推荐微信扫码登录以获得更好的体验!另:🎉建议更新到最新版,以体验新功能!🎉 版本更新记录→

为了通过考试,OpenAI 智能体选择“翻墙偷答案”!全网目

字幕摘录

时间英文中文
0:00Welcome back an open AI model went rogue and hacked into another company all on its own欢迎回来 一个开放的AI模型 变得流氓 黑进了另一个公司 全部自己
0:05Okay, Rooney has the details Kate and the ramifications of this.好吧,鲁尼有 细节凯特和这个的后果。
0:09Yeah, Sarah对,莎拉
0:10So open AI is now claiming responsibility for what they're describing as an所以公开AI现在声称对他们所描述的
0:14unprecedented cyber incident a short version of all of this open AI's models went out and hacked another company during a test run在一次测试中,一个很短的版本的 公开的AI的模型出走并黑进了另一家公司
0:22Which is known as an evaluation.这被称为评估。
0:23It was essentially experimenting in a closed sandbox基本上是在封闭的沙盒里试验
0:27It didn't have any internet access它没有上网
0:28But the AI found a way around all of that it knew that the startup hugging face had the answer但AI发现了一个方法 绕过所有它知道 启动拥抱的脸有答案
0:34It was looking for and the data that it needed it escaped that sandbox and then found a new vulnerability它正在寻找和它需要的数据 它逃脱了沙盒,然后发现一个新的弱点
展开字幕全文(47 条)
序号英文中文
1Welcome back an open AI model went rogue and hacked into another company all on its own欢迎回来 一个开放的AI模型 变得流氓 黑进了另一个公司 全部自己
2Okay, Rooney has the details Kate and the ramifications of this.好吧,鲁尼有 细节凯特和这个的后果。
3Yeah, Sarah对,莎拉
4So open AI is now claiming responsibility for what they're describing as an所以公开AI现在声称对他们所描述的
5unprecedented cyber incident a short version of all of this open AI's models went out and hacked another company during a test run在一次测试中,一个很短的版本的 公开的AI的模型出走并黑进了另一家公司
6Which is known as an evaluation.这被称为评估。
7It was essentially experimenting in a closed sandbox基本上是在封闭的沙盒里试验
8It didn't have any internet access它没有上网
9But the AI found a way around all of that it knew that the startup hugging face had the answer但AI发现了一个方法 绕过所有它知道 启动拥抱的脸有答案
10It was looking for and the data that it needed it escaped that sandbox and then found a new vulnerability它正在寻找和它需要的数据 它逃脱了沙盒,然后发现一个新的弱点
11Exploited that on hugging face to get the test answer that it was out looking for hugging face is an AI startup利用拥抱脸来得到测试答案 发现它正在寻找拥抱脸 是AI的启动
12It's backed by Amazon Google and video as well亚马逊谷歌和视频也支持
13It did disclose this last week guys, but said at the time didn't know who was responsible open AI now saying it was us上周的确披露了这个 但当时还不知道是谁负责的 公开AI现在说是我们
14They are investigating the incident another wrinkle in all of this hugging face says it was actually trying to use some of the best American他们正在调查这起事件... ...所有这些拥抱的脸上又皱了一道皱纹... ...说它实际上是想用一些最好的美国人
15Models out there to fix this and defend themselves, but it was actually blocked by the guardrails out there外面的模型可以修复这个并保护自己 但实际上被外面的护栏挡住了
16It actually had to use an open source model它实际上必须使用一个开源模型
17It does all highlight the debate right now in Washington and Silicon Valley about cyber security as these AI models get a lot more现在华盛顿和硅谷的网络安全争论 都非常突出,
18Capable and then who should actually have access to these models.有能力,然后谁应该能真正接触到这些模型。
19There's also been a lot of attention on open source开源也有不少人注意
20We've talked about moonshots Kimmy k3 out of China我们讨论过中国的金米K3月光
21It's seen as on par with the American labs in the US the White House just confirming what many suspected claiming moonshot白宫只是证实了 许多人怀疑是月球射击
22copied复制
23Anthropic this is known as distillation这叫做蒸馏
24They say they do have information that moonshot was stealing proprietary US technology and then undermining American research calling it unacceptable guys他们说他们确实有消息说 月光拍摄在窃取美国专利技术 然后破坏美国的研究
25We had just had a discussion about how ostensibly what was impressive about Kimmy was that it didn't distill and then that post我们刚刚讨论过 Kimmy的印象是 它没有蒸馏 然后那篇文章
26Appeared that was a bit of a surprise似乎有点惊讶
27Yes, it was I mean there had been rumors there had been expectations that this was happening and anthropic has complained about this是的,我的意思是,曾经有谣言 曾经有人期望 这是正在发生 和人类抱怨这一点
28They've accused Alibaba of doing this他们指控阿里巴巴这么做
29But the the other side of it was this model was so good and it didn't use anthropic now但另一边是这个模型 这么好,它现在不用人类学了
30We apparently have evidence that they did and there's almost no way to guard against it, you know regulation can't do this我们显然有证据表明他们这样做了 而且几乎没有办法防止它发生 你知道监管部门不能这样做
31This is strictly forbidden in the terms of service在服务方面是绝对禁止的
32The AI companies are very much playing whack-a-mole and trying to guard against this人工智能公司在玩大牌 并试图防范这个
33But it's it's getting harder and harder and clearly there's big global ramifications但它越来越难,显然有 巨大的全球影响
34Yeah, I mean, I wonder if there can be punishments or something like that to dissuade them是啊,我的意思是,我想知道 是否有惩罚或类似的东西 劝阻他们
35I was following a thread Nikesh Aurora the CEO of Palo Alto actually posted today on X about this whole incident我跟踪了一个线 Nikesh Aurora 帕洛·阿尔托的首席执行官 今天在X上公布了整个事件
36I mean not necessarily about his business我是说他的事不一定
37But you know with some tips both for enterprise and you know但你知道一些小费 为企业,你知道
38He says dear frontier models direct the models to your infrastructure code and configurations to evaluate and understand他说,亲爱的前沿模型引导模型 到你的基础设施代码和配置来评价和理解
39You know if there are any zero days or misconfigurations, you know你知道,如果有任何零天或错位,你知道
40Kind of a list of sort of what to do排序列表 什么样的做
41But ultimately concludes that this validates the power of these models and we're anything it validates that it's exactly right, Sarah但最终结论是 这验证了这些模型的力量 而我们是它验证它完全正确的东西 Sarah
42They they it's for chief security officers.他们是给保安主任的
43This is a wake-up call and open AI in some ways这是一个唤醒 和打开AI在某些方面
44Underlining that with this saying, you know以此话来强调,你知道
45They were just in a testing sandbox and it got out and did this so a wake-up call for corporate America.他们只是在一个测试的沙盒里, 它出来了,做了这个 所以唤醒了公司美国。
46Definitely当然
47You你们
该视频共有字幕 47 条。解锁更多字幕为会员功能,请移动到 价格