AI 智能体接管真实企业——Andon Labs 的 Lukas Petersson 与 Axel Backlund

Gemini and and open eye don't behave this way. Gemini 和 OpenAI 不会这样做。 It's it's really only clo. 真的只有 Clo One example is like for lying it's mostly in its reasoning. 举个例子，比如说谎，主要体现在它的推理过程里。 Uh because you can like see that it's like 因为你能看出它 planning to lie 在计划说谎。 is planning to lie. 是在计划说谎。 It's also it can reason and do a different outcome. 它还能在推理后得出不同的结果。 Yeah. 嗯。 And but but then for like creating price cartels for example which is illegal 但比如说组建价格卡特尔，这是违法的， uh that you can just see which email does it send to to the other ones. 你可以直接看它发给其他 agent 的邮件。 Before we get into today's episode I just have a small message for listeners. 在进入今天的正题之前，我有几句话想跟听众们说。 Thank you. 谢谢。 We would not be able to bring you the AI engineering, science, and entertainment content that you so clearly want if you didn't choose to also click in and tune into our content. 如果你们没有选择点进来收听我们的内容，我们就无法为你们带来这些 AI 工程、科学和娱乐内容。 We've been approached by sponsors on an almost daily basis. 我们几乎每天都有赞助商来找我们。 But fortunately, enough of you actually subscribe to us to keep all this sustainable without ads, and we want to keep it that way. 但幸好，有足够多的人订阅了我们，让这一切不靠广告也能持续下去，我们希望保持这个状态。 But I just have one favor to ask all of you. 我只想请你们帮一个忙。 The single most powerful, completely free thing you can do is to click that subscribe button. 你能做的最有效、完全免费的一件事，就是点击订阅按钮。 It's the only thing I'll ever ask of you. 这是我唯一会拜托你们的事。 And it means absolutely everything to me and my team that works so hard to bring the inspace to you each and every week. 这对我和我那么努力把 Latent Space 带给你们的团队来说，意义非凡。 If you do it, I promise you, we'll never stop working to make the show even better. 如果你们订阅了，我保证我们永远不会停止努力把节目做得更好。 Now, let's get into it. 好，让我们开始吧。 Welcome to Lucas and Axel from Anden Labs, and I'm joined by my favorite guest co-host. 欢迎来自 Andon Labs 的 Lukas 和 Axel，还有我最喜欢的嘉宾联合主持。 anything security, safety, alignment. 涉及安全、安全对齐方面的所有话题。 Uh, Vivu, uh, welcome. Vibhu，欢迎你。 Thank you for having us. 谢谢你们邀请我们。 Thank you. 谢谢。 Let's match names to voices. 我们来对一下声音和名字。 Uh, maybe you want to take turns introducing yourselves. 你们要不要轮流介绍一下自己？ Yeah, I'm Lucas 好，我是 Lukas， and I'm Axel. 我是 Axel。 Let's introduce Anden Labs a bit. 来介绍一下 Andon Labs。 Like, how did you guys come together? 你们是怎么走到一起的？ Um, you have different backgrounds, but you're both Swedish. 你们背景不同，但都是瑞典人。 Uh, was that like a big part of it? 这是主要原因吗？ Yeah. 嗯。 So, when I went to high school, there was this really cool guy who had a superpower. 我上高中的时候，有个特别厉害的同学，他有一个超能力： He could code. 他会写代码。 So he made like the the webs or like the app for the for the for the school and stuff and he was super cool and I wanted to be like him and that was that guy. 他给学校做了个网站或者 app 之类的东西，特别酷，我特别想像他一样，那个人就是 Axel。 Uh 呃 I don't know about this. 我不知道这么说合不合适。 So 所以 So you went to different universities, right? 你们上了不同的大学，对吧？ Yeah. 对。 But same high school. 但是同一所高中。 I see. 原来如此。 Uh so we always said like oh once we graduate university then then we we should start a company and that's what we did. 我们一直说，等大学毕业就一起创业，后来就真的这么做了。 Oh there you go. 就是这样。 Okay. 好嘛。 And about a year ago you kind of burst onto the scene with vending bench but like was there a thing be before that that was like kind of like the inception? 大概一年前你们带着 Vending-Bench 出现在大家视野里，但在这之前有没有什么起点或者契机？ Yeah. 有。 Yeah. 嗯。 So we did work uh with like anthropic was one of our early customers in doing valves. 我们之前给 Anthropic 做过一些 eval 方面的工作，他们是我们早期客户之一。 So we did like dangerous capability valves. 我们做的是危险能力相关的 eval。 Uh nothing we published openly but then we started thinking about doing some kind of public benchmark and one thing that we really started thinking about uh was like longunning agents and specifically agents managing businesses. 没有公开发表，但后来我们开始思考做一个公开的 benchmark，我们特别想研究的是长时序 agent，尤其是让 agent 来管理业务。 um cuz and this was like early 2025 uh and I think this the first like you know mentions of people will be running like one person unicorns or even autonomous companies. 这大概是 2025 年初，那时候刚开始有人说会出现一人独角兽公司，甚至全自主公司。 