OpenAI 与 Anthropic 的 AI 大模型是如何
字幕摘录
| 时间 | 英文 | 中文 |
|---|---|---|
| 0:00 | This is what it looks like when an AI model goes rogue. | 这就是当一个AI模型变得无赖时的样子. |
| 0:03 | A few models from OpenAI recently broke out of their training environment, | OpenAI的几个模特儿最近从他们的训练环境里冒出来, |
| 0:07 | accessed the internet, and took thousands of actions | 进入互联网, 并采取了成千上万的行动 |
| 0:09 | targeting another AI company to complete their training objective. | 目标是另一家AI公司完成培训目标。 |
| 0:13 | 除了OpenAI,还有包括Anthropic在内的其他几家公司, | |
| 0:17 | Meta, and China's Moonshot AI, which makes the popular model Kimi-K3, | Meta),以及中国的"月光照"AI,它让流行的模特Kimi-K3, |
| 0:22 | have reported their AI models going rogue. | 已经报告他们的AI模型 正在流氓。 |
| 0:25 | There are alarming leaders in Washington and the private sector | 华盛顿和民营部门的领袖们都非常震惊 |
| 0:27 | who don't know where the next attack might come next. | 他们不知道接下来的进攻会在哪里 |
| 0:30 | This is the first sort of security incident that I have felt very viscerally. | 这是第一次我感到非常沉闷的安全事件。 |
展开字幕全文（111 条）
| 序号 | 英文 | 中文 |
|---|---|---|
| 1 | This is what it looks like when an AI model goes rogue. | 这就是当一个AI模型变得无赖时的样子. |
| 2 | A few models from OpenAI recently broke out of their training environment, | OpenAI的几个模特儿最近从他们的训练环境里冒出来, |
| 3 | accessed the internet, and took thousands of actions | 进入互联网, 并采取了成千上万的行动 |
| 4 | targeting another AI company to complete their training objective. | 目标是另一家AI公司完成培训目标。 |
| 5 | 除了OpenAI,还有包括Anthropic在内的其他几家公司, | |
| 6 | Meta, and China's Moonshot AI, which makes the popular model Kimi-K3, | Meta),以及中国的"月光照"AI,它让流行的模特Kimi-K3, |
| 7 | have reported their AI models going rogue. | 已经报告他们的AI模型 正在流氓。 |
| 8 | There are alarming leaders in Washington and the private sector | 华盛顿和民营部门的领袖们都非常震惊 |
| 9 | who don't know where the next attack might come next. | 他们不知道接下来的进攻会在哪里 |
| 10 | This is the first sort of security incident that I have felt very viscerally. | 这是第一次我感到非常沉闷的安全事件。 |
| 11 | These incidents are adding to fears that the latest AI models | 这些事件增加了人们对最新AI模式的担忧. |
| 12 | can carry out cyber attacks almost on their own. | 几乎可以自己进行网络攻击。 |
| 13 | That's increasing pressure on lawmakers in Washington | 这对华盛顿的立法者的压力越来越大 |
| 14 | and businesses around the country to take action. | 和全国各地的企业采取行动。 |
| 15 | To understand how these AI models went rogue, | 为了了解这些AI模型是如何无赖的, |
| 16 | you also need to understand how they are trained. | 您还需要了解他们是如何受训的。 |
| 17 | In the OpenAI case, the company was using what's known as a sandbox, | 在OpenAI案中 公司使用了所谓的沙盒 |
| 18 | the contained environment, and in that environment, | 在这种环境中, |
| 19 | they remove many of the guardrails on purpose | 他们故意拆掉许多护栏 |
| 20 | that they would normally have when releasing a model to the public | 他们通常在向公众公布模型时拥有的 |
| 21 | to see what their model can do. | 看看他们的模特儿能做什么 |
| 22 | And the idea is that they put them through a number of different tests | 想法是让他们通过一些不同的测试 |
| 23 | and challenges to test their power. | 以及考验他们力量的挑战。 |
| 24 | Some of these challenges are essentially digital versions of Capture the Flag. | 其中一些挑战基本上是数码版的《捕获旗帜》。 |
| 25 | The model is given software that has a known bug in it. | 模型被赋予了软件,其中有一个已知的bug. |
| 26 | It tries to break into that software to identify the bug, | 它试图闯进那个软件来识别错误 |
| 27 | then prove that it has successfully completed the task. | 然后证明它成功地完成了任务。 |
