Nvidia 黄称 Anthropic 的人工智能安全反垄断豁免“完全没有必要”

Nvidia 首席执行官黄仁勋 (Jensen Huang) 在 CNBC 上拒绝了 Anthropic 提出的人工智能安全协调反垄断豁免，认为现有法律已经管辖前沿实验室。

Nvidia CEO Jensen Huang has sharply rejected Anthropic's proposal for an antitrust waiver that would let AI companies coordinate on safety efforts without running afoul of competition law, calling the idea "completely unnecessary" in a CNBC interview broadcast Tuesday.

"The fact that we need new laws, new antitrust laws, or new regulations, so that these companies could do their fundamental engineering and do it properly before they release products — that is just completely unnecessary," Huang said on CNBC's "Mad Money," according to CNBC. “我们有大量的法律。我们有大量的法规来管理产品的可靠性和功能。”

The comments put the world's most valuable chipmaker squarely against the most consequential governance idea to emerge from the frontier labs' safety debate — and extend a public feud that has been building all week.有关此故事的更多背景信息，请参阅我们正在进行的最新人工智能开发。

人类实际提出了什么

豁免的想法来自 Anthropic 首席执行官达里奥·阿莫迪 (Dario Amodei)。 In his September 12 essay calling on the industry to pace frontier development, Amodei acknowledged an awkward legal reality: some forms of coordination between rival labs — sharing safety evaluations, comparing internal test results, agreeing on capability thresholds — sit close to conduct that antitrust law traditionally treats as collusion.他提出的解决方案很狭窄：美国政府应该进行调解，或发布有限的豁免，以便竞争对手之间的安全对话显然是合法的。

Days later, The Information and then Bloomberg, Reuters and TechCrunch reported that OpenAI, Anthropic and Google DeepMind had been in talks for weeks on coordinating AI safety efforts — talks OpenAI's chief global affairs officer Chris Lehane confirmed on Tuesday, adding that OpenAI supports legislation to open frontier labs to independent verification organizations.

换句话说，阿莫代担心的协调已经发生了。是否需要法律保护现在成为业界分歧的问题。

黄的论点：现有法律就足够了

黄的反对意见是不需要排除，因为现行法律并不妨碍负责任的工程。在他看来，豁免框架把问题倒退了：实验室不需要竞争监管机构的许可，就可以在发布之前仔细测试他们的模型——他们需要在与其他行业相同的产品可靠性规则下进行工程设计。

这是一个背后有明显业务逻辑的视图。 Nvidia supplies the majority of the GPUs training large language models, and anything that slows the frontier — coordinated pauses, pacing commitments, capability thresholds that gate releases — slows demand for its chips.黄一直拒绝接受经济放缓的说法。 At Dreamforce on Monday, he told Salesforce chief Marc Benioff that speed versus safety is "a false choice" and that the industry "could definitely have both at the same time," CNBC reported.

他并不是唯一一个反对的人。 PCMag 报道称，Nvidia 和 Meta 的首席执行官都反对 Anthropic 的放缓计划，马克·扎克伯格也加入了支持持续快速发展的行列。 The lineup matters: the two companies with the most capital deployed against fast AI progress — one selling the compute, one spending it — are now the public face of opposition to the labs' proposed self-governance regime.

政策争论的分歧

这场争端在一个微妙的时刻来到华盛顿。 Congress has shown little appetite for new AI legislation, leaving the labs to negotiate governance among themselves — precisely the arrangement critics warn about.一封回应阿莫迪文章的公开信被疯传，认为安全规则上的行业协调相当于监管捕获：“所有这些都是在当前前沿实验室的帮助下编写的，因为没有其他人足够了解技术细节来编写它。”

黄的干预为豁免的支持者增添了第二个更严重的担忧：即使政府批准了一项，该行业的主要供应商支持其最大客户之间的卡特尔相邻安排的出现也会立即引起反垄断审查。联邦贸易委员会此前曾通过前主席莉娜·汗 (Lina Khan) 的公开评论表示，现行法律不存在人工智能豁免权——黄现在奇怪地认同这一立场，尽管出于相反的原因。