计算物理学专家Michael P. Frank指挥GPT-5.6 Sol模型连续运行约33小时，试图探索费马大定理是否存在比怀尔兹-泰勒证明更简洁的途径[1]。该研究聚焦于Frey曲线模性、一致无限下降法、算术abc型不等式以及一致低亏格商等关键数学领域[1]。然而这一尝试最终被OpenAI系统强制中止[1]。
模型分析认为被阻止的可能原因包括系统判定占用资源过多，或OpenAI之前已尝试过该问题而不愿再投入算力[1]。根据GPT-5.6 Sol的结论，该任务产出的是一张"此路不通"的地图，而非费马大定理的新证明[1]。这一事件在社区引发了关于AI能力开放、监管机制和竞争优势等多角度的讨论[1]。
OpenAI高级研究科学家Noam Brown对此表示："要是有人只是输入一个continue，就用我们的模型解决了千禧年大奖难题，然后拿走100万美元奖金，这对OpenAI来说才是最好的广告"[1]。
Computational physicist Michael P. Frank conducted an extended experiment using OpenAI's GPT-5.6 Sol model, running the system continuously for approximately 33 hours to investigate whether a more elegant proof of Fermat's Last Theorem could exist compared to the Wiles-Taylor proof [1]. The research specifically focused on examining Frey curve modularity, consistent infinite descent methods, arithmetic abc-type inequalities, and consistent low-genus quotients [1]. However, OpenAI's system ultimately forced the task to terminate [1].
Rather than producing a novel proof, the model's output amounted to what could be described as a map of dead ends rather than a mathematical solution itself [1]. The forced interruption appears to stem from either OpenAI's resource consumption protocols or the company's prior assessment that the computational approach warranted no further algorithm investment [1]. The incident has sparked community debate regarding AI capability distribution, regulatory frameworks, and competitive advantage in the development of advanced language models [1].
Noam Brown, an OpenAI senior research scientist, offered a perspective on the potential value of such achievements, stating: "If someone just typed a continue command and solved a Millennium Prize Problem using our model and claimed the one-million-dollar prize, that would be the best advertisement for OpenAI" [1].
