IT之家 10 月 7 日消息,OpenAI 在一批包含 722 份手稿、涵盖 372 个结果家族的文件中,公布了某款未发布的前沿模型应对的多项长期存在的数学难题。
IT之家注意到,这一举措延续了其一系列突破性成果,这些成果既给数学界部分人士留下了深刻印象,也引发了不安,同时还带来了关于研究伦理和学术行为的疑问。据负责以负责任的方式传达这些成果、新成立的精英数学家独立咨询小组 AGMAI 介绍,此次公布的内容包含对“数百个”未应对难题的解答。
数周以来外界一直在期待这批成果,不过在此之前,OpenAI 始终没有说明模型究竟应对了哪些难题,也没有给出确切的发布时间。该公司曾在九月份称,自家模型“已经应对了一百余个长期悬而未决的公开难题,覆盖数学的绝大多数分支领域”。本次对外发布的文稿以及相关资料还包含一部分模型推理过程摘要、算力消耗估算值,以及尝试求解过的难题数量统计数据。OpenAI 称,一项普通成果所消耗的算力,大致相当于 ChatGPT Pro 连续思考三小时的算力水平。
数学与人工智能咨询小组(Advisory Group on Mathematics and Artificial Intelligence,AGMAI)于九月下旬发布首批建议。该小组呼吁人工智能实验室尽可能借助成熟的学术渠道及时对外发布数学研究成果,同时应当披露所使用模型的名称、提示词、算力开销等详细信息。小组还恳切规劝人工智能企业,“不要把发布数学成果当作推广自家模型的营销手段”;该机构认为,这类做法会给整个数学社群带来严重损害。
OpenAI 是这样描述本次发布的工作流程的:
针对本次发布,笔者将成果发布在一处 GitHub 代码仓库当中,同时制定了文稿修订与引用相关的处理规范。笔者还会继续探寻其他由学术社群托管的发布思路,使之符合该咨询委员会提出的各项指导意见。对于后续版本的成果发布,笔者承诺会进一步完善文稿质量,优化引用情况、数学内容阐述以及成果展示形式,便于人们理解。
数学家还需要对这批成果开展评估、消化吸收,因而这批研究带来的完整影响或许尚需一段时间才可以显现。
此前 OpenAI 以及 Anthropic 等竞争对手实验室已经产出数量迅速增长的一批数学成果,整个数学界仍处在消化阶段,其中甚至包含一道千禧年大奖难题的相关结果,该难题属于数学领域最知名的未解难题。
今年各家人工智能实验室快速闯入数学这门学科,尤其是它们对外发布消息的方式,已经激起激烈争论,议题涉及科研范式、伦理规范,以及企业应当如何向人类数学家致谢;人工智能平台正是建立在人类数学家的已有工作之上,并有可能依托这些前人工作生成新结果。
本文来源:https://www.ithome.com/1/010/137.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Thanks to IT Home netizens Hu Nan Wu Yanzu, Wu Ying Ge's Left Hand, Yu Gong Riding a Horse, and Fo Li is Me for submitting clues!
IT Home October 7 news: OpenAI has published the answers to multiple long-standing math problems addressed by an undisclosed frontier model in a batch of files containing 722 manuscripts covering 372 result families.

IT Home notes that this move continues a series of groundbreaking achievements that have both impressed and unsettled parts of the mathematical community, while also raising questions about research ethics and academic conduct. According to AGMAI (Advisory Group on Mathematics and Artificial Intelligence), a newly formed elite mathematician independent advisory group responsible for responsibly communicating these findings, the published content includes solutions to "hundreds" of previously unsolved problems.
The outside world has been anticipating these results for weeks, though until now OpenAI had not specified which problems the model addressed or provided an exact release date. The company stated in September that its model "has already solved over a hundred long-standing open problems across most branches of mathematics." The published manuscripts and supporting materials also include summaries of the model's reasoning process, computational cost estimates, and statistics on the number of problems attempted. OpenAI stated that a typical result consumes roughly equivalent compute to ChatGPT Pro thinking continuously for three hours.
The Mathematical and Artificial Intelligence Advisory Group (AGMAI) released its initial recommendations in late September. The group called on AI laboratories to promptly publish mathematical research findings through established academic channels and disclose detailed information including model names, prompts, and computational costs. The group also strongly advised AI companies not to "treat publishing mathematical results as a marketing tool to promote their own models," arguing that such practices cause serious harm to the entire mathematics community.
OpenAI described the workflow for this publication as follows:
"For this release, I published the results in a GitHub repository and established guidelines for manuscript revision and citation. I will continue to explore other academically hosted publication approaches that align with the advisory committee's recommendations. For future result releases, I commit to further improving manuscript quality, optimizing citations, mathematical content exposition, and presentation formats to facilitate understanding."
Mathematicians still need to evaluate and digest these results, so the full impact of this research may take time to become apparent.
Previously, rival labs including OpenAI and Anthropic have produced a rapidly growing body of mathematical results, and the entire mathematics community is still in the process of digesting them—including results related to one of the Millennium Prize Problems, the most famous unsolved problems in mathematics.
This year, AI laboratories have rapidly entered the field of mathematics, and especially the way they announce their findings has sparked intense debate, touching on research paradigms, ethical standards, and how companies should acknowledge human mathematicians—upon whose prior work AI platforms are built and who may generate new results based on their contributions.
Source: https://www.ithome.com/1/010/137.htm For copyright issues, please contact: 13951431913@139.com or submit a comment on the right.
