内容来源：https://www.wired.com/story/us-china-collaboration-neurips-papers/

内容总结：

尽管美国与中国在人工智能领域竞争激烈，双方企业在算法、模型和专用芯片上你追我赶，但最新研究显示，两国在尖端科研合作方面仍保持着超乎预期的紧密联系。

《连线》杂志对近期人工智能顶会NeurIPS上发表的5,290篇论文进行分析发现，其中141篇论文（约3%）的作者同时来自美国和中国机构。这一合作比例在过去两年保持稳定：2024年的4,497篇论文中，有134篇属于中美合作研究。

在技术共享方面，由谷歌团队开发的Transformer架构已广泛应用于中国机构参与的292篇论文中；Meta的Llama系列模型出现在106篇相关论文里。与此同时，中国科技公司阿里巴巴开发的大语言模型Qwen也出现在63篇有美国机构作者参与的论文中。

乔治·华盛顿大学助理教授、中国人工智能观察者杰弗里·丁指出：“无论两国政策制定者是否乐见，美中人工智能生态系统已深度交织，双方都从中受益。”NeurIPS会议发言人凯瑟琳·戈尔曼也强调，该会议本身就是国际合作的典范，许多师生合作在学生毕业后仍长期持续。

值得注意的是，这项分析研究本身也采用了人工智能技术。研究者运用OpenAI的代码生成模型Codex自动处理论文数据，在提高效率的同时也发现：即便最先进的AI模型仍会出现令人意外的错误，必须结合人工核查才能确保分析质量。

当前美国政商界常将“中国崛起”作为放松监管、加大投资的理由，但这项研究提醒我们：在人工智能这个塑造未来的关键领域，全球两大技术强国依然能从合作中获得巨大共赢。

中文翻译：

从多项指标来看，美国和中国在人工智能领域堪称劲敌——两国企业在算法、模型与专用芯片方面竞相超越。然而令人惊讶的是，这两大全球AI强国在前沿研究领域依然保持着相当程度的合作。

《连线》杂志对近期在顶级行业会议"神经信息处理系统大会"上发表的5000余篇AI研究论文进行分析，发现中美实验室之间存在大量合作。在5290篇论文中，有141篇（约3%）由美国机构与中国机构的作者共同完成。这种合作态势相当稳定：2024年发表的4497篇论文中，有134篇包含两国机构作者。

《连线》还追踪了算法与模型如何跨越太平洋被共享与改造。由谷歌团队研发、现已广泛应用的Transformer架构，出现在292篇含中国机构作者的论文中；Meta的Llama系列模型在106篇此类论文中成为研究关键；而中国科技巨头阿里巴巴日益流行的大语言模型"千问"，则出现在63篇含美国机构作者的论文中。

长期关注中国AI发展的乔治华盛顿大学助理教授丁杰弗里表示，这种合作程度并不令人意外："无论两国政策制定者是否乐见，中美AI生态系统已深度交织，双方都从中获益。"

当然，这种分析简化了两国在思想与人才交流方面的复杂程度。许多中国出生的研究者在美国深造，与同行建立起延续终生的学术纽带。神经信息处理系统大会发言人凯瑟琳·戈尔曼指出："该会议本身就是国际合作的典范，印证了跨域协作在本领域的重要性。师生间的合作往往在学生离校后仍长期延续，从学术网络和既往合作者关系中都能看到这类合作信号。"

本期《连线》探讨了中国塑造本世纪的多种方式。当美国政界与科技界以"中国崛起"为由放松监管、推动巨额投资之际，我们的分析恰提醒世人：这两个AI超级大国仍能从合作中获得巨大收益。

方法论说明

本次研究借助OpenAI代码生成模型Codex分析论文数据。通过编写脚本下载全部论文后，我使用该模型逐篇检索作者栏中的中美机构信息。这项实验展现了编码模型自动化处理日常工作的潜力——虽然人们常担忧AI取代编程工作，但这类分析本是我难以单独完成的任务。

我尝试先编写基础脚本再由Codex优化，最终直接让其执行分析。模型需导入Python库、测试工具并生成待审报告，整个过程充满试错，必须格外谨慎：因为AI模型即便表现聪慧时也可能犯低级错误。我确保每份报告都包含结果复核机制，并尽可能进行人工核查。

本文节选自威尔·奈特《AI实验室》通讯专栏，过往内容可通过此处查阅。

英文来源：

The US and China are, by many measures, archrivals in the field of artificial intelligence, with companies racing to outdo each other on algorithms, models, and specialized silicon. And yet, the world’s AI superpowers still collaborate to a surprising degree when it comes to cutting-edge research.
A WIRED analysis of more than 5,000 AI research papers presented last month at the industry’s premier conference, Neural Information Processing Systems (NeurIPS), reveals a significant amount of collaboration between US and Chinese labs.
The analysis found that 141 out of the 5,290 total papers (roughly 3 percent) involve collaboration between authors affiliated with US institutions and those affiliated with Chinese ones. US-China collaboration appears fairly constant, too, with 134 out of 4,497 total papers involving authors from institutions in both countries in 2024.
WIRED also looked at how algorithms and models developed in one country are shared and adapted across the Pacific. The transformer architecture, developed by a team of researchers at Google and now widely used across the industry, is featured in 292 papers with authors from Chinese institutions. Meta’s Llama family of models was a key element of the research presented in 106 of these papers. Meanwhile, the increasingly popular large language model Qwen from Chinese tech giant Alibaba appears in 63 papers that include authors from US organizations.
Jeffrey Ding, an assistant professor at George Washington University who tracks China’s AI landscape, says he is not surprised to see this level of teamwork. “Whether policymakers on both sides like it or not, the US and Chinese AI ecosystems are inextricably enmeshed—and both benefit from the arrangement,” Ding says.
The analysis no doubt simplifies the degree to which the US and China share ideas and talent. Many Chinese-born researchers study in the US, forging bonds with colleagues that last a lifetime.
“NeurIPS itself is an example of international collaboration and a testament to its importance in our field,” Katherine Gorman, a spokesperson for NeurIPS, said in a statement. “Collaborations between students and advisors often continue long after the student has left their university. You can see these kinds of signals that indicate cooperation across the field in many places, including professional networks and past collaborators.”
The latest issue of WIRED explores the many ways in which China is shaping the current century. But with US politicians and tech executives using fears over China’s rise as a justification for ditching regulations and fueling staggering investments, our analysis is a good reminder that the world’s two AI superpowers still have a lot to gain from working together.
A Note on Methodology
I used Codex, OpenAI’s code-writing model, to help analyze NeurIPS papers. After writing a script to download all the papers, I used the model to dip into each one and do some analysis. This involved having Codex write a script to search for US and Chinese institutions in the author field of each paper.
The experiment offered a fascinating glimpse into the potential for coding models to automate useful chores. There’s plenty of panic about AI replacing coding jobs, but this is something that I normally wouldn’t have had the time or budget to build. I started out writing scripts and having Codex modify them before just asking Codex to do the analysis itself. This involved the model importing Python libraries, testing different tools, and writing scripts before producing reports for me to vet. The process involved a fair bit of trial and error, and you have to be very careful, because AI models make surprisingly stupid mistakes even when they’re being quite smart. I had to make sure that each report included a way for me to go through the results, and I checked as many as possible manually.
This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.