OpenAI 纳维-斯托克斯方程证明方法被指与官方叙事不符 学界验证尚需数年 智能体贡献同样成谜OpenAI's Navier-Stokes Equation Proof Method Said to Diverge From Official Narrative


AI free token/AI免费积分
OpenAI 纳维-斯托克斯方程证明方法被指与官方叙事不符 学界验证尚需数年 智能体贡献同样成谜
感谢IT之家网友 愚公骑马、咩咩洋、华南吴彦祖、Autumn_Dream、肖战割割、LeoFize、小星_14 的线索投递!

IT之家 10 月 17 日消息,OpenAI 今年 9 月公布了一份长达 166 页的纳维-斯托克斯方程(Navier–Stokes equations)相关证明,称其内部 AI 平台在约 88 小时内完成了对这一千禧年难题的求解。

不过,随着数学界进一步审查相关工作,研究人员开始对证明使用的方法、与已有研究的关系,以及大规模 AI 智能体究竟发挥了多大作用提出质疑。

数学家贡萨洛 · 曹-拉博拉(Gonzalo Cao-Labora)指出,OpenAI 可能自己都没看懂就急着发论文了。

他表示,OpenAI 的证明细节与西班牙数学家迭戈 · 科尔多瓦(Diego Córdoba)和路易斯 · 马丁内斯-佐罗亚(Luis Martínez-Zoroa)此前提出的方法展现出不同的路线,二者的相似之处可能更多停留在表面上。

参与审读的多组专家均向《科学美国人》表示,越来越多迹象表明,OpenAI 的证明并没有遵循“西班牙路线”。这意味着 OpenAI 在发布该项成果时似乎并未完全理解其模型究竟干了什么。

IT之家注:纳维-斯托克斯方程用于描述流体运动,是克雷数学研究所设立的千禧年大奖难题之一。OpenAI 于 9 月 8 日发布相关研究,称其内部模型完成了数学证明,并同时公开了使用 Lean 形式化验证语言编写的证明。OpenAI 表示,其研究团队使用约 1 万个 AI 智能体,经过约 88 小时得到结果,随后又花费约 17 小时完成形式化及验证工作。

Lean 可以检查形式化证明中的逻辑推导是否符合设定规则,但借助形式化验证并不意味着证明所应对的难题一定符合研究人员原本关注的科学难题,也不意味着证明的思路容易被人类理解。OpenAI 公布的证明长达 166 页,数学界当下仍在审查其具体方法及结论。

此次争议之一在于,OpenAI 的研究与科尔多瓦、马丁内斯-佐罗亚此前探索的研究路径究竟有多大关联。

《科学美国人》指出,OpenAI 的证明利用了纳维-斯托克斯方程中的外部力项,构造出流体速度在有限时间内无界增长的情形。

数学家对这一结果的意义存在不同看法:它符合克雷数学研究所原始难题表述允许的一种情形,但并未应对不施加外部力时,三维纳维-斯托克斯方程的解是否始终保持光滑这一更受关注的难题。

在研究成果的归属维度,纽约大学数学家特里斯坦 · 巴克马斯特(Tristan Buckmaster)质疑 OpenAI 的研究可能受益于他与 Anthropic 员工莱文特 · 阿尔珀格(Levent Alpöge)使用公开模型开展的相关研究。

对此,OpenAI 否认使用了巴克马斯特的相关数据,并在 9 月 10 日更新的官方说明中表示,经过调查,巴克马斯特此前两个月在 Codex 中输入的提示词不可能借助任何方式影响此次研究,涵盖借助模型训练产生影响。但截至当下,双方对研究方法及成果的关系仍存在争议。

另一项尚未得到明确答案的难题,是这些智能体协作究竟给出了多少贡献。OpenAI 官方称其研究涉及约 1 万个并行运行的智能体,累计发送约 270 万条消息,使用约 1300 亿个输出 Token。不过,这些数字只能反映计算规模和运行过程,不能直接说明多智能体架构对最终成果的实际贡献。

OpenAI 研究员诺姆 · 布朗(Noam Brown)也指出,当下研究人员并不清楚如此大规模的智能体集群究竟带来了多少额外收益。由于使用数千个智能体开展平台性对照实验的成本很高,研究人员尚未充分比较单个智能体完成同一任务所需的时间,也难以准确衡量 1 万个智能体相较于 1000 个智能体的长处。布朗估计,多智能体协作架构对这次成功的贡献可能不到 10%。

在他看来,由于 AI 智能体工艺新颖且易受广泛关注,容易取得超出实际贡献的认可;真正关键的因素可能是底层模型本身的能力,以及模型可以长时间持续处理任务、并行探索多种解法的特性。

10 月初,OpenAI 又公开了约 400 项数学研究结果。数学界普遍表示仅消化这些内容就需要数年。

这些成果以相对松散的形式集中发布在 GitHub 上,同样又引发部分数学研究人员不满。他们担心,大量成果在缺乏充分解释、整理和学术交流的情况下集中发布,会降低数学研究成果的可读性,也可能模糊严谨的研究验证与大规模自动生成结果之间的区别。

当下,数学界仍在评估 OpenAI 的相关研究。此次争议涉及的不只是 AI 能否生成形式化证明,还涵盖研究方法是否真正应对了目标难题、成果如何归属,以及多智能体平台的实际贡献应当如何衡量。

相关阅读:

本文来源:https://www.ithome.com/1/014/370.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

OpenAI released a 166-page proof of the Navier–Stokes equations in September, saying its in-house AI platform solved the millennial puzzle in about 88 hours.

