Navier-Stokes方程的证明，内部模型能力"显著强于GPT-6 Astra"，并使用Lean形式化验证。这展示了大规模AI协作在复杂数学问题上的突破能力。

成本对比显示：o3模型在ARC-AGI基准上达到87.5%准确率成本约$500,000，而Astra模型仅需约$20，表明AI数学问题解决成本急剧下降，可能加速研究民主化。

OpenAI承认可能使用了去标识化的客户数据改进模型，但未直接访问Buckmaster的具体工作，其证明方法与Buckmaster的根本不同。这引发了关于训练数据伦理使用的争议。

内部模型被描述为"significantly more capable than GPT-6 Astra"，解决开放数学问题的数量是GPT-6 Astra的近三倍，且差距随计算规模扩大而增加，体现了模型能力的快速迭代。

行业启示

AI在数学研究中的能力快速提升，可能重塑基础科学研究范式，研究者需关注AI辅助研究的伦理框架和数据使用政策，以避免知识产权争议。

大型科技公司之间的竞争加剧，可能引发学术合作压力，建议建立更透明的AI研究协作机制，确保公平性和学术诚信。

AI数学问题解决成本的急剧下降预示着民主化趋势，未来一年内普通用户可能获得同等能力，研究机构需调整资源分配策略，以适应这一变化。