OpenCode 创始人：为什么开放模型最终会占据多数

This is a very impactful technology. I want to make sure that everybody in the world can have access to it Having this be in the hands of a few companies a few entities that dictate the rules around it. I think historically we've seen that it's not something that ever works out. Every stack of it needs to be open-source Dax, thanks for being here at your house. I thought we weren't doing intros. Oh, okay, okay. I see what you were doing. My beautiful house, yes. 这是一项影响极其深远的技术。我希望确保世界上的每个人都能使用它。如果它掌握在少数几家公司、少数几个制定规则的实体手中，历史已经反复证明，这样的局面从来不会有好结果。整个技术栈的每一层都必须开源。 Dax，谢谢你在家里接受我们的采访。 ——我还以为我们不做开场介绍。 ——哦，好吧，好吧，我明白你刚才在做什么了。

Um, so you were at I hadn't been there since before COVID, so it was a while. Oh, wow. The city smells weird now. Okay. It does? Okay. Do you not think there's a weird smell in San Francisco? Because when I- Like, when I was hanging out with you, or, like, the city in general? No, but just, like… And not a metaphor, it's not a- not anything. I'm literally talking about the smell. I feel like, uh, the whole time I was there, there was just, like… 对，我漂亮的房子。嗯，所以你去了… …我从疫情前就没再去过，已经很久了。 ——哇。 ——那座城市现在闻起来怪怪的。 ——是吗？你不觉得 San Francisco 有一股怪味吗？因为当我… … ——你是说和我待在一起的时候，还是整座城市？ ——不是，就是… …这不是隐喻，也不是在暗指什么。我说的真的是气味。我觉得我在那里时，从头到尾都好像有一股…

And I don't remember this from before. Yeah. I feel like it's a new thing. Maybe you're just, maybe you're just used to the smell. The smell is bad? I think it's a bad smell, yeah. Okay. What does it smell like? I don't know. Just, like, wet. It smells like when you wash your- When you're in the, in the rain? Like… No, it just… It wasn't raining. Okay. I don't know. You don't think the city has a weird smell? No, I think it smells great. I brought this up to a few… …我不记得以前有这种味道。 ——对，我觉得这是最近才有的。也许你只是闻习惯了。 ——那味道很难闻吗？ ——我觉得挺难闻的。 ——像什么？ ——不知道，就是有点潮湿，像你洗… … ——像淋雨时的味道？ ——不是，当时也没下雨。 ——好吧。 ——我也说不清。你真不觉得这座城市有怪味？ ——不，我觉得闻起来很好。

You think it smells great? Yeah. I love it. I, uh, I brought this up to a few people after, and they were all like, "Yeah, it smelled weird the whole time." And the people that were there thought it would smell weird. ——后来我还问了几个人… … ——你觉得很好闻？ ——对，我喜欢。我后来问了几个人，他们都说：‘对，那几天一直闻起来怪怪的。 ’而当时住在那里的人也觉得气味很奇怪。

Oh. I don't know. Huh. I've never thought that. Is that a thing people say? Do you think New York smells? Yes. Okay. I think it smells terrible. So you know what? Maybe it's just that, because the whole time I was in New York, I was like, "I don't know what everyone's talking about. It doesn't smell like anything here." Yeah. So I guess- We have different senses. But that's just, that probably does mean there is a smell in San Francisco. Yeah. It's just there isn't… 哦，我不知道。我从没这么想过。大家会这么说吗？你觉得 New York 有味道吗？ ——有。 ——好吧，我觉得它闻起来很糟。所以也许就是这个原因，因为我在 New York 的整个期间都在想：‘我不知道大家在说什么，这里明明什么味道都没有。 ’——对。 ——所以我猜… …我们的嗅觉不一样。但这可能恰好说明 San Francisco 的确有一种味道。

Maybe I'm just used to it 'cause I live there now. Yeah. How was the rest of your San Francisco time? It was fun. We were only there for a couple days. Uh, I was telling you earlier, it feels like I stepped into a real-life version of the internet, 'cause I only interact with tech on the internet. That's the only time I'm doing tech stuff. When I'm out here in Miami, nobody knows what I do. Nobody wants to know what I do. ——是啊。只是… …也许我住在那里以后已经习惯了。 ——对。你在 San Francisco 剩下的时间过得怎么样？ ——挺有意思的。我们只待了几天。我之前跟你说过，感觉像是走进了一个现实版的… …互联网，因为我平时只在网上接触科技，那是我唯一会做科技相关事情的地方。在 Miami，没人知道我是做什么的，也没人想知道。

Uh, so I feel like I'm alone here, but then in San Francisco, it just feels like, oh, I'm, like, in Twitter, but, like, in real life. Yeah. Uh, to the point where, like, people, like… Like, our, our team was walking around, and then we walked by someone, and they were like, "Oh, it's an OpenCode team." And I'm like, "This is so weird." It's like, it feels like everyone that I know is in this small location, 'cause, 'cause San Francisco's not big, right? 所以我在这里会觉得自己有点孤单；但到了 San Francisco，就像是走进了现实中的 Twitter。甚至会发生这种事：我们团队走在街上，路过一个人，对方突然说：‘哦，是 OpenCode 团队。 ’我心想：‘这也太奇怪了。 ’感觉我认识的所有人都集中在这么小的一块地方，因为 San Francisco 并不大，对吧？

It's like a, it's physically a small place. Uh- Yeah. So yeah, it just feels like I can, like, look out to the city, and I'm like, "Oh, all the Twitter users," like, this is where they live. Would you ever move to San Francisco? We've talked about it. I think ultimately it's a no. Um, I, on the plus side, the plus side of known a bunch of people there is, you know, I'm sure every day, every day of the week I could hang out with someone. 它在地理上就是个很小的地方。 ——对。 ——所以我会觉得，只要望向这座城市，就像能看到：‘哦，所有 Twitter 用户都住在这里。 ’你会搬去 San Francisco 吗？ ——我们讨论过，但最终答案大概是否定的。好处当然是我在那里认识很多人，一周里的每一天大概都能约到人，社交生活应该会非常丰富。

Like social life would probably be really great. Yeah. Um. I get enough of the tech stuff online, like I don't think I want it in person. Um, I also think it's too cold for me. ——对。 ——但科技方面的交流，我在线上已经够多了，不觉得还想把它带进线下生活。而且我觉得那里对我来说太冷了。

San Francisco is literally too cold for me. I know. It's supposedly great weather. Yeah. California has great weather, but I like being in the heat, so Yeah. Yeah, it's good. It's nice out today though. It's so beautiful. Miami's great most of the year. It's like two months in the summer where it's bad, like it's too hot. Um, but most of the year it's like this. San Francisco 对我来说真的太冷了。 ——我知道。大家不是都说那里天气很好吗？ ——对，California 天气很好，但我喜欢炎热的地方。 ——嗯。今天天气确实不错，太美了。 ——Miami 一年里大部分时间都很棒，只有夏天大概两个月会热得难受，其他时候基本都像今天这样。

Yeah, yeah. So you like having the separation between like work and personal life, your social life, because I'm in San Francisco now and there's, it's all blended together. Yeah. I think so. I think there's this feeling, and this, this is true, like if you are in San Francisco it is the heart of tech. There's a lot of benefits from being that close. Anything that's going on is gonna happen there. You can go to these things, you can meet people. ——对。所以你喜欢把工作和私人生活、社交生活分开？因为我现在住在 San Francisco，这些完全混在一起。 ——我想是的。我觉得人们会有这种感受，而且这也确实是真的：如果你在 San… … Francisco，它就是科技行业的中心。

So that always feels like, oh, I should be able to take advantage of that. But it's this weird thing for me where I'm like, the whole point of tech is that it enables. Global scale, like the internet enables you to launch stuff to millions of people from anywhere in the world. Uh, so I feel like to natively be thinking of that mindset, you really should be able to do that without physically being in a location, right? 离中心这么近有很多好处，任何新动向都会在那里发生，你可以参加活动、认识人，所以总会觉得自己应该利用这种优势。但对我而言，这里有一种奇怪的矛盾：科技的意义本来就在于扩大规模，互联网让你能从世界任何地方把产品发布给数百万人。那么，如果真正以这种原生的互联网思维来做事，你理应不必亲自待在某个特定地点，对吧？

Um, so that definitely is a benefit to being there. But I also like the fact that we're able to do something that's like deeply in the SF zone, like AI and coding and none of us live in, live in San Francisco. Um, I'm, I'm like, I'm really proud of that. Not saying that like one is better than the other. 所以身处那里当然有优势。但我也很喜欢我们能够做一件完全属于 San Francisco 核心领域的事情——AI 和编程——而我们没有一个人住在 San Francisco。

It's just, uh, I think it's fun for us to embrace the fact that, yeah, in 2026 you can do everything on the internet and have like massive region reach massive people without having to physically be anywhere. That is an interesting point because now it seems like there's this new narrative of like, you must be in San Francisco or you'll be left behind. Yeah. And I feel like OpenCode is a great example that that obviously isn't true. 我对此真的很自豪。不是说哪种选择一定更好，只是我觉得，我们很乐于接受这样一个事实：到了 2026 年，你可以完全依靠互联网做成事情，触达极其广泛的地区和大量人群，而不必身处任何特定地点。 ——这很有意思，因为现在出现了一种新叙事：你必须在 San Francisco，否则就会掉队。 ——对。 ——OpenCode 显然是一个很好的反例：你不在 San Francisco，也做得非常出色。

Like you're doing so well not being in San Francisco. Yeah, it's, uh, like I said, I can't deny that there's not a benefit to being there. Like if you're starting a company and you have the option to be in San Francisco, it's probably what you should do. Um, and obviously it's not like any other place in the world has OpenAI and Anthropic but I was saying this the other day, there's so many AI products that I use that aren't based in San Francisco. ——对，就像我刚才说的，我不能否认在那里有优势。如果你要创业，而且有条件去 San Francisco，那可能确实是更合适的选择。当然，世界其他地方也没有 OpenAI 和… … Anthropic。但我前几天也说过，我使用的很多 AI 产品并不来自 San Francisco，其中一些用户量最大、经常登上新闻的产品也不在那里。

Um, some of the biggest ones that are like getting a lot of users and like kind of in the headlines, a lot of them aren't in San Francisco. So, uh, yeah, I, I, I can't pretend like the numbers don't swing towards sf, but the counter examples are also pretty interesting to look at. So you just wrote this article that exploded on X Twitter about Miami. 所以我不能假装数据没有明显偏向 San Francisco，但那些反例也很值得研究。 ——你刚写了一篇谈 Miami 的文章，在 X，也就是 Twitter，上引发了巨大反响。

Um, tell me about like the current Miami scene right now. Texting. Yeah, so I moved here five years ago and that was during like peak crypto, so I wasn't part of any of that. I moved here 'cause my wife's from here. We decided to come back down and settle closer to our family. 跟我说说 Miami 现在的科技圈吧。 ——对，我五年前搬来，当时正值 crypto 最火热的时候。我并不是因为那股浪潮来的，而是因为我妻子来自这里，我们决定搬回来，住得离家人近一些。

Uh, but that was around the time where there was a wave of like, we're moving Silicon Valley to Miami and like all these VCs were, came over and then crypto was booming and I was like, okay, this is not why I'm moving to Miami, but hey, there's some tech stuff here. 但那时正好有人喊着要把 Silicon Valley 搬到 Miami，很多 VC 也来了，crypto 又在蓬勃发展。我当时想：好吧，这不是我搬来 Miami 的原因，不过既然这里有些科技活动，

Let me go check it out. And I would try going to the events and it was like such a miserable experience. It would just be people that learned about NFTs the week before that had already started a company doing something and like they were giving a talk. Like they, they, it just like was a lot. It was very fake and amateur. Um, and I was like, okay, maybe. This will change at some point. 那就去看看。我试着参加了几场活动，体验非常糟糕：有人一周前才刚听说 NFT，就已经成立了一家做相关业务的公司，还站上台发表演讲。整个场面过于虚假，也很业余。

But I went to a few of those things and at that point I was like, okay, I'm just gonna do my own thing and not really engage. 我想，也许以后会有变化。但参加几次之后，我就决定还是做自己的事，不再参与其中。

Um, and then as you, as you would expect, that like Wave died down, a bunch of the people that moved here went back. A bunch of VCs that came here, went back. Uh, and nothing like super lasting or anything, nothing big came out of it. Uh, I'm starting to see that again a little bit because of some of the taxes that are being, uh, proposed in California and New York. 后来果然如你所料，那股浪潮退去了。很多搬来的人又离开了，很多来的 VC 也回去了，没有留下什么特别持久的东西，也没有诞生什么重要成果。现在，因为 California 和 New York 提出了一些税收政策，我又开始看到类似的苗头。

Mm-hmm. Uh, a bunch of people are moving down to Miami and Florida again, and there's this other narrative of, oh, we're gonna make Miami Tech a thing, but it's probably gonna go the same way. It's just gonna be like these things where everyone's kind of pretending to do stuff. Uh, I think it doesn't have to be that way. Uh, what I wrote my article was it's actually very simple. If you, if we really want to make Miami Tech an actual thing, it's definitely possible. 很多人再次搬到 Miami 和 Florida，于是又出现了… …另一种叙事：要把 Miami Tech 做起来。但这一次很可能还是重演旧路，大家只是装作在做事。我认为事情不必如此。我在文章里写的是：如果我们真的想让 Miami Tech 成为现实，其实办法很简单。

You don't, we don't need like a thousand companies here. We just need like five great founders here and like start small and like build. You have to build, like in the, in the beginning of anything like this,you have to do it with people that don't necessarily even need it. Um, I think right now a lot of the events are trying to cater towards people that are earlier in their career or people trying to get something out of it. 这里不需要一千家公司，只需要五个出色的创始人，从小处开始，真正去建设。任何这类生态在早期，都必须先由那些甚至并不依赖它的人来推动。现在很多活动试图服务职业生涯较早、或希望从中获得些什么的人。

And we wanna be able to do that eventually, but when we don't have anything, we usually can't serve that audience well. Um, so yeah, I think there's some great founders that I know here. Um, we're all very focused on our business businesses, so we don't have time to do this community thing. 我们最终当然希望服务这些人，但在生态本身还一无所有时，通常没法真正服务好他们。我认识一些很棒的创始人，但大家都专注于自己的公司，没有时间做社区建设。

Um, but someone is particularly motivated, I think they can make it happen. Uh, it just can't look like copying San Francisco 'cause San Francisco's like much later. stage. Yeah In that whole process. Yeah. I know that was one of your tips. Don't copy San Francisco. Yeah. Um, what else, what else can be done besides not copying San Francisco? Um, yeah, I don't know. Like, I think the only thing I said is just being really honest with ourselves. 不过如果有人特别有动力，我觉得这件事可以做成。只是不能照搬 San Francisco，因为 San Francisco 在整个发展过程中已经处于很后面的阶段。 ——对。我记得你的建议之一就是不要复制 San Francisco。除此之外还能做什么？ ——我也不确定。

I think another thing that we try with, with every single one of these waves. There's way too much. 我文章里另一个重点，大概就是要真正诚实地看待自己。我觉得每一轮浪潮里，我们都过度… …

Fake it till you make it. I like fake it till you make it. I like that concept and you have to do it, but it's done in a way where it's so obvious to everyone that you're faking where we'll have like one little thing that like, is kind of interesting and everyone will blow up. Blow it up. Like it's the biggest thing ever. Or like told you guys, like Miami Tech is definitely a thing and everyone is just looking at it being like, you we're actually just broadcasting. 强调‘先假装成功，直到真的成功’。我喜欢这个理念，有时你也必须这么做，但现在的做法往往假得所有人都看得出来：这里刚出现一点稍微有趣的东西，大家就把它吹成史上最大事件，然后说‘早就告诉你们了，Miami Tech 肯定成了’。

That's not a thing. 'cause we're, we're trying so hard to get people to believe it. Yeah. 旁观者只会觉得，你们如此拼命让别人相信它存在，恰恰是在向外界广播它其实还不存在。

Um, so yeah, these are just like little like marketing slash pr slash perception things, but uh, I do think it is necessary to, to get those things right for any of that to, for perception actually change. Yeah. Um, we have plenty of retired billionaire, I'm not retired is exaggeration, but people that have, that are billionaires that have done built their company and then that, that are moving here for like, the later part of their life. Um, that doesn't really mean much. 所以这些只是一些营销、公关和认知层面的小问题，但我确实认为，想让外界观感真正改变，就必须把这些事情做好。这里有很多可以说是退休了的亿万富豪——‘退休’有点夸张——他们已经建成公司，到了人生后半程才搬来这里。但这其实说明不了什么。

Like, uh, of course there's tons of software engineers that are here. Every company in the world hires software engineers. It doesn't matter what field you're in, if you're an oil company, you hire software engineers. If you're a marketing company, you hire software engineers for a place to actually be considered. Oh, there's tech there. It means that there's companies and people there that are influencing the rest of the space. Uh, Miami flat out does not have that at all. 这里当然也有大量软件工程师。全世界每家公司都会雇软件工程师：石油公司会雇，营销公司也会雇。一个地方要真正被称为‘有科技产业’，意味着那里有公司和个人正在影响整个行业。 Miami 目前完全没有做到这一点。

Um, the way we build software, uh, who we look at to learn how to do it, who influences us, it's also coming from San Francisco. 我们如何构建软件、向谁学习、谁在影响我们，这些依然都来自 San Francisco。

Um, and we can't claim anything is happening here until that changes, at least a little. Yeah. What were some of the responses on Twitter? I feel like it was a much bigger debate that I would think from. Yeah, it's funny 'cause I, I wrote it thinking like, oh, no one will care. 'cause uh, my audience is like, mine's on, on X is like the whole world. I'm like, there's some people in Miami that follow me. I'm like writing it for like my Miami community. 在这种情况至少发生一点改变之前，我们不能声称这里已经形成了什么。 ——Twitter 上有哪些回应？我感觉它引发的争论比我预想的大得多。 ——挺有意思，我写的时候还以为没人会关心。因为我在 X 上的受众来自全世界，虽然也有一些 Miami 的人关注我，但我原本只是在写给本地社区看，结果传播得远远更广。

Um, but it like went much broader than that. Uh, I think the two responses were, there's a bunch of people being like, yeah, Miami sucks. It's never gonna happen. Um, which is like kind of a not related point. Like I, I was saying, I was agreeing like, it hasn't happened yet, and here's some things that we could do to maybe make it a little more of a thing. 回应大致有两类：有一群人… …说：‘Miami 很差，永远不可能成功。 ’这其实有点偏离主题，因为我本来就在承认它还没有成功，只是提出一些也许能让它更像一个真正科技生态的做法。