However, as the mathematical community further scrutinized the work, researchers began to question the methods used to prove it, their relationship to existing research, and the role of large-scale AI agents.

The mathematician Gonzalo Cao-Labora points out that OpenAI may be eager to publish a paper without understanding it itself.

The details of OpenAI's proof, he said, show a different route than the methods previously proposed by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa, whose similarities may be more stuck on the surface.

Multiple groups of experts who participated in the review told Scientific American that there were growing indications that OpenAI's certification did not follow the “Spanish route.” This means that OpenAI didn't seem to fully understand what its model did when it published the result.

IT House Note: The Navier-Stokes equation, used to describe fluid motion, is one of the Millennium Prize puzzles set up by the Clay Mathematical Institute. OpenAI released the research on September 8, saying that its internal model completed the mathematical proof, and at the same time disclosed the proof written in Lean formal verification language. OpenAI said that its research team used about 10,000 AI agents, after about 88 hours to get the results, and then spent about 17 hours to complete the formalization and verification work.

Lean can check whether the logical derivation in the formal proof is in line with the set rules, but formal verification does not mean that the problem solved must be in line with the scientific problem that the researcher was originally concerned about, nor does it mean that the idea of proof is easy to be understood by humans. The proof published by OpenAI is 166 pages long, and the mathematical community is still reviewing its specific methods and conclusions.

One of the controversies is how relevant OpenAI's research is to the research paths previously explored by Córdoba and Martínez Zorroa.

Scientific American pointed out that OpenAI's proof uses the external force term of the Navier-Stokes equation to construct a situation where the fluid velocity increases unbounded for a limited time.

There is a difference of opinion among mathematicians as to the significance of this result: it corresponds to a situation permitted by the original puzzle formulation of the Clay Institute of Mathematics, but does not address the more interesting puzzle of whether the solution to the three-dimensional Navier-Stokes equation always remains smooth when no external force is applied.

In the attribution dimension of the research findings, New York University mathematician Tristan Buckmaster questions whether OpenAI's research could benefit from his work with Levent Alpöge, an Anthropic employee, using publicly available models.

OpenAI's Navier-Stokes Equation Proof Method Said to Diverge From Official Narrative

In response, OpenAI denied using Buckmaster's relevant data, and said in an official note updated on September 10 that after investigation, the prompts Buckmaster had entered in Codex two months earlier were unlikely to affect the study in any way, including through model training. However, as of now, the relationship between the research methods and results is still controversial.

Another puzzle that has yet to be answered is the extent to which these agents contribute to collaboration. OpenAI official said its research involved about 10,000 agents running in parallel, sending a total of about 2.7 million messages and using about 130 billion output tokens. However, these figures only reflect the computational scale and operational process, and do not directly account for the actual contribution of the multi-agent architecture to the final result.

OpenAI researcher Noam Brown also pointed out that it is not clear how much additional benefits such a large cluster of agents would bring. Due to the high cost of using thousands of agents to conduct platform-based controlled experiments, researchers have not yet fully compared the time required for a single agent to complete the same task, and it is difficult to accurately measure the strengths of 10,000 agents compared to 1,000 agents. Brown estimates that multi-agent collaboration architectures may have contributed less than 10% to this success.

In his view, because the AI agent process is novel and easy to be widely concerned, it is easy to obtain recognition beyond the actual contribution; the real key factor may be the ability of the underlying model itself, and the characteristics of the model that can continue to process tasks for a long time and explore multiple solutions in parallel.

In early October, OpenAI published the results of about 400 more mathematical studies. The mathematical community generally says that it takes years just to digest this content.

These results were centrally published on GitHub in a relatively loose form, which also caused some math researchers to be dissatisfied. They fear that centralizing the publication of a large number of results without adequate explanation, collation, and academic communication will reduce the readability of mathematical research results, and may also blur the distinction between rigorous research validation and large-scale automatic generation of results.

At present, the mathematical community is still evaluating the relevant research of OpenAI. This controversy involves not only whether AI can generate formal proofs, but also whether the research method really responds to the target problem, how the results are attributed, and how the actual contribution of the multi-agent platform should be measured.

Read Next:

This article is from: https://www.ithome.com/1/014/370.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry