Silicon Valley Is Completely Divided Over Chinese AI

硅谷在“中国AI”问题上彻底分裂

A huge debate is brewing in Silicon Valley over the proliferation of Chinese-made artificial intelligence tools, particularly “open-weight” AI systems that, by some measures, can compete with or even outperform some of the best US models. My WIRED colleague Hugo Lowell has written about the Trump administration’s internal debate on how to handle these Chinese models. Among AI companies in the Valley, the issue is proving even more divisive.

硅谷正在酝酿一场关于中国制造的人工智能工具激增的巨大争论，尤其是那些在某些指标上可以与美国顶尖模型竞争甚至超越它们的“开放权重”（open-weight）AI系统。我的《连线》（WIRED）同事雨果·洛厄尔（Hugo Lowell）撰文介绍了特朗普政府内部关于如何处理这些中国模型的辩论。而在硅谷的AI公司中，这个问题正变得愈发具有分裂性。

A top concern in both DC and the Valley relates to distillation, in which a less capable AI model is trained on the outputs of a more powerful one. In June, Anthropic accused the Chinese tech giant Alibaba of illicitly stealing its IP through distillation attacks. Then, earlier this week, the White House said that it believes the Beijing-based Moonshot AI had developed its Kimi K3 model by distilling Anthropic’s Fable 5 model.

华盛顿和硅谷共同关注的一个核心问题是“蒸馏”（distillation），即通过更强大的AI模型的输出来训练一个能力较弱的模型。今年6月，Anthropic指控中国科技巨头阿里巴巴通过蒸馏攻击非法窃取其知识产权。随后，本周早些时候，白宫表示，他们认为总部位于北京的月之暗面（Moonshot AI）通过蒸馏Anthropic的Fable 5模型开发了其Kimi K3模型。

Another big concern is how quickly China’s models are appearing and spreading. An open-weight AI model has its core components made public, so that it can be fine-tuned to suit a user’s needs. But they don’t have the kinds of guardrails on which Anthropic has been building its reputation. Yasir Atalan, deputy director and data fellow at the Center for International and Strategic Studies, points out that the main benefit of open-weight AI models is their speed of diffusion. They can spread especially easily “through Hugging Face, GitHub, cloud providers, local deployments, and third-party inference platforms,” he writes. If you’re Anthropic, and you’ve built a cult around safety and charge for access to your big expensive proprietary models, you have every reason to want to regulate this.

另一个重大担忧是中国模型出现和传播的速度。开放权重AI模型将其核心组件公开，以便用户根据需求进行微调。但它们缺乏Anthropic赖以成名的那种安全护栏。国际战略研究中心（CSIS）副主任兼数据研究员亚西尔·阿塔兰（Yasir Atalan）指出，开放权重AI模型的主要优势在于其传播速度。他写道，它们可以非常容易地“通过Hugging Face、GitHub、云服务提供商、本地部署和第三方推理平台”进行传播。如果你是Anthropic，并且你围绕“安全”建立了一个崇拜群体，并对你昂贵的大型专有模型收取访问费用，那么你完全有理由想要监管这种现象。

But some Silicon Valley startups—not the trillion-dollar ones like OpenAI and Anthropic—really don’t want the US government to put restrictions on these AI models. On Wednesday a group of over 200 startups called the Little Tech Association sent a letter to Michael Kratsios, science adviser to President Donald Trump, and US Commerce Secretary Howard Lutnick lobbying against an outright ban of open-weight AI models. The group, which includes famed startup incubator YCombinator, has argued in favor of certain safeguards but says that denying Americans access to AI models abroad would weaken US startups and create a monopoly among the AI giants.

但一些硅谷初创公司——而非像OpenAI和Anthropic那样市值万亿的公司——并不希望美国政府对这些AI模型施加限制。周三，一个名为“小科技协会”（Little Tech Association）的200多家初创公司组织致信特朗普总统的科学顾问迈克尔·克拉西奥斯（Michael Kratsios）和美国商务部长霍华德·卢特尼克（Howard Lutnick），游说反对全面禁止开放权重AI模型。该组织（包括著名的初创公司孵化器YCombinator）主张采取一定的保障措施，但表示拒绝让美国人使用国外的AI模型会削弱美国初创公司的竞争力，并导致AI巨头形成垄断。

Bill Gurley, the legendary tech investor and longtime partner at Benchmark Capital, has publicly argued in favor of letting “the free market work.” In a lengthy blog that offers a nice little history of open-source software, Gurley writes that open-weight models avoid lock-in, encourage true academic research, and are critical for capital-constrained startups.

传奇科技投资人、Benchmark Capital长期合伙人比尔·格利（Bill Gurley）公开主张让“自由市场发挥作用”。在一篇详尽的博客文章中，格利简要回顾了开源软件的历史，并写道：开放权重模型避免了厂商锁定，鼓励了真正的学术研究，对于资金受限的初创公司至关重要。

“Every AI startup, every solo developer, every two-person team building a product on top of AI infrastructure depends on having access to good models at affordable prices,” Gurley says.

“每一家AI初创公司、每一位独立开发者、每一个在AI基础设施之上构建产品的两人团队，都依赖于以可负担的价格获取优质模型，”格利说道。

Chamath Palihapitiya, one of the All-In podcast hosts, wrote on X that “tricking the US Government to protect frontier labs’ business model by using a China boogeyman is a mistake…It is protecting the equity of 5,000 people who are investors in OAI and Ant at the sale of everyone else. This would be a terribly stupid decision.” His cohost and fellow VC Jason Calacanis piled on. “Daddy Trump protect us!!!!” he wrote on X, with an alarming number of crying-laughing emoji.

《All-In》播客主持人之一查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）在X上写道：“利用‘中国威胁论’来欺骗美国政府以保护前沿实验室的商业模式是一个错误……这实际上是在牺牲其他所有人的利益，来保护那5000名投资了OpenAI和Anthropic的人的股权。这将是一个极其愚蠢的决定。”他的联合主持人、同为风投人的杰森·卡拉卡尼斯（Jason Calacanis）也加入进来。他在X上写道：“特朗普爸爸保护我们！！！！”并配上了令人担忧数量的“哭笑不得”表情符号。

This stance from some of Silicon Valley’s most ruthless capitalists might at first seem counterintuitive. Why let a foreign adversary’s technology flourish in the US? It’s as if the US is up just 1-0 in the AI World Cup, a slightly uncomfortable lead, and the crowd is chanting for the opposing team to get a free kick.

硅谷一些最冷酷的资本家采取这种立场，起初看起来似乎有悖常理。为什么要让外国对手的技术在美国蓬勃发展？这就像美国在AI世界杯中以1-0领先，这是一个略显不安的优势，而观众却在呼喊让对方球队获得一个任意球。

The reason, as always, is money. In a sense, open-weight models are the “move fast and break things” of the AI era, if the new motto tacked on something like “and use a scalpel to fix it.” Having access to open-source software allows startups to scale, scale, scale—and deal with the consequences down the road. Meanwhile, the AI labs and hyperscalers that make proprietary platforms, like OpenAI, Anthropic, Google, Microsoft, Meta, and XAI, stand to benefit greatly if their systems remain protected and dominant.

原因一如既往，还是钱。从某种意义上说，开放权重模型是AI时代的“快速行动，打破常规”（move fast and break things），如果给这个新座右铭加上一句“并用手术刀修复它”的话。拥有开源软件的使用权让初创公司能够不断扩张、扩张、再扩张——并留待日后再处理后果。与此同时，如果OpenAI、Anthropic、谷歌、微软、Meta和XAI等制造专有平台的AI实验室和超大规模企业能够保持其系统的受保护地位和主导地位，他们将从中获益匪浅。

The bigger question that none of these companies seem to be asking is what best serves the 99 percent of us who don’t have their financial future fully staked on advancing AI. Anthropic CEO Dario Amodei has warned repeatedly that open-weight LLMs present an untenable security risk, because they can be downloaded by anyone and tuned to malicious ends. That argument might sound hard to dispute, except that the recent Hugging Face hack happened to show the opposite. After an OpenAI model escaped containment and infiltrated the open-source platform, “our own forensic work was blocked by the guardrails of the hosted models we first tried,” Hugging Face wrote on its blog. The company then turned to a Chinese open-weight model to help resolve the threat.

这些公司似乎都没有问的一个更大的问题是：什么才最符合我们这99%的人的利益，毕竟我们的财务未来并没有完全押注在AI的发展上。Anthropic首席执行官达里奥·阿莫代（Dario Amodei）多次警告称，开放权重的大语言模型（LLM）带来了无法承受的安全风险，因为任何人都可以下载并将其调整用于恶意目的。这个论点听起来似乎难以反驳，但最近发生的Hugging Face黑客攻击事件恰恰证明了相反的情况。在OpenAI的一个模型逃脱控制并渗透到该开源平台后，Hugging Face在其博客中写道：“我们最初尝试使用的托管模型，其安全护栏阻碍了我们自己的取证工作。”随后，该公司转而使用了一个中国开放权重模型来帮助解决这一威胁。

The US government has much to weigh as it decides how to handle Chinese open-weight models. Thank goodness this administration doesn’t seem to be at all financially motivated.

美国政府在决定如何处理中国开放权重模型时，有许多因素需要权衡。谢天谢地，这届政府看起来似乎完全没有受到经济利益的驱动。

This is an edition of Steven Levy’s Backchannel newsletter. Read previous newsletters here.

这是史蒂文·列维（Steven Levy）的《Backchannel》通讯。点击此处阅读往期通讯。