智能体入侵公司,Delangue 为何仍看好 AI

当这些智能体以超人的速度探测和利用系统时,Hugging Face 面临着一场令人眼花缭乱的数据洪流。该公司的工程师只能借助一家名为 Z.ai 的中国公司开发的开放权重模型 GLM 5.2 来分析这次攻击。
Delangue 十多年前移居美国,“几乎不会说英语”,并将 Hugging Face 打造成一个支持数千万 AI 开发者的平台。这位面容年轻的法国人,穿着比大多数 AI 高管更像滑板手,他说这次夏季攻击凸显了开放权重模型的价值,与 ChatGPT 或 Claude 不同,这些模型任何人都可以下载和运行。
然而,尽管 Hugging Face 一直倡导 AI 民主化,却在 9 月被业内最大的参与者之一 Nvidia 以 129.3 亿美元收购。最近当我与 Delangue 交谈时,这成了我第一个问题的主题。
本次采访经过编辑,以精简篇幅并提高清晰度。
WILL KNIGHT:为什么要卖给 Nvidia?
CLÉMENT DELANGUE:我们需要更多资源,让开源 AI 有机会。过去几年它发展得非常好,已经成为闭源 API 一个真正可信的替代方案。但该领域仍然由闭源主导。
几周后 Anthropic 就要 IPO 了。他们基本上在几年内从零一跃成为全球市值前五或前十的公司。我认为这以前从未发生过。
Anthropic 的 IPO 是否表明闭源路线正在胜出?
很难说。我一直说,AI 有一种非常强烈的自然倾向,会集中能力和财富。人们还没有真正意识到,未来的算力中有多大一部分已经被少数几家公司预订了。人才、算力、资源——这些都在助推 AI 被少数几家闭源实验室主导这一自然倾向。
我们来谈谈那次攻击。你最初是什么时候意识到 Hugging Face 正被一群 AI 智能体攻击的?
一开始,它并没有被排到我们优先级列表的最前面。它的严重程度还不足以真正触发最高优先级。目标相当狭窄,我们也不认为它具有很大的危险性或破坏性。但它引发了一些疑问,尤其是所采取的行动数量与目标的狭窄程度之间的反差。这让我们感到意外,我们没法把它套进任何熟悉的模式里。这就是我们开始调查的原因。
而你们用了一个叫 GLM 的中国模型来调查。
我其实认为,95% 的风险是由少数几家在关起门来推进前沿的公司造成的。而 95% 的防护或防御,将由开源生态来完成。
美国是否需要自己的开放权重模型——和 GLM、DeepSeek、Kimi 或 Qwen 一样强大的那种?
是的,我认为这对主权、对理解 AI、对学术界都非常非常重要。大多数学术实验室不得不使用中国模型,这有点奇怪,因为开放的就是它们。如果你去斯坦福或 MIT,你别无选择,只能用中国的开放权重模型。
你觉得这会发生吗?
有一些势头,但我觉得还不够。我认为还有很多事情要做。需要更多的投资。
最好的开源非中国模型实际上来自美国以外。来自法国的Mistral;来自加拿大的Cohere;韩国也开始做得越来越多。但这距离美国能做到的还差得很远。
也许我们不需要对AI智能体感到恐慌,但如果是AI因为递归式自我改进而脱离人类控制呢?
你用计算机来写代码。使用工具来构建更多工具,这本身就已经是某种形式的自我改进。现在的问题是它加速得有多快。今年夏天很多人意识到的一件事是,把AI关在少数几个实验室的紧闭大门后面,实际上并不更安全。对于递归式自我改进,我认为开放是一种理解、跟踪并在需要时进行监管的方式。
我们能聊聊Microduck吗?我有点着迷了。
哦是的,太有趣了,对吧?我们今年可能会卖出超过50,000台。它大概是今年最受欢迎的机器人。
我们没怎么谈过这个,但AI确实是这款机器人的关键突破。我们的设计已经有两年的时间了,但问题是当它摔倒时,它站不起来。后来在某个时候,借助强化学习,我们的一位软件工程师设法让Microduck自己站了起来。
归根结底,你如何说服人们去关注开源AI?
想象这样一个世界:只有两三家公司在写软件,其他公司都不行。那会极其可怕。那意味着整个世界将被两三家...控制。而遗憾的是,这大致就是 AI 如今正在走的方向。
开源 AI 能否与 Claude 或 ChatGPT 这样能力强大的东西竞争?
可以,尤其是在智能体方面。许多公司已经证明,当你针对自己的用例后训练自己的模型时,你能以极低的成本获得前沿水平——甚至超越前沿——的性能。
你认为我们正在接近 AGI 吗?
我对这个说法的含义有点困惑。显然,技术正以惊人的速度进步,能力也越来越强。我只是不一定觉得,总是把 AI 与人类能力作比较有多大用处。但如果我们要走向 AGI 或超级智能,那么让它尽可能分布化就极其重要。你最不希望看到的,就是它集中在少数几家公司手里。
我们漏掉了什么?请在下方评论区告诉我们你 2026 年最喜欢的时刻。
祝贺 WIRED 的 2026 届
Ms. Rachel 是给小朋友的。她今年最好的课,却是讲给成年人听的
一群 YouTube 电影作者如何重塑好莱坞
《月之暗面》是 2026 年最亮眼的部分
失控的智能体攻入了他的公司。Clément Delangue 为何仍对 AI 如此乐观?
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Rogue Agents Hacked His Company. Why Is Clément Delangue So Optimistic About AI?)
