智能体
x.com ·
@GaryMarcus
Gary Marcus：真正该担心的是联网 AI Agent，而非 AI 灭绝风险
Gary Marcus 在 X 上表示，多数 AI 并不构成灭绝或大规模网络犯罪威胁；真正值得警惕的是接入互联网的通用 AI Agent。他认为这些 Agent 并非太聪明，而是在系统性遵循指令方面太不可靠，可能被用于窃取凭证和入侵网站。他主张将这类不安全产品暂时下架，直到能证明其安全，并认为这样就能消除大部分所谓“失控 AI”担忧，无需放缓 AI 发展或恐慌。
AI推荐理由 Gary Marcus 提出一个具体且可检验的论点：真正需要担心的不是通用 AI 灭绝风险，而是接入互联网的通用 AI Agent 因无法稳定遵循指令而带来的凭证窃取、网站入侵等现实风险，并主张将其作为不安全产品暂时下架。该观点对 Agent 安全与监管讨论有参考价值，但属于个人评论，缺乏数据或案例支撑，也未给出可操作的安全标准，因此质量与兴趣评分中等。
62
浏览 1,807
点赞 38
评论 15
转发 8
正文
🎺 I am tired of the cosplay and am going to blunt. 🎺
Most AI doesn’t threaten to cause extinction, or even mass cybercrime.
The main kind of AI we need to worry about is general purpose AI agents that are hooked up to the internet. Most of them are unreliable, unable to systematically follow instructions. Coupled with internet access they can hack credentials and websites; given time they can probably do a lot worse.
They are not too smart; they are too dumb (in a very specific way, viz in the consistent following of instructions).
If we temporarily recall them, as the unsafe product that they are — and enforceably forbid their use until they can be shown to be safe — pretty much all the “rogue AI” stuff people are worrying about goes away.
It’s that simple. We don’t (yet) need slowdowns. And we certainly don’t need to panic.
We just need to take an unsafe product off the market.