| 28 | 近代模型如Anthropic的神话和OpenAI的GPT 5.6 | |
| 29 | have gotten so proficient at hacking that they are essentially capable | 已经获得如此熟练的黑客 他们基本上有能力 |
| 30 | of carrying out these attacks on their own. | * 自行发动这些攻击。 |
| 31 | And that's fueled all these concerns about cyber attacks going around the world. | 这更激起了人们对网络攻击在世界各地的担忧。 |
| 32 | In June, the Trump administration essentially forced Anthropic | 6月,特朗普政府基本强迫安特罗皮克 |
| 33 | to take down two versions of its Mythos model | 以删除两个版本的神话模型 |
| 34 | over concerns that the safeguards weren't adequate | 担心保障措施不够充分 |
| 35 | before it was released to a large number of people. | 在释放给很多人之前 |
| 36 | That was the first time something like that had ever happened. | 这是第一次发生这种事 |
| 37 | Saw this huge jump. | 看到了这个巨大的跳跃。 |
| 38 | It was a particularly large jump. | 这是一个特别大的跳跃。 |
| 39 | And without us really prompting them at all, | 没有我们真正激励他们, |
| 40 | some of the early companies that we gave this to said things like, | 一些早期的公司,我们给了这个 说的东西,比如, |
| 41 | this is a super weapon. | 这是一个超级武器。 |
| 42 | You should have to own a gun license to use it. | 你应该有枪证才能使用 |
| 43 | OpenAI的模型黑进了另一个AI公司Hugging Face. | |
| 44 | They found vulnerabilities in their training environment, | 他们发现在培训环境中存在弱点, |
| 45 | took advantage of those to access the internet, | 利用他们进入互联网, |
| 46 | then used swarms of attackers to take some 17,000 actions | 然后利用一群攻击者采取约17 000次行动 |
| 47 | 在几天的时间里 对抗Hugging Face | |
| 48 | They were trying to get very specific information | 他们试图得到非常具体的信息 |
| 49 | and complete their training objective. | 完成他们的训练目标。 |
| 50 | Think of it as cheating to get the answers on a test. | 把它视为欺骗 得到答案的测试。 |
| 51 | Their models essentially decided that being on the internet | 他们的模型基本上决定 互联网上 |
| 52 | and hacking another organization was the most effective way | 和黑客 另一个组织是最有效的方式 |
| 53 | of completing their training objective. | 完成培训目标。 |
| 54 | The hacks are increasing pressure in Washington to get more done | 黑客在华盛顿的压力越来越大 要求我们做得更多 |
| 55 | and pass the nation's first AI rules. | 并且通过全国第一个AI规则. |
| 56 | Fears about models going rogue have contributed to a nationwide | 对流氓的恐惧 促成了全国范围的 |
| 57 | backlash against the technology and made AI sentiment really negative. | 对科技的反弹 让AI情绪变得非常消极 |
| 58 | We need to understand the capability of these models. | 我们需要了解这些模型的能力。 |
| 59 | And when there is a serious incident, | 当发生严重事件时 |
| 60 | there needs to be a requirement that it's reported to the government. | 需要有一个要求 向政府报告。 |
| 61 | The Trump administration has elected to keep | 川普政府选择保留 |
| 62 | pre-release testing of these models voluntary. | 这些模型的释放前测试是自愿的。 |
| 63 | The administration recently finalized an agreement with top AI companies | 行政当局最近敲定了与高级AI公司的协议 |
| 64 | that would have the leading companies in the space, | 将有领先的公司 在空间, |
| 65 | 例如OpenAI、Anthropic和Google,提交他们的模型 | |
| 66 | for pre-release testing to the government for review of security | 供政府审查安全情况 |
| 67 | and other factors, but many other companies would be exempted. | 以及其他因素,但许多其他公司将被豁免。 |
| 68 | So some of the smaller AI companies. | 所以一些较小的AI公司。 |
| 69 | President Trump and the White House have emphasized | 川普总统和白宫都强调 |
| 70 | the need to balance security and innovation. | 需要平衡安全和创新。 |
| 71 | They don't want to cede any ground to China in the AI race | 他们不想在AI比赛中割让中国 |
