Cognition 的 Scott Wu 谈 AI 编程代理与富足时代

Cognition 已融资超过 10 亿美元。首位 AI 软件工程师 Devon。微软和 Cognition。Cook Mission 和梅赛德斯-奔驰、高盛、梅赛德斯-奔驰、花旗、戴尔、Sonandonder、NASA、美国海军和美国陆军。我们迄今已融资超过 25 亿美元，最新估值 260 亿美元。不到 3 年，你做到了 5 亿美元营收。太疯狂了。不到 3 年做到 5 亿美元营收。

Cognition has raised over a billion dollars. First AI software engineer Devon. Microsoft and Cognition. Cook Mission and Mercedes-Benz, Goldman Sachs, Mercedes-Benz, City, Dell, Sonandonder, NASA, the US Navy, and the US Army. We've raised over 2.5 billion to date at a last valuation of 26 billion. In less than 3 years, you got to 500 million in revenue. That's crazy. Half a billion in revenue in less than 3 years.

我们的客户在过去 6 个月左右，使用量也增长了大约 11 到 12 倍。一旦你解决了 AGI，你就接管整个世界，然后只有一个实体统治世界。我只是觉得那不是我们正确的未来。人类能做的，AI 都能做。也许你可以有一个 AI 工程师伙伴，帮你干活。但那正是我们第一次相信这是可能的时刻。我认为富足时代是真实的，而且很快就会到来。

Our customers in the last 6 months or so have also grown their usage something in the range of like 11 to 12x. Once you've solved AGI, you just take over the entire world and then there's just one entity that rules the world. I just don't think that's the right future for us. There is nothing that humans do that AIs can't do. Maybe you can just have like an AI engineer buddy that can just do work for you. But that was like the very first moment that we believed it was possible. I think the abundance era is real and I think we'll be soon.

Devin，欢迎来到 Sorcery。

Devin, welcome to Sorcery.

嗯，很高兴来到这里。

Yeah, great to be here.

你经常遇到这种情况吗？

Do you get that a lot?

嗯，是的。是的。比如我在大楼里坐电梯，下来的时候人们就会问：“你是 Devon 吗？”然后，你知道，其实我不是 Devon，但我明白你在问我什么。

Um, yeah. Yeah. Like I'll be in the elevator in my building and I'll just come down and people be like, "Are you Devon?" And then, you know, like, well, I'm actually not Devon, but yes, I understand what you're asking me.

哦，天哪，真是的。嗯，好吧，Scott，欢迎来到 Sorcery。我们在巴黎，这很酷。我们来参加 Ray Summit。这是个大型 AI 峰会。我想这是他们的第二或第三届。你会上台演讲。

Oh gosh, damn. Um, well, Scott, welcome to Sorcery. We are in Paris, which is pretty cool. We're here for the Ray Summit. It's like this huge AI summit. I think this is like their second or their third year. You're going to be on stage.

是的。是的，我明天上台。我一小时前刚落地，直接赶过来了。

Yeah. Yeah, I'll be on stage tomorrow. I just landed an hour ago. Came straight here.

嗯，你能来太好了。我很期待聊聊当下自主编码智能体的 Cognition 状态。这很紧迫。你知道吗？

Well, it's awesome to have you on. I'm excited to talk through the state of cognition in autonomous coding agents today. It's very pressing. Did you know that?

我确实知道。是的。我的意思是，我现在甚至都跟不上它了。所以，不，事情发展得太疯狂了。我的意思是，即使对我们内部来说，最疯狂的事情之一就是我们大量使用 Devon 来构建 Devon。事实上，我们现在几乎完全用 Devon 来构建 Devon。大约 95% 的代码是由 Devon 写的。但疯狂的是，即使在过去的六个月里，你知道，如果你问人们，哦，AI 编程，是的，已经进行了三四年了，这是真的，确实如此，但即使在过去的六个月里，我想我们的数字是，我们交付的代码总量大约增长了 7 倍。所以就像我们交付的 PR 数量、代码量。显然还有字面输出，你知道，在如何实际衡量你从中获得的生产力方面，有不同的方法。但我认为在实践中，我们能够做更多的事情，获得更多的杠杆，即使在这段时间里也是如此。模型一直在变得更好。我想，在产品体验方面，已经有越来越多的东西被充实起来。所以，是的，不，事情在飞速发展。太疯狂了。

I did actually. Yeah. I mean, I can't even keep up with it at this point. So, no, things have been going like crazy. I mean, even for us internally, like one of the craziest things is obviously we use a ton of Devon to build Devon. In fact, we almost exclusively use Devon to build Devon at this point. Like 95% of the code or something is written by Devon. But the crazy thing is even in the last six months, you know, if you ask people, oh, like AI coding, yeah, it's been going on for, you know, three, four years now, which is true, it has been, but even in the last six months, I think the number for us was like our total amount of code shipped is roughly 7xed in that time. And so it's like the number of PRs we've shipped, the amount of code. And obviously literal output, you know, like there's different ways in terms of how you actually measure the productivity that you get out of it. But I think in practice like we're just able to do so much more and get so much more leverage in even this period of time. Like the models just keep getting better. I think there's more and more that's been kind of fleshed out, I would say, with the product experience. And so yeah, no, things are flying. It's crazy.

那么，你如何衡量它的生产力？

So, how do you measure the productivity of it?

是的，这是个好问题。关于这个，我想说几点。首先，我认为显然字面上的 token 并不是正确的答案。我的意思是，很多人讨论过这个，你知道，我认为 token，你知道，我有点说 token 最大化时代，你知道，大约从 2026 年 1 月持续到 5 月。

Yeah, it's a great question. So, a couple things I'd say on this. First of all, I think obviously literal tokens is just not the right answer really. I mean, lots of people talked about that, you know, I think the token, you know, I kind of say the token max era, you know, lasted from January to May of 2026, roughly.

安息吧。

RIP.

嗯，是的，安息吧，你知道，那是一段有趣的时光。但是不，我的意思是，我认为我们现在已经到了这个点，显然你关心的是结果和你实际交付的东西。所以我们看到我们交付了多少，我们看到我们做了多少。显然我们关心的是，你知道，我们通过产品达到了什么？我们实际向客户和用户交付了什么？我认为这对每个企业来说大致相同。在衡量实际 ROI、生产力等方面，没有什么秘密技巧，也没有捷径，但很大程度上归结为实际关注你推动的每一个结果。所以对我们来说，你知道，显然我们有所有业务 KPI，我们跟踪并遵循它们。你知道，当我们考虑我们的产出时，我们考虑很多的是，你知道，我们增长这些数字有多快？而不是你实际看的任何 token 或花费之类的东西。

Um, yeah, RIP, you know, it was a fun time. But no, I mean I think we're now at this point where obviously you know what you care about is outcomes and like what you're actually delivering. And so you know we see how much we're shipping. We're seeing how much we're doing. Obviously what do we care about is like you know what are we getting to with the product? Like what are we actually delivering to our customers and our users and I think that's roughly the same for every business. There's no kind of like secret trick or there's no kind of like easy way out in terms of measuring actual ROI, productivity and so on, but a lot of it comes down to just actually looking to each of the outcomes that you're driving. And so for us it's you know obviously we have all the KPIs of the business that we track and that we kind of follow. And you know, when we think about our outcome output, like a lot of what we think about is obviously it's like, you know, how quickly are we growing those numbers? As opposed to anything that we actually, you look at in terms of tokens or or spend or something like that.

你有没有想过你会增长这么快？

Did you ever think you'd be growing this fast?

