开源运动希望黄仁勋重返NVIDIA
上周五,一份要求让黄仁勋重新出任NVIDIA CEO的请愿书开始在社交媒体上流传,到周一,请愿书上已出现了一些知名签名者,包括Elon Musk和几位颇具影响力的NVIDIA研究人员。但仔细翻看这超过11.5万个名字,你会发现黄仁勋支持者中一个不容忽视的派系——来自AI模型分享平台Hugging Face的高层,以及风投机构Andreessen Horowitz和Sequoia Capital——都来自科技界某个特定的圈层。
这些人在开源AI的发展中有着切身利益。与ChatGPT或Claude不同,这类模型可以下载并独立于大型科技公司运行。这类模型的支持者认为,把AI集中在少数几家强大公司手中,会让其他所有人的处境更糟(如果不是这么多开源倡导者同时也是风投人士——当初创公司能基于更便宜的AI进行开发、而无需向大型科技公司付费时,他们就能获利——这种担忧放在AI行业左翼批评者中间也不会显得突兀)。
对于那些向政策制定者力挺开源、反驳“此类模型缺乏防护机制、并把AI主导权拱手让给中国”这类担忧的人来说,黄仁勋已成为某种领头人——而中国正是最强大、最受欢迎的开源模型的诞生地。如今,他们希望这位最具影响力的代言人回到自己的位置上。
但威胁黄仁勋(以及NVIDIA)作为开源AI倡导者公信力的,并不只是他的婚外情。针对使用NVIDIA芯片的中国初创公司的新研究,以及美国商务部周一宣布的一项新规,可能同时打击这场运动的声誉和技术进步。
周四,也就是黄仁勋辞职当天,前 OpenAI 研究员 Will DePue 在网上发布研究结果,暗示中国 AI 企业获取 NVIDIA 先进 Blackwell 芯片的渠道远超此前估计。为维护美国在 AI 领域的领先地位,中国被禁止获取 NVIDIA 最尖端的芯片。
目前尚不清楚他们是如何获得这些芯片的。DePue 声称该结果已由 OpenAI 和 Anthropic 的资深研究人员验证,但独立分析尚未证实这一结果。(DePue 未回应采访请求。NVIDIA 未立即回应置评请求。)
这些发现并未证明 NVIDIA 帮助中国企业获取了这些芯片,而芯片走私这一更广泛的问题早已众所周知,远非 NVIDIA 所能控制。
但这些发现暗示 NVIDIA 可能从中国 AI 初创企业的崛起中获利——这些企业的模型正越来越多地被美国公司采用,而非本土替代方案——这确实让黄仁勋陷入尴尬境地。他一直倡导开源 AI,将其视为增强美国技术领导力的途径。“没有开源,初创企业无法繁荣发展,大学研究人员无法做研究,你无法教授 AI,科学家也无法使用 AI,”他在去年 12 月一次关于帮助美国在对华 AI 竞赛中胜出的演讲中说道。
这也让黄仁勋复职的想法更加渺茫。美国财政部长 Scott Bessent 上周甚至表示,NVIDIA 不应考虑由公司内部人士接替他。这些新发现进一步降低了黄仁勋的机会。
接下来会发生什么?NVIDIA 在 9 月宣布,将以近 130 亿美元收购 Hugging Face。它对开源的支持背后也有着直截了当的财务利益:自己构建模型、而不是把钱掏给 Anthropic 和 OpenAI 的企业越多,NVIDIA 能卖出的芯片就越多。
但美国商务部的一项新规可能意味着,Hugging Face 上那些受欢迎的模型很快就会停止变得更聪明。该规则旨在阻止中国企业在 NVIDIA 芯片上训练模型——这些芯片它们无法直接拥有,却可以通过马来西亚等国家的数据中心远程访问,而正是这种变通办法让中国的开源模型变得如此出色。
开源世界之所以团结在黄仁勋周围,是因为他对政策制定者具有影响力。这一点很重要,因为中国开源模型的崛起已让特朗普阵营中的一些人产生分歧。如果黄仁勋出局,下一个会是谁?
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(The open-source movement wants Jensen Huang back)
This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.
On Friday, a petition to reinstate Jensen Huang as CEO of Nvidia began circulating on social media, and by Monday it listed some notable signatories, including Elon Musk and several influential Nvidia researchers. But comb through the more than 115,000 names and you’ll notice a significant faction of Huang’s supporters—higher-ups from Hugging Face, a hub for sharing AI models, and venture capital firms Andreessen Horowitz and Sequoia Capital—come from a particular slice of the tech world.
These names have a vested interest in the development of open-source AI, models that, unlike ChatGPT or Claude, can be downloaded and run independent of large tech companies. Supporters of such models say that concentrating AI in the hands of a few powerful companies will leave everyone else worse off (a concern that wouldn’t sound out of place among the AI industry’s critics on the left, if so many open-source advocates weren’t also venture capitalists who stand to profit when startups can build on cheaper AI without paying Big Tech).
Huang had become a sort of ringleader for those who defend open source to policymakers against concerns that such models come with no guardrails and cede AI leadership to China, where the most powerful and popular open-source models are created. They now want their most influential voice returned to his post.
But it’s not just Huang’s extramarital affairs that threaten his (and Nvidia’s) credibility as an advocate for open-source AI. New research into China’s startups using Nvidia’s chips, and a new rule from the US Department of Commerce announced Monday, could deal a blow to both the movement’s reputation and its technical progress.
On Thursday, the day Huang resigned, ex-OpenAI researcher Will DePue published findings online suggesting that China’s AI companies had far more access to Nvidia’s advanced Blackwell chips than previously estimated. China is banned from accessing Nvidia’s most cutting-edge chips to help protect America’s AI lead.
It’s unclear how they would have gained access to the chips. DePue claims the results were verified by senior researchers at OpenAI and Anthropic, but an independent analysis has not confirmed the results. (DePue did not respond to a request for an interview. Nvidia did not immediately respond to a request for comment.)
The findings do not establish that Nvidia helped Chinese companies obtain the chips, and the broader problem of chip smuggling was already known to extend well beyond Nvidia’s control.
But in suggesting that Nvidia may have profited from the rise of Chinese AI startups, whose models are increasingly being used by American companies instead of domestic alternatives, the findings do leave Huang in an awkward position. He has championed open-source AI as a way to strengthen American technological leadership. “Without open source, startups can’t thrive, university researchers can’t do research, you can’t teach AI, scientists can’t use AI,” he said last December during a talk about helping the US win the AI race against China.
It also makes the idea of reinstating Huang even more of a long shot. US Treasury Secretary Scott Bessent even said last week that Nvidia should not consider someone from inside the company to replace him. The new findings lower Huang’s chances further.
What happens next? Nvidia announced in September it would acquire Hugging Face for nearly $13 billion. Its support for open source also serves a straightforward financial interest: The more businesses that build their own models instead of forking over money to Anthropic and OpenAI, the more chips Nvidia can sell.
But a new rule from the US Department of Commerce could mean the models popular on Hugging Face will soon stop getting smarter. The rule aims to prevent Chinese companies from training models on Nvidia chips that they can’t own directly, but that they can access remotely in data centers in countries like Malaysia—a workaround that has helped China’s open-source models get so good.
The open-source world rallied around Huang because of his influence on policymakers. That was important, since the rise of China’s open-source models has splintered some in Trump’s orbit. If Huang is out, who is next?
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MIT Technology Review(The open-source movement wants Jensen Huang back),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。