Uh, there's few responses saying, no, it is a thing, it's already working, you know, kind of refuting that, which makes sense. Like, I get why people feel that way. Uh, and there was a funny one where there were a few people that were like, yeah, I tried moving to Miami. I had like a horrible experience and I left. 另一些回应则说：‘不，它已经是了，而且已经运转得很好。 ’他们是在反驳我的判断，我也理解为什么有人会这么认为。还有一种挺好笑的回应，有几个人说：‘我试过搬去 Miami，经历糟透了，所以离开了。 ’

And then there's like a bunch of arguments that were happening there. Um, but again, like I don't like care pretty much. I don't care about all three of those responses. Uh, I'm, I just wrote it for the people that are trying to do stuff here. Yeah. If you're gonna spend time and energy on it, here's some guidelines that help you kind of get the most out of it. Uh, but yeah, I mean, I might be wrong, but that's kind of how I saw things. 然后下面又吵成一团。但坦白说，这三类回应我基本都不在意。我写文章只是为了那些正在这里认真做事的人：既然你要投入时间和精力，这里有一些原则，能帮助你更有效地利用这些投入。当然，我也可能是错的，但这就是我对现状的观察。

Yeah. For someone who's listening and they're maybe thinking of moving to Miami, like, tell me some of the benefits. Or she moving from San Francisco to me. Yeah. ——如果听众正在考虑搬去 Miami，能说说这里有哪些好处吗？比如从 San Francisco 搬来。

I'll explain why I like it here. Me and Liz like kind of went through this yesterday. That's actually why I wrote the article. Like me and her had a conversation about this whole dynamic. Um, and she like pointed out a few things. So it's, what I like about Miami is that it is weird. You, you don't think about it in this way. 'cause everyone thinks Miami South Beach clubs nightlife, but it's extremely family friendly. Um, everyone I know here has kids. 我可以解释为什么我喜欢这里。我和 Liz 昨天刚梳理过这些事情，那次谈话正是我写那篇文章的原因，她也指出了几点。我喜欢 Miami 的地方在于，它和人们的印象很不一样。大家想到 Miami，通常会想到 South Beach、夜店和夜生活，但这里其实极其适合家庭生活。

Uh, the neighborhood I live in, super safe. 我认识的人几乎都有孩子，我住的社区非常安全。

I have a house with a yard, I can walk to the downtown, get food, tons of kids everywhere. Um, it's for people that are like, not in that early, super early stage in their life, but like kind of maybe in their thirties, like settling down. Um, and for me, at that stage of life that I'm in, uh, that's a part of the reason why F doesn't really appeal to me. Like of course I know there's people with kids and it's probably great for kids in, in its own way, but. 我有一栋带院子的房子，可以步行去市中心吃饭，四周到处都是孩子。这里很适合那些已经不在人生最早期、可能三十多岁、正准备安定下来的人。对处于我这个人生阶段的人来说，这也是 San Francisco 不太吸引我的一部分原因。当然，那里也有人养育孩子，而且可能以自己的方式很适合孩子，但整体上确实有一种年轻、拼命冲刺的氛围。

There's just a lot of, uh, young grindy mentality there. Yeah. Uh, there's nothing wrong with that. 这本身没有错。

Um, I think working hard is important, but yeah, if you're at a certain stage in your life and you wanna kind of be comfortable, any these other things that you need to need to worry about, um, yeah, I think it's a, it's a great place to be. 努力工作很重要。但如果你到了某个人生阶段，希望生活舒适一些，同时还要兼顾其他事情，这里就很适合。

Um, obviously taxes are thing, um, um, that's like, you know, if you are a founder and you're looking at a potentially big outcome, it's pretty crazy what a difference it makes living somewhere like this versus somewhere that is like more aggressive tax wise. Um, I don't think that's like a primary motivator. 税收当然也是因素。如果你是创始人，预期公司可能取得很大的结果，那么住在这里和住在税收更激进的地方，最终差距会大得惊人。不过我不认为它应该是首要动机。

I don't think that, I don't think you should live your life based on where you get tax, the lease. I think that's a kind of dumb way to live, but that's a benefit. Um. And then for me, I just like the weather. Like I like the weather. Uh, a lot of people hate it. I was, that was another response. A lot of people were like, Miami Tech will never happen. It's too hot. Like, what do you mean it's too hot? Wait, it's not that hot. Yeah. I mean it's, it is really great right now. 我不觉得人生应该完全按照哪里税最少来安排，那是一种很愚蠢的生活方式，但低税确实是好处。对我而言，我也真的喜欢这里的天气，虽然很多人讨厌它。还有一种回应是：‘Miami Tech 永远做不起来，天气太热了。 ’什么叫太热？ ——等等，也没那么热吧。

Uh, but also like you can't work when it's too hot. That's the other thing. That's the another thing that I hate every single time. People talk about where's the best place to live a company, build a company. ——对，今天就非常舒服。当然有时确实很热。但还有人说，太热就没法工作。每次有人讨论在哪里创业最合适，我最讨厌的就是这种论调。

They're always like, you can't do it here. There's like too much fun stuff everywhere. Or like the weather is, it's like none of these things matter when you're building a company. I'm not like, oh, I couldn't work hard enough 'cause I just had to go to the club. Like, you know, it's just like not a thing that happens at all. Um, I spend most of my time in my house walking around my neighborhood. 他们总说：‘这里不行，周围好玩的东西太多了’，或者‘天气不合适’。创业时这些事情根本不重要。我不会因为‘不得不去夜店’而无法努力工作，这根本不是现实中会发生的事。

Uh, I think I work as many hours as anyone anywhere else, but, uh, it's, it's like, not that whether your company is successful or not. It's like these things aren't that, that relevant. Um, yeah. 我大部分时间都在家里、或在社区散步。我工作的时长不比任何地方的人少。公司能否成功和这些因素没有多大关系。

It does seem funny that I saw this online where people are like, there's too many fun things to do in Miami. I wouldn't get work done. It's like, how weak is your willpower? I'm not that, you're just like, you have to be in San Francisco because if there's some fun things to do in Miami, you wouldn't be able to work. That's, but I, I also don't understand that. 网上有人说：‘Miami 好玩的东西太多，我肯定没法工作。 ’这听起来很好笑。你的意志力到底有多薄弱，才会因为 Miami 有些娱乐活动，就必须搬到 San Francisco 才能工作？

'cause what I was saying earlier was if I had moved to San Francisco every single day of the week, there is an event I can go to to hang out with people that I know. That to me is like a thousand times harder to say no to than like the vague notion of going to a club. 而且我也不明白这种逻辑。就像我之前说的，如果搬去 San Francisco，我一周里的每一天都有活动可以参加，能和认识的人见面。对我来说，拒绝这些具体邀请，比拒绝‘也许可以去夜店’这种模糊念头难一千倍。

Like, I don't know what, what do people think? Why did the people think Miami's so fun? Like Yeah. The beach. It's so true. You guys have beaches too. Yeah, we do. Yeah. Good weather right now. Yeah. And yeah. And in San Francisco and whenever people talk about is, I'm like, so are you saying that San Francisco's extremely like, unpleasant to be in? Like what? It, it can't be that. Yeah. I'm sure it's fun to be in San Francisco too, right? It's quite fun. 我也不知道人们为什么觉得 Miami 特别好玩。是因为海滩吗？你们那里也有海滩。 ——对，我们也有，而且现在天气也不错。 ——对。每次大家这样谈 San Francisco，我都想问：你是在说 San Francisco 极其不宜居吗？总不能是这样吧，那里肯定也很好玩。

Yeah. I think it is actually kind of hard to focus in San Francisco. Yeah. There's always some event or something happening, so, yeah. ——确实挺好玩的。 ——我觉得在 San Francisco 反而有点难专注，因为总有活动发生。

. So it's uh, again, like when you're starting a company, your goal is to find product market fit. Yeah. Which is really, really hard and is like so binary. None of these, like little optimizations you make, determine whether you find that or not. Like nothing's gonna save you from finding it. Uh, your company won't be successful unless you find it. The fact that you are like not distracted in a weirdly specific way doesn't really change that outcome at all. Yeah. 所以说，创业的目标是找到 product-market fit，而这件事非常非常难，结果也近乎非成即败。你做的这些小优化，并不能决定你是否找到 product-market fit，也没有什么能让你绕过它。找不到，公司就不会成功。你只是在某个非常具体的方面少受一点干扰，并不会真正改变最终结果。

I wanna hear more about like why you decided to launch OpenCode 'cause I remember you kind of talking to me about it before it launched and then it's been incredible just seeing how well it's done. Yeah. You know what it felt like, it really felt like, by the way, you know what's funny? Uh, our company, like our legal entity is 16 years old. Yeah. Which, uh, people always laugh at when they hear that 'cause that there's like, no one's ever heard of that. 我想多听听你为什么决定推出 OpenCode。我记得它发布前你就跟我聊过，而现在看到它发展得这么好，真的很惊人。 ——你知道吗，这件事感觉… …顺便说个好笑的事情，我们公司的法律实体已经存在 16 年了。

Like, no one's heard of a company that just lets kept the same investors in the same company. Like just trying itself. Yeah. How did this happen? 大家听到都觉得很荒唐，因为几乎没人听说过一家公司让同一批投资人一直留在同一个实体里，就这样不断尝试。 ——这到底是怎么发生的？

Um, yeah, I mean, like, uh, so I, I became a, uh, so I wasn't one of the original founders. I became a co-founder, uh, later. So I joined five years ago. Um, but the company's been around for 16 years and it's just isn't, then it's like long. Very long slog of like taking swings and every swing like, kind of works better than your previous swing, but not as good, but, but not good enough to actually do the thing. 我不是最初的创始人，而是后来成为 co-founder 的，五年前才加入。但公司已经存在 16 年了。整个过程就是一场极其漫长的跋涉：不断尝试，每一次都比上一次更有效，却又始终不足以真正做成那件事；然后凭借近乎无限的耐心继续尝试，并想办法让公司存活下来。

And just kind of like infinite patience on continuing to continue to keep trying find ways to stay alive. Um, so when we saw Claude Code come out, uh, that was the first time anyone on our team really adopted AI tooling. 所以当我们看到 Claude Code 发布时，那是团队里第一次有人真正采用 AI 工具。在此之前，我们试过各种产品，也试过 Cursor，但都没有留下来。

Like we tried all stuff that was coming out before then we tried Cursor nothing stuck, and we didn't, we didn't really like any of it and we were kind of unsure about AI coding in general, but Claude code came out and we were like, oh, this is like, we get it. This helped us get it. Uh, it really stuck, like it's the first tool that our team adopted and we were like, okay, now that we understand this space, um, is there something we can do given our experience? 我们一直不太喜欢那些产品，对 AI coding 整体也有些怀疑。但 Claude Code 出现后，我们突然明白了：‘原来是这样，这让我们真正理解了。 ’它确实用起来了，也是团队采用的第一个这类工具。于是我们想：既然已经理解这个领域，凭借自己的经验，能不能做些什么？

Um, so Claude code is obviously very tied to Anthropic, is closed source. Our expertise is open, doing open-source and also, uh, doing stuff with, with a, has a long tail how we refer to it. Uh, but, and long tail in this case means supporting every single LLM supporting every single infra provider's, like a super difficult, tedious task, something that open-source does really well. Claude Code 显然与 Anthropic 深度绑定，而且是闭源的。我们的专长则是开源，以及处理我们所谓的‘长尾’问题。在这里，长尾意味着支持每一个 LLM、每一家基础设施供应商；这是极其困难又繁琐的工作，而开源非常擅长解决它。

Uh, so weirdly, we're kind of looking at cloud code beyond. We really like this. There's another form of product that can exist that we feel, we felt like we were so uniquely positioned to do. Um, 'cause we've been doing open-source for a while. We knew how to bootstrap an open-source community overnight. Uh, it really felt like very soon after we started working on it, it really felt like, oh, this is a thing we were meant to do. 很奇妙的是，我们看着 Claude Code，会想：我们真的很喜欢它，但还可以存在另一种形态的产品，而我们恰好拥有极其独特的条件去做。我们做开源已经很久，知道如何在一夜之间启动一个开源社区。开始开发后没多久，我们就强烈感觉到：这就是我们注定要做的事情。

There's really nobody else out there that has a better shot of taking a swing like this. 几乎没有谁比我们更适合来尝试这一把。

And it's like very, it's so rare to find yourself in that situation. And again, it took 16 years Yeah. To find it. Yeah. So I wanna talk more about OpenCode and specifically like why did you decide to make it open-source? You talked a little bit about your expertise, but I wanna talk more about that. Yeah. So I think there's like, uh, there's a few different reasons. So the first one is mentioned, um. For a coding agent that can work with anything you need just a ton of help. 一个人很少会发现自己正处于这种位置。而且我们花了 16 年才找到它。 ——我想更深入谈谈 OpenCode，特别是你为什么决定把它做成开源。你刚才提到了一些团队专长，但我想再展开。 ——原因有几个。第一个刚才说过：对于一个要兼容所有东西的 coding agent，你需要大量外部帮助。

You can't in-house a team that can actually build this because every single week there's an issue like, hey, in, in Azure, in the Australian data center, they actually require this extra parameter to be set. Otherwise, you know, LLM cache prompting caching doesn't work properly. Um, that is not a situation that our team, our very small team, is gonna find, be able to debug, reproduce and fix. That is a type of thing that a contributor will find. 仅靠内部团队无法真正构建它，因为每个… …星期都会冒出这种问题：‘在 Azure 的 Australia 数据中心，必须额外设置某个参数，否则 LLM 的 prompt caching 就无法正常工作。

Reproduce the, say someone that's actually using OpenCode in that environment. Um, so to build something that can kinda like, get into every little crevice of the world, it almost always has to be open-source so you can kind of get the help from, from the community for those things. That's one reason. The second reason, this is kind of a first principle for me. Uh, I don't believe that the tools we use to build software remain closed source long term. ’这种情况不是我们这个很小的团队能够发现、调试、复现并修好的；真正能发现并复现它的，是恰好在那个环境里使用 OpenCode 的贡献者。所以要构建一个能够进入世界每个细小角落的产品，。几乎总得开源，才能借助社区的力量解决这些问题。这是第一个原因。第二个原因对我来说更像一条第一性原理：我不相信我们用来构建软件的工具会长期保持闭源。

Um, I think there's plenty of room for proprietary options. Uh. We have like a lot of history of that, right? We have proprietary editors. You know, JetBrains is massively popular. People love JetBrains, they make a great product. 当然，专有产品有充分的生存空间，历史也证明了这一点。比如专有编辑器 JetBrains 非常流行，很多人喜欢它，他们做出了很好的产品。

Uh, I don't think it's open-source I think your opinion gets paid for it. And there's a great market, massive company that's serving that, but 70% of the market is owned by open-source editors. Uh, so if you look at every component of tech databases used to be proprietary. Now mostly open-source, still proprietary options, but mostly open-source. Um, every single tool you run locally on your computer to do dev work. The default is an open-source option. 我认为它不是开源的，用户需要付费。它拥有很好的市场，也是一家服务该市场的大公司。但编辑器市场有 70% 被开源产品占据。如果观察科技栈的每一层，会发现数据库过去是专有的，如今大多已经开源，虽然仍有专有选项，但默认趋势是开源。你在本地计算机上运行的每一种开发工具，默认选项通常都是开源的。

So when it came to coding agents, and if we're saying that, you know, the future software development involves coding agents, uh, you know, the bulk of that should be an open-source option. Um, so I feel like a really good position to have. And we're looking around and we're like, why is no one taking this position yet? Um, there's so many coding agent companies, they're doing this, but nobody's taking this like obviously valuable, uh, real estate. 因此，如果未来的软件开发会依赖 coding agent，那么其中的大部分也应该有一个开源选项。我觉得这是一个非常有价值的位置。我们环顾四周，发现明明有那么多 coding agent 公司，却没有人占据这块显而易见的领地，于是我们就走了进去。

So we just kinda like stepped in there and the first couple of months we were like, the only thing we need to make sure is we're identified as the open-source option. Um, and that's kinda what we were focused on. So we can kind of claim that territory. Yeah. So you were saying historically most dev tools end up open-source Yeah. Like a default option. Can you talk a little bit more about that? 最初几个月，我们唯一需要确保的事情就是让大家把我们认作… …开源选项。我们专注于这一点，从而占据这块领域。 ——你刚才说，从历史上看，大多数开发工具最后都会出现默认的开源选项，能进一步解释吗？ ——我觉得年轻一些的人可能根本不记得了。

Yeah, and I think people especially that are a bit younger, maybe don't even remember this, but when I first, when I first started in, um, in tech, I was on a Microsoft stack I built using.net. And these are things that aren't open-source or weren't uh, and you had to pay for them, like you have to pay for your editor, you had to pay for a license to deploy these things. Um, and a lot of the market used them. That was like a huge thing. 我刚进入科技行业时使用 Microsoft 技术栈，用 .NET 开发。那时这些东西不是开源的，或者说当时还不是，而且你… …必须为它们付费：编辑器要付费，部署也要购买许可证。当时很大一部分市场都在使用这些产品，初学编程的人从专有工具开始也很常见。

And it was not uncommon for someone to start programming to learn on these proprietary things. If you go back even before then, this is, people probably would find this crazy, but uh. Compilers used to be, uh, proprietary. We had to pay for a thing. The compiler code or the proprietary compilers, um, database is another example. You know, there's a time where if you wanted to build a product that needed a database, you had to pay Oracle. That was a default option. 再往前追溯，今天的人可能会觉得难以置信：编译器过去也是专有软件，是需要付费购买的。数据库也是另一个例子。曾经如果你要构建一款需要… …数据库的产品，默认选择就是向 Oracle 付费。这些领域通常都是这样开始的，因为最初需要商业力量去提出概念并真正造出产品。

Uh, what happens in these cases are that's how it starts. 'cause you need a commercial effort to even conceptualize these concepts and build them. But eventually the customers of these products end up being massively successful, like way more successful than a company that makes the database. Uh, so if you think about Oracle for example, you know, they made the famous or data in everyone needs that Amazon needed, that Facebook needed that. 但最终，这些产品的客户会成长为极其成功的公司，甚至远比制造数据库的公司更成功。以 Oracle 为例，他们做出了著名的 Oracle Database，所有公司都需要它，Amazon 需要，Facebook 也需要。

Uh, so, you know, these companies became so massive, so much bigger than a company making a database that they're looking at their bill being like, why are we licensing. Millions and millions and millions of dollars to this other company we're so big that we can afford to just do our own thing just to get out of the situation. Um, so eventually there's incentive for open-source efforts to get funded. 后来这些公司变得非常庞大，比数据库厂商大得多。它们看到账单会问：为什么我们每年还要向另一家公司支付数百万、数千万美元的许可费？我们已经大到完全可以自己做一套，从这种关系里摆脱出来。