这是个好问题。关于这个，我一直在想的是，一方面，我认为相对于企业，顺便说一句，不只是我们。就像我们的客户也看到了这一点，我们的客户在过去 6 个月左右，对 Devon 的使用量也增长了大约 11 到 12 倍。我想说的是，是的，一方面，你只是想想过去的企业，想想历史上所有科技公司的增长，显然这很疯狂。另一方面，你知道，当我问这个问题时，比如第一性原理的问题：好吧，现在有多少人在使用编码智能体？他们用它做什么？他们从中得到了多少？而我们认为他们会使用什么，以及他们在三年后会使用多少？我认为粗略的答案，特别是如果你看，你知道，如果我们稍微回拨一下，比如一年前，粗略的答案大约是零。

It's a good question. So, what I always think about with this is I think on the one hand, I think relative to businesses and by the way, it's not just us. It's like our customers have seen this where similarly our customers in the last six months or so have also grown their usage in something in the range of like 11 to 12x of Devon. And the way I'd say it is yeah on the one hand I think you just like think about businesses of the past you think about the growth of all these tech companies in history and obviously it's pretty crazy. On the other hand you know when I kind of asked the question like the first principles of question: Okay, how many people are using coding agents now? How much are they using it for? How much are they getting out of it? Versus like what do we think they're going to be using and how much of it are they going to be using in, you know, three years or something. I think the rough answer, especially if you look, you know, if we rewind a bit like a year ago, the rough answer a year ago was roughly zero.

粗略的回答是，两三年后，基本上世界上每个软件工程师，而且很可能远不止软件工程师，因为我认为会有更多人能够编写代码、开发产品等等。所以当你把曲线上的这两个点连起来，然后外推必须发生什么时，我认为必须出现大规模采用和大规模增长。我认为这在实践中往往表现为：当然，有病毒式传播本身，有很多不同的人发布大量优质内容来教育人们、展示如何有效使用这些工具，然后还有所有这些公司必须发生的真正的组织变革。

And the rough answer for you know two three years from now is roughly every software engineer in the world and probably a lot more than all just the software engineers because I think a lot more people will be able to produce code and produce products and so on right. And so then when you take those two points on the curve and you just extrapolate out what has to happen, I think there has to be massive adoption, massive growth. I think what that looks like in practice often is like, of course, there's the virality itself, there's a ton of great content that I think a lot of different folks are putting out in terms of educating people and showing people how to use the tools effectively, and then there's real organizational change that has to happen in all these companies to make that happen.

太疯狂了。在所有类别中，我很久以前采访过 Mayfield 的 Naveen，他列出了增长最快的三个类别。我知道这很明显，当你身处其中、整个市场都在随之而动时，很容易忘记这些事情。

It's insane. Like, of all the categories, there was an interview I did way back when with Naveen from Mayfield, and he laid out the three fastest growing categories. And I know it's really obvious, it's quite easy to kind of forget about these things when you're in it, like when the whole market is moving with it.

但令人疯狂的是，这个新类别在不到 3 年内就达到了 5 亿美元营收。太疯狂了。不到 3 年就做到 5 亿美元营收。

But it's insane how fast this new category in less than 3 years you got to 500 million in revenue. That's crazy. Half a billion in revenue in less than 3 years.

太疯狂了。

That's insane.

是的。我知道，我是说，我们非常幸运，而且我认为，显然，AI 整体在过去几年里发展得极其迅猛，但我们的想法是，我们必须以这样的速度前进。坦率地说，如果我们只是每年增长 3 倍之类的，我们就会在采用速度上落后于市场。我认为这更多地反映了这项技术的意义有多大。就像每个企业或一般产品开发者，构建产品、销售产品最难的部分就是向人们解释他们为什么应该关心。比如，嘿，我有这个薪资优化工具，我有这个基于使用量的计费工具，我有这个那个，但这就是为什么它真正有意义，这就是为什么你应该考虑它。而 AI 能自己写代码、为你构建自动驾驶产品，好处在于，我认为人们为什么应该关心这一点是显而易见的。所以很大程度上只是解决如何把技术落地并交付给人们的所有实际问题。

Yeah. I know it's, I mean, I think we're very fortunate, and I think that the, obviously, AI as a whole has moved like crazy in the last few years, but the way that we think about it is like we have to move this fast. If we frankly only 3xed year-over-year or whatever it is, we'd be lagging the market in terms of how fast all this adoption is happening. And I think it's more a reflection of just how meaningful the technology is. It's like everybody in enterprise or just building products in general, the hardest part about building a product, selling a product, all that is just explaining to people why they should care. You know, it's like, hey, I've got this payroll optimization thing. I've got this usage based billing thing. I've got this whatever, but here's why this is really meaningful and here's why you should be thinking about this. The nice thing about this with AI that can write code by itself and build self-driving products for you is, I think it's kind of obvious why you should care. And so a lot of it is just figuring out how you solve all the practical problems of getting the technology there and then getting it out to people.

你们已经达到 5 亿美元营收，而且你们的客户包括——我来列举一下，因为到目前为止这些都是相当大的客户：高盛、梅赛德斯-奔驰、花旗、戴尔、Sander、NASA、美国海军和美国陆军。你们迄今已融资超过 25 亿美元，最新估值 260 亿美元。我想回到过去，因为 Cognition 在一家名为 Windsurf 的公司的救赎故事中扮演了重要角色。你能带我们回到那个时刻，并谈谈并购在增长中扮演的角色吗？

You're at half a billion in revenue and you have customers that entail—I'm going to list them out because these are some pretty big customers so far. Goldman Sachs, Mercedes-Benz, City, Dell, Sander, NASA, the US Navy, and the US Army. And you've raised over 2.5 billion to date at a last valuation of 26 billion. I want to go back in time because Cognition played a pretty big role in a redemption story of a company called Windsurf. So can you bring us back to that moment and how M&A has played a role in the growth?

是的，那是一段有趣的时光。其实也就是 11 个月前。而且，顺便说一句，这很有话题性，因为我觉得最近 Cursor 也有类似的消息。我们当时和其他人差不多同时听到这个消息：这家叫 Windsurf 的公司，正在打造一个非常棒的 IDE，很多个人和公司都在用，团队中有一部分人要加入 Google。所以这几乎算是一种“收购式招聘”，团队中的一些研究员要去 Google。而背后还有一家公司，仍然拥有所有客户，仍然有一个准备就绪的市场团队，还有大量的平台工程、产品本身等等所有细节。但显然，它正在努力决定该怎么办。这个消息是那个周五下午宣布的。我们在 Cognition 内部讨论这件事，一开始还开玩笑，觉得挺有意思，是不是——然后当我们真正认真思考时，我们觉得，这对我们来说真的很有趣，因为如果你想想我们的处境，我们一直专注于研究、产品本身，真正在智能体产品上取得进展。我认为随着时间的推移，越来越清楚的是，真正优秀的编码 IDE 产品——你坐在里面查看代码本身的工具——与智能体产品——那种几乎可以委派给它的工程师——你显然希望它们紧密结合，并且在理解你的确切代码库方面共享大量相同的信息。你希望相同的知识能贯穿你如何进行工作和管理会话的许多方面，这很常见。你自己查看代码，然后发现你想去做某件事，于是你把它交给一个远程智能体。所以产品上有大量协同效应。同样在市场推广方面，Windsurf 拥有出色的企业客户吸引力，已经有一个团队，在向这些客户销售和实际交付方面拥有丰富经验。所以这是一个非常出色的部署工程体系，基本上你需要的所有要素都有。于是我们当晚就联系了 Windsurf 团队。那是周五晚上，然后基本上在那个周末我们就把整个收购谈妥了。所以到周一早上，我们已经准备好发布公告。有趣的公告——我和 Jeff，说实话我们整个团队，那个周末都没怎么睡。但我们把所有事情都整合好了，是的，我们能够做到这一点真是太棒了。我认为这对我们公司来说很有意义。我认为那显然也是一个非常优秀的团队，只是需要弄清楚该怎么做。而且我认为在过去一年里，我记得我说过，对我们来说这就像花生酱和巧克力。真的很有趣。