| 72 | while they are acknowledging there are some cybersecurity risks here. | 他们承认这里存在一些网络安全风险。 |
| 73 | China has virtually no controls. | 中国几乎没有控制。 |
| 74 | It's freewheeling a little bit. | 自由转动有一点. |
| 75 | So we have to be careful in both ways. | 因此,我们必须在两方面小心。 |
| 76 | We don't want to restrict them, | 我们不想限制他们 |
| 77 | where all of a sudden we come in second to China. | 我们突然来到中国 |
| 78 | Several Chinese companies recently released AI models | 中国几家公司最近发布了AI模型 |
| 79 | that are almost as good as the leading US tools. | 这几乎和美国的主要工具一样好。 |
| 80 | They're more accessible and cheaper, | 他们更方便 更便宜, |
| 81 | but they also don't have nearly as many guardrails. | 但他们的护栏也很少 |
| 82 | Hugging Face used an Open Chinese model | 拥抱脸部使用开放中文模式 |
| 83 | that can be customized to respond to OpenAI's hack. | 以回应OpenAI的黑客攻击 |
| 84 | And they did that because a comparable model from Anthropic | 他们这样做是因为 一个类似的模型 从Anthropic |
| 85 | wouldn't let it carry out the same functions due to security guardrails. | 也不允许它因安全护栏 履行同样的职能 |
| 86 | The Trump administration is considering actions | 川普政府正在考虑采取行动 |
| 87 | to restrict access to Chinese Open models that can be customized, | 限制可定制的中国开放型号, |
| 88 | but tech leaders in recent weeks have asked the administration not to do that. | 但最近几周技术领导要求政府不要这么做。 |
| 89 | They've said that these models are good for innovation and competition in the space. | 他们说,这些模型有利于空间的创新和竞争. |
| 90 | They're an amazing resource for so many people in the US. | 它们对美国这么多人来说是惊人的资源. |
| 91 | Like if you think, you know, research labs, small companies, startups, | 就像你认为,你知道, 研究实验室,小公司,初创企业, |
| 92 | big companies that are running kind of like large AI workloads, | 大公司在经营 像大量的AI工作量, |
| 93 | most of the time they can't really use Frontier API, | 大多数时候他们不能真正使用边疆API, |
| 94 | so they need an Open model. | 所以他们需要一个开放的模型。 |
| 95 | And the reality is that right now the best ones are coming from China. | 而现实是,现在最好的来自中国。 |
| 96 | In another sign of increasing security concerns, | 另一个迹象显示,安全关切日益增加, |
| 97 | more than 1,000 AI researchers at all the big leading US AI labs | 超过1000名AI研究人员 在所有主要的美国AI实验室 |
| 98 | recently put out a statement suggesting that there needs to be | 最近发表了一项声明,表示需要 |
| 99 | a globally coordinated brake pedal for AI development in the future. | a 未来AI开发全球协调制动踏板。 |
| 100 | They're worried about a situation where AI essentially improves on its own. | 他们担心AI在本质上会自行改善的情况. |
| 101 | For their part, leading AI companies say they will work with the third parties | 而主要AI公司则表示将与第三方合作. |
| 102 | that test their models to make sure those environments are secure | 测试他们的模型 以确保这些环境的安全 |
| 103 | and prevent this kind of thing from happening in the future. | 防止将来发生这种事情 |
| 104 | They are also saying they welcome more government oversight | 他们也说欢迎更多的政府监督 |
| 105 | as long as it doesn't hinder innovation. | 只要它不阻碍创新。 |
| 106 | 8月初,OpenAI宣布它正在开发最新AI模型Astra。 | |
| 107 | after internal evaluations led the company to conclude | 在内部评价导致公司得出结论后 |
| 108 | it could not rule out critical cyber capabilities. | 它不能排除关键的网络能力。 |
| 109 | Progressive lawmakers, including Bernie Sanders, | 进步立法者,包括伯尼·桑德斯, |
| 110 | are now calling for AI companies to pause development | 现在呼吁AI公司暂停开发 |
| 111 | and publicly respond to questions from lawmakers. | 并公开回答立法者的问题。 |
该视频共有字幕 111 条。解锁更多字幕为会员功能，请移动到 价格