And if you look at databases, kind of the history of most of them, you put a lot of these like open-source tools, that's like roughly the history, history of them. 于是，资助开源项目的动力就出现了。回顾数据库以及许多开源工具的历史，大致都是这样的过程。

Eventually, uh, the company being the proprietary thing just can't outcompete the incentive to provide a free option. Yeah. Another thing about OpenCode being open-source is that anyone can kind of jump in and contribute. Yeah. So for someone who's just finding out about OpenCode 'cause so many people are learning more and more about it every day, it seems like, like what is your advice for someone who wants to jump in and get involved? Don't contribute. 最终，一家公司依靠专有产品获得的竞争力，抵不过市场提供免费选项的动力。 ——OpenCode 开源也意味着任何人都能参与贡献。对于刚认识 OpenCode 的人——现在似乎每天都有越来越多人了解它——你会给想参与的人什么建议？

Well, it's funny because I, I just talked to all this stuff about how, uh, how how important our positioning is around like, you know, get, get contributions. Yeah. Uh, that was pre, it's funny, we're part of the problem that was pre. Uh, that was before the situation where everyone had access to coding agents and could just prompt anything. So us and every open-source project is now dealing with tons of prs being open, like probably every couple minutes. ——不要贡献。说起来很有意思，因为我刚才一直在讲我们的定位有多依赖大家贡献。那是以前，在… …人人都能使用 coding agent、随手 prompt 任何东西之前，‘欢迎贡献’当然成立。但现在我们和所有开源项目都面临大量 PR 涌入。

I think we probably get a new PR open every couple minutes to it's OpenCode repo. Wow. And it's not very thorough. OpenCode 仓库大概每隔几分钟就会收到一个新 PR。 ——哇。 ——而且通常并不严谨。

It's always people that run to an issue prompt an answer and just the LM tells 'em it's fixed. They maybe don't even run it. They just open the PR or the LM opens the PR on their behalf. Uh, so we're all like kind of underwater trying to deal with this. Um, so if you do wanna be a contributor, you want to improve OpenCode uh, please don't just prompt and send. 人们总是跑到一个 issue 里，让 LLM 生成答案；LLM 告诉他们已经修好，他们可能连运行都不运行，就直接提交 PR，甚至由 LLM 代替他们提交。我们所有人都快被淹没了。所以如果你确实想成为贡献者、改进 OpenCode，请不要只是 prompt 一下就发过来。

I'm begging you, not just me, that's everyone that's in the open-source community, like managing open-source products, projects, some of them have shut down to external computers now, which is kind of a shame. Yeah. 我恳求你，这不只是我的请求，也是所有维护开源项目的人共同的请求。有些项目现在甚至完全关闭了外部贡献，这很可惜。

Um, that said, despite how crazy it is there. We'll always be people that find a way to stand out. So stand up quite through the noise. Yeah. Um, we still have fantastic contributors. We still have, uh, people that kind of can see, they kind of pay attention to what our team is doing, where're spending our time. They kind of see the direction we're going in and they just figure out ways to be useful. Right? Like the thing with open-source, you're creating work for yourself. 不过，尽管情况如此混乱，总会有人找到办法从噪声中脱颖而出。我们仍有非常出色的贡献者。他们会留意团队正在做什么、时间花在哪里、产品准备往哪个方向走，然后自己找到有用的切入点。开源的特点是，你得主动为自己找到工作。

We're not gonna product manage for you. We're not gonna tell you like, here's what you need to do. 我们不会替你做 product management，也不会告诉你：‘这是你该做的事情。 ’

You have to pay attention and, and find a way to deliver something useful for us. And people continue to do that. Like I said, few prs open every minute, every couple weeks we come across like a really, really great contributor. Um, and oftentimes we end up hiring them. I would say probably, uh, every single person that we've hired on a team that I didn't know already from previous work, uh, has been through our community. 你必须持续观察，找到能为我们交付真正价值的方式。人们也一直能做到。虽然每几分钟就有 PR 出现，但大概每隔几周，我们仍会遇到一个非常出色的贡献者，而且常常最终会雇用他们。团队里凡是我在过去工作中不认识、后来才招聘的人，几乎都来自社区。

Um, so yeah, people find a way to stand out, but it's harder and harder given how much noise there is. 人依然能脱颖而出，只是噪声越来越大，让这件事变得越来越难。

So your first tip for anyone who wants to get involved with OpenCode don't contribute. Yeah. You know what the other mess up thing is? Sometimes they contribute using Claude code. And I'm like, why are you, this is like an insult. This is like an insult to injury, you know? Uh, you can give us a slot PR and then you're not gonna even do it without our tool. Like, I can see this, this stupid hilarious whoever and Claude like, Claude commit. 所以你给想参与 OpenCode 的人的第一条建议是：不要贡献。 ——对。更过分的是，有时他们还用 Claude Code 来贡献。我会想：你这是在侮辱我们，简直是伤口上撒盐。给我们一个糟糕的 PR 也就算了，居然还不用我们的工具。

Yeah. And I'm just like, man, I think they're dudes just to hurt us there. There's a period of time where people on our team were kind of convinced that our competitors were like sending uh, people to like open prs to like waste our time. 我能从那种可笑的 Claude commit 里看出来。真的，我怀疑他们就是故意来伤害我们的。有一阵子，团队里有人甚至怀疑…

Um, 'cause yeah, we've gone through a phase where we were a bit more liberal with merging things and yeah, every, it was constantly break stuff, so yeah, we had to become a lot more conservative and people get so mad at us, they're like, oh, you didn't merge my pr It took you guys three months to merge. A very basic pr. But it's, I think people don't understand how, how difficult it's especially against serving millions of people. …竞争对手是不是在派人提交 PR，故意浪费我们的时间。因为我们曾经有一段时间对合并比较宽松，结果代码不断被弄坏。后来我们不得不保守得多，于是人们又非常生气，说：‘你们为什么不合并我的 PR？一个很基础的 PR，居然花了三个月。 ’但大家可能不理解，尤其当产品服务数百万人时，这件事有多难。

Any small thing that breaks affects so many people that we just are super careful. Yeah. You talked a little bit about Claude and Anthropic and you talked before about how. Anthropic has maybe different values than OpenCode that are maybe in conflict. Can you talk a bit more about this? Yeah. I think the thing that a lot of people think, they think that I, uh, it just, this is so funny 'cause there's another aspect of the internet. 任何一个很小的故障都会影响大量用户，所以… …我们必须极其谨慎。 ——你提到 Claude 和 Anthropic，也曾说过 Anthropic 的价值观可能与 OpenCode 不同，甚至存在冲突。

Um, I always, I often come across comments that are summarizing what they think are my views on things. Um, and people think that I like have this crazy extreme view on, on Anthropic I think I have a disagreement with them, but I also kind of can respect the place they're coming from. I think they're looking at their work as we're working on something that is very impactful, just true. Um, and it's gonna be, there's potential for misuse for it. 能进一步谈谈吗？ ——很多人以为我对 Anthropic 持有某种非常极端的看法。这也是互联网很有意思的一面：我经常看到评论替我总结他们认为我持有什么观点。我确实和 Anthropic 有分歧，但也尊重他们的出发点。他们把自己的工作看成是在开发一种影响非常深远的技术，这一点没错。

Um, there's potential for, uh, large negative impact because AI gets developed and they're very concerned about safety and their point of view is,okay, given that we feel that way, we need to be very, uh, aggressive about. How you can use these models, who has access to them? 这种技术可能被滥用，AI 的发展也可能产生巨大的负面影响；他们非常关注安全。因此，他们的立场是：既然我们认为风险如此严重，就必须非常严格地控制模型如何使用、谁能访问。

Anyone that is doing any slightly, uh, different than what Anthropic trick we wants you to do, um, you know, they're gonna clamp down on, and I, I'm, I'm not saying that's coming from a malicious or shameful place. I think it's coming from a genuine place of we're very cons. We're, we're worried and we're kind of scared about this thing. 只要有人做了稍微偏离 Anthropic 期待的事情，他们就会收紧限制。我并不是说这种做法出于恶意或可耻的动机，我认为它来自真诚的担忧：他们真的对这项技术感到忧虑，甚至害怕。

Um, I think what's in conflict for me is I have values that I think can also be respected, which is, this is a very impactful technology. I want to make sure that everybody in the world can have access to it. Um, it needs to be in the hands of as many people as possible. It needs to be in the hands of people that are, have conflicts with each, with each other. 与之冲突的是，我也有一套同样值得尊重的价值观：这是一项影响极其深远的技术，我希望确保世界上的每个人都能使用它。它需要掌握在尽可能多的人手里，需要同时掌握在彼此存在冲突的人手里，也需要掌握在相互竞争的公司手里。

And it's be in the hands of people, companies that are footing each other because that, to me is the only way to guarantee, uh, something stable. Long term. 对我来说，这是保证长期稳定的唯一方式。

I think almost all stability comes from entities that are like locked in, like they're kind of aggressively locked against each other and that kind of provides the stability. Having this be in the hands of a few companies or a few entities that dictate the rules around it. I think historically we've seen that it's not something that ever works out. 我认为，几乎所有稳定都来自相互制衡的实体：它们彼此强力锁定，从而形成稳定。如果这项技术掌握在少数几家公司、少数几个制定规则的实体手中，历史已经反复证明，这种安排不会有好结果。

Yeah. Um, I don't care who you are, I don't care how big of a person you are, when you have that much power and capability, it just, it always corrupts, it's like every single dictator that ever was, uh, looked on historically as a bad person initially had thoughts like this of like, there, maybe they were even, well attention, maybe they thought that I could be the one that uses this power for good. Um, and I don't think that ever really, really works out. 无论你是谁、品格多么高尚，当你拥有如此巨大的权力和能力时，权力最终总会腐化，就像… …历史上每一个后来被视为恶人的独裁者，最初可能也怀有类似想法。他们甚至可能本意良善，相信自己能利用权力做好事，但我不认为这条路最终会成功。

So for me, uh, it's totally in conflict with, with the safety goal because I'm willing to accept the fact that I'm helping, I'm gonna help, uh, LLMs get into the hands of people that wanna do bad things. Uh, and they're gonna probably do bad things with them. 所以我的立场确实和安全目标冲突，因为我愿意接受一个事实：我会帮助 LLM 进入那些想做坏事的人手里，而他们很可能会真的用它做坏事。

I'm accepting that it's a cost we're gonna pay. I think the alternative is, is actually worse. Um, and so I, I can respect Anthropic's position. Yeah. I bet they, if they heard me talk about it in this way, they probably respect my position too. It's just, it's actually an incompatible, you can't actually have both things. Yeah. I wanna talk about this a little more. 我接受这是我们必须付出的代价，因为我认为另一个选择其实更糟。我尊重 Anthropic 的立场。我猜如果他们听我这样解释，也可能尊重我的立场。只是两者确实不兼容，无法同时完全实现。

'cause, so there's a difference between open-source AI is like a general term and then open-weight models and things like that. Yeah. So you believe that Absolutely. Like the dev tools should be open-source Yeah. For all the reasons you just said. Mm-hmm. ——我想再深入一点。大家会笼统地说‘开源 AI’，但 open-source 和 open-weight models 其实并不相同。你坚信开发工具应该开源，理由就是刚才那些。

And then, and then let's talk about like open-weight models. 'Cause I know that that's a big debate. Like what are your thoughts on this? Yeah, so I think, uh, I think every stack of it needs to be open-source If this is a thing people use to do work, produce software or whatever, um, having that be essentially controlled, any aspect of it, uh, is not great. It's okay for there to be proprietary options. 那我们再谈谈 open-weight models，这是一个争议很大的话题。你怎么看？ ——我认为整个技术栈的每一层都需要开源。如果人们用它工作、生产软件或做其他事情，那么任何一个环节受到实质控制都不理想。

It's great. There's proprietary models that are maybe technically the best. It's great there's proprietary tooling around it, but it can be a world where it's, it's just only that we need to also have the, uh, like the free, uh. Free version, but when I say free, I don't mean by cost-wise, I mean just like freely accessible. Um, so yeah, I, I strongly believe in open-weight models as well. 专有选项存在完全没问题，也很好：可以有技术上最强的专有模型，也可以有围绕它的专有工具。但世界不能… …只有专有产品，我们还必须有自由的版本。我说的 free 不是价格免费，而是可以自由获取。所以我也非常支持 open-weight models。

I think there needs to be, uh, not needs to be, I just think there is just natural incentive in the world to fund open-source models and they'll kind of continue to exist whether people want them or not. Yeah. You feel like there are natural incentives? I feel like it's just so hard because obviously there's not a great, there's not a great pla like plan or way to monetize open-weight models, right? 我认为，世界上天然存在资助开源模型的动力，因此无论一些人是否愿意，它们都会持续存在。 ——你觉得这种动力是天然存在的？这件事看起来很难，因为 open-weight models 显然没有一种很好的…

Yeah. So I think there's, I think a lot of people focus on morality when they talk about open-source. Um, I don't like to focus on that. I think, I mean, to be fair, I, I'll give credit some morality argument. There's, uh, I think a good moral reason to make technology that better is fully accessible. Um, but that's never enough to make sure it exists and make sure it like continues to be sustainable. It has to be like real structural reasons for these things to exist. …商业计划或变现方式，对吧？ ——对。很多人谈开源时会集中讨论道德，但我不喜欢只从这个角度论证。公平地说，道德论证确实有价值：让改善世界的技术能够被所有人充分使用，本身有很好的道德理由。但道德从来不足以保证某样东西存在，更不足以保证它持续发展；必须有真正的结构性原因支撑它。

Uh, so from my point of view, the reason open way models will exist is because. A world where there's just two companies that you buy models from is not a world that many large companies with lots of money and power wants. Um, and countries too, right? Like there's adversarial countries that can't use OpenAI and Anthropic you know, they're gonna do their own thing. 在我看来，open-weight models 必然存在，是因为一个全世界只能向两家公司购买模型的局面，并不是那些拥有巨量资本和权力的大公司愿意看到的。国家也一样，对吧？有些彼此对立的国家无法使用 OpenAI 和 Anthropic，它们自然会做自己的模型。

There's very large companies in the US that are, uh, that don't wanna be spending all of their money through like a single vendor, right? 美国的大公司也不愿意把所有资金都花在单一供应商身上。

There are, uh, inference companies like, like, uh, like Baseten that need open-source models to, to exist for, uh, for their business to, uh, selling inference. Um, so I think there's a lot of, just like with any tech, it's very hard for the pure commercial options to like prevent open-source from popping up. So generally I think that's true. I think there's, there's some nuances with open-source, uh, sorry, with, uh, LLMs and complicate it. Um. 还有 Baseten 这样的 inference 公司，它们的业务要成立，也需要开源模型存在。所以和任何技术一样，纯商业产品很难阻止开源方案涌现。总体而言，我认为这个判断成立。当然，LLM 的开源还有一些让问题更复杂的细节。

Obviously the monetary like investment to produce these things is, is really large. Yeah. But also the access to the hardware to, uh, produce these things is very large. Yeah. Um, sorry, the access is very limited. So Yeah. The big, uh, US AI companies plus, like the big tech companies are vacuuming up all GPU capacity. Mm-hmm. Um, they're kind of like spending billions and billions of dollars. 显然，制造这些模型需要的资金投入极其庞大，训练所需硬件也很难获得——抱歉，应该说供给非常有限。美国的大型 AI 公司和大型科技公司正在吸走几乎所有 GPU 产能，投入数十亿、数百亿美元。

The amount of resources being like, available for open-source efforts, you know, that might be a thing that, uh, that's a limiter. So that is something I'm concerned about. But I think long term there is, there's, there's always an incentive to make it exist. I don't, I think the, when it happens, the order happens in how long it takes, uh, that's up for question. Yeah. I'm concerned about this too, because I've kind of heard. 因此，开源项目能够获得多少资源，可能会成为真正的限制因素。这确实让我担忧。但从长期看，总会有让开源模型存在的动力。问题不在于会不会发生，而在于以什么顺序发生、要等多久。 ——我也担心这一点。人们常说 open-weight 能让 AI 民主化，让…

Right. The argument of open-weight and how they can democratize ai, give everyone access, but then it's like the, who owns like the compute power is actually who really might be Yeah. Holding the power of ai, right? Not just the weights, but actually who's going to be having all of these data centers. …每个人都能使用，但真正掌握 AI 权力的，也许是拥有算力的人。不只是权重，而是谁拥有所有这些 data center。

Yeah. So I think that's an interesting part too, is obviously the hardware you need to train these models and that's, you know, you can't have an open-source model unless someone makes an investment. Um, the inference side I think is actually the fun part because it really showcases why open-source is very good. ——对，这也很有意思。训练模型确实需要硬件；除非有人投资，否则开源模型不会凭空出现。但 inference 这一侧反而更有意思，因为它充分展示了开源为什么有价值。

Um, when someone produces an open-source model, anyone can host it. Uh, there, there can be like a hundred companies competing to host that model. You're not forced to buy it from any single person. You need that to exist to drive costs down. Inference costs are crazy. Um, the margins on doing inference are also really crazy, like OpenAI and, uh, Anthropic make a lot of margin when they. 一旦有人发布开源模型，任何人都可以托管它，甚至可能有一百家公司竞相提供同一个模型的托管服务。你不必被迫向某一个供应商购买。这种生态必须存在，才能把成本压下来。 Inference 成本高得惊人，利润率也高得惊人。

When they sell you a token, uh, for that to ever go down, there needs to be like this rich ecosystem of companies competing against each other. OpenAI 和 Anthropic 每卖给你一个 token，都会获得很高的利润。要让价格真正下降，就必须有丰富的公司生态彼此竞争。

Uh, so yeah, there's like a, again, even here because the GPUs are so limited, hardware's so limited. Uh, it might be possible in the short term, these smaller inference players struggle. Um, I hope that's not the case. Uh, we depend on 'em a lot. Um, they need to exist and thrive and compete with each other for prices to go down and price going down is very important. So they would struggle because of that competition you're talking about? 不过即便如此，由于 GPU 和硬件如此紧缺，短期内小型 inference 服务商仍可能举步维艰。我希望不会如此，因为我们高度依赖它们。它们必须存在、成长并相互竞争，价格才能下降，而降价非常重要。 ——它们会因为你刚才说的竞争而陷入困境吗？

