AI Interview Wiki
主题综述

开源 AI 作为抗租金基建

主题综述

更新日志

主流共识

一、"开源还不够好"这个论证，在这批访谈里已经基本消失了。

Glean 的 Arvind Jain 把企业侧的数字说死：

"90% or greater of use cases can now be fully handled by many, many different models, including open source models."
「90% 或更高的用例现在可以由许多不同的模型完全处理，包括开源模型。」
Arvind Jain · 20VC: Why OpenAI and Anthropic Won't Win the App Layer

Benchmark 的 Ev Randle 用一个消费侧的土办法量到同一个结论：

"There's nothing that my mom actually asks of her AI products that needs to be done by the frontier or even a near frontier model."
「我妈妈实际上对她的 AI 产品有的任何需求都不需要用最前沿的，甚至是接近最前沿的模型来完成。」
Ev Randle · Benchmark's AI Bets

二、开源真正被买的是"控制权"，几乎每个人都用了同一个词组——own your destiny。

Fireworks 的 Lin Qiao 把它当公司第一性原理：

"Because openness gives control to the user. Think about open models, right? Once the model is released, you have the full control of the weights. You can change it however you want. It's yours."
「因为开放性赋予用户控制权。想想开源模型，对吧？一旦模型发布，你就拥有完全的权重控制。你可以随心所欲地更改它。它属于你。」
Lin Qiao · 20VC: Are OpenAI and Anthropic Overvalued?

Arvind Jain 在买方那一侧看到的是同一句话的镜像——而且他强调这不是新愿望，是等了很多年的旧愿望终于有了可执行的选项：

"There's no enterprise that we talk to Which is okay with saying that, hey, look, I can get my work done with OpenAI or with Anthropic and I'm good. Everybody wants to make sure that they are in control of their destiny."
「我们与的每一位企业都不满意地说，嘿，我可以通过 OpenAI 或 Anthropic 完成我的工作，我就可以了。每个人都希望确保自己掌握自己的命运。」
Arvind Jain · 20VC: Why OpenAI and Anthropic Won't Win the App Layer

连把绝大部分需求押在前沿模型上的 Sierra，选择也一样——只是他们把"掌握命运"划在了微调层而不是预训练层：

"So today we have a set of our own proprietary fine-tuned models, but these are fine-tunes on top of open weights models. … And I think it's important that you are in control of your own destiny enough and that you don't tell yourself a story that you need to go further than you actually need to do."
「到今天，我们有一套我们自己专有的微调模型，但这些是基于开放权重模型的微调。……我认为，重要的是你要在自己的命运中掌握足够的控制权，而不是告诉自己需要走得比实际上更远。」
Clay Bavor · 20VC: Open Models vs Frontier Models

三、没人认为"开源 = 免费"。 这条共识很朴素但很重要，因为它限定了抗租金能抗掉的到底是哪一层：

"open source inference also costs money. It's not like open source means free. Someone still has to run the GPUs. Someone still has to build the data center."
「开源推理也要花钱。开源并不意味着免费。总得有人运行 GPU，总得有人建数据中心。」
Ev Randle · Benchmark's AI Bets

分歧在哪

分歧一：抗租金的机制是价格，还是谁有权替你设边界

这是本次综述里最新、也最被低估的一条裂缝。a16z 的 Matt Bornstein 在最新一期里当面把它拆成两半问 vLLM 的 Simon Mo：

"There's almost two pieces to this, right? There's like the cost thing where it's like the closed models are too expensive. And then there's sort of the control thing where I want to sort of be in control of my infrastructure and in control of the model, right, if I need to extend it or put on my own guardrails or anything."
「这几乎有两个部分，对吧？一是成本——闭源模型太贵了。二是控制——我想掌控我的基础设施、掌控这个模型，如果我需要扩展它、或者加上我自己的 guardrail。」
Matt Bornstein · The Engine Powering Open-Source AI
"I think right now the open source drive is coming from the cost point of view. … When AI just came, companies were a lot more afraid of getting their data outside of their own control and model companies training with their data. But that sort of is a fear that's no longer there."
「我认为现在的开源驱动来自成本的角度。……当 AI 刚出现时，公司对此非常担心他们的数据会在自己的控制之外，而且模型公司使用他们的数据进行训练。但这种恐惧现在不复存在了。」
Arvind Jain · 20VC: Why OpenAI and Anthropic Won't Win the App Layer
"a lot of the anthropic models are banning frontier AI research. And then when we're studying GPU kernels, even as an invalid memory access error, we are triggering the red line."
「很多 anthropic 的模型禁止前沿 AI 研究。于是当我们研究 GPU kernel 时，哪怕只是一个非法内存访问错误，我们都会触发红线。」
Simon Mo · The Engine Powering Open-Source AI
"a lot of our developers within Infrax and for VLM are like retreating from using Fable 5 because you have a two-hour job and you trigger the red line, which is false positive, and then you have to lose all of your work."
「我们 Infrax 和 vLLM 的很多开发者正在从 Fable 5 上撤退，因为你跑一个两小时的任务，触发了红线——一个误报——然后你所有的工作都没了。」
Simon Mo · The Engine Powering Open-Source AI