Yeah, it's a fun time. I mean, it was all of 11 months ago. But we—and by the way, it's quite topical because I think recently there's similar news with Cursor. We at the time kind of heard this news roughly around the same time as everyone else, that this company Windsurf, which is building a really great IDE that a lot of folks used—individuals, companies, and so on—some of the team was going to Google. So there was kind of like an acqui-hire deal almost, where a number of the researchers on that team were going to Google. And there was a company kind of behind that that still had all the customers, still had a ready-to-go go-to-market team, a lot of the platform engineering, the product itself, all the details there. But obviously it was trying to figure out what to do. And that was announced that Friday afternoon. And we at Cognition were talking about that internally, kind of joking about it at first, like, interesting, is there—and then as we actually thought about it seriously more, we were like, actually this is really interesting for us, because if you think about where we're at, we've always been focused on, obviously, the research, the product itself, really making progress with the agent product. I think as time goes on, it's become more and more clear that the really good coding IDE products—the tool that you're sitting in and using to look at the code itself—versus the agent product, which is kind of that almost the engineer that you can delegate—you obviously want those to be very hand-in-hand, and a lot of the same information on understanding your exact codebase. You want the same knowledge to persist across a lot of how you even go and do your work and manage sessions is very common. You're looking at the code yourself, and then you figure out there's something you want to go do, and so you hand that off to a remote agent. So ton of synergies in the product. Similarly in terms of go-to-market, Windsurf had amazing enterprise traction, already had a team that had a lot of expertise in going and selling to these folks and figuring out how to actually deliver. So a really amazing deployed engineering motion, all the pieces that you would need basically. And so we got in touch with the Windsurf team that evening. So it was Friday evening, and then basically over the course of that weekend we worked out the entire acquisition together. So by Monday morning we had the announcement ready to go. Fun announcement—both myself and Jeff and honestly our entire team, we all got very little sleep that weekend. But we put everything together, and yeah, it's great that we were able to do that. I mean, I think it made a lot of sense for us as a company. I think it was also obviously a really great team that was just kind of figuring out what to do. And I think over the last year, I remember saying it's like peanut butter and chocolate honestly for us. It's been a lot of fun.

把团队合并到一起是什么感觉？

What was it like merging the teams together?

是的，那确实是个真正的过程。顺便说一句，这个过程中棘手的一点是，当时 Cognition 总共只有 35 到 40 人。我们有一项有意义的业务，有真实的增长势头。当时我们的营收运行率在 7000 万到 8000 万美元左右。但那是一个非常小的团队。另一方面，Windsurf 因为已经扩展了市场推广并建立了相关体系，当时大约有 200 人。所以显然有很多细节需要理清。在最初的几个月里，我的意思是，就在合并当天或那一周，我们所有人都只是想着：好，我们先把船稳住。确保大家知道，客户知道，我们会为他们服务，我们会持续发布产品，我们会把产品两边都照顾好，并且做好。然后就往前冲。在最初的几个月里，显然很多工作只是弄清楚我们如何进入一个良好的状态。然后，就像所有事情一样，花些时间去构建这些东西、做这些事情会容易得多。所以我们有很多有趣的小型团建活动，我觉得这有助于大家互相了解。显然还有大量的战略会议，把各种细节都敲定下来。然后很快，在去年年底，我们到了一个需要找更大的旧金山办公室的阶段，显然要尽快，因为从战术上讲，我们需要一个能容纳所有人的办公室。所以到去年年底，我们找到了那个办公室，然后能把所有人放在一个空间里，一起工作。我想说，这些就是让一切变得超级顺利的重要里程碑。

Yeah, that was a real process. And by the way, one of the things that was tricky about that process is Cognition at the time was all of 35 or 40 people. And we had a meaningful business with real traction. We were doing in the range of 70 or 80 million of revenue run rate at the time. But it was a very small team. Windsurf, on the other hand, because it had scaled out go-to-market and set all that up, was around 200 people at the time. So a lot of details to figure out obviously. And so for a few months, I mean, the immediate day of or week of, all of us were just like, okay, let's just right the ship. Let's make sure folks know, customers know, that we're going to be there for them, we're going to be shipping, we're going to be taking care of both sides of the product and doing that well. And let's just go. Over the first few months, obviously, a lot of it was just figuring out how we get into a good state. And then, as with all things, it's much easier to go build these things and do all this with some time. So we had a lot of little fun offsites, which helped, I think, for people getting to know each other. Ton of strategy sessions obviously, working out all the details and so on. And then pretty soon, at end of last year, we got to a point where we were looking for a bigger SF office space obviously as soon as possible, just as a tactical matter, because we needed an office that could fit everybody. And so by the end of last year, we got that office space and then were able to put everyone in one space and all work together. And those were, I would say, the big milestones that made things super smooth.

你们是如何推进产品整合的？

How did you work through the product integration?

当然，这有很多步骤。我会这么说：不要过度强求。所以在当时，就像是，看，有一个 Windsurf 产品。两个产品都很棒。我们两个都做，在每个产品上做明显的下一步。然后当我们有重叠的部分时，就自然地一起构建。所以并不是一个月后 Windsurf 就消失了，就这样。在过去的一年里，我们一直在构建所有这些，现在 Devin Cloud 和 Devin Desktop 已经非常完美地配对在一起。你可以从一个开始会话，然后交给另一个。你可以在同一个系统里使用它们，等等。但这更像是一个渐进的过程。我觉得对我们来说有趣的一点是，很多使用任一产品的用户本来就在自然地寻找另一面。比如我们有很多人使用 IDE 框架，显然在我们现在讨论的这个时间点，也就是 2025 年底、2026 年初，正是异步智能体开始真正火起来的时候。所以这些人都在想，好吧，我的策略应该是什么？显然，能和一个单一的团队合作，而且这个团队已经理解你的代码库，把所有细节都搞清楚并部署好了，这很棒。所以这很好。反过来也一样，对于很多采用云智能体的用户来说，能说“好的，本地端也在那里，供那些仍然想使用本地形态的开发者使用”，这非常好。所以我觉得简单的说法是，与其强行推动，说“好，在接下来的 30 天或 60 天内我们必须整合这些产品”，不如更多地让实际使用情况和我们正在构建的新功能、我们正在进入的下一个范式，让这些来逐渐地把产品拉在一起，而不是强行整合。

It was a lot of steps for sure. One way I'd put it is, not trying to overly force it. So for the time being, it was like, look, there's a Windsurf product. Both those are great. Let's work on both of those and do the obvious next steps on each of these things. And then as we had pieces of overlap, just naturally building them in together. So it wasn't like one month later Windsurf is gone and that's it. Over the course of the last year, we've been building all these things, and now it's like Devin Cloud and Devin Desktop are super nicely paired together. You could start sessions from one and hand them off to the other. You could work with them all in the same systems, whatever. But it was much more of a gradual thing. And I think one of the things that was interesting for us was a lot of the users of either product were already naturally looking for the other side. So we had a ton of people, for example, who were on the IDE framework, and obviously right around this time that we're talking about, late 2025 early 2026, is right around when these asynchronous agents were really starting to get big. So all these folks were thinking about, okay, what should my strategy be? Obviously it's great to be able to work with a single team and somebody that already understands your codebase and has all the details figured out and deployed. So that was nice. And then on the reverse side, similarly, for a lot of folks who were adopting the cloud agent, it's very nice to say, okay, yeah, the local side of it is there too for any of the developers who still want to use the local form factor and who want to have that. So I think the simple way to put it is, rather than force it and say, okay, over the next 30 days or 60 days we have to integrate these products, it was much more of letting the actual usage and the new features we're building, the next paradigms we were getting to, letting those pull the products together more gradually rather than forcing it.

