中国积极发展开源人工智慧,不仅对矽谷构成挑战,更成为北京提升全球影响力的重要筹码。然而,随著开源模型能力增强,北京正努力在维持透明度优势与防范潜在风险之间取得平衡。由于人工智慧安全事故的影响可能跨越国界并阻碍全球技术发展,各国政府与顶尖实验室必须在灾难发生前主动应对这些安全挑战。
关于开源与闭源模型安全性的争论过于简化,事实上许多滥用事件也发生在设有防护机制的闭源模型上。开源系统在加强网路安全防御方面能发挥关键作用,例如 Hugging Face 曾利用开源模型成功抵御攻击。专家认为,为防御者配备优良的开源工具,配合对网路攻击的严厉惩罚与强制通报机制,是提升整体安全性的有效途径。
运行强大的开源模型通常需要庞大的运算资源,这使得云端服务供应商成为监控可疑活动的重要防线。尽管美中两国之间存在不信任,但在人工智慧安全领域的对话仍不可或缺。决策者应将人工智慧防御视为关键基础设施,促进情报共享与标准化,因为忽视这些共同威胁而仅专注于技术竞赛,最终将对各方造成损失。
China's active embrace of open-source artificial intelligence is posing a challenge to Silicon Valley and becoming a significant asset for Beijing's global influence. However, as open models become more capable, Beijing is striving to balance the benefits of transparency with the need to prevent potential risks. Since AI safety failures can easily cross borders and hinder global technological progress, governments and leading laboratories must proactively address these safety challenges before a catastrophe occurs.
The debate over the safety of open versus closed models is often oversimplified, as many instances of abuse actually occur with closed models that have safeguards in place. Open systems can play a crucial role in strengthening cybersecurity defenses, illustrated by Hugging Face utilizing an open model to successfully counter an attack. Experts argue that equipping defenders with superior open tools, combined with strict penalties for cyberattacks and mandatory incident reporting, is an effective way to improve overall safety.
Running powerful open models generally requires substantial computing resources, making cloud service providers a critical line of defense for monitoring suspicious activities. Despite the existing distrust between the United States and China, dialogue regarding artificial intelligence safety remains essential. Policymakers should treat AI defense as critical infrastructure by promoting shared intelligence and standardization, because ignoring these mutual threats in favor of sheer technological competition will ultimately result in losses for all parties.