他把这条推到了一个结构性判断上——moderation 是个永远解不完的问题，所以这个迁移不会因为降价而停止：

"If moderation is never solved, in the future people will go to open-weight by default because that is where you know for sure you can control your guardrail for trusted use cases."
「如果 moderation 永远解决不了，未来人们会默认走向 open-weight，因为那是你能确定自己可以为受信任的用例控制 guardrail 的地方。」
Simon Mo · The Engine Powering Open-Source AI
"Once the model is open, you can put all kinds of guardrails specialized to your business around. I would say to all models, it doesn't matter if open or closed, you should put your own guardrail around it. … A model provider will infuse their own judgment, their own taste into the model training process. You cannot guarantee it matches yours."
「一旦模型是开放的，你就可以围绕它加上各种针对你业务的 guardrail。我想对所有模型都这么说，不管开源还是闭源，你都应该加上自己的 guardrail。……模型提供方会把他们自己的判断、自己的品味注入模型训练过程。你无法保证它和你的一致。」
Lin Qiao · 20VC: Are OpenAI and Anthropic Overvalued?
"So today, 90% of our workflow is on open-source. And again, the main reason was for latency to really optimize our voice agents."
「所以今天我们 90% 的工作流程都是基于开源的。而且主要原因是为了降低延迟，以真正优化我们的语音代理。」
Jesse Zhang · Decagon's Playbook for Building Enterprise AI Applications

搭档 Ashwin Sreenivas 顺手否掉了"便宜=更笨"这个前提：

"when we fine-tune smaller, dumber models, it's that they're just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models. So we end up getting all three things. It is better at the task, it is cheaper, and it is faster."
「当我们对更小、更笨的模型进行微调时，它们只是没那么通用，但在我们希望它们完成的特定任务上，它们实际上超越了大型的、聪明的、最先进的模型。所以我们三样都拿到了：任务上更好、更便宜、更快。」
Ashwin Sreenivas · Decagon's Playbook for Building Enterprise AI Applications

分歧二：租金到底会不会被压掉

"these models are starting to get so capable at the frontier that you might not want to release them anymore. … At the same time, open models are getting stronger and stronger. So I think as a proprietary model builder, you're kind of starting to get squeezed in."
「这些模型在前沿变得如此强大，以至于你可能不想再发布它们了。……与此同时，开源模型正变得越来越强。所以我认为作为一个专有模型的构建者，你有点开始受到挤压。」
Igor Babushkin · Ep 92: xAI Co-Founder Unpacks the Future of Model Development

Arvind Jain 给了这条判断一个可证伪的时间表：

"I believe that majority of enterprise workloads will actually be on open source models in three years for sure."
「我相信大多数企业工作负载在三年内肯定会基于开源模型。」
Arvind Jain · 20VC: Why OpenAI and Anthropic Won't Win the App Layer
"it's not this zero-sum game, at least yet, in terms of like, well, either it's all going to be open source or all going to be a frontier model from one of the three frontier providers. It turns out that like there's use for all of it and it's all growing very, very quickly."
「这不是一种零和游戏，至少目前不是——不是说要么全是开源，要么全是三家前沿供应商的前沿模型。事实证明所有这些都有用，而且都在非常非常快地增长。」
Ev Randle · Benchmark's AI Bets
"if at any point it seems like capabilities are actually hitting an absolute ceiling, And distillation continues as it has historically. And the open source actually gets, you know, 95% as good as wherever the ceiling of capabilities tops out. That's a really scary situation for the Frontier Labs. … you'd have a much, much greater impact from open source depressing their ability to have pricing power and charge a premium for the tokens that they're producing."
「如果在任何时候能力似乎真的触到了绝对上限，而蒸馏像历史上那样继续，开源真的做到了能力上限的 95%——那对前沿实验室是个非常可怕的处境。……开源会大大压制他们的定价能力、压制他们为自己生产的 token 收取溢价的能力。」
Ev Randle · Benchmark's AI Bets
"open weights models will be cheaper because you're kind of avoiding some of the margin stack in, you know, the hosted frontier models. Okay, but what is the fundamental input? It's GPU capacity, it's power, that's still constrained."
「开放权重模型会更便宜，因为你避开了托管前沿模型里的一些利润堆叠。好的，但基本的输入是什么？是 GPU 能力，是电力，这仍然受到限制。」
Clay Bavor · 20VC: Open Models vs Frontier Models