在这期间你是怎么获得建议的？你是边做边摸索，还是有向谁请教？

How were you getting advised during this? Were you just making it up as you go or were you like who were you turning to?

这是个好问题。很多都是边做边摸索。

It's a good question. A lot of making it up.

好的。

Okay.

我觉得对我们来说，关于 Cognition 的一点，说实话，是我们团队里很多人都是前创始人。我们最初的 50 到 60 人里，我想大约有 30 个人在之前创办过公司。

And I think for us, one of the things about Cognition, honestly, is that a lot of our team are former founders. Our first 50 or 60 people, I think like 30 of us had founded a company before this.

所以，显然，我们很多人都是喜欢有雄心、有创业精神、从第一性原理思考的人。但没错，我的意思是，这真的是一个团队的努力。我特别记得那个周末。有一个时刻，最初的讨论是我和 Russell 代表我们这边，Jeff 和 Graham 代表 Windsor 那边。但一旦我们过了最初的讨论，问题就变成了：好吧，但我们到底要做什么？然后那个周六，业务的每一个部分都参与进来了。那是我最喜欢的日子之一，老实说，是我人生中最有趣的日子之一。因为我们的技术团队，我们有 Steven 和 Walden 等人，他们是工程和产品负责人，我的联合创始人，他们去和各自的团队坐下来了解：WinSurf 的情况如何？我们怎么推出 Wave 11？我们需要确保按时交付哪些具体功能，我们承诺过的事情显然要兑现？我们开始走向整合的第一步是什么？同样，在所有市场推广的事情上，我们的人也在跨团队协作，让这一切发生。我觉得我们所有人都有一种感觉，就是我们就是喜欢解决这类问题。我想有时候人们会描述成：哦，我只想安静地打造我的产品，而建立公司的其他一切事情，我只是忍受一下，但那不是我的菜。我觉得对我们所有人来说，建立公司的体验本身就是很多乐趣所在。所以那对我们来说是一个非常有趣的周末。

And so, obviously, a lot of us together were people who liked being ambitious and entrepreneurial and thinking from first principles. But yeah, I mean, it was a real team effort. I specifically remember that weekend. There was a point where the initial discussions were with myself and Russell from our side, and then Jeff and Graham from the Windsor side. But then, as soon as we got past the very first initial discussion, there was a question of, okay, but what are we actually going to do? And so then there was every single part of the business that Saturday. It was one of my favorite, honestly, one of the funnest days of my life. Because we had our technical team, we had Steven and Walden and so on, who are engineering and product leaders, my co-founders, going and sitting with their team to understand, okay, what's going on with WinSurf? What do we do to get out Wave 11? What are the specific features we need to make sure to deliver on time, the things we've promised folks that we obviously want to live up to? What are the next first steps on how we would start getting towards an integration? Similarly, on all the go-to-market stuff, we had our folks working across teams to make that happen. There was just a feeling, I think, for all of us, of like, we just love figuring this kind of stuff out. I think sometimes people describe it as, oh, I just want to build my product in peace, and everything else about building a company is just something I'll put up with, but it's not my cup of tea. I think for all of us, the experience of building the company is a lot of the fun of it. So it was a really fun weekend for us.

哇。是啊。我可以想象。我敢肯定每周都有别的事情。

Wow. Yeah. I would imagine. I'm sure it's something else every week.

是啊，很有趣。我的意思是，我们有过各种各样的……是的，从那以后我们还有很多更有趣的星期。所以总是一团糟。

Yeah, it's fun. I mean, we've had all sorts of... Yeah, we've had a lot more fun weeks since that one too. So it's always a mess.

是啊。说到并购的背景，Cursor 和他们的交易，他们与 SpaceX 的 600 亿美元交易，并购传闻四起，还有各种有趣的故事在发生，但 Cognition 仍然保持独立。以至于你们在独立日去了华盛顿特区。那么，你们为什么在独立日去华盛顿特区？这里的华盛顿特区策略是什么？

Yeah. Well, speaking on the backdrop of M&A with Cursor and their deal, their $60 billion deal with SpaceX, M&A is floating around and there are all those kinds of fun stories that are happening, but Cognition remains independent. So much so that you spent Independence Day at DC in DC. So, why were you guys in DC for Independence Day? And what's the DC play here?

是的，当然。不，我的意思是，这现在真的很热门，而且我认为特别是因为代码在某种程度上是最成熟的垂直领域。我想很多人把代码视为对整个行业和其他垂直领域将要发生的事情的预期。而且，是的，我认为有一种说法是：哦，为了获胜，你必须自己成为一个实验室，或者你必须去卖给一个实验室。显然，我们一直相信保持独立有很多价值和力量。你可以从我们与人们的合作方式中看到这一点。我们完全模型中立。我们与所有不同的提供商合作：OpenAI、Anthropic、Google 等等。而且，在我们与公司和我们合作的所有组织的合作方式中也是如此。所以华盛顿特区对我们来说真的很令人兴奋，有几个原因。第一，因为坦率地说，你谈到那些需要软件的地方，他们有想构建的东西，但没有足够的软件工程师来构建它们。华盛顿特区和整个政府可能是最大的一个点，对吧？所有这些组织都有很多事情……我觉得一句话就是：想象一下去车管所，它居然能正常运作。为什么它不运作？因为所有那些过时的软件。所有那些疯狂的程序。因为网站不太好用，或者你无法提前看到这些东西并追踪所有这些。政府软件不一定要很糟糕。所有这些都可以变得很好。而且我认为这对我们来说是一个非常重要的事情去努力。我认为另一个原因，为什么这对我们如此重要，坦率地说，我认为 AI 本身在未来几年将成为一个非常根本的政策问题。你想想互联网的早期，例如，有所有这些讨论：好吧，会发生什么？会不会有少数公司控制整个互联网，或者其他什么？显然，我们最终有了开放的互联网，现在世界上所有的企业都建立在互联网上。但总有一天，世界上所有的企业都会建立在 AI 上。而且我认为我们应该尽早而不是推迟思考这个问题。所以我们也想为那个对话贡献我们的想法。

Yeah, for sure. No, I mean, it's really topical right now, and I think especially because code is, in some ways, the most mature vertical. I think a lot of folks look to code as an expectation of what's going to happen in the rest of the industry and all these verticals. And yeah, I think there's been a narrative out there that, oh, in order to win, you have to be a lab yourself, or you have to go sell to a lab. And obviously, we've always believed that there's a lot of value and a lot of power in being independent. And you see that in how we work with folks. We're completely model neutral. We work with all the different providers: OpenAI, Anthropic, Google, and so on and so forth. But also just in our approach with companies and all the orgs that we work with. So DC is really exciting for us for a couple of reasons. I think one, because frankly, you talk about places that need software and have things they want to build but not enough software engineers to build them. DC and government in general might perhaps be the single biggest point of that, right? All these orgs have so many things that... I feel like the one-liner is: imagine going to the DMV and it actually worked. Why doesn't it work? It's because there's all this archaic software. It's all these crazy processes. It's because the website doesn't really work, or you're not able to see these things and track all these things in advance. And government software doesn't have to be bad. All these things can be good. And I think it's a really important thing for us to work towards. I think the other reason why it's so important to us is frankly, I think AI itself is just going to be a pretty fundamental policy issue over the next few years. And you think about the early days of the internet, for example, there were all these discussions about, okay, what's going to happen? Are there going to be a handful of corporations that control the entire internet, or whatever? Obviously, we ended up with the open internet, and now all the businesses of the world are built on the internet. But there will come a day where all the businesses of the world are built on AI. And I think it's something that we should be thinking about sooner rather than later. And so we wanted to give our thoughts to that conversation as well.

