← 返回 Avalaches

白宫指示Anthropic暂停其最强大的人工智慧模型Fable 5与Mythos 5的外国存取权限,旨在保护敏感技术。然而此举可能反而成为竞争对手中国的助力,推动对中国开源人工智慧的需求。Trump政府以「国家安全考量」为由,要求Anthropic在向外国人提供其最前沿模型之前须取得美国批准,Anthropic随即遵从,实质上切断了全球其余地区对其高度抢手技术的存取。

数日后,中国人工智慧实验室智谱(Zhipu)发布了GLM-5.2。尽管该开源模型尚未达到美国最尖端模型的水准,但它提供了强大的程式编码效能、更低的价格,以及最关键的——零地理限制。该模型迅速进入Arena.ai的Code Arena前端编码排行榜,基于社群投票在顶级人工智慧编码任务中排名第二,仅次于Fable 5。出口管制逻辑对软体的适用方式不同于晶片:晶片是实体物品,其跨境流动可以被减缓;但模型权重——即决定人工智慧如何回应的数值参数——只是一个档案。当这些权重可被任何人下载时(如中国开源模型的情况),禁令效力减弱,对强大人工智慧的存取也更难监控。在华盛顿实施封锁后的数日内,Anthropic指控中国科技巨头Alibaba非法存取其人工智慧模型,但未具体说明涉及哪些模型。

Elon Musk预测中国将在2027年第一季前追上Fable的能力水准,智谱创办人、清华大学计算机科学教授暨中国人工智慧领域顶尖人物唐杰在X平台上简短回应:「不需要那么久。」Musk谨慎地区分了基准测试表现与实际应用价值之间的差距——这一差距确实重要。但随著每一个发布周期,这一差距正在逐渐缩小。

The White House directed Anthropic to suspend foreign access to its most powerful AI models, Fable 5 and Mythos 5, aiming to protect sensitive technology. However, the move may instead act as a tailwind for rival China, boosting demand for Chinese open source AI. The Trump administration ordered Anthropic to obtain US approval before providing its most cutting-edge models to foreign nationals, citing "national security concerns," and Anthropic complied, effectively cutting off access to its highly sought-after technology for the rest of the world.

Days later, Chinese AI lab Zhipu released GLM-5.2. Though not as advanced as the most cutting-edge US models, the open source model offers strong coding performance, lower prices, and crucially, zero geographical restrictions. It quickly entered Arena.ai's Code Arena Frontend coding leaderboard, ranking second behind Fable 5 based on community votes among top AI models at coding tasks. Export control logic applies differently to software than to chips: chips are physical and their cross-border movement can be slowed, but model weights—the numerical parameters determining how AI responds—are simply a file. When those weights can be downloaded by anyone, as with China's open models, bans become less effective and access to powerful AI harder to monitor. In the days following Washington's block, Anthropic accused Chinese tech giant Alibaba of illicitly accessing its AI models, though it did not specify which.

Elon Musk predicted that China would be able to match Fable's capabilities by the first quarter of 2027. Jie Tang, Zhipu's founder, a computer science professor at Tsinghua University and a top name in Chinese AI, responded briefly on X: "won't take that long." Musk was careful to distinguish benchmark performance from real-world usefulness—a gap that truly matters. But that gap is narrowing with every release cycle.

2026-06-26 (Friday) · 056b9e35812bfa3ca4e34c854304212766dde48c