Yeah. I mean, if, if, I mean, I, I shared this thing the other day where, uh,like Google, Amazon, all, all the big hyperscalers, these are companies that, like no company in the world makes as much money as 'em, like these, these companies have so much cash that they can direct to anything. They're directing all of it to buying and setting up inference infrastructure. Um, so it's very hard. It's like, yeah, you can go raise, like if you're a very impressive company. ——对，我前几天分享过一件事： Google、Amazon 以及所有大型 hyperscaler，都是全世界最赚钱、现金最充裕的公司。它们正把这些资金几乎全部导向购买和建设 inference 基础设施。

Raising a billion dollar dollars is a crazy number to raise. These companies have tens and tens of billions of dollars to spend every year,so it's very hard to try to get in there and, and do inference at the same scale. Yeah. I wanna talk more about this tweet. I saw hundreds of people replying to this. Uh, can you explain a little bit about like, what's going on in this, in this tweet, like quick refresher? 就算一家非常优秀的创业公司能融资 10 亿美元，这已经是疯狂的数字；但这些巨头每年可以投入数百亿美元，所以想进入这个市场并以同等规模提供 inference，极其困难。我还想谈谈你那条 tweet，下面有数百人回复。能快速回顾一下发生了什么吗？

Yeah, so basically we've been hearing from, uh, our partners, from people we work with, just people we know in the industry, that there is a massive crunch on not just GPUs, but like the whole stack of stuff that is needed to deploy in physical capability. ——我们不断从合作伙伴、业内熟人和其他合作方那里听说，供应紧张的不只是 GPU，而是部署物理算力所需的整套资源都在严重短缺。

Um, everything from like the physical hardware to the labor to deploy it. Uh, there there's a, there's a crunch and because people, it's one of these things where it kind of manifests itself. Uh, not everyone's freaked out, everyone's hoarding, everyone's trying to get as much as they they can, which makes it make, makes the situation even worse. Um, so we're hearing about this and we're like trying to dig into it. Is it overblown? 从实体硬件到负责部署的人力，全部都短缺。而且这种情况会自我强化：每个人都开始恐慌、囤积，想尽可能多地抢占资源，于是短缺进一步恶化。我们听到这些消息后，也一直试图弄清它究竟被夸大了还是真实存在。

Is it real? Every single person we talk to, uh. Is in this hoarding mode. Um, and they're like, it's probably good to secure capacity 'cause we don't even know if it's gonna be available in six months. 我们谈过的每一个人都处于囤积模式，而且都说最好提前锁定产能，因为甚至不知道六个月后还有没有资源可用。

Mm-hmm. Um, so we're starting looking more into it. 'cause as a business, how do we prepare? Um, if token prices double because of this, that's not a good thing for us, not a good thing for our users. What can we do to get ahead of that? Um, and in that process we kind of came across, uh, just details on how much money these big comp, these, these big tech companies are spending, uh, on all this. 所以我们开始进一步研究：作为一家公司，该如何准备？如果 token 价格因此翻倍，对我们和用户都不是好事。我们能提前做什么？研究过程中，我们看到了大型科技公司在这套基础设施上的投入规模。

I think there's an estimate that, um, debt wise is probably gonna be 1.5 trillion borrowed to invest in building out this infras infrastructure over the next,I dunno, five years, whatever it is. Uh, it's getting to the point where like, yeah, the GPU's scarce, but it's even hard to get money, right? If you're an inference provider and you wanna do a build out, you might even have a customer, you might have a customer saying, I want to give you guys a hundred million dollars. 有估算认为，未来大概五年里，为建设这些基础设施而借入的债务可能达到 1.5 万亿美元。现在不只是 GPU 稀缺，连资金都难获得。如果你是一家 inference 服务商，想扩建基础设施，手里甚至可能已经有客户承诺支付 1 亿美元，但你仍需要借钱完成建设；如今即便能证明客户存在，贷款也越来越难。

You need to borrow money to go do that. Build out. It's even getting hard to borrow that money even. Saying you have a customer. So yeah, everything is like really, I never say like, this is my career. Yeah. Like, this is a really unique thing. Um, so I'm going to do it for the first time. 我的职业生涯里从没见过这种情况，真的非常独特，所以我也是第一次公开谈它。

So how does a business prepare if tokens are go token cost is going to double win so and so months? Yeah. I don't know. I'm not saying they will, I'm not saying that maybe they'll just stay flat. Maybe they won't go down. Yeah. Uh, like, like we want 'em to, maybe they'll go up. Um, I don't know, to be honest, we're trying to figure that out right now. I think we're trying to do, we're trying to like overbuy currently. 如果 token 成本会在几个月后翻倍，一家公司该怎么准备？ ——我不知道。我不是说它一定会翻倍，也许价格只是不再下降，也可能上涨。坦白说，我们现在也在寻找答案。

Yeah. So we're trying to build up more capacity than we strictly need so we could scale to whatever we need six months down the road. Okay. Um, I'm hoping this might not be true, but the other thing that I said was, uh, I have seen shortages in my lifetime, never in, in our industry. Um, they kind of all go this way. Where, if you remember like during COVID, Everyone started hoarding toilet paper and like just stuff for your like just paper towels and stuff. 当前的做法是尝试超额购买：建立比眼下严格所需更多的产能，以便六个月后无论增长到什么规模，都还有余量。我希望下面这种判断不成立，不过我还说过… …我一生见过不少短缺，但从没在我们这个行业见过。短缺通常会以类似方式发展。还记得疫情期间，所有人都开始囤卫生纸、厨房纸等日用品吗？

Yeah. And there was this period of time where it felt like, oh, we have a crazy shortage of all this stuff and I feel like it was gonna be bad. But then all of a sudden things kicked into gear and we had an oversupply of all that stuff. 'cause you know, it wasn't actually used as much as we expected. 有一阵子仿佛这些东西出现了严重短缺，局面会越来越糟。但没过多久，生产突然全面启动，最后反而供过于求，因为真实消耗没有大家预期得那么多。

So typically shortages are followed by gluts. It's like a well-studied phenomenon that we can look at this historically it happened just 'cause the whole world mobilizes take advantage of, uh, serving the shortage. Um, that said, producing paper towels or producing wheat, it's a lot simpler than producing A GPU, which is like very complex. Many different countries involved, many different resources involved, politics involved. 所以短缺之后通常会出现过剩，这是历史上有充分研究的现象：全世界都会动员起来满足短缺，从中获利。但生产纸巾或小麦，远比制造 GPU 简单。 GPU 极其复杂，牵涉多个国家、多种资源和政治因素。

So I don't know if it'll have that same exact outcome, but, uh, I'm hoping it does, but yeah, it might be different. Yeah. 因此我不知道这一次是否会得到完全相同的结果。我希望会，但它也可能真的不同。

This leads into something else I'm really curious about. I feel like as a founder, you kind of have to make bets on where things are going. Yeah. But how do you do that when things are moving this fast? I think for a lot of, I mean the, the, the GPU thing aside, I think generally when I find that, I mean I complain about this all the time. I open my Twitter feed every single post prediction, prediction, prediction. 这引出了另一个我很好奇的问题。创始人似乎必须判断未来走向、做出押注，但当变化这么快时，你怎么做到？ ——先把 GPU 问题放在一边。

Everyone's just predicting six months down the road, one month down the road, it's gonna be like this, it's gonna be like that, it's gonna be like that. Twitter 上这类内容让我一直抱怨：每一条帖子都在预测，预测六个月后、一个月后会变成这样或那样。

But, and I get why people feel that way 'cause we're in a phase of great change. So you are trying to really like see if you can think many, many months out. It is so hard to do a prediction because typically a good prediction involves getting not just one assumption, right? It's like stacking six, six assumptions correctly on top of each other and having each one of them be right at the right time. So most of the predictions you see are like wildly wrong. 我理解为什么大家会这样，因为我们正处于巨变期，都想看看自己能否预见几个月后的世界。但做出好预测实在太难，因为它通常不只是一个假设正确，而是要把六个假设层层叠加，并让每一个都在正确时间成立。所以你看到的大多数预测都会错得离谱，这件事本质上就极难。

'cause it's just innately very, very difficult. I think as a company we don't try to do predictions. 作为一家公司，我们并不尝试做预测。

Like we don't try to like really orient ourselves around this is gonna be the outcome or this is this future scenario we're gonna live in. We try to look at, because we talked towards the territory, what is a useful territory to occupy? In that territory. We try to provide some value every single day. Like, how can I help someone today? What's something we can ship today that that helps people? And just kind of take it a day at a time. 我们不会围绕某个预想结果、或某个未来情景来组织公司。我们会问：什么是一块值得占据的领域？在这个领域里，我们每天都尽量创造一点价值：今天怎样帮助一个人？今天能发布什么真正对人有用的东西？然后一天一天往前走。我不知道六个月后会怎样。

I don't know where things are gonna be six months down the road. If you asked me where we are today, six months ago, I would've not known that. I didn't even think we're gonna be in the inference business. 如果六个月前问我今天会在哪里，我也完全答不上来；那时我甚至没想过我们会进入 inference 业务。

I don't think we're gonna be thinking about GPUs. I don't know about any of this. Um, but you know, here we are. I don't know if we're gonna be six months on the road, but, uh, yeah, we try to get, try to stay grounded and, uh, just we have a lot of people that are telling us what we should be doing. 'cause they're trying to use their product and they complaining so we can just listen to them and fix their issues. 我当时完全没想到我们会讨论 GPU，对这一切也一无所知，但现在我们就在这里。六个月后会走到哪里，我还是不知道。我们尽量保持脚踏实地。很多用户会告诉我们该做什么，因为他们在使用产品，也会抱怨；我们只要听他们说，然后解决问题。

Yeah. I wanna talk more about inference, but before, I wanna go back to what you were saying earlier about AI and access. Mm-hmm. ——我还想深入谈 inference，但在那之前，我想回到刚才 AI 与访问权的话题。

Um, some people make the argument that AI should actually be in the hands of a few, like open-weight models, for example, because it's growing increasingly powerful, right? Like they liken it to a weapon now. And I'm curious your thoughts on that. Yeah. So if you liken it to a weapon. Here's a little PR advice for everyone working in AI. stop likening it to a weapon because every time you do this, okay, think about you're a normal person that's not in tech. 有人认为 AI，特别是 open-weight models，随着能力越来越强，反而应该只掌握在少数人手里。他们如今甚至把 AI 比作武器，你怎么看？ ——如果你要把它比作武器，那我给所有 AI 从业者一条公关建议：别再这样比了。想象一下，你只是个不懂科技的普通人，一群 San Francisco 的技术宅却说：‘我们造了一件武器，像核弹一样。

And then a bunch of nerds in SF are like, we have a weapon. It's like a nuke. ’

They're gonna come to your house and cut your head off. Like that's what's gonna happen. Uh, so like it's bad PR move to say we're building nuclear weapons. It's on a scale of nuclear weapons. Uh, you attract a lot of negativity, you attract a lot of government scrutiny. Like, just like, it's not a good, it's not a good way to position things. I also don't think it's true. I think AI is gonna be really powerful. 普通人的反应会是：那就去你家把你处理掉。这就是会发生的事情。宣称自己正在制造核武器、影响规模堪比核武，是糟糕的公关策略，只会招来大量负面关注和政府审查。这不是一种好的定位，而且我也不认为它是真的。

I think it's gonna be really useful and it's gonna impact a lot of parts of the world. Uh, I think there's a little bit of wanting to feel like you're working on a Manhattan project, uh, by people in this space. So it's a little overstated, I think. Um, so yeah, on one hand I'm like, it is overstated. On the other hand, if it is true, it definitely should not be in the hands of the field. Um, like I said earlier, it's uh. AI 会非常强大、非常有用，并影响世界许多领域，但这个行业里有些人似乎很想让自己觉得正在参与… … Manhattan Project，所以多少有些夸大。一方面，我认为这种说法被夸大了；另一方面，就算它是真的，AI 也绝不应该只掌握在少数人手里。

As much as you believe, you want to believe that you have good morality, you have good judgment, you have, you can make the right call In these situations that are, that are difficult, it is very, very hard. 正如前面所说，无论你多么相信自己道德高尚、判断正确，能够在困难情境下做出正确选择，现实都极其艰难。

Um, the right thing only happens at scale. It's like very, it's very hard to rely on one person to always be making the right decision for the whole world. For decades and decades and decades. Uh, you really need chaos and competition and diversity. Uh, and again, this is also a very old concept. You know, it's, uh, people want stability and they, they think stability looks like this. 只有在足够大的规模上，正确的结果才更可能出现。依赖某一个人在几十年里始终替全世界做出正确决定，几乎不可能。你真正需要的是混乱、竞争和多样性。这也是一个非常古老的观念。

But typically when you have a really flat stability, it's punctuated by crazy moments of like earth shattering chaos. 人们想要稳定，并以为稳定是一条平直的线，但这种表面上的平稳，通常会被突如其来、天翻地覆的混乱打断。

And it goes back to stability. That's not, that's not what real disability looks like. What real disability looks like is constant chaos like this. Mm-hmm. That's like always a little uncomfortable, but never like turning into someone dropping a nuke. Right. Um, and if you look like, you know, if you, if you do wanna make a nuclear weapon comparison, that kind of is what the world looks like. Post nuclear weapons, a lot of. Countries have them. 然后又回归平稳。这不是真正的稳定。真正的稳定更像持续的小幅混乱：它总让人有点不舒服，却不会突然演变成有人投下核弹。如果一定要用核武器作比较，核武出现后的世界某种程度上正是如此。

Mm-hmm. Not all of them we would consider are good actors, but it's so far created this weird stability in the world that we really haven't seen before because it's this constant tension that nobody feels happy with, but nobody can actually make the big move. Uh, so I, I think I would definitely encourage people working in this space to kind of look at just these like fundamental dynamics about just like chaos and like what, what actually provides stability. 许多国家拥有核武，其中并非所有国家都被我们视为可靠参与者；但到目前为止，它反而创造了一种此前未曾见过的奇特稳定。这种持续的张力让所有人都不舒服，却也让任何一方都无法迈出毁灭性的一步。所以我会强烈建议这个领域的人研究这些关于混乱与稳定的基本动力。

It's not like, it's not this rosy thing that you'd imagine. Stability actually really looks pretty messy most of the time. Yeah. I hadn't thought about that. It's interesting. Um hmm. 真正提供稳定的，并不是想象中那种美好的画面；稳定在大多数时候看起来其实非常混乱。 ——这个角度我以前没想过，挺有意思。

Yeah, I hadn't, I like that, by the way. This is like, uh, these aren't my ideas. These are like the most cliche ideas that everyone that's trying to get into reading. These are the first topics they, they pick up. Those are my questions for you. Like, it's like, I wanna be an intellectual, I'm gonna read. The same tale. I'm gonna read the Black Swan, I'm gonna read s Ski in the Game. I did all that 10 years ago. Uh, and I went through that whole phase. 顺便说一句，这些都不是我的原创思想，而是最常见的经典观点。每个刚开始认真阅读、想让自己变得更有思想的人，最先都会接触这些主题：‘我要成为知识分子，我要读 Nassim Taleb，我要读《The Black Swan》，我要读《Skin in the Game》。

But you know, the ideas in it are, are true. Uh, and you know, these are, these are the things that uh, like he documents these dynamics really well. ’我十年前也经历过这个阶段。不过书里的观点确实成立，Taleb 对这些动力的记录也非常好。

He goes over historical examples and I can just kind of see that playing out again. Uh, yeah. And then it seems like people are talking more and more about open-weight models and like the everyone versus a few debate. Yeah. The, the open one, uh, the open-weight model thing is interesting because I feel such like a weird inversion with them. 他列举了很多历史案例，而我能看到那些规律再次上演。 ——现在大家似乎越来越多地讨论 open-weight models，也就是‘人人可用’与‘少数人控制’的争论。 ——open-weight model 这件事有一种很奇怪的倒置感。

'cause right now, China, the country that we would think of as very closed or very repressive or very like top down, they're the only ones really releasing open-weight models. 我们通常认为 China 是一个非常封闭、压制、由上而下的国家，但现在真正持续发布 open-weight models 的主要就是 China。

And there's like, like I said with all this, I don't wanna make it about morality. I think the world can be a great place. Even with everyone operating selfishly, I think if you structure incentives correctly, or if you structured systems correctly, everyone can be selfish and still do good. That's kinda what's happening with the Chinese open-weight model, is they're not doing it to be like, here's a gift for the world. 正如我一直说的，我不想把这一切归结为道德。我认为，只要激励和制度设计正确，即便每个人都从自身利益出发，世界仍可以变得很好。 China 的 open-weight models 正是这样。

They're doing it because they're behind open and Anthropic overhead. And for them to actually even get noticed, they have to do things that are open-weight. It's like the only angle of attack. 它们不是把模型当成送给世界的礼物，而是因为在 OpenAI 和 Anthropic 面前处于落后位置。想获得关注，唯一的进攻角度就是开放权重，

Try to commoditize what, uh, the American companies are doing. So I'm not like compliment complimenting them on their morality. I'm saying it is weird that, uh, a country that we have the perception of being very closed off, uh, is doing this open-source thing that is allowing for even companies in the US to flourish. You know, it helps us flourish, helps your company flourish. Um, you know, the open-source models coming from China are very important. 把美国公司正在做的事情商品化。所以我不是在赞扬它们的道德，而是说：一个被我们视为封闭的国家，却正在做开源的事情，甚至让美国公司得以成长，这很奇特。它帮助我们成长，也帮助你的公司成长。来自 China 的开源模型非常重要。

Um, there's US labs now working on them too, which is great. 现在美国的 labs 也开始研发这类模型，这是好事。

Uh, so yeah, I think it's, uh, it, it's weird that it's coming from, it feels inverted. It feels weird that the US. So closed off and we're all like, depending on China, which it's just like a weird, uh, dissonance I think for everyone. Yeah. It does seem like there's going to be more and more good open-weight models coming out of America or like different countries like Mytral and things like that though. 所以这一切确实很反常：美国如此封闭，我们反而都依赖 China，形成一种让所有人都觉得别扭的认知失调。 ——不过 America 以及其他国家似乎会发布越来越多优秀的 open-weight models，比如 Mistral。

Yeah, I think the US was a little delayed because, uh, again, this all comes back to can you convince funding to show up for these things? ——对，美国起步稍晚，原因最终还是能否说服资本为这些项目提供资金。

Um, I think US funding didn't understand the model for the business model for these open-source models Um, I think China doing it anyway now, there's like a thing to point to like, hey, like, you know, here's how their company works. You can kind of see how it works. And I funding for uh, US open-source models showed up. Um, again, nothing like the scale that Anthropic or OpenAI have, but you know, something's better than nothing. 美国的投资者起初不理解开源模型的商业模式。China 先做了起来，现在终于有了可以指给别人看的案例：‘看，这家公司就是这样运转的。’于是美国的开源模型也获得了资金。当然，规模仍远远比不上 Anthropic 或 OpenAI，但有总比没有好。