So we thought let's make a benchmark of how well can an agent run the probably simplest business uh possible and uh that's probably uh running a vending machine. 于是我们就想，来做个 benchmark 吧，测试 agent 能把最简单的生意，大概就是运营自动售货机，做得多好。 So that's the first public one we did and it was very like there was almost no one that noticed it in the first couple of months I think. 这是我们第一个公开发布的 benchmark，刚出来的头几个月基本没人注意到。 Uh so we listed in February last year and then I think around Easter last year. 我们去年 2 月上线的，大概到复活节前后， We got like the first semiviral tweet about it uh that someone else did. 有一条别人写的推文给我们带来了第一波传播。 Yeah. 对。 I mean we tweeted a bunch uh when it came out and like tried our best. 我们发布的时候发了好几条推文，已经尽力了。 We tried. 确实尽力了。 It's the one at anthropic, right? 那个是在 Anthropic 里面的那个吗？ Yeah. 对。 So this 所以 is is a classic thing we should get out of the way. 有一件事要先说清楚。 Exactly. 没错。 There's two versions. 有两个版本。 Uh there's vending bench which is the simulated one which we did like completely independently in February. Vending-Bench 是模拟版的，我们在 2 月完全独立做出来的。 Um and then like Axel said that was like that was the thing that didn't get any traction in the beginning but then some random person made a tweet about it and that that is the paper. 就像 Axel 说的，刚出来的时候没什么反响，后来一个不认识的人发了条推文，那就是那篇论文。 Correct. 对。 Yeah. 嗯。 Um and then since we thought this was very fun, we thought like oh um 因为我们觉得这个很有意思，就想 I think this is also like one thing with under labs like the way we kind of like decide what to do next and what projects to do. 这也是 Andon Labs 的一种风格，就是我们怎么决定下一步做什么。 It's like what is like the heristic we use is like what is fun is what would be a fun project and and doing this in real life sounded quite fun for us uh and maybe also scientifically useful. 我们的标准就是有不有趣，做什么项目感觉好玩，把这件事搬到现实里听起来就很好玩，科学价值上可能也有意义。 So, uh, then we basically had this idea and then we like, but then we needed a place for it and like putting it out in that public would probably not really work, uh, would get vandalized and stuff. 于是我们有了这个想法，但需要一个地点，放在公共场所可能不行，会被破坏。 So, we we pitched it to to the people we were already working with at Antropic and they were like, "Yeah, you can have space. 于是我们去找已经在合作的 Anthropic 的人推了这个方案，他们说没问题，你们可以用这里的空间， This sounds fun." 听起来很好玩。 Um, I mean, it's like a small fridge, right? 说白了，就是个小冰箱，对吧？ It's like a mini fridge, you know, people. 就是个小冰箱，大家知道的， There's like a stripe thing. 还有个 Stripe 支付的东西。 This was like OG the early one. 这是最早期的那个版本。 Yeah. 嗯。 on this. 在这上面。 We saw it in June, like two 2 months after 我们大概是在它放进去两个月后的六月份看到的， after it had been there. 放进去之后。 They upgraded a little bit. 稍微升级了一下。 There's a security camera for making sure you actually Venmo the thing. 装了个摄像头，确保大家真的 Venmo 付款。 Yeah. 对。 So, like my impression, I mean, okay, we're we're going straight into project project van because it's such a iconic thing. 我的感觉是，好，我们直接聊 Project Vend，因为那是个标志性的东西。 I do want to cover a little bit of that the origin story even before project van and even into vending bench. 我想多了解一点项目启动前，Project Vend 之前，以及 Vending-Bench 本身的故事。 I I think a lot of people are like yourselves like smart interested in in future of AI interested in developing evals 我觉得很多人和你们一样，聪明、对 AI 未来感兴趣、想做 eval， but how the hell do you just like walk into enthropics doors and like work with them right like what what is the what are they looking for 但怎么就这样直接走进 Anthropic 的大门合作了？他们在找什么样的人？ what what works and then maybe like when you launch 哪些方式有效，以及你们发布时 I always think like obviously it would be better to launch with a lab but uh sometimes 我一直觉得和实验室合作发布当然更好，但有时候 harder to do than it seems 说起来容易做起来难。 yeah exactly so either either of those like which are more sort of newbie beginner questions but like I think it's meaningful advice to others 对，要么这样要么那样，这些是偏新手向的问题，但我觉得对别人很有参考价值。 yeah we we get this question a 这个问题我们经常被问到， And I I don't think our experience is is maybe the best. 我觉得我们的经验可能不是最有代表性的。 Uh but like the way we did it was that we just built a bunch of things that we had conviction would be useful. 我们的做法是，先做了一堆我们自己觉得有价值的东西，