看到新政府如此拥抱技术真是太棒了。他们在过去一年里做的事情令人难以置信。但从你保持独立的立场来看，你提到了几个原因，但你认为从长远来看，为什么对你来说，尽可能独立地把这家公司做大，比在另一个保护伞下更重要？

It's awesome to see the new admin just embrace technology so much. The amount that they've done in the last year has been incredible. But from your standpoint of being independent, like you mentioned a couple reasons, but why do you think in the long run it's more important for you to build this company as big as you can independently than under another umbrella?

是的，不，我的意思是，我想说几件事。我认为有一种世界，人们有时会谈论，那里有完全的递归超级智能，然后一切都由一个单一实体拥有，或者也许有两个实体什么的。但然后他们有……就像一旦你解决了 AGI，一切都会崩溃。你接管整个世界，然后只有一个实体统治世界。我只是不认为那是我们的正确未来。而且我认为 AI 会达到那种能力水平，而且我认为我们会看到它做出越来越疯狂的事情。比如，我一直觉得，就纯粹的知识工作而言，人类能做的任何事情，AI 都能做。

Yeah, no, I mean, I think a few things I'd say. I think there's a world that folks sometimes talk about where there's the full recursive superintelligence and then everything is all owned by a single entity, or maybe there's two entities or something. But then they have the... it's like everything all collapses once you've solved AGI. You just take over the entire world and then there's just one entity that rules the world. I just don't think that's the right future for us. And I think AI will get to that level of capabilities, and I think we'll see it do even crazier and crazier things. Like, I've always felt that in terms of just pure knowledge work, there is nothing that humans do that AI can't do.

你知道，部分因为人类就是人类，我们的大脑就是计算机，就像 AI 也是计算机一样，对吧？我的意思是，它们之所以叫神经网络，就是因为受到了大脑运作方式的启发。所以我认为我们会解决所有这些问题，但我觉得我们希望以这样的方式来实现：人们能够掌控自己的智能，能够获得所有这些能力等等。而且我觉得不仅如此，我认为这是可以做到的。

You know, partly because humans are humans, our brains are just computers, you know, like the AIs are too, right? And I mean, they're literally called neural networks because they were inspired by how our brains operate. So I think we will solve all these problems, but I think we want that to be done in a way that people can control their own intelligence, can have access to all these things and so on. And I think more than that too, I think it can be done.

这显然是你想押注的任何使命的重要部分。但我认为还有很多空间，而且我认为每个人——无论是个人还是企业——都有很大的需求，他们想要真正独立的工具，想要在产品层和价值层真正创新的东西，而不仅仅是纯粹的智能层。

Which is obviously an important part of any mission that you want to bet on. But I think there is a lot of room and I think there is a lot of appetite from everyone out there, you know, individuals and businesses and so on, for truly independent tools that they can use and for things that really innovate at the product and the value layer rather than just the pure intelligence layer.

关于这个话题，Christian Garrett，我问了他们一些问题。137，他们是超级粉丝。我们刚刚和 Justin Fishner Wolson 做的访谈里提到了你。Justin Christian 也是。他们很棒。所以 Christian 问：为什么企业想要第三方提供商或开源，而不会完全依赖实验室的编码工具？

I mean, so on this topic, Christian Garrett, I asked them for some questions. 137, they're big fans. We talked about you in their interview we just did with them, um with Justin Fishner Wolson. Justin Christian also. So they're great. So Christian asks: why do enterprises want third party providers or open source and won't solely rely on the labs' coding tools?

是的，是的。我认为有几个原因。首先，显然企业的运作方式是他们想要长期合作伙伴关系。坦率地说，很难知道长期会发生什么，对吧？比如六个月或十二个月后谁会有最好的模型。我认为这非常合理，你不想教会所有工程师或整个团队使用某一套特定的产品和工具，然后发现，哦，实际上人们不再用这个了，因为大家都说另一个更好，对吧？但另一个重要的原因是，正如我们所说，对每家公司、每个团队来说，他们关心的不是用了多少 token，而是创造了多少价值，对吧？我认为有一个参与者去帮助他们实现这些价值并将其变为现实是非常重要的，对吧？在实践中，这关乎你如何组织团队、如何思考规划、如何思考规格或设计、如何做用户研究。所有这些在 AI 时代都应该不同，对吧？这不像把工具扔过墙那么简单。这是你的聊天机器人，试试吧，希望它对你有用。对吧？有很多基本问题需要你深入思考和解决。而这些显然需要组织变革，对吧？所以从他们的角度来看，他们想和一个在这方面思考很多、真正理解这一点并且乐于帮助他们处理所有细节的人合作。

Yeah. Yeah. I mean I think several reasons. First of all, obviously the way enterprises work is they want to have long-term partnerships. And frankly, it's just hard to know what's going to happen in the long term, right? Like who even is going to have the best model in six months or 12 months or whatever. And I think that's very reasonable, you know, like you don't want to teach all of your engineers or your entire team how to use one particular suite and one particular product and then find out, oh, actually it turns out people aren't using this one anymore because everyone says this other one is better or something, right? But I think the other reason that it's important too is because as we said, for every company, for every team, what they care about is not how many tokens they're using, it's how much value they're driving, right? And I think it's very important for there to be a player that is going and helping them drive all that value and turning that into reality, right? And I think in practice it's how you organize your teams, how you think about planning, how you think about specs or design, how you do user research. All of these things should be different in the era of AI, right? It's not as simple as throwing the tool over the wall. Here's your chatbot. Try that out. Hopefully, it's good for you. Right? It's like there's a lot of fundamental questions that you have to think through and work through. And these obviously require organizational change, right? So from their perspective, they want to work with somebody who's thinking a lot about that, who really understands that and is excited to go help them with all those details.

如果你正在构建 AI 领域的下一个大事件，你需要了解 MongoDB，这个开发者喜爱的数据库平台，专为你运行的智能体而构建。MongoDB 通过 Voyage AI 的向量搜索和嵌入，实时存储、搜索和推理你的数据，全部在同一个系统中。无需单独的管道，无需拼接 10 种不同的工具。这就是为什么 75% 的财富 100 强公司和领先的 AI 原生初创公司都运行在 MongoDB 上。从你的第一个用户扩展到数十亿向量。访问 mongodb.com/ai 了解更多。那就是 mongodb.com/ai。再见。Assembly AI 是数百万开发者构建的语音 AI 基础设施层。他们构建了业界最好的语音转文字、语音智能体和语音理解模型，为 Granola、Hen、Ashby 和 ClickUp 等公司提供关键基础设施。他们的语音转文字模型在准确性和质量上领先业界，他们的语音理解模型帮助你超越转录，通过从语音数据中揭示洞察、识别说话人、突出关键信息。你可以今天在 assemblyai.com/sourcy 开始，并获得 50 美元的免费额度来构建语音 AI 产品。那就是 assemblyai.com/sourcy。

If you're building what's next in AI, you need to know MongoDB, the database platform developers love and built for the agents you're running. MongoDB stores, searches, and reasons over your data in real time with vector search and embeddings from Voyage AI, all in the same system. No separate pipelines, no stitching together 10 different tools. It's why 75% of the Fortune 100 and leading AI native startups run on MongoDB. Build and scale from your first user to billions of vectors. Go to mongodb.com/ai to learn more. That's mongodb.com/ai to learn more. Bye. Assembly AI is a voice AI infrastructure layer millions of developers build on. They build the industry's best speech-to-text, voice agent, and speech understanding models that serve as critical infrastructure for companies like Granola, Hen, Ashby, and ClickUp. Their speech-to-text models lead the industry in accuracy and quality, and their speech understanding models help you go beyond transcription by uncovering insights, identifying speakers, and highlighting key information from voice data. You can get started today at assemblyai.com/sourcy and get $50 of free credits to start building voice AI products. That's assemblyai.com/sourcy.