And what's also interesting is I think, uh, there is a world where some great innovation happens 'cause these companies have less money and we've already seen that some of the Chinese open-source models, like they've done. Really impressive things to train models with. Less money, less compute. Um, so that pressure is gonna be interesting. Uh, so yeah, we'll see. Um, like I said, the incentives are there. 另一个有意思的可能是，这些公司因为资金更少，反而会产生很出色的创新。我们已经看到一些 China 的开源模型用更少资金、更少算力完成训练，成果非常令人佩服。这种压力会带来什么很值得观察。

Uh, it's a very delicate situation right now, but so far they're being produced and open-source models are great. Like they've worked really well. 总之，激励确实存在。当前局面非常微妙，但开源模型一直在产出，而且它们真的很好用。

They're very close to, uh, the closed source ones. And I think even if you're a big fan and you're only ever gonna use one of the closed source ones, it's totally fine. Like I said, everything. There's definitely space for that. Um, even if that's a case, you should still kind of want there to be the open-source option. 'Cause it just creates a healthier ecosystem. 它们已经非常接近闭源模型。即便你是闭源模型的忠实拥护者，今后也只打算用其中一种，这完全没问题；正如我说的，市场当然有专有产品的空间。但即便如此，你仍应该希望开源选项存在，因为它会创造一个更健康的生态。

Yeah. I was reading that, um, open models used to be something like 18 months behind and now it's something like three months. And so it's, the gap is narrowing. ——我读到过，开放模型过去大约落后 18 个月，如今只落后三个月左右，差距正在缩小。

And I'm curious, like, where do you think that's going in the future with this, with this gap between open and closed models? Yeah, I mean it's, that's definitely the feeling that I have too. And it's weird. I, I, it feels like years have passed, but I've only been looking at space for less than a year. Uh, but when I was first starting looking at the space, it felt like there'd be a period of time where, you know, Anthropic or open night would drop a model. 你觉得未来开放与闭源模型之间的差距会如何变化？ ——我的感受也确实如此。说来也怪，感觉已经过了很多年，但我关注这个领域其实还不到一年。

It would just be totally dominant for six months. And the moment the open-source models put something close, like a new version would come out and they did that to stay ahead. 刚开始关注时，Anthropic 或 OpenAI 每发布一个模型，都会完全统治市场六个月；等开源模型刚追到接近的位置，它们就再发布新版本，把领先优势重新拉开。

The past coup the first couple months of this year though, have been interesting because a bunch of open-source models came out. And then like the, the big, the US labs. Had their like drops also like within the same timeframe and they all still seem pretty close. So it feels like that like TikTok thing that's happening is like very, very, very close now. Um, I'm definitely not saying the open-source models are better. 但今年最初几个月很有意思：一批开源模型集中发布，美国大型 labs 也在差不多同一时间发布新模型，可它们看起来依然很接近。这种你来我往的节奏如今已经贴得非常紧。我绝不是说开源模型更好，我依然认为 Anthropic 和 OpenAI 的模型…

Mm-hmm. Um, I still very much think Anthropic models and open models are, are the best, but man, is it easy not to use them these days? Like I use open-source models all the time, uh, and I don't miss the closed source ones that much and that wasn't the case until like February. So OpenCode OpenCode You guys kind of made a bet on people switching between models and wanting to freely switch between them, but can you talk about this and why you kind of focused on this? …是最好的。但现在不用它们竟然如此容易。我一直在使用开源模型，并不会特别想念闭源模型；直到二月，情况还完全不是这样。 ——OpenCode 某种程度上押注了用户会在不同模型之间切换，也希望能自由切换。你们为什么把重点放在这里？

Yeah, so this was kind of one of the reasons why we thought a thought tooling that was model agnostic could be interesting because we saw that this is a very competitive space. You know, OpenAI puts out a model OpenAI has to respond and Anthropic to respond. And there's just gonna be this constant battle of, uh, who has the best model or even like models that do better under certain sort, use cases. Of course, all, all the, all the other models as well. ——我们觉得 model-agnostic 的工具可能有价值，其中一个原因就是看到了… …这个领域竞争极其激烈。一家公司发布模型，OpenAI 和 Anthropic 就必须回应。大家会不断争夺‘最强模型’，而且不同模型可能在特定 use case 下表现更好，当然还有其他所有模型。

Given this amount of chaos, I was like, I don't wanna switch my whole tooling stack every single time I want to try to use another model. Mm-hmm. Um, which is kind of why we, we did OpenCode uh, 'cause I'm like, everything about it changes stays static. 在这种混乱下，我不想每次尝试新模型都更换整套工具链。这也是我们开发 OpenCode 的原因：除模型之外，其他一切都保持不变。

The only thing that changes it is the model that I'm, that I'm selecting. Um, so for us, the more competition there is, the more back and forth there is, the better it is for us. 'cause uh, yeah, I think people kind of get sick of. Using Claude Code one week and developing their workflows around that, them being like, oh, I heard Codex is a lot better. And then you're starting from scratch when you switch over to the Codex tool. 唯一变化的只是我选择的模型。对我们而言，竞争越激烈、领先者来回更替越频繁，就越有利。因为人们会厌倦这种状态：这一周使用 Claude Code 并围绕它建立工作流，下一周听说 Codex 好得多，切去 Codex 工具后却要从头开始。

Not everyone cares, but for the people that do, you know, we, we have that option. Yeah. So it's some people caring and switching and then others just kind of like staying on one model in OpenCode Yeah. 不是每个人都在意，但对在意的人来说，我们提供了另一种选择。 ——所以有些人重视切换，另一些人则一直停留在 OpenCode 里的同一个模型。

I think a lot of people not just OpenCode I think a lot of people just pick their tools and, and stay put. Yeah. And that's again, totally fine. Yeah. Um, there's something wrong with that, but uh, you know, for a lot of people they do want to have the try the newest thing. Um, it's very easy when new model drops to boot up OpenCode and just try it. You might not like it. You might go back to other model, but you got a chance to try it. 我觉得不只是 OpenCode，大多数人都会选定工具后保持不变，这完全没问题。但也有很多人确实想尝试最新模型。新模型发布时，打开 OpenCode 就能直接体验；你可能不喜欢，然后切回原来的模型，但至少试过了。

Uh, the other thing is I don't use just a single no matter what the model is my day-to-day is not just using a single model I use, I usually flip between two. Um, I use one for most of my tasks and, and a different one when I'm trying to bet on like longer tasks or more asynchronous work or bigger tasks. I'm gonna like kind of background. Different models make sense for different use cases. So that's for me why personally I couldn't just use, you know, a single tool. 另外，我自己无论用什么，都不会只用一个模型。日常工作里我通常会在两个模型之间切换：一个处理大多数任务，另一个用于耗时更长、更适合异步执行，或规模更大、可以放到后台完成的任务。不同模型适合不同 use case。这就是我个人无法只使用单一工具的原因。

That makes sense. Yeah. Um, let's talk a little bit more about inference. Um, I think we've kind of talked about it before, but I wanna hear your thoughts on choosing an inference provider. Yeah. ——有道理。我们再谈谈 inference。我知道以前聊过，不过我想听你谈谈如何选择 inference provider。

I think this was, uh, this was kind of why going into building OpenCode I didn't think we were gonna be doing so much with Inference. 'cause in my head I was like, oh yeah, there's, if you wanna use Kimi K whatever, go, go pick any Infras provider and use it. And the first couple months of OpenCode we had so many angry issues being like, fuck you, your tool sucks. It doesn't work. And we'd go into it and it would be because. 开始构建 OpenCode 时，我没想到后来会在 inference 上做这么多工作。当时我想，如果要用 Kimi K2 之类的模型，随便选一家基础设施供应商就行了。但 OpenCode 最初几个月收到了大量愤怒的 issue：‘去你的，你们的工具太差了，根本不能用。

They pick some random inference provider that doesn't really serve on the model correctly. ’我们进一步调查后却发现，是因为用户随便选了一家 inference provider，而它没有正确地提供模型。

And so the end user, they don't know that it's the inference provider is a problem. Uh, 'cause that's the same perception I did. It's just hosting a model, like how can that be variability? But there is a lot of variability. Um, so people kept having bad experiences and I, they would blame us for it. 最终用户不知道问题出在 inference provider，因为我以前也有相同误解：不就是托管一个模型吗，怎么可能有这么大差异？但差异确实很大。用户不断遇到糟糕体验，却把问题归咎于我们。

So we're like, okay, I guess we need to do the work of finding good inference providers and just picking them for people and like having a single place where they can just choose a model and we'll make sure it's routed to, to a good provider. Um, and we've learned a lot about the space. We're not experts in doing inference at all. 于是我们意识到，必须主动找到优质 provider，替用户做选择，让他们可以在一个地方直接选择模型，并确保请求被路由到可靠的 provider。这个过程中我们学到了很多。我们完全不是 inference 专家，

We rely on our partners to do that. Um, but we've learned a lot in terms of how difficult it is when a new model comes out. You don't just like download it and like throw it on your hardware. There's a bunch of tuning you can do. There's a bunch of work with the model maker, with Nvidia to kind of get everything working correctly. There's different optimizations you can make depending on the workload. 真正的 inference 依赖合作伙伴完成。但我们逐渐理解，新模型发布后，绝不是下载下来扔到硬件上就完事了。这里有大量 tuning 工作，需要与模型开发者、Nvidia 协作，才能让所有东西正常运行；还要根据 workload 做不同优化，而 coding workload 和其他 workload 又非常不同。

You know, uh, coding workloads looks so different than other kinds of workloads. Uh. So there's like dedicated smart people that just spend all their time focused on that. Yeah. 有一批聪明的专业人士把全部时间都花在这些事情上。

Um, so yeah, for us it was just trying to find those people, communicating to them what we care about, giving them feedback on when things don't go right. Uh, and kind of codeveloping like a quality product. Yeah. Obviously like your uptime inference is so important. How do you choose an inference provider? Like what's the most important thing when you're looking for one? Yeah. I think for us, number one thing is, uh, that the model is not lobotomized. 所以我们要做的是找到这些人，向他们说明我们重视什么，在出问题时反馈，并共同开发出高质量的产品。——Inference 的 uptime 显然非常重要。你如何选择 provider，最看重什么？——对我们来说，第一重要的是模型不能被‘脑叶切除’。

Like a lot of these, uh, when you go to try something, these weirder inference providers, the model will just be stupid. Which, uh, is really bizarre. It's a really bizarre experience. It's like, it's like model X dumb, but it's really because it's not deployed, right? So it's obviously number one. Yeah. Deployed correctly. Uh, two is speed. Speed is like a very interesting thing, um, especially for coding agents. When you're using a slow model, you kinda get used to it. 尝试一些比较奇怪的 inference provider 时，模型会突然变得很笨，这是一种极其诡异的体验。你会以为‘模型 X 原来这么差’，其实只是部署方式不正确。所以第一点显然是正确部署。第二点是速度。速度对 coding agent 尤其重要：使用慢模型时你会逐渐习惯，可一旦换到快模型，感觉完全像另一种产品。

Then when you use a fast model, it's like, feels like a totally different thing. It's like addicting to use fast models. 高速模型甚至让人上瘾。

Uh, so we, we really appreciate in infras providers that can offer us speed. There's just like an innate like cost and speed trade off where if you wanna make it cheap, you kind of have to make it slower. So you kind of just get delicate balance. Um, and of course uptime's big, but uptime we're a little bit forgiving on uh, 'cause we'll have multiple providers for every model. And it's a new space. Like it's a new space. 因此，我们非常看重 provider 能否提供速度。成本和速度之间存在内在权衡：要把价格压低，通常就得牺牲速度，必须找到微妙平衡。 Uptime 当然也重要，不过我们对此稍微宽容，因为每个模型都会配备多家 provider，而且这是一个新领域，所有人都第一次摸索如何做 inference。

Everyone's figuring out how to do inference for the first time. This isn't gonna be the level of time we expect from like, like compute, which the cloud compute has been a thing for like decades now. 我们不能期待它立刻达到 cloud compute 那样的 uptime；cloud compute 已经发展几十年，稳定性远高于 inference。

It's way more stable than, than inference. Yeah. Okay. I want to read a couple of your inference tweets to you. Are you ready? Okay. So you had one, you said strongly believe open models will eventually become majority of inference. Doesn't mean proprietary ones disappear, but if I take a step back and think about incentives they exist. Yeah. Can you explain a little bit what you meant here? 我想读几条你谈 inference 的 tweet，准备好了吗？其中一条说：‘我坚信，开放模型最终会占据 inference 的大多数份额。这不意味着专有模型消失；但如果退一步观察，支撑开放模型的激励确实存在。

Yeah. So I think there's one perspective where people think, most of the market will always want the smartest option and will pay for the smartest option. it's not completely wrong. ’能解释一下吗？ ——有一种观点认为，市场上大多数人永远会选择最聪明的模型，也愿意为它付费。这个观点并非全错。

I think that the rough thinking behind that makes sense. But as the gaps between the smartest option and the second smartest option shrink, and as willingness to spend on AI right now we're in a phase where every company is like, we have to be AI first. spend whatever you need on ai. We're starting to see people being like, Hey, we're spending a million dollars a month on coding agents. This is crazy. we gotta optimize this. 背后的大致逻辑是合理的。但随着第一名与第二名的差距缩小，再加上企业愿意为 AI 支出的预算发生变化——现在每家公司都说‘我们必须 AI-first，需要多少钱就花多少钱’——已经开始有人意识到：‘我们每个月在 coding agent 上花 100 万美元，这太离谱了，必须优化。’

so at some point there's gonna be some reckoning where there's not gonna be unlimited budgets for this thing. maybe the second best model is good enough given it's 10 times cheaper. 所以某个时刻一定会出现一次清算，预算不可能永远无限。也许第二好的模型已经足够，而且便宜十倍；也许不是。

Maybe not. And this is also a funny thing. one of our customers, their CTO was telling us, 'cause they're using OpenCode not just with their coding, not just with their engineers, their marketing team is also using it and they're like, for the things our marketing team is using, The things that our marketing team is doing, they don't need opus. 还有个有意思的例子：我们有一家客户不仅让工程师使用 OpenCode，营销团队也在用。他们的 CTO 对我们说：‘营销团队做的那些事情根本不需要 Opus。 ’

they keep picking opus. picking ultra high thinking and spending a crap ton of money. and they're like, we want to give them cheaper models that are close to as smart or good enough for them. so I think there's, gonna be a bunch of spend optimization at one point. I don't think there's just infinite desire of we need to have the smartest model at all times or otherwise we're not gonna be competitive. ‘但他们一直选 Opus，还打开 ultra-high thinking，烧掉一大笔钱。我们想给他们更便宜、能力接近或至少足够用的模型。 ’因此，未来一定会出现大量支出优化。

I think that's a false narrative right now. so yeah, I think there's like a lot of incentives to, make you cost-effective You always wanna do that. 我不相信企业会永远认为必须时刻使用最聪明的模型，否则就会失去竞争力；这是一种错误叙事。让成本更有效率的激励始终存在。

so yeah, I think open-source is gonna be a big part of it. I think it's, gonna be a long time before I'd say the majority, and even it being the majority is kind of a bet. But if you look at other markets, the open-source messy, crazy, chaotic part of the world ends up being the majority. I mean, iOS, Android's a perfect example. You know, apple invented this concept. They're the ones that made it super big. They're iconic. Everybody knows the iPhone Everybody has an iPhone. 所以开源会是其中很重要的一部分。距离它真正占据多数可能还很久，而且‘最终占多数’本身也是一种押注。但观察其他市场，开源世界虽然混乱、无序，却常常最终成为多数。 iOS 和 Android 就是完美案例：Apple 发明并做大了智能手机概念，品牌极具标志性，人人都知道 iPhone，身边似乎人人都有 iPhone；可真正占据全球多数份额的是 Android。

Yet the majority of the market is Android. It's very counterintuitive, right? 这很反直觉。

So it wouldn't be surprising if models went the same way. Okay. Yeah. This question might be too vague, but we have this like huge demand for inference, open models, all these things happening. Then we have this, this shortage. And I know you're pretty careful about trying to predict the future, but what do you think might happen? Yeah, I mean, there, there's some scenarios. 所以模型市场沿同一路径发展也不奇怪。 ——这个问题可能太宽泛了：一边是 inference 和开放模型的巨大需求，一边又是供应短缺。我知道你对预测未来很谨慎，但觉得可能发生什么？

So we had, uh, one of our investors, he had like this interesting way of framing it, which is, uh, demand for tokens is going exponential, but production of, uh, the GPU capacity is linear. Like, we haven't had like exponential improvements on how fast we can deploy, uh, uh, inference. So if you agree with this perspective on things, the only conclusion there is token prices are gonna go up because the linear can't keep, keep up with exponential. ——有几种情景。我们的一位投资人有个很有意思的表述：token 需求呈指数增长，而… … GPU 产能只能线性增长。我们并没有让 inference 部署速度实现指数级提升。如果接受这个视角，唯一结论就是 token 价格会上涨，因为线性供给跟不上指数需求。

Um, assuming there's like infinite exponential and like the linear, like the obviously intersection point, but, uh. 当然，这是假设需求会一直指数增长，也忽略了两条曲线最终如何相交，但… …

So I think there's a world where token prices can go up. I, but I also think there's a world where, uh, there is a temporary thing where this supply demand thing isn't great, but eventually we do end up overbuilding Um, I think generally it's impossible to under build, or no, sorry, strategically, you never want to under build because if you under build, you basically go to zero because you miss out on the thing. 确实存在 token 价格上涨的情景。但也可能只是暂时的供需失衡，最终我们会建设过量产能。从战略上看，所有人都不愿建设不足，因为一旦少建，就可能错过整个机会，业务直接归零。所以人人都会倾向于过度建设，这意味着某个时刻我们很可能拥有远超需求的产能。

So everyone's optimizing for overbuilding, so it's very likely that we'll have a lot more capacity than we need at some point. I don't know when that is. That might take years. But, uh, at which point I think we'll see the crisis come down and things being, uh, a competitor. But in the short term, I think it's just complete chaos. Next question. I've been working as a software engineer for a while now. Am I going to be automated? Um. What does it feel like in San Francisco? 我不知道什么时候会发生，也许要几年。到了那时，价格应该会下降，市场也会恢复竞争。但短期内，我觉得就是彻底的混乱。 ——下一个问题：我做软件工程师已经一段时间了，会被自动化取代吗？ San Francisco 现在是什么氛围？

What's the vibe? I think at this point, I've lived in San Francisco for one year, and I think I'm just friends with too many people that are very AGI pilled because I feel, I feel quite worried sometimes and I, I I'm worried I'm too in the bubble of people that are Yeah. Thinking, you know, coding is dead and things like this. Okay. That, yeah, that makes sense. I think my view as people that know me know it's a lot more conservative, uh. We as a team build a coding agent. ——我在那里住了一年，可能认识了太多深信 AGI 的人，所以有时会相当担心，也担心自己过度陷在一个人人都觉得 coding 已死的泡沫里。 ——这很合理。熟悉我的人都知道，我的看法保守得多。

We try to use our coding agent a lot, uh, and we do use our coding agent a lot. And the results are not very good as a team. Like if we step back and look at it as a team, I'm, I'm not even talking about like, oh, the LM said did the wrong code. 我们团队自己构建 coding agent，也尽量大量使用它，事实上确实用得很多。但从团队整体结果来看，并不好。我甚至不是在说 LLM 写错了代码这种局部问题。

I'm just taking a step back. And it's very, very, very difficult for our team not to just completely fuck up our code base to a point where all this annoying stuff is happening. Um, I think that discipline has always been a big part of software engineering. 'cause when you are tasked with building a feature and you're looking at, how can I do this? There is the easy, hacky way that feels bad. 只要退一步看，就会发现团队很难避免把 codebase 彻底弄坏，积累出一大堆烦人的问题。我认为纪律一直是软件工程的重要部分。当你被要求构建一个功能、思考如何实现时，会看到一种简单但 hacky、让人感觉不好的方案，也会看到绝对正确但可能耗时太久的方案，以及两者之间许多不同选择。

There is like the absolute correct way, which maybe takes too long, and there's like a bunch of little options in, in between.

