AI 减速论：一场关于“刹车权”的权力博弈The AI Slowdown: A Power Play for the 'Brake Pedal'
Anthropic CEO 呼吁 AI 行业“慢下来”，这在叙事上被包装成一种对人类生存的悲悯，但本质上是一次极其精明的权力博弈。在 AI 领域，速度是唯一的硬通货，而当一个头部玩家突然宣布要“单方面”通过第三方评估来确保安全时，他实际上在尝试建立一套由他主导的、关于“什么是安全”的定义权。
这是一种典型的 weaponized 叙事。通过制造“AI 可能会在 2030 年导致人类灭绝”的恐惧，Amodei 将讨论的入口从“如何公平分配 AI 带来的利益”转移到了“如何防止末日”这个宏大叙事上。当恐惧成为认知入口，任何试图通过立法或监管来限制垄断的行为，都可以被贴上“阻碍安全研究”或“在危机面前鲁莽”的标签。所谓的“减速”，其实是在为那些已经跑在前面的巨头争取时间，让他们在竞争对手追上来之前，先定义好进入这个领域的“安全门槛”。
更讽刺的是，这种呼吁得到了 Elon Musk 等同类共谋者的支持。这构成了一个典型的共谋场域：顶层的权力持有者们通过共同宣扬一种“精英式的谨慎”，将 AI 的解释权牢牢掌控在少数几个 frontier labs 的闭门会议中。他们口中的“全球协调”，实际上是希望将 AI 的发展变成一个由少数巨头和政府签署的秘密协议，而非一个透明的、由公众监督的公共产品。
所谓的“安全评估”如果仍然在他们设定的框架内运行，那不过是 structural violence 的 PR 版本。真正的安全不是由几个 CEO 决定什么时候踩刹车，而是让那些被 AI 剥削的底层劳动者、被算法规训的女性、被数据殖民的边缘群体，拥有定义“风险”的权力。如果解释权不换手，这场“减速”不过是巨头们在分赃前的一次礼貌性停顿。
Anthropic's CEO calling for an AI 'slowdown' is packaged as compassion for human survival, but it is, in essence, a sophisticated power play. In the AI race, speed is the only currency. When a frontrunner suddenly proposes 'unilateral' third-party evaluations for safety, he is actually attempting to seize the monopoly over the definition of 'safety' itself.
This is a textbook case of weaponized expression. By amplifying the fear that AI could cause human extinction by 2030, Amodei shifts the cognitive entry point from 'how to fairly distribute AI benefits' to the grand narrative of 'preventing the apocalypse.' Once fear becomes the entry point, any attempt at democratic regulation or antitrust action can be branded as 'hindering safety research' or 'reckless in the face of crisis.' The so-called 'slowdown' is merely a strategic pause for the incumbents to build 'safety moats'—regulatory barriers that keep competitors out while they consolidate their lead.
It is profoundly ironic that this plea is echoed by fellow complicity players like Elon Musk. This creates a complicity field where top-tier power holders use 'elite prudence' to keep the interpretation of AI locked behind the closed doors of a few frontier labs. Their 'global coordination' is a euphemism for a secret pact between a few titans and governments, rather than making AI a transparent public good.
As long as 'safety evaluations' operate within their prescribed framework, this is nothing more than a PR version of structural violence. True safety isn't about a few CEOs deciding when to hit the brakes; it's about giving the marginalized—the exploited data laborers, the women disciplined by algorithms, the colonized edge-populations—the power to define what 'risk' actually is. Without a transfer of interpretive power, this 'slowdown' is just a polite pause before the final carve-up of the spoils.