所以你住在硅谷。你相信那个“6 个月到奇点”的说法吗？你对整个奇点怎么看？

So you're based in Silicon Valley. Do you believe the whole 6 months to singularity thing? What's your take on the whole singularity?

六个月太短了。是的。我的意思是，关于递归这件事，我要说的是：显然有某种版本的递归，就像我甚至刚才说的，Devon 写了 Devon 的大部分代码。所以，另一方面，我认为关于人类是否会被变得无关紧要，所有科学问题和社会问题是否会在六个月内被终极智能解决，我不太相信。我认为有几个原因。首先，世界上有很多实际问题并不一定纯粹靠智能就能解决，对吧？有组织上的事情要做，有需要时间的过程要搞清楚，有硬件方面的限制，比如你能获得多少 GPU 算力等等。所以我认为这些事情显然存在。其次，我认为我们在模型训练领域看到越来越多的是，今天的氛围几乎是你几乎可以教模型任何东西，只要你知道你试图衡量什么或试图得到什么结果。所以基准分数就是明显的例子，每次有新模型人们都会发布基准。有点反直觉的是，我们几乎到了可以解决任何基准的地步，对吧？因为拥有基准意味着什么？意味着你已经定义了任务。你已经明确了成功或失败的样子。你已经给出了一堆例子来说明任务是什么以及正确的行为应该是什么。而事实是，如果你这样做，那么你就可以教模型去做那件事。你知道，强化学习是有效的。

Six months is very short. Yeah. I mean, so look, the recursive thing here's what I'd say: there is obviously some version of the recursive thing which is literally, I mean I even just said that like Devon writes most of the code of Devon. So it's like, on the other hand, I think the question of whether humans will be rendered irrelevant and all human like all science problems and all societal problems will be solved six months from now with the ultimate intelligence, I don't quite believe that. And I think there's a few reasons for why that is. For one, there's a lot of practical problems out there in the world that are not actually necessarily just purely intelligence soluble, right? There's just organizational things that have to be done, there's processes that take time that have to be figured out, there's hardware components of like how much GPU compute can you get and so on. So I think all those things obviously exist. For two, I think we see more and more that in the model training landscape, the vibe today is almost like you can teach the model pretty much anything as long as you know exactly what you are trying to measure or what kind of outcome you're trying to get out of it. And so benchmark scores for example are obvious, every time there's a new model people release benchmarks. The thing that's kind of almost counterintuitive is we're kind of getting to the point where you can solve basically any benchmark, right? Because what does it mean to have a benchmark? It means you've already defined the task. You've clarified what success or failure looks like. You've given a bunch of examples of what that task looks like and what the right behavior should be. And the truth kind of is, well, if you do that, then you can teach the model to go do that. You know, RL works.

这真的令人难以置信，它居然这么有效，对吧？但显然，当今世界上的所有任务并不都能完全符合这种描述。还有很多事情是非常模糊的，没有清晰的定义，很难说什么算成功或失败，而且时间跨度很长，需要很长时间才能知道当初的决定是否正确。所以如果你让它写一个简单的程序去执行某个任务，运行程序然后说“好，它得到了正确结果或没有”，这非常容易。但如果你让它做一个 10 年的战略决策，那就很难在等待 10 年看到结果之前，获得关于你做得对还是错的数据。所以我认为我们最终会解决所有这些问题，明确地说。而且我认为我们会……

It's incredible how much it works, right? But obviously not all of the tasks of the world today cleanly fit into that description. And then there are lots of things that are super amorphous, that don't really have clear definitions, that it's hard to say what counts as success or failure, that have really long time horizons where it takes a long time to find out whether it was the right decision. So if you're asking it to write a simple program that goes and does a thing, it's super easy to run the program and say, okay, it got the right result or it didn't. But if you're asking it to make a 10-year strategic decision, well, it's kind of hard to get the data on whether you did that right or wrong until you've waited 10 years out and seen that. And so I think we will solve all these things, to be clear. And I think we'll...

我真的不认为存在 AI 长期无法解决的智能问题。

I really don't think that there are intelligence problems that AI will not solve in the long term.

我只是不认为这会在六个月内就全部结束。

I just don't think of that as like it's going to be all over in six months.

好吧。

Okay.

我可能错了。我不知道。也许真的就结束了，但你知道，如果真的结束了，那我们也度过了美好的时光。

I could be wrong. I don't know. Maybe it is all over, but you know, if it is all over, we had a good time until then.

是的，有一种……我实际上觉得那里有一种虚无主义，就像……

Yeah, there is a certain... I actually think that there's a certain nihilism there, which is like...

我同意，在某个点上，AI 几乎就像有一个事件视界。在某个点上，AI 会变得如此聪明，以至于我们甚至很难推理那个世界会是什么样子。人类还会有欲望吗？我认为我们仍然会有欲望。我们仍然有激情和想要表达的东西。我认为我们会有，但谁知道呢，也许 AI 会为我们解决所有这些问题。所以事件视界确实存在。过了那个点，就很难预测了，但你只能保持理智。你必须做出合理的决定。你必须去构建。也许一切都会在六个月内结束，但我目前不这么认为。

I agree, there's some point at which there's almost like this event horizon in AI. There's some point in which the AI gets so smart that it's hard for us to even reason about what that world will look like. Will humans still have desire? I think we still will have desire. We still have passions and things that we want to express. I think we will, but who knows, maybe AI will just solve all of those problems for us or something. So the event horizon does exist. And past that point, it's hard to predict, but you just got to stay sane. You got to make reasonable decisions. You got to build. And maybe it all ends in six months, but I don't currently think that's the case.

顺便说一句，我觉得这些人大多数已经疯了。我不知道你有没有读你 Colossus 文章的开头部分，但硅谷科技界的其他人是怎么回事？我读了那部分，然后在下面评论，我就想，这什么……我无法忘记那部分。谢谢你，Jeremy。

By the way, I think most of these people are already insane. I don't know if you read the beginning part of your Colossus piece, but what is up with the rest of the Silicon Valley tech community? Like, I read that and then I commented below and I'm like, what the... I can't unread that. Thank you, Jeremy.

那里发生了什么？

What was going on there?

确实有一些这样的情况。是的，我的意思是，这很疯狂，你知道。显然，就像我们说的，这些公司以前所未有的速度增长。这些能力显然……我认为 AI 范式确实比手机或互联网更上一层楼。它不仅仅是一项人们每天使用的技术。我认为它将从根本上重塑我们的生活。但是的，你只能保持理智。

There's some of that for sure. Yeah, I mean, it's crazy, you know. Obviously, it's like we're saying, these companies are growing faster than ever before. These capabilities are clearly... I think it's true that the AI paradigm is on another level compared to the mobile phone or the internet. It's not just a technology that people use every day. I think it will fundamentally reshape our lives. But yeah, you just got to stay sane.

所以 Devin 做的第一个任务是启动 MongoDB。我想知道你为什么选择那个。

So the first task that Devin did was spin up MongoDB. I want to understand why you picked that.