And depending on the scenario you're in, you, you, you, you evaluate all those, you understand the business, you understand your goals. You try to like synthesize all that together and make a decision on what you want to do. When you're approaching that from point of laziness, you kind of always pick the easy hacky version and you pay for that later. You pay for that six months down the road. 你会根据具体情景评估所有选项，理解业务和目标，把这些因素综合起来，再决定采用什么方案。如果从偷懒的角度出发，你几乎总会选择简单的 hacky 版本，然后在六个月后付出代价。

A code base set is fun to work in a product that's fun to work on becomes a miserable product. Uh, down down the road. Even before AI was a thing, we all struggled with this. We all struggled with, nobody was really happy with the code bases they were working in. They started out great, but eventually they sucked. 一个原本开发体验愉快的 codebase、一个原本很有意思的产品，后来会变成让人痛苦的东西。即便 AI 出现以前，我们也一直在和这个问题斗争，没有人真正对代码… …库感到满意。刚开始时都很好，最后却往往变得糟糕。

Um, most teams, there were always a couple people on the team that were like begging the rest of the team to like put that extra mile in, like think, like, think a little bit harder or like, it's okay if things take one more day if we can do a little bit more, right? Um, they were kind of just like kind of holding things together from completely collapsing. 大多数团队里总有一两个人不断恳求其他人多走一步、多想一点：‘如果多花一天就能做得更好，那也没关系。 ’这些人几乎是在勉强维持系统，避免它彻底崩坏。

The interesting thing, interesting thing about coding agents, like the one that I build is that it encourages, kind of brings out our worst tendencies. Coding agent——包括我自己构建的产品——有意思的地方在于，它会鼓励、甚至放大我们最糟糕的倾向。

Um, that hacky solution, you know, isn't right. Doesn't feel that bad to ship 'cause you're not the one like putting it in there physically. And there's, there's this psychological effect where. It's ugly, but I don't have to look at it like, I just told Ellen to do it and LLMs did this ugly thing and it's done. All of a sudden it's done. Like the thing I have to do is done. So I think that's really impacting our judgment in a lot of ways. 那个你明知不正确的 hacky 方案，交付时却没那么难受，因为不是你亲手把它写进去的。这里存在一种心理效应：代码很丑，但我不必直视它；我只是让 LLM 去做，LLM 产出了一堆丑陋的东西，可任务突然就完成了。

And we as a team, like I said, I'm not trying to like make big predictions about the whole world or give you advice for your company. I think I can just talk about what we're, we as a team we're struggling with,which is how do we continue to be good engineers and build systems that are productive to work in, that are reliable, long term, uh, that are fun to work on. Long term codings agents are making that very, very hard. It's a lot harder than it used to be. 于是我认为这正在从很多方面影响我们的判断。就团队而言，我不是要预测全世界，也不是给你的公司提建议，只能谈我们自己正在面对的困难：我们怎样继续做优秀的工程师，构建长期高效、可靠、让人乐于维护的系统？ Coding agent 让这件事变得非常非常困难，比以前难得多。

Um, even some of the stuff that's novel and like interesting I think are kind of backfiring. So one thing that we saw recently is, uh, you know, we are very excited by the fact that our designer can now ship stuff. 甚至一些新颖、有趣的能力也可能适得其反。比如最近我们很兴奋地发现，设计师现在也能直接发布改动。

Like when he see, 'cause designers just see stuff that's broken that the average engineer doesn't in the past. They have to go like, beg the engineer to go fix it. Now the designer can just go fix it and your and your product becomes more and more polished. But they also definitely can't tell when they're shipping in like a bad solution. They can't tell when they're like shipping something in a way that makes a whole system more brittle. 设计师会看到普通工程师注意不到的视觉问题。过去他们必须求工程师修复，现在可以直接自己改，产品也能越来越精致。但他们确实无法判断自己是否提交了一个糟糕方案，也不知道某种改法是否会让整个系统更脆弱。

So even though we're getting this great capability, we're also now trying to understand like, is it worth it? 所以我们获得了很棒的新能力，却也开始追问：它真的值得吗？

Like, 'cause now all this code kind of sneaks in that the engineering team isn't aware of that. Then later they discover, um, so how do we like even work in this way? How do we enable our designer to like, contribute in this way while having guardrails to make sure that there isn't rot that's growing over time? Uh, it's not criticism like the designers and job is not to understand that stuff. It's engineering team's job. 因为这些代码会在工程团队不知情时悄悄进入代码库，直到后来才被发现。我们该如何用这种方式协作？怎样让设计师继续贡献，同时设置 guardrails，确保系统不会随时间逐渐腐烂？这不是在批评设计师，他们的职责本来就不是理解这些工程问题，责任在工程团队。

So yeah, right now we're definitely, coding agents are productive. We enjoy using them, but we are nowhere close to having figured out. 所以目前 coding agent 确实提高了产出，我们也喜欢用，但距离真正弄明白… …

Here's how we, uh, use 'em effectively and here's what our jobs need to look like to create the guardrails and, and do all this stuff. We don't, we don't know. We're still figuring that out right now. Yeah. Gi and to answer your original question, given we're so lost and in there, yeah. We're pretty far from us not having to do anything because we're still so much, we're so, we're, so we're still struggling with this so much. 如何有效使用它、岗位应该怎样调整、如何建立 guardrails，还差得非常远。我们不知道，仍在摸索。回到你最初的问题：既然我们自己还如此迷茫、仍在艰难处理这些问题，那么距离工程师什么都不用做还远得很。

Uh, I think it's really hard to get that level of honesty from people that are excited about these tools. 'cause we all feel productive and we're excited about them. 即便是对这些工具最兴奋的人，也很难保持这种坦诚，因为我们都感觉自己效率提高了，也都很兴奋。

It's really hard to admit that. Yeah. We maybe overstated it or like all this productivity we claimed, if we sit down and analyze it, it's like more incremental than, you know, an order, order magnitude or whatever. Uh, the other thing I'll say is, and I've said this before. We just need to look at output. We focus so much on here's how we build stuff. Here's how we're using coding agents. I have seven Claude codes going out at once. 要承认自己可能夸大了它，或者坐下来分析后发现所谓生产力提升只是渐进式，而不是一个数量级的飞跃，真的很难。另一个我反复说的观点是：只看输出。大家过度关注自己如何构建、如何使用 coding agent，或者‘我同时跑着七个 Claude Code’，这些都不重要，而且把因果关系倒过来了。

Um, none of that stuff matters is backwards.

What matters is, is there a small team absolutely crushing it? Are they absolutely crushing bigger teams? Do we have a competitor that's like completely killing us? Uh, if that's true, then it's worth looking at how they're doing it. I haven't seen that. I haven't seen like a total shift in like power dynamics in this way. Um, there's not a team out there that's like leveraging coding agents way better than us, that are just completely crushing us. 真正重要的是：有没有一支小团队取得压倒性成果，彻底击败大团队？有没有竞争对手因为使用这些工具而把我们打得毫无还手之力？如果有，才值得研究他们如何做到。但我还没有看到这种权力结构的全面改变，也没有哪支团队把 coding agent 用得远胜于我们，并因此完全碾压我们；我们现在依然做得不错。

Um, like we're still doing okay.

So until that starts to happen, it's very unlikely that these tools are actually. Capable of what people are calling you about. Yeah. I think people just don't want to admit that, right? 'cause they don't, everyone's saying like, look how much more productive I am with these tools. So people are scared to look stupid and say, am I really that much more productive than three, four engineers or whatever? Yeah, maybe. Yeah. And this is the thing I talk about a lot. 在这种现象真正出现之前，这些工具大概率还没有达到人们宣称的能力。 ——大家只是不愿承认，对吧？所有人都在说‘看这些工具让我提高了多少效率’，于是没人愿意显得愚蠢，去问自己是否真的抵得上三四个工程师。 ——也许吧。我经常谈到这一点：真正想要‘判断正确’的人非常少。

It's, uh, it's so rare to find people that really wanna be right. I really, really wanna be right. Like I, I care about being right 10 years down the road. 我非常想判断正确，而且在意的是十年后仍然正确。

Um, if I'm like technically wrong a bunch of ways in the short term, that's fine. Um, I think right now it's hard to find people that like, because to, to do that you have to, uh, really criticize yourself a lot. 'cause you really want, if you really wanna be right in the end, you have to like question every assumption you have question every bias that you have. 短期内在很多细节上错了也没关系。现在很难找到同样想法的人，因为要做到这一点，就必须持续严厉审视自己。想在终局上正确，就得质疑自己的每一个假设和偏见。

Um, and in trying to be right in the end, I am landing on a place that's very critical of LLMs because I don't see the stuff that is claimed to the degree that is claimed. 为了最终判断正确，我目前落在了一个对 LLM 非常批判的位置，因为现实中的效果远没有达到宣传中的程度。

I, I then I defeat the retort I get is. Well, that's just gonna be better a year from now, or like Yeah, you said that before and things got better. Um, of course things get better, but fundamentally the problems I'm struggling with haven't changed. And so thinking about architecture and system design and like all my skills of being a programmer for the past 10, 15 years are still very much in play. I'm still using them every single day. 别人会反驳：‘一年后它就会更好’，或者‘你以前也这么说，但它已经进步了。 ’能力当然会进步，可我面对的根本问题并没有改变。对 architecture 和 system design 的思考，以及过去十到十五年积累的 programmer 技能，依然每天都在发挥作用。

Um, yeah. And I, I have a tough time really imagining, and I could definitely be wrong. 我确实很难想象它们会突然失效，当然我也可能错。

Uh, maybe there is a massive shift and everything I'm saying goes away and there are companies operating that's already true, but so far they're not beating us. So I'll worry when that starts to happen. Yeah. One thing you said that I wanna touch on. This is something that is really concerning to me, that I feel like when I am just writing code and not using ai, I'm thinking about it quite hard. 也许真的会发生巨变，让我说的一切都不再成立，也许已经有公司那样运作。但到目前为止，它们还没有击败我们。等这件事开始发生，我再担心。

Yeah. And then if I'm using an AI tool, it is, you've, you've kind of talked about this before, it's really easy to just give into the temptation to hit, like accept changes or whatever. Um, it's really easy to just outsource your thinking, right? It's like your brain wants to go do. Thats, and I do worry about like the, the long-term effects of just like, you know, these solutions that aren't the best, but you're kind of putting them in. ——你刚才提到一点，也是我很担忧的：不使用 AI、自己写代码时，我会思考得非常深入；但一旦使用 AI 工具，就会像你以前谈过的那样… …很容易屈服于诱惑，直接点‘accept changes’。把思考外包出去太容易了，大脑本来就想偷懒。我担心长期不断接受那些并非最佳的方案，会产生什么影响。

It feels good and like curious your thoughts on what that could happen over time. It sounds like your team's not doing that, but I think, um, many people are right now. I, I mean our team is doing that 'cause we're, it's all, it's like a,it's like a personal struggle, right? I think even for myself, I'm like, it's so easy. As I said, it's just so easy to just incrementally every day just get a little bit lazier. 这样做当下感觉很好。听起来你的团队没有这样，但现在很多人都在这么做。 ——我们团队也在这么做，因为这其实是一场… …每个人自己的斗争。即便对我来说，每天只是比昨天再懒一点点也太容易了。

Yeah. And you look, you look back a year and you're just a totally different person how you're operating. Um, and it's, it's so hard. 'cause like now when I go and do something manually, it feels something I like. Never complained about before now feels like so slow or so tedious. Um, and sometimes you do have to do those things. So yeah. 一年后回头看，你的工作方式可能已经完全变了。现在当我必须手工完成某件事时，以前从不抱怨的工作突然显得极其缓慢、繁琐，可有些事情就是必须亲手做。所以这确实很难。

To me it's like a, it's like a, it's such a weird thing because it feels like I'm struggling with an addiction almost. It feels like there's this thing that, it's a complex thing. I, I get a lot of good out of it. It also harms me in some certain ways. I'm like struggling to figure out my relationship with this thing. Uh, it's so weirdly like intimate and I'm like, it's so weird that programming in my career is looking like this now. 对我来说，这种感受甚至有点像在和成瘾斗争。它非常复杂：我从中得到很多好处，它也在某些方面伤害我，而我还在努力弄清自己该和它建立什么关系。这种关系莫名地很私密。想到自己的 programming 生涯如今变成这样，真的很奇怪。

Like we're all, we all just got hit by this thing. Yeah. 我们所有人突然都被这件事击中了。

And we're figuring out where our relationship with it should be. Yeah. And there's so much fear in the air. You've talked about this, but there's so many people trying to like, convince themselves. Or other people that they're like not gonna fall behind in some ways. Yeah, like I think you said in an interview that people are just so worried 'cause there's big changes here and they feel like they're like racing to keep up all of the time. 大家都在寻找与它相处的方式。 ——空气中还有很多恐惧。你以前提到过，很多人试图说服自己或别人，证明自己不会掉队。你在一次采访中似乎说过，面对巨变，人们非常担忧，总觉得自己必须不停追赶。

Yeah, and I think a lot of it is, uh, if you, again talking about like really introspecting and being honest with yourself, uh, here's like the emotions behind it. ——对。如果真正向内审视并诚实面对自己，就会发现背后的情绪是这样的：

The emotions are things are changing when things change. We all feel like there's gonna be winners and losers. You wanna be part of the winners and you know, it is not maybe obvious how you, what like meaningful real thing can do to like, you know, win in this shift. So what people end up doing is kind of performatively talking like what they imagine the winners are like. So there's like this whole world where you know, okay, this shift is gonna happen. 事情正在变化，而变化意味着会有赢家和输家。你想成为赢家，却不清楚究竟该采取什么真正有意义的行动来赢得这次转变。于是，人们会表演性地模仿自己想象中的赢家：既然巨变要来了，赢家大概会极端深入地使用 AI，

All everyone's like, oh, the winners are probably like using AI super hardcore.

They're probably like using it for everything. They're probably running a bunch of 'em in parallel. They're probably doing all this advanced workflow. So everyone's kind of emulating that 'cause they think that. That's what the winners are going to look like. Yeah. The two things that are wrong with that is, one, you don't know if that's what the winners are gonna look like, and two, like either way emulating it isn't gonna help you get there. 把 AI 用在一切事情上，同时并行运行许多 agent，还采用各种高级 workflow。于是大家都去模仿，因为觉得未来赢家就是这样。但这里有两个错误：第一，你根本不知道赢家是否真的长这样；第二，无论如何，模仿表象都不会让你成为赢家。

So, and there, there's a lot of that in the air. Uh, I don't blame anyone for it. Like, I I, but like I understand the motions behind it. It's, uh, yeah, it's really hard, uh, to not do that. But, uh, I mean that, that, that's what kind of creates a lot of noise. Yeah. That's a good point. 现在到处都是这种现象。我不责怪任何人，因为我理解背后的情绪，克制这种冲动非常困难。但它确实制造了大量噪声。

You see the, you see the people talking about like, I am great at using this AI tool, and then you're trying to read their articles, kind of copy them, figuring out, not you, but Yeah, yeah, yeah. People are doing this or, or I have this urge too, like, oh, look how they're doing this productive thing with ai. Like, gotta learn this right now, kind of thing. Yeah. It's funny 'cause I was thinking about, I was on the way here. 你看到有人说‘我特别擅长使用这个 AI 工具’，就去读他们的文章、尝试照着做。不是说你，而是很多人这样，我自己也有这种冲动：‘看，他用 AI 做得这么高效，我现在就得学会。 ’——对。今天过来的路上，我正好想到 Mitchell Hashimoto。

I was thinking about, uh, uh, Mitchell Hashimoto, he's the, uh, founder of Terra, uh, founder of HashiCorp. They made like a bunch of really impactful dev tools. 他是 Terraform 和 HashiCorp 的创始人，做出了许多影响深远的开发工具。

Uh, a big part of like my earlier career, like using them. Super, super successful guy. Uh, then now he's working on Ghostie, like in his like pseudo retirement. Another very great polished product and he talks a lot about how he like, you know, he's not like anti, he's obviously a extremely good programmer, extremely competent product person. Like, just very, very talented overall. And he's definitely using AI to, to build all these things now. 我职业生涯早期大量使用过那些工具。他极其成功，现在处于半退休状态，正在开发 Ghostty，又是一款非常优秀、精致的产品。他并不反 AI，自己显然也是极其优秀的 programmer 和产品人才，整体能力非常强；如今他也确实在用 AI 构建这些东西。

And I'm like, I'm looking at him, I'm like, yeah, I still can't compete with him. Like nothing I can do today with AI is going to get me to be better than him. 我看着他会想：对，我依然竞争不过他。今天无论怎样使用 AI，都不会让我突然变得比他更强。

'cause he is just so good at his job, so experienced, so talented. AI is only helping him like get even further away from like the average person. So there's no world where I had I, all I have to do is adopt AI harder than him and somehow I'm better than him now. It's not that like his fundamental skills are very good. I need to work on my fundamental skills to get them that good. Uh, and the AI is like kind of irrelevant to that whole process. 因为他的专业能力、经验和天赋都太强，AI 只会帮助他进一步拉大与普通人的距离。不存在这样一个世界：我只要比他更激进地采用 AI，就能突然超过他。真正重要的是，他的基本功非常扎实；我需要磨炼自己的基本功，达到同样高度。 AI 与这个过程本身几乎无关。