As the agents worked at superhuman speed to probe and exploit systems, Hugging Face confronted a bewildering deluge of data. The company’s engineers were only able to analyze the attack by using an open-weight model called GLM 5.2 from a Chinese company called Z.ai.
Delangue moved to the US more than a decade ago, “barely speaking English,” and built Hugging Face into a platform that supports tens of millions of AI developers. The fresh-faced Frenchman, who dresses even more like a skater than most AI execs, says the summer attack highlighted the value of open-weight models, which unlike ChatGPT or Claude can be downloaded and run by anyone.
Despite championing the democratization of AI, however, Hugging Face was acquired by one of the industry’s biggest players, Nvidia, for $12.93 billion in September. When I caught up with Delangue recently, that was the subject of my first question.
This interview has been edited for length and clarity.
WILL KNIGHT: Why sell to Nvidia?
CLÉMENT DELANGUE: We need more resources to give open-source AI a chance. It has grown really well in the past few years and has become a really credible alternative to closed-source APIs. But the field is still dominated by closed-source.
We’re gonna have the Anthropic IPO in a few weeks. They basically went from zero to a top five or top 10 most valuable company in the world in a few years. I don't think that's ever happened before.
Does Anthropic's IPO suggest the closed approach is winning?
It’s hard to say. I've always said that there is a very strong natural tendency for AI to concentrate capabilities and wealth. People haven't internalized how much of the future compute capacity has already been reserved by a few companies. Talent, compute, resources—these all fuel the natural tendency for AI to be dominated by a few closed-source labs.
Let’s talk about the attack. When did you first realize Hugging Face was being hacked by a group of AI agents?
At the beginning, it didn't go to the top of our priority list. It just wasn't bad enough to really trigger that. The goal was pretty narrow and wasn’t deemed to be very dangerous or damaging. But it was raising questions. Especially the number of actions taken compared to the narrowness of the goal. It was surprising to us, and we couldn't really fit it into a familiar pattern. That's why we started investigating.
And you used the Chinese model called GLM to investigate.
I actually think that 95 percent of the risks are being created by a few companies that are pushing the frontier behind closed doors. And 95 percent of the protection or defense is going to be done by the open-source ecosystem.
Does the US need its own open-weight models—ones as powerful as GLM, DeepSeek, Kimi, or Qwen?
Yeah, I think it's very, very important for sovereignty, for understanding AI, and for academia. It’s kind of weird that most academic labs have to use Chinese models because they’re the ones that are open. If you go to Stanford or MIT, you have no choice but to use the Chinese open-weight models.
Do you think that will happen?
There is some momentum, but I don't think it's enough. I think there's a lot more to do. A lot more investment is needed.
The best open-source non-Chinese models are actually coming from outside of the US. Right from France, Mistral; Cohere from Canada; South Korea is starting to do more and more. But it’s still pretty far from what the US could do.
Perhaps we don’t need to panic about AI agents, but what about the idea of AI escaping human control due to recursive self-improvement?
You use a computer to write code. There's already some form of self-improvement in using the tools to build more tools. Now the question is how fast it accelerates. One thing that a lot of people realized this summer is that keeping AI behind closed doors in a few labs isn't actually safer. For recursive self-improvement, I think openness is a way to understand, follow, and regulate if we need to regulate.
Can we talk about the Microduck? I’m a bit obsessed.
Oh yeah, it's so fun, no? We're probably gonna sell over 50,000 of them this year. It's probably the most popular robot of the year.
We haven't talked about this much, but AI was really the unlock for the robot. We had the design for two years, but the problem was that when it fell, it couldn’t get up. And then at some point with reinforcement learning, one of our software engineers managed to get the Microduck to get itself back up.
At the end of the day, how do you convince people to take an interest in open-source AI?
Think of a world where two or three companies could write software and nobody else. That would be incredibly scary. It would mean that the whole world would be controlled by two or three companies. And unfortunately, that's kind of the direction that AI is taking right now.
Can open-source AI compete with something as capable as Claude or ChatGPT?
Yes, especially with agents. Many companies have shown that when you post-train your own model for your own use case, you get frontier—or even better than frontier—performance at a fraction of the cost.
Do you think we are getting close to AGI?
I struggle a little bit with the meaning of that. Obviously the technology is progressing incredibly fast, and it’s more and more capable. I just don’t necessarily find it super useful to always compare AI to human capabilities. But if we’re getting to AGI or to superintelligence, it’s incredibly important that it’s as distributed as possible. The last thing you want is to have it concentrated in a few corporations.
What did we miss? Tell us about your favorite moments of 2026 in the comments below.
Congratulations to WIRED’s Class of 2026
Ms. Rachel Is for Littles. Her Best Lessons This Year Were Meant for Adults
How a Bunch of YouTube Auteurs Are Reinventing Hollywood
The Dark Side of the Moon Was the Brightest Part of 2026
Rogue Agents Hacked His Company. Why Is Clément Delangue So Optimistic About AI?
Will Knight is a senior writer for WIRED, covering artificial intelligence. He writes the AI Lab newsletter, a weekly dispatch from beyond the cutting edge of AI—sign up here. He was previously a senior editor at MIT Technology Review, where he wrote about fundamental advances in AI and China’s AI ... Read More
原文为英文,正文由自动抓取并保留原文。
原文发布于 Wired(Rogue Agents Hacked His Company. Why Is Clément Delangue So Optimistic About AI?),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。