是的。关于 MongoDB，顺便说一句，这不是故意的或什么的。我们当时真的在玩……我想那时候还没有“编码智能体”这个术语，但就是关于拥有一个编码智能体会是什么样子的想法。那是 2023 年底。所以当时还不存在任何这类东西。但我们做了这些小东西。有点好笑。我的一个联合创始人 Stephen，联合创始人 Walden。我们都想，“好吧，我们要做我们自己的开发版本，就像 AI 开发人员，对吧？”所以我们的 Slack 里有一个 dev Steven，然后还有一个 dev Walden。然后最终，显然，我们把所有这些想法整合成一个单一产品，那就是 Devin，自然是所有这些的结晶。但我们有每一个这样的东西，人们在构建它们，只是随便玩玩，试图让它们做很酷的事情，对吧？比如如果你能真正测试你的代码，而不仅仅是 LLM 补全你的代码，你能做什么？如果你把它连接起来，让它能在终端中实际运行命令，或者有一个基本的、超级基本的浏览器使用功能，你能做什么？我们当时在摆弄这些，然后我们实际上自己也需要 MongoDB。所以 Walden 在设置 MongoDB，然后，你知道，有时候当你试图设置一个开发工具或数据库或任何这些东西时，你运行一堆命令，粘贴它告诉你要做的事情，然后它就是不工作，出现一些错误，比如“哦，这个端口”之类的。也许是因为你有一个随机的依赖包版本不对，导致一切都乱了。也许是因为你电脑上开着某个其他进程占用了它之类的。但不管什么原因，Walden 就是没能把 MongoDB 设置好。你做了你总是做的事，就是谷歌搜索错误信息，看看它说什么，然后尝试 Stack Overflow 上那个人说的命令，但没用。在 AI 出现之前的日子里，你不得不那样做。他花了一些时间在上面，然后他只是说，“是啊，我不知道。”所以那时他基本上有他的 Devin，他说，“好吧，Devin，去试试让它工作吧。”结果这变成了一个非常好的用例，因为如果你想想你需要什么，你需要第一点，就是关于你可能遇到的所有不同错误以及每个错误的原因的百科全书式的知识。然后第二点，你需要实际运行和诊断事物的能力。只有错误信息而没有其他东西是一回事，但能够去运行命令、去查看其他正在运行的进程、去做你需要做的任何其他事情是另一回事。显然这两点，百科全书式的知识和实际运行命令的能力，这正是编码智能体的定义。Devin 只是运行了一堆东西，几分钟后就修复了它，然后给 Walden 发了一条消息，说：“嘿，我搞定了。它现在起来了。”他一开始真的不相信。

Yeah. So MongoDB, it wasn't on purpose or anything, by the way. We were literally playing around with... I guess there wasn't a term for a coding agent back then, but like the idea of what it would look like to have a coding agent. This is like end of 2023. So there weren't any of these that existed. But we made these little things. It's kind of funny. One of my co-founders, Stephen, co-founders, Walden. We were all like, "Okay, we're gonna make the dev version of ourselves, like the AI dev, right?" So there's like a dev Steven in our Slack, and then there's like a dev Walden in our Slack. And then eventually, obviously, we took all those ideas and made it into a single product, and that was Devin, naturally, as the culmination of all these. But we had each of these and people were building them and just messing around and trying to get them to do cool things, right? Like what can you do if you can actually test your code and not just like LLM completion your code? What can you do if you hook it up so that it can actually run commands in the terminal or have a basic, super basic browser use thing? And we were messing around with these, and then we actually just needed MongoDB ourselves. So Walden was setting up MongoDB, and then, you know, it's like sometimes with these when you're trying to set up a developer tool, or a database, or any of these things, you just run a bunch of commands, you paste in the things that it told you to go do, and then it's just not working and there's some error, like, oh, this port or whatever. Maybe it's because you have one random dependency package on the wrong version and that messes up everything. Maybe it's because you have some other process that's open on your computer that's hogging it or whatever. But for whatever reason, Walden was just not getting MongoDB set up. And you do what you always do, which is you just Google the error message and you see what that says and you try that command of whatever this person on Stack Overflow said, and it didn't work. Back in the days before AI, you had to go do that. And he spent some time on it and he was just like, "Yeah, I don't know." So then he had his Devin at the time, basically, and he's like, "All right, Devin, just go try and make it work." And it turned out to be a really great use case, because if you think about what you need, you need number one, you just need encyclopedic knowledge of all the different errors that you could run into and what would cause each of them. And then number two, you need the ability to actually run and diagnose things. It's one thing to just have the error message and nothing else, but it's another thing to be able to go run commands, to go look at the other running processes, to go and do whatever else you need to go do. And obviously those two things, the encyclopedic knowledge and the ability to actually go and run commands, that's literally what a coding agent is. And Devin just ran a bunch of stuff for a few minutes and then fixed it and then sent Walden a message like, "Hey, I got it. It's up now." And he didn't really believe it at first.

我觉得我们当时还留着那段的视频录像，因为实在太难以置信了。我记得那天晚上我睡不着。我们就在想，也许它真的能行，你知道吗？也许你可以有一个 AI 工程师伙伴，直接帮你干活。但那确实是我们第一次相信这是可能的。

I think we still had the video recording of that because it was so unbelievable. I remember I couldn't sleep that night. We were just like, maybe it does work, you know? Maybe you can just have an AI engineer buddy that can just do work for you. But that was the very first moment we believed it was possible.

太棒了。感谢 MongoDB 给了工程师们可以追求的东西。

That's amazing. Thank you MongoDB for giving something engineers can go after.

是啊。感谢那些复杂但可解的文档页面。那正好是 2023 年 12 月所需要的临界点。

Yeah. Thank you for the complicated yet solvable documentation pages. That was right in the cusp of what it needed to be in December of 2023.

天哪。你觉得现在人们应该问但还没问的最大问题是什么？

Oh my gosh. What do you think the biggest question people should be asking right now that they're not?

你知道有趣的是什么吗？我觉得其实有一个重要的问题要问，那就是我们会用富足来做什么。我其实不太清楚工业革命期间到底发生了什么，你知道，我当时不在场，但我猜有一个阶段是大家都在种地，然后到了某个点，大家突然觉得，哇，我们居然可以大规模生产了。显然，从那以后我们进化了，我们有了各种想要的东西，有了我们在乎的东西，有了自我表达，有了各种知识工作等等。我觉得中间显然有一个阶段，富足已经存在，但还在摸索那些东西。我有点觉得我们在 AI 上就要进入这样一个阶段了。我这么说是因为，说到软件，显然人们首先想到的是，好吧，我怎么才能更高效？就像我现在做的所有工作，我可以用三分之一的时间做完。这是真的，你可以做到，这很神奇，真的很酷。显然这已经价值巨大了。但我认为真正的解锁不在于效率，而在于容量。就像，如果我们能构建这么多东西，我们能做什么？我们举过例子，也许车管所在软件富足之后真的能正常运转了，或者你登录银行账户、查看医疗记录，所有这些都会成为非常流畅的产品体验。但我其实觉得还有更多的东西。软件正在吞噬世界，这是一句名言，它早就成真了。我们到 2031 年或什么时候回头看时会发现，即使在你还得手动编写每一段软件、每一行将要运行的代码的时候，软件吞噬世界、把所有流程变成软件就已经是正确的了。想象一下，当你可以即时生成软件去做任何事情时，那会有多好，能做多少事。一次性软件将会存在，自动驾驶软件将会存在。所以任何你想去做或想变成现实的事情，任何以某种形式涉及使用计算机的事情，最终都是某种软件。而当所有软件都自己驱动时，你想做什么？你想创造什么？所以这真的是我脑海中首先浮现的问题，就是我们将用这些富足来创造什么？我觉得我们开始触及这个问题了。