Yeah. I feel like that's a great point. It's like comparative advantage. Mm-hmm. ——这个观点很好，像 comparative advantage。

I saw a Dario interview recently where he talked about, so even if AI is writing 95% of a code, what does that 5% human part? And if you're better at that than anyone else, than you have this like possibly very large comparative advantage to anyone else. Yeah. And I think the, the thing that gets lost there is that 5% can be coding. It can be that you're very, very, very good at programming. 我最近看了 Dario 的一次采访。他说，即便 AI 写了 95% 的代码，剩下那 5% 人类工作究竟是什么？如果你在这 5% 上比所有人都强，就可能拥有相对别人非常巨大的 comparative advantage。 ——对。这里常被忽略的是，那 5% 仍然可以是 coding，可以是你极其擅长 programming。

Even if AI is writing 90% of the code, uh, someone that's a very, very good programmer still is much better than you if they're better than you. 即便 AI 写了 90% 的代码，一个真正优秀的 programmer 如果本来就比你强，最后依然会远胜于你。

Right. So, uh, it doesn't have to be, I think people kind of, I, I just see stuff get repeated all the time where I know no one's thinking about it. They're like, it's gonna be just like being a manager. It's like, we're all gonna be managers now. We're gonna be managing agents. I'm like, I was the middle manager for a little bit. It is nothing like that. Skillset set is nothing like managing agents. 它不必变成其他工作。我总看到一些人未经思考就重复：‘未来就像当经理，我们都会变成 manager，负责管理 agents。 ’我做过一段时间 middle manager，这两件事完全不同。

Like we're not, you're not a manager when you're telling an agent with do, it's like a totally different thing. I do say that a lot. It's like you'd be like a manager and I've never been a manager, so I'm thinking, what, what does that mean? 管理 agent 所需的技能与管理人毫不相似；告诉 agent 做什么，并不意味着你成了 manager。 ——我也经常听到这种说法。没做过 manager 的人会想，这到底是什么意思？

Being a manager is all about motivating people and getting people excited to work on stuff and like run through walls every day that like, you don't need to do that when you motivate your agent. Yeah. It's like, that's like not a thing. Like you're just like, prompting is nowhere near as hard as being like a, a manager of real people. Yeah. Um, yeah, so I, I think, I don't think that's, we're gonna be managers or be managing agents. 管理人的核心是激励他们，让他们对工作产生热情，愿意每天克服巨大困难。你完全不需要这样‘激励’agent。 Prompting 的难度远低于管理真实的人。所以我不认为未来我们的工作就是当 manager、管理 agents。

I think it's, my job still feels roughly the same. Like Yeah. The details of what I'm doing every single day or like what I'm typing into a computer are different. Yeah. 我的工作整体上仍与过去相似；每天具体做什么、往计算机里输入什么确实不同，

But the painful parts of my job, when I say painful, I mean like where I struggle, where I exert a lot of energy, where like stuff is on my mind. The stuff that I'm thinking about. Exact same stuff I was thinking about 10 years ago. So, yeah. Even if this, this 5% that we're talking about. I don't think we're all gonna be like prime project managers or like human managers, whatever it is. Yeah. I don't think people are thinking as much about that 5% too. 但真正让我痛苦的部分——需要挣扎、消耗大量精力、一直在脑中思考的事情——和十年前完全相同。所以即便只剩我们所说的 5%，我也不认为大家都会成为纯粹的 project manager 或 people manager。

Like, I hear people sometimes, like, I think the rationale behind just hitting accept changes or like not being so in the code anymore or, or really like not having to learn hard things. Like I think some people genuinely believe, like some people genuinely believe that you don't need to learn hard things anymore because AI is becoming so intelligent. ——我觉得人们也没有认真思考这 5%。比如直接点… … ‘accept changes’、不再深入代码，甚至不再学习困难的东西，背后的逻辑似乎是：有人真的相信，AI 变得如此聪明，以后已经不需要学习难题了。

And I feel like that sounds kind of great in some ways 'cause you're like, oh, I don't have to go do this really hard studying. Like, I don't have to keep doing my gpu Yeah. Kernel engineering study group or whatever. Um, but then on the other hand, like if that doesn't come true or like then you're just setting yourself up to not be learning hard things in the future. Yeah. No. Which seems bad. Yeah. And it's like the fundamental nature of the world hasn't changed. 这个想法在某些方面听起来很诱人：不用再做艰苦学习，不用继续参加 GPU kernel engineering 学习小组。但另一方面，如果这件事没有成真，或者之后… …你只是让自己失去了继续学习难题的能力，那就很糟糕。

Like from a like microscopic level nature's about competition. It's about like trying to do stuff a little better than people around you and that. Is true in society. It is true in like the animal world. It's, it's just always true. So I don't know where this thing of like, oh, we're all just gonna be vacationing all the time comes from, it's just like there's gonna be people that want to do things and like to like, compete with them to like, try to do something similar. ——对。世界的基本性质并没有改变。从微观层面看，自然界就是竞争，就是努力把事情做得比周围的人好一点；社会如此，动物世界也如此，一直都是如此。所以我不知道‘未来所有人只需要一直度假’这种想法从哪里来。总会有人想要做事情；要和他们竞争、取得类似成果，你就必须非常优秀。

You're gonna have to be really good. Um, it might look like the details of what that is might look different. The people that were like exceptional at their time a hundred years ago, their day-to-day look very different than mine. But I bet their mindset and their effort and their energy and their motivations, I bet all that look pretty similar. So I don't think this stuff really changes that much. Um, were all still gonna have to do hard things. 具体工作内容也许会变化。一百年前最杰出的人，日常生活与我完全不同，但我敢说，他们的思维方式、付出的努力和精力、内在动机都与今天非常相似。所以我不认为这些东西会发生太大变化，我们仍然必须做困难的事。

Uh, also gonna have to exert energy. I don't know how you can live a life. That's, to me, that's kind of the point of life. Like it's exert a bunch of energy, you know, like, yeah. Otherwise, how do you feel alive or feel like you're living a life worth living? You know? It's, it's, it's, uh. It's a very, like, sad, even if it's true, it's like a very sad situation. You've talked in the past about how being a founder is figuring out how to stay motivated every single day. 也仍然必须投入精力。我不知道完全不投入精力的人生该怎么过。对我来说，这本来就是生命的意义：投入大量精力。否则怎么感到自己活着，怎么觉得过的是值得的人生？即便‘什么都不用做’真的实现，那也会是非常悲伤的局面。

Is that right? Can you talk more about that? Yeah. Like I said earlier, uh, this company's been around for 16 years,so we know a little bit about trying to stay in the game forever. Uh, yeah. It's just, it's this crazy thing. I mean, everyone, it is one of those things where when you start a company, you'll go read advice about starting companies and you'll come across as advice. It's very old advice. So basic, which is just stay alive. Stay alive for as long as you can. ——你以前说过，做 founder 就是想办法让自己每天保持动力，对吗？能多谈谈吗？ ——正如前面说的，我们公司已经存在 16 年了。所以对于怎样长期留在牌桌上，我们确实知道一点。创业时你会读各种建议，其中一定会遇到一条古老又最基础的建议：活下去，尽可能长久地活下去。

Uh, if you stay alive for long enough, you then you get rich all of a sudden. But that's, it's like, it's like kind of bra in that way, or like, you're just successful all of a sudden. All I have to do is stay alive. And it sounds so simple. 只要活得足够久，某一天就会突然成功、突然有钱。听起来像一句陈词滥调：我只要活下去就行了，多简单。

It sounds so stupid. It is both the simplest thing ever. 'cause we all know what that means. But it's the hardest thing ever because staying alive, staying motivated to continue doing work every single day, when some days it feels like you're doing everything wrong, when you take swings that are like complete misses, it's very difficult. 它既是最简单的事，因为每个人都明白含义；也是最困难的事，因为所谓活下去，意味着每天都要保持动力、继续工作，即便有些日子觉得自己做错了所有事情，即便一次尝试彻底落空。

Um, but, and the truth of the world is there's like a few people that are both very smart and very lucky and like they kind of work on something. 这非常难。现实中当然有少数人既聪明又幸运，碰巧做了某件事，

It's a right timing, it's a right idea, or they're like smart enough to see it and it's like an instant success. And I think we see a lot of those stories, but the rest of us aren't that smart and aren't that lucky. Uh, a way you can compensate for that is by being in the game for super long. 'cause eventually out of sheer like numbers, like something lines up correctly for you. 时间和想法都恰到好处，或者他们聪明到足以提前看见机会，于是立刻成功。我们经常看到这些故事，但其余人没那么聪明，也没那么幸运。补偿办法就是在游戏里待得足够久，因为仅凭次数积累，最终也会有某件事恰好与你对齐。

Um, we've seen so many companies like come into this space over the years, get further than we did, but eventually die and kind of, we still continue to exist. So yeah, it is this weird thing where. 多年来，我们看过许多公司进入这个领域，走得比我们更远，最后却死掉，而我们依然存在。所以这是一件很奇怪的事。

Every company has its own unique trajectory, but for us it's been this game of how do we stay motivated and extended every single day and believing that if we can, if we're able to do that, there's gonna be some kind of outcome eventually. Yeah. So you talk about taking a lot of different swings. Yeah. Um, and staying motivated. But now with the swing of OpenCode I mean, it's doing so incredibly well. So how does that change the motivation? 每家公司都有独特轨迹；对我们而言，核心问题一直是如何每天保持动力、延长公司的生命，并相信只要能继续，总会在某个时刻得到结果。 ——你谈到许多次尝试和保持动力，但 OpenCode 这一次做得实在太好了。这会怎样改变动力？ ——有意思的是，我觉得公司的内在 DNA 并没有改变。

Yeah, it's uh, it's funny 'cause I think our inherent DNA of the company has not changed. 'cause now we're thinking about things like, okay, this is doing well and we're gonna obviously gonna see this out and we're gonna kind of continue down this path. But we really want this to be a forever company. There's very few companies that become forever companies. And what we mean by that is no matter what you get interested in the future, you do it to the same company. 现在我们会想：好，这件事发展顺利，我们当然会坚持到底、继续沿这条路走。但我们真正想打造的是一家‘永续公司’，而能做到这一点的公司极少。我们的定义是：未来无论对什么产生兴趣，都在同一家公司里完成。对我们而言，OpenCode 会是我们在 dev tools 领域做的最后一件事。

Um, so for us, this is gonna be the last thing we ever do in dev tools. Uh, we've been doing stuff in dev tools. This is, and trying to look for something that can hit the whole dev tools market. 我们一直在 dev tools 市场寻找一个可以覆盖整个市场的产品。

We found that thing, whether or not this success, this is successful or not, whether or not this is like massively successful kind of, uh, successful at a mediocre level. This is our final dev tools thing ever. So if you like me in dev tools, like you're not gonna see me again. Half of this, uh, 'cause now we're thinking about, okay, the next thing we do needs to be bigger. 现在已经找到了。无论它最终大获成功，还是只取得中等程度的成功，这都是我们最后一次做 dev tools。所以如果你喜欢我做开发工具，以后可能不会再看到我做下一款。因为我们现在正在思考：下一件事必须更大。

You know, we're, so we're trying to set up the company so that can we build something that is now for the whole world, not just for developers, not just for programmers. I don't know what that product is gonna be. 能不能把公司准备好，下一次为全世界构建产品，而不只面向开发者和 programmer？我还不知道那会是什么产品。

Uh, this thing that we're kind of talking about, it might happen five years down the road, I don't know. But we're not like gonna do this and retire. There's gonna be more stuff we're interested in. There's gonna be more stuff that pulls our attention. There's something we're gonna get excited about one day and not be able to stop thinking about. And we wanna structure the company in a way where it makes sense to continue to do that. At our current company. 我们谈的事情也许五年后才发生，我不知道。但我们不会做完 OpenCode 就退休。未来还会有其他事引起兴趣、吸引注意；某天总会有一件事让我们兴奋到无法停止思考。我们希望把公司组织成一种形态，让自己届时仍能在当前公司里继续做它。

Yeah. Something else you said that I'd love to hear the story of this is that, so you've had different acquisition offers that you said no to and you posted one in your company Slack, like, Hey, we're getting software, and everyone ignored it. Your team ignored. Yeah. I was super proud of our team for this. Um, obviously with how crazy AI is, there's like lots of money flying around. There's opportunities to like get a life changing amount of money. ——你还提过一件很有意思的事：你们拒绝了不同的收购邀约。有一次你把… …某个 offer 发到公司 Slack，说‘有人想收购我们’，结果所有人都忽略了，团队完全没回应。 ——对，我非常为团队感到自豪。 AI 现在如此疯狂，到处都是钱，也不断出现能让人一夜改变人生的机会。

Like that's, it's not is in a time like this, those offers are kind of showing up. Um, and yeah, I like, you know, I talked to, uh, this company, uh, their CEO called me and mentioned like, Hey, would you guys be interested in joining us? We'd like to acquire you. And posted that in our team Discord. 在这样的时期，收购 offer 自然会找上门。某家公司的 CEO 给我打电话说：‘你们有兴趣加入我们吗？我们想收购你们。 ’我把这件事发到了团队 Discord。

And like, everyone's ignored me. They just kind of kept talking about work. 'cause everyone was so into, we're building what we're doing, excited by like the, the, the amount of users we were getting all stuff they wanted to do with them. And yeah, they just didn't care that much. And it's this bizarre thing where when I first became a founder, I was like, man. I would love to get acquired one day like that. That would be like the ultimate outcome. 所有人都无视了我，继续讨论工作，因为大家完全沉浸在正在构建的东西里，为快速增长的用户和想为他们做的一切而兴奋，根本没那么在意收购。说来很奇怪，我刚成为 founder 时会想：有一天能被收购就太好了，那是终极结果。

To be able to sell your company and make a bunch of money that's like why I'm doing this. And it would feel ridiculous to ever be in a position where you don't really care about that. 卖掉公司、赚很多钱，这正是我创业的原因。我从没想过自己会走到一个根本不在乎收购的位置。

But I think when you're, when you've been doing stuff like this for so long, you've just been on the hunt for these opportunities where you can build something for so many, so many, so many people. And when you find it now you're like, okay, now I get to do the thing that I've been looking for for all this time. Selling it at that point. It just feels like, like the saddest thing ever. 但当你做这种事情做得足够久，一直寻找一个能让自己为非常非常多人构建产品的机会，终于找到时，就会觉得：我终于可以做过去一直在寻找的那件事了。此时卖掉它，感觉会是最悲伤的选择。

Like all it feels like all the dreams you had for the past, you know, decade or whatever you were trying to build something, it feels like they all die when you do that. 过去十年为了构建某种东西而怀有的所有梦想，似乎都会在出售那一刻死去。

It feels like everything, you spent all those, all that time swinging and missing, it feels like all that was pointless. Um, so yeah, it's a very like, counterintuitive thing. I'm not saying my feeling that is infinite. I'm sure someone wants to overpay to buy our company. Like of course I'm gonna take it. I'm not like super unrealistic in that way, but uh. I think right now it's like, this is the thing we've been looking for. 你会觉得此前所有尝试和失败都变得毫无意义。所以这是一种非常反直觉的感受。我不是说这种心态永远不会改变。如果有人愿意用离谱的高价收购，我当然会接受，我没那么不现实。但就现在而言，OpenCode 正是我们一直寻找的东西。

So yeah. As a founder, like this is what you're looking for. This is because it's kind of why you play the game. This is not the point where you, where you leave. Yeah. 作为 founder，这就是你进入这场游戏想要找到的机会，并不是应该退出的时刻。

And so you talked about how now obviously the market is hundreds of millions of developers. It's, it's so big. But you talked about this a little bit earlier, but how do you focus on when there's so many people telling you what they want? There's so many different developers and there's all this speculation on the future of what IDs will even look like. Like how do you decide what your users like want when there's so many? 如今面对的是数亿开发者，市场显然极其庞大。但有这么多人告诉你他们想要什么，开发者群体又如此多样，大家还在猜测未来 IDE 会变成什么样，你如何判断用户真正需要什么？ ——我总是告诉创业者，不要尝试为别人构建产品，因为只有极少数天赋异禀的人能做到。

Yeah, I think, uh, I always tell people if you're founding a company, don't try to build a product for somebody else because only like a very gifted person can do that. It is a lot easier just to build a product for yourself because you are the user. You understand it. People always try to like, identify a gap in some industry and try to like build a product for the industry that sucks. Like it's, it's, it's very hard to do. 更容易的方式是为自己构建，因为你自己就是用户。你理解自己的需求。人们总想在某个行业里找一道缺口，再为那个行业构建产品，但这通常很糟糕，也极难做到。

Uh, what's nice is the whole team is the programmers. We're building tools for programmers, very easy for us to use our own products and understand what we like, what we don't like. Um, and then, so that's like primarily where our motivation comes from in terms of what we prioritize. 我们的好处是整个团队都是 programmer，正在为 programmer 构建工具，所以很容易亲自使用产品、理解喜欢和不喜欢什么。产品优先级的动力主要来自这里。

Um, and of course we spend a lot of time talking to our users or paying customers, like big companies that are using it. Um, when you do that, these things stop feeling like predictions. They just feel like massive holes that are like screaming at you. Um, so like we're, we're, you know, we're, we're kind of working on a new phase of our product now, and I think when we launch it, and if it goes well, people kind of look at it in hindsight. 当然，我们也会花大量时间与用户或付费客户交谈，包括使用 OpenCode 的大公司。一旦这么做，产品方向就不再像预测，而像一个个对你大声呼喊的巨大缺口。现在我们正在开发产品的新阶段。如果发布成功，人们事后回看，也许会觉得……

Like, whoa, they like saw this crazy thing coming. They like totally like, guessed it or predicted it, but it's not really what happened. It's that we talked to a few, like, maybe like three CTOs that kind of all said roughly the same. They, they said different, specific things, but we pay attention. They were kind of all three saying the same thing. Effectively telling us exactly what we needed to do. Right. So, uh, yeah, that, that part's really fun. ‘哇，他们预见到了这么疯狂的趋势，完全猜中了未来。 ’但真实过程并非如此。我们只是和大概三位 CTO 交谈，他们具体说的事情不同，但仔细听就会发现，三个人实质上都在表达同一个问题，几乎直接告诉了我们应该做什么。这个过程很有意思：从用户的话里提炼出他们真正的意思。

It's kind of teasing out what people actually mean. Um, so yeah, like we try to talk to our users. 所以我们确实会持续和用户交流。

Like I said, a lot of every piece of feedback is useful in that, uh, it is a real data point, um, but understanding the root of the issue is what you need to do, and no one's gonna tell you what the root of the issue is. That makes sense. Yeah. Awesome. Okay. I wanna talk about a couple more things. Yeah. We're pretty good on time. Every time a dog barks or something, I'm like, oh, no, we'll edit it out. Yeah. I wanna talk about your, I wanna talk a little bit about Zen. 每一条反馈都有价值，因为它都是一个真实 data point；但真正需要做的是理解问题根源，而没有人会直接告诉你根源是什么。 ——有道理。我们时间还很充裕。我想再聊几件事。每次狗叫之类的声音出现，我都会担心。 ——后期剪掉就行。 ——我想谈谈 Zen。

Mm-hmm.

Um, you had this launch video that was essentially you walking around in your backyard. Yeah. And I thought this was great because first of all, it did very well. Um, but second most launch videos are these incredibly overproduced people are putting thousands of dollars into these things. Yeah. Um, so can you talk a little bit about your decision of making your launch video as you walk around your backyard? 它的发布视频基本就是你在自家后院边走边说。我觉得很棒：一是效果非常好，二是如今大多数发布视频都制作得过度精致，人们为此花费数千美元。为什么决定在后院散步时拍发布视频？

Yeah. The, the thing, uh, the thing I like doing, so I, uh, am super into marketing now. It's like a, like at the company, marketing is assigned to me. ——现在我非常投入 marketing，公司里的 marketing 工作归我负责。

It's like my responsibility, this is so funny for me to say out loud because a couple years ago I knew nothing about marketing. I'd be like, what the hell? That's like someone else's job. Like, that's not, I'm a programmer. I like build, build products. Uh, but at some point I got interested in marketing just through necessity. Like we needed to get our stuff out there. And it's, it took me a lot to wrap my head around it. 亲口说出来都觉得好笑，因为几年前我对 marketing 一无所知，还会想：‘这到底是什么？那是别人的工作。我是 programmer，只负责构建产品。 ’但后来因为现实需要，我们必须让外界看到产品，我开始对 marketing 感兴趣。

But the area that I landed, this is what works for me, is. 理解它花了很久，最后找到了一套对我有效的方法。

As someone that is very online, I am constantly ingesting other companies. Marketing. I'm seeing what everyone's doing. And patterns like emerge very quickly. Like trends emerge very, very quickly. One company will do a thing, all of a sudden everyone is kind of doing the same thing. 作为一个高度在线的人，我一直在接收其他公司的 marketing 内容，因此很快就能看出模式和趋势：一家公司做了一件事，突然所有人都开始照着做。

Um, it may not seem that obvious to like just someone that's naturally experiencing it, but given that I'm looking at it from that lens, I can immediately tie they're doing that because they're copying this company, this company's copying that company with like this exact twist. 普通人自然接触这些内容时也许看不明显，但我从 marketing 的视角观察，立刻就能看出这家公司在模仿谁、另一家公司又在同一套路上加了什么具体变化。

Uh, so I try to stay on top of all of those if like, whatever the meta is, right? Mm-hmm. And then I try to do these exact opposite. And I'm like, how? Because like if I just do another one of those things, if they spent, you know, $10,000 on it and I spent $10,000 on mine, I'm gonna get at best the same outcome as them. Realistically a worse outcome. 'cause they've already done it before. 所以我会跟踪当下所有流行的 meta，然后尝试做完全相反的事情。因为如果别人花 1 万美元做一支同类视频，我也花 1 万美元，最好也只能获得相同结果；更现实的情况是更差，因为别人已经先做过了。

Uh, so I try to understand like what the current expectation is and try to like subvert it in some way. 因此我会先理解当前预期，再尝试以某种方式颠覆它。

Uh, and so yeah, there was a period of time where uh, every like tiny launch was like getting this, I mean this kind of treatment. Uh, and it was always for something like really boring, uh, or like, just like most basic launch. And it just felt, and there was a time where that felt really cool 'cause it was like nothing in tech was well produced. Then all of a sudden stuff was being well produced and that was awesome. 有一阵子，每一个极小的产品发布都会得到这种大片式制作，而且发布的往往是极其无聊或基础的功能。最初这确实很酷，因为科技行业过去几乎没有制作精良的内容，突然出现高质量视频让人眼前一亮。

But then everyone was just kind of doing the exact same thing and it felt a little cliche. 但随后所有人都开始重复同一种形式，它就变得有些俗套。

And I can see the great thing about, uh, the something posted online is it's not a Instagram whether or not succeeded, whether or not it succeeded. You can see how many views it got. You can see the likes it got. And you can see the measurement of they, they probably spent like 50 K on this and they got not a great outcome. Yeah. Uh, yeah. 线上内容有个好处：它不像传统广告，是否成功完全可见。你能看到播放量和点赞数，也能判断对方可能花了 5 万美元，却没有得到很好的结果。

So, and I was like, I think the interesting for us to do at this point is go the exact opposite and do the most down to earth thing where it's just me talking on the phone and kind of explaining what we're doing and why. 所以我觉得当时最有意思的做法就是走向完全相反的一端，制作最朴实的内容：只有我对着手机，解释我们在做什么、为什么做。

Keep it short, don't be salesy. Just kind of say what our motivation is for doing this and be so like direct, uh, and do it in my backyard walking around and yeah, it got like, you know, it did really well and yeah, I spent 30 minutes doing it. Uh, literally $0. Yeah. So, yeah, I'm not saying every single thing we do is gonna be low budget. Like the point isn't copying the thing, it's understanding the sentiment that no one's articulating and kind of doing something for it. 视频保持简短，不要有推销感，只直接讲清我们的动机，然后在自家后院边走边录。最后效果很好，我只花了 30 分钟，成本真的为零。当然，我不是说今后所有内容都要低预算。重点不是复制某种形式，而是理解那种尚未被说出口的情绪，并做出回应。

Yeah.

And I do feel like you are really authentic on Twitter, which seems really rare these days. Like I feel like everything you're posting is like what you're currently thinking now. Like you, you kind of talk about how you're thinking about like heuristics for predicting the future and, and you're very much like, I don't wanna say build in public, 'cause I feel like that has become kind of cliche at this point. 我觉得你在 Twitter 上非常真实，这在今天很少见。你发的似乎都是当下真正在思考的东西，比如分享自己如何建立预测未来的 heuristic。我不想说 build in public，因为这个词如今也有点陈词滥调，但你确实经常这样做。

Um, but you're doing that often and like what is your advice for people who like, want to start doing that? Yeah, so the, it's finally you say build in public because. 对想开始的人有什么建议？ ——你提到 build in public 很有意思，因为… …

I think our company is one of the best examples of build in public. It's funny 'cause a phrase is very, it's kind of lame these days. Yeah. Uh, because like everything, there's good advice, there's a good idea. The easiest way to implement a good idea is to do it wrong. 我觉得我们公司是 build in public 最好的例子之一，只是这个词如今显得有些俗气。很多好建议、好想法，最容易的执行方式恰恰是把它做错。

So if you see, if you see someone, now build in public is a great idea and it works, but most people do it wrong and they do it in a performative way, or like, everyone kind of copy pay something and it kind of becomes like, not the right, like it kind of misses the point of the whole thing. Build in public 本身是好想法，也确实有效，但多数人会以一种表演性的方式做错它，或者所有人复制粘贴同样的内容，最后完全错过了这个理念的重点。

But for us, what build in public means is every single day we spend, you know, at least eight hours a day with our brains on exerting some amount of energy trying to like, learn something or trying to implement something. Um, all building in public is, is developing a habit of sharing that at the end of the day, sharing what you worked on, sharing, uh, a thought you had or an idea you had or like something you learned or something you were wrong about that you now understand. 对我们而言，build in public 的含义很简单：每天至少有八个小时，我们的大脑都在消耗精力、学习或实现某样东西；所谓公开构建，只是养成习惯，在一天结束时分享你做了什么、产生了什么想法、学到了什么，或者过去错在哪里而现在终于理解了。

Um, and just getting into a habit of doing that, you know, a couple times a week. 每周这样做几次。

The, it's not the goal, the goal isn't to like directly d drive people to your product. It's not like, 'cause I, I, I see this a lot where people are like, today while working on link to my product, I did it. It's like, it's, it's, you can tell why they're doing it. Yeah. Um, but for us it's a way to keep ourselves motivated. 'cause it's fun to talk about stuff that we, uh, built. It attracts people that we like into a world. 目标不是直接把人导向产品。我经常看到有人写：‘今天在开发… … ’，然后立刻附上产品链接。谁都能看出他们为什么这么做。对我们来说，分享能帮助自己保持动力，因为谈论刚做出来的东西很有趣；它也会把我们喜欢的人吸引进这个世界，很多后来招聘或认识的人都来自这里。

Like a lot of people we hired or people we've met through doing this. Um, you know, we learn, we, we post something and people tell us stuff that they know about it. Uh, so it's just a very valuable thing to do and it's not something that changes your life overnight, but if you do it consistently over weeks, over years, uh, it's how you can, you know, have a lot of reach and a lot of influence. And that's all we did. 我们也会学习：发出一个想法后，人们会告诉… …我们他们知道的相关信息。所以这是非常有价值的做法。它不会一夜改变人生，但如果连续几周、几年坚持，就能逐渐获得广泛触达和影响力。我们做的也只有这些：五六年来，每一周都持续分享。

We've just been doing this every single week for five or six years now. Uh, and everything we built and have been able to do is on, on top of this. So, yeah, like we very much believe in building public. We think almost nobody does it. 后来构建并实现的一切，都建立在这件事之上。所以我们非常相信 build in public，也认为几乎没人真正这样做。

Uh, 'cause it, it is a little unnatural. And the other thing that's unique about us is all river. Our code's different source anyway, so why not? Yeah. Talk about it, right? So, uh, yeah, I, I think it's a great way to operate and, and if you do it right, it, there's a lot of nice outcomes that you get from it. It does seem like a lot of founders get it wrong where they're trying to learn in public, but it's like they're shill their product. Like, here's the link. 因为它多少有些违背直觉。我们另一个特别之处是，反正所有代码都开源，那为什么不谈呢？如果做对了，这确实是一种很好的运作方式，也会带来许多不错的结果。 ——很多 founder 似乎把它做错了。他们想公开学习，实际却只是在推销产品、贴链接。

Yeah. Like, what else do you think people are doing wrong when they're trying to like, build in public? Yeah. 除此之外，人们 build in public 时还会犯什么错？

I think it's, uh, you don't, I think what people see is they'll kind of, I mean, it's great to have, uh, role models. Like if you're first getting into this, identify people that you're like, I would like to be like them one day, or I'd like to be in their position one day. Um, it's good to have that, but you need to understand that you are not going to get there overnight. So copying what they're doing is just gonna be superficial. It's gonna be stupid. 人们往往会找 role model。刚开始时，找到一个让你觉得‘有一天我想成为这样的人、到达他的位置’的对象当然很好。但你必须明白，自己不可能一夜抵达那里，所以照抄对方现在的做法只会流于表面，甚至显得愚蠢。

If you do get there, you're gonna have your own unique journey of, of getting there. Uh, and it's not gonna be something that anyone could prescribe to you. It's not gonna be an instruction that someone else gave. Uh, understand the outcome you want. But yeah, you have to discover your, your, your own pathway to it. Uh, the only thing that I can say is talk about interesting stuff that you've done or learned every single week, like genuinely interesting. 即使最终真的走到那里，你也必然拥有自己独特的… …旅程。没有人能为你规定路线，它也不会来自别人的操作说明。你可以理解自己想要的结果，但必须亲自发现通往它的路径。我唯一能给的建议是：每周谈谈自己真正做过或学到的有趣事情，必须是真有意思。

Um, don't try to force it through any kind of lens of, uh, your product or like some other goal you have. 不要强迫所有内容都套进产品或其他目标的框架。

Uh, I think like it's, this is another thing that I think people get a little bit too fixated on the whole productivity mindset of being like, I'm a founder, I'm working on a product. I need to grind on the product every single day. Like, I need to like, go directly at what I'm building and every minute not spend on that is wasted. Uh, I don't believe in that. I think you need to meander. I think you need to like. 我觉得人们还会过度执着于 productivity mindset：‘我是 founder，正在做产品，必须每天拼命开发；每一分钟都要直接投入当前产品，否则就是浪费。 ’我不相信这种方式。你需要漫游，需要有一些奇怪的日程，需要偶尔拿一天去做极其反常的事情。

Have like a weird schedule or one day you like do something really, really odd. I think these things are just good for creativity and I think they, uh, it's very easy to get caught up in this. I need to spend every minute as efficiently as possible. Another way to phrase that is you're spending every minute as boring as possible. Um, and that doesn't lead to anything good or interesting that other people want to hear about. 这些都有利于创造力。人太容易陷入‘必须把每一分钟都用到最高效率’的思维。换个说法，这就是把每一分钟都过得尽可能无聊，而无聊不会产生任何好东西，也不会产生别人想听的有趣内容。所以，去做有趣的事，你自然就有有趣的东西可谈。

Uh, so yeah, just do interesting things and you'll have interesting stuff to talk about. Yeah. I love that. That is definitely not a San Francisco take, I think, where everyone's like, how can I be as efficient as possible in every moment? Yeah. It's funny, a lot of these things are like, I'm just kind of rephrasing, I would argue like cliche advice. ——我很喜欢。这确实不像典型的 San Francisco 观点，那里每个人都在问如何让每一刻尽可能高效。 ——有意思的是，很多时候我只是在换一种方式表述那些堪称陈词滥调的建议。

'cause it's, this is the thing that I've like just learned over and over in my career. It's all the stuff that's ultimately the best advice. It's not hidden. It's like stuff that is been said a million times. Yeah. It's very, very simple, but it's so hard to actually live by. Yeah. And the reason I say that is there's always this thing that goes around where, uh. It's like, oh, the daily routines of Cecil people. 我的职业生涯反复教会我：最终最好的建议并不隐藏在什么地方，而是已经被说过一百万遍。它们极其简单，却很难真正照着生活。比如总有人传播‘成功人士的每日作息’，你会发现很多历史上的伟大人物，日常生活都非常曲折、松散而奇怪。

And a lot of them have these weird meandering day to day like historical,historical figures that are like the grays and like people that, that have lived. Yeah. They weren't like, you know, like snorting modafinil every single day. They had just like these weird, odd, quirky routines. And I think it's, I think we kind of learned from that because I think if you're doing any creative work, being a founder is creative work. It's all about creativity. 那些伟大的历史人物并不是每天靠服用 modafinil 强行运转，而是有各种奇特、古怪的生活规律。我觉得我们应该从中学习。任何 creative work 都需要创造力，而做 founder 本身就是 creative work，所以不能过一种过度僵化的生活。

Uh, yeah. You kind of, you can't have this like rigid lifestyle. Yeah. That makes sense. Love it. Um, what was I gonna say that I was gonna talk about how, you were talking about how advice. A lot of this advice is not hidden. Uh, yeah. But I do think right now, because founders and software engineers, and there is this like palpable fear in the air because of ai. So it feels like people are listening to advice even more than they normally would, right? ——有道理。我刚才想说什么来着… …你… …提到很多好建议并不隐藏。我觉得现在因为 AI，founder 和软件工程师都能感到空气中真实存在的恐惧，于是人们比平时更渴望听建议。

Mm-hmm. Like people want to be comforted right now. People want to like, listen to a podcast and think, okay, like, this is gonna be my strategy for being okay. Yeah, yeah, yeah. I think the, the issue I, I, I totally understand that poll. I do some form of that in my life as well. 他们想获得安慰，想听完一档播客后说：‘好，这就是让我安然度过变化的策略。 ’——我完全理解这种吸引力，自己生活里也会做某种类似的事。

Um, but this kind of goes back to what I was saying is do you want to be right right now, or do you want to be right in the end? In this case, being right means feeling good. Like, do you wanna feel good right now or do you want to actually feel good at the end of all this? Um, if you're a founder, some aspect of what you're doing is making an asymmetric bet. Like it's a bet where you're probably wrong, but if you're right, the upside is like massive. It's like huge. 但这又回到我刚才的问题：你想现在就觉得自己正确，还是想最终真的正确？在这里，‘正确’就是感觉良好。你想立刻感觉好一点，还是想在一切结束后真正拥有好结果？如果你是 founder，所做的事情里一定包含 asymmetric bet：你很可能错，但一旦正确，上行空间极其巨大，成功与失败的结果可能相差一百倍。

It's like, it's like a hundred x difference than being wrong. They have a downside of being wrong.

Uh, you, you are not gonna find this in a podcast. Like this isn't gonna be information that is, uh, advice, that's like, I, I think like first principles are out there, like basic things a way to like operate are out there. But in terms of what you're actually doing day to day, uh, what you're working on, what you're trying to get a swing at, uh, yeah. 你无法从播客里找到这项押注。第一性原理、基本的行动方式可以公开获得；但你每天究竟做什么、正在研究什么、要在哪件事上进行尝试，不会有人通过一条建议告诉你。我也不可能说出一句话，就以那种方式改变你的人生。

Like you're not gonna, I'm not gonna tell you, I'm not gonna be able to tell you like something that, that changed your life in that way. You gotta find your own unique asymmetric thing that everyone else is ignoring or even like, explicitly thinks is wrong. 你必须找到属于自己的独特 asymmetric bet：一件被所有人忽视、甚至被大家明确认为错误的事情。

Um, I mean, if you look at our, our company, we have a very conservative view on AI and coding. Every other coding company or like any AI company doing anything in coding has a very like futurist accelerationist view of things. And we have a very conservative view. Uh, it's possible we're wrong, but on the off chance that we're right, we're gonna seem like the very reasonable company that composition everything correctly, right? Um, so that's our asymmetric. 看看我们公司：我们对 AI 与 coding 的看法非常保守，而几乎所有其他 coding 公司、所有涉足 coding 的 AI 公司，都持有未来主义、加速主义的立场。我们可能是错的；但万一我们是对的，就会成为那家看起来最理性、对一切做出正确判断的公司。这就是我们的一项 asymmetric……

Kind of one of our asymmetric positioning, uh, is very counter to what most of the industries, how the most industry is operating. Uh, you have to find something like that for yourself. And it's not easy, you know, it is just a show up overnight. But it's kind of what I was saying, be weird, do weird things like have a weird day. These things will kind show up. Yeah. Love it. That's, thanks for being here. Yeah. It was fun for chatting. Thanks for coming to my house. 定位：它与行业主流运作方式截然相反。你必须为自己找到类似的东西。这并不容易，也不会一夜之间出现。但就像前面说的，保持古怪，做些古怪的事，偶尔过一天古怪的生活，这些机会会逐渐浮现。——很喜欢。谢谢你来。——聊得很愉快。谢谢你来我家。