You know what's kind of interesting? I think there's actually an important question to ask about what we would do with abundance. I don't actually know what went down in all the industrial revolution, you know, I wasn't there, but I assume that there was some phase where everyone was farming, and then there was a point where it was like, oh wow, we can actually mass-produce. Obviously, in the time since then, we evolved, there are different things that we all want. There are things that we care about. There's self-expression. There's all of these kind of knowledge work and so on that's come about since then. I think there was a middle period obviously where the abundance was there but it was like figuring out those things. I kind of think we're about to enter something like that in AI. The reason I say that is because with software, obviously the first thing that people think about is, okay, how can I be more efficient? It's like all this work that I'm doing now, I can do that work in a third of the time. And it's true, you can, it's amazing, it's really cool. Obviously that's worth a lot already. But I think the real unlock is not in efficiency but in capacity. It's like, what can we do if we can build so much more? We gave the example of maybe the DMV is going to actually work in the post-software abundance, or maybe all of your logging into your bank account or checking your medical record, all of this would be really smooth product experiences. But I actually think there's just so many more things. Software is eating the world, this is a famous line, it was true already. Here's how we'll look back on this in 2031 or something. It was true already back when you had to manually write every single piece of software, every line of code that was going to run, and yet still it was correct for software to eat the world and to turn all these processes into software. Imagine how good it is and how much more you can do where you can just generate software on the fly to go do anything. Single-use software is going to exist, self-driving software is going to exist. So anything that you want to go do or want to turn into reality, anything that involves using a computer in some form, is ultimately some kind of software. And when all the software drives itself, what do you want to do? What do you want to create? So that's the first thing that comes to my mind, honestly, is what are we going to create with all this abundance? And I think we're starting to get to that question.

那么最后一个问题。在象棋中击败彼得·蒂尔是什么感觉？

So last question. What was it like beating Peter Thiel in chess?

哦，我没有在象棋上击败他。不，我不会建议你跟彼得·蒂尔下象棋。我觉得那不是什么好主意。你得挑你能赢的游戏。但我确实向拿破仑挑战了一场扑克赛。我想当时两者都有，但实际发生的只有扑克。那是在 Cognitive 最早的一轮融资期间。所以这基本上就是种子轮。我们到了那个点，我们跟他们说得很清楚——顺便说一句，他们确实很厚道，他们在跟我们谈判时也一直非常坦诚、非常透明——但我们到了那个点，我们说，看，我们想去做这件事，这些是对我们有意义的数字，他们说，这些是对他们有意义的数字。我们足够接近，让人觉得这事应该能以某种形式达成，只是问题在于我们最终落在什么数字上。我说，你知道，有意思，你这么喜欢扑克，不如我们直接打一场一对一扑克赛，让赢家决定我们到底签谁的条款。事后看来，那会涉及公司大约 1% 的股份，那会是很大一笔。但我在拿破仑有点低落的时候提出了这个建议，然后彼得否决了。

Oh, I didn't beat him in chess. No, I would not recommend playing against Peter Thiel in chess. I don't think that's a good idea. You gotta pick games that you can win. But I did challenge Napoleon to a poker match. I think there was somehow both, but it was just the poker thing that actually happened. There was a round that was one of the earliest rounds of Cognitive. So this is like the seed basically. We got to a point where we were clear with them—to their credit, by the way, they were always super upfront and super heads-up about how they negotiated with us too—but we were at the point where we said, look, we want to go do this, here are the numbers that would make sense for us, and they were like, here are the numbers that would make sense for them. We were close enough that you felt like it should work in some form, there's just a question of what we land on. I was like, you know, it's interesting, you're such a big fan of poker, why don't we just play a heads-up poker match for that, and we could have the winner decide whose terms we actually signed. In retrospect, it would have been for like 1% of the company or something, which would have been a lot. But I proposed that when Napoleon was kind of down, and then Peter shut it down.

哦。

Oh.

彼得，你知道，我真的不觉得这是做这件事的正确方式。

Peter, you know, I don't really think that's the right way to do this.

然后他就那么说了，所以那事最终没有发生。我们还是谈妥了，达成了协议，一切顺利，从那以后我们就一直合作。

Then he said that, and so that didn't end up happening. We worked it out anyway, came to a deal, and it was all good, and we've been working together ever since.

未来 12 个月你最期待什么？

What are you most looking forward to in the next 12 months?

是的。说实话，我真的很兴奋，因为我们正在接近这个富足时代。我觉得在接下来这段时间里，我们会看到很多这样的进展——不仅仅是作为 AI 从业者，也是作为旁观者，作为消费者。我期待 AI 能在我必须做出的所有艰难个人决策中陪伴我。我期待有 AI 来处理我生活中的所有这些琐碎细节。我期待有 AI 基本上能让我把所有时间都花在我关心的事情上，并处理好其余的事情。我觉得富足时代是真实的，而且我们很快就会到达那里。

Yeah. I'm honestly just really excited about us getting closer to this era of abundance. I think we'll see a lot of that happen in the next—not just as somebody in AI, but as a spectator also, as a consumer. I'm excited for AI to be there for me for all of my tough personal decisions that I have to go make. I'm excited to have AI handle all these different little details in my life. I'm excited to have AI that basically just allows me to spend all of my time and focus on the things that I care about and take care of the rest. I think the abundance era is real, and I think we'll be there pretty soon.

太棒了。以如此积极乐观的方式结束。非常感谢你，Scott。我真的很感激。

Amazing. Such a positive optimistic way to end it. Thank you so much, Scott. I really appreciate it.

谢谢邀请，也许我们会在凡尔赛宫得到一些热门观点。

Thanks for having me, and maybe we'll get some hot takes at Versailles.

是啊。好的。非常感谢。好的，太棒了。非常感谢整个 Raise 团队，举办了一场精彩的活动。

Yeah. Okay. Thank you so much. Okay, cool. Huge thank you to the entire Raise team for an incredible event.

感谢 Brex、MongoDB 和 Assembly AI 让这次行程和系列节目成为可能。如果你喜欢这次对话，你一定会喜欢 Ray 系列的其余部分，包括 BlackRock 的 Tony Kim、Cognition 的 Scott Wu、Cerebras 的 Andrew Feldman、SambaNova 的 Rodrigo Liang、Lumen 的 Michael Hurston、MongoDB 的 CJ Desai，以及更多精彩内容，比如我们在一个秘密地点录制的热评，你可以在 X、YouTube 和 Instagram 上找到。订阅 Sorcery 的 YouTube 频道，获取更多与塑造 AI 的人们的对话，并加入免费通讯。你也可以在 sorcery.vc 付费订阅，每周获取关于 AI、机器人、企业软件、消费、半导体——我说过 AI 了吗？再说一遍 AI——以及即将到来的所有大事，比如融资公告和科技界的所有重大新闻。谢谢，再见。

And thank you to Brex, MongoDB, and Assembly AI for making this trip and series possible. If you enjoyed this conversation, you're going to love the rest of the Ray series with Tony Kim from BlackRock, Scott Wu from Cognition, Andrew Feldman from Cerebras, Rodrigo Liang from SambaNova, Michael Hurston from Lumen, CJ Desai from MongoDB, and many, many more like our hot takes that we did at a secret location that you can find on X, YouTube, and Instagram. Subscribe to Sorcery on YouTube for more conversations with the people shaping AI and join the free newsletter. You can also do paid at sorcery.vc for weekly insights on AI, robotics, enterprise software, consumer, semiconductors, did I say AI? AI again, and everything that's coming next like funding announcements and all big things in tech. Thank you. Bye.