他还顺手指出了一件在"开源 vs 闭源"框架里常被忽略的事——美国实验室没有动机自己制造这个价格压力：

"are they going to compete with themselves and drive price pressure on the frontier models by developing and releasing open weights models that are of similar capability? If I was running that business, that's not something I would do."
「他们会通过开发和发布能力相近的开放权重模型来跟自己竞争、对前沿模型施加价格压力吗？如果我是经营那家企业的人，我不会做那样的事。」
Clay Bavor · 20VC: Open Models vs Frontier Models

分歧三：抗的是谁的租——如果开源供给由中国实验室主导

Arvind Jain 直接把这条摆到了台面上，并且认为真正的分界线根本不是开闭源：

"The question is going to be, are they okay with the Chinese model or not? That's the only question here. It's not open source versus closed source."
「问题是，他们是否对中国模型感到满意？这就是唯一的问题。不是开源对封闭源。」
Arvind Jain · 20VC: Why OpenAI and Anthropic Won't Win the App Layer
"At least for now, it looks like effectively the Chinese government is subsidizing at least a large subset of these models. And that subsidy or surplus is effectively just being passed on to U.S. enterprises for adopting these models."
「至少目前看来，中国政府实际上在补贴这些模型中的很大一部分。而这种补贴或盈余实际上只是被转嫁给了采用这些模型的美国企业。」
Elad Gil · Baseten CEO Tuhin Srivastava on the AI Inference Crunch
"it does put the Chinese labs in a preferred position where they also can exhibit some control. So, for example, they might stop releasing their open weights in the future. Everyone relies on them. That's not a great thing for the U.S. economy. They could also change the licenses to those models."
「这也使中国实验室处于一个优越的位置，让他们也可以施加一些控制。例如，他们可能未来会停止发布他们的开放权重。每个人都依赖于他们。这对美国经济来说不是一件好事。他们也可能改变这些模型的许可证。」
Igor Babushkin · Ep 92: xAI Co-Founder Unpacks the Future of Model Development
"I think that would be a massive loss if there are five companies You know, five different labs in China that are creating open source models. And we're struggling to get one set up. So it's necessary. I also think it's inevitable."
「如果中国有五家不同的实验室都在开发开源模型，而我们还在努力建立一个，那将是一个巨大的损失。所以这是必要的。我也认为这是不可避免的。」
Tuhin Srivastava · Baseten CEO Tuhin Srivastava on the AI Inference Crunch

分歧四：政治论证 vs 经济论证——同一批人，两套完全不同的语言

"the regulators are now moving in and very ironically, oddly, bizarrely talking about trying to ban open source, which is probably the safest thing that could possibly happen in AI because if AI is this all-powerful thing, then the last thing you want is it in the hands of one person or one company."
「现在，监管机构开始介入，具有讽刺意味的是，他们竟然在讨论禁止开源。这可能是人工智能领域最安全的事情，因为如果人工智能真的如此强大，你最不希望的就是它掌握在一个人或一家公司手中。」
Ben Horowitz · Ben Horowitz on How a16z Was Built
"And so if you believe that, then I think what you want is open source. And I think if you want regulatory capture or monopoly for yourself, you want to shut that down."
「如果你相信这一点，那么我认为你想要的就是开源。我认为，如果你想进行监管俘获或垄断，你就会想要阻止开源。」
Ben Horowitz · Ben Horowitz on How a16z Was Built
"the reason that you can get a Android phone for $10 and can get on the Internet so cheaply, right, is that basically all the software is free. I mean, imagine if there was an open source and, you know, operating system providers used to charge $100 and you'd be paying that on client and maybe on the back end"
「你能以 10 美元的价格买到安卓手机，并且能如此廉价地上网，原因基本上是所有软件都是免费的。想象一下，如果操作系统供应商过去常常收取 100 美元，那么你会在客户端以及后端支付这笔费用。」
Chris Dixon · Chris Dixon on How to Build Networks, Movements, and AI-Native Products

然后他指出 AI 与操作系统的类比在成本结构上断掉了：

"it's just the thing with AI that's different than operating systems, like with operating systems and databases, you just needed a bunch of coders sitting around. With AI, you need massive capital expenditure to train the models. So I just don't know. I think it's an unknown question long term."
「AI 与操作系统不同之处在于，像操作系统和数据库，你只需要一堆程序员坐在那里。对于 AI，你需要大量的资本支出来训练模型。所以我不知道。我认为从长远来看，这是一个未知的问题。」
Chris Dixon · Chris Dixon on How to Build Networks, Movements, and AI-Native Products

他愿意接受的最好结局，其实是一个降级版的抗租金——开源永远落后一点点：

"a possible outcome, which I think is a pretty good outcome, is open source is just always a little bit behind, like the way open AI is now releasing older models."
「一个可能的结果，我认为这是一个相当好的结果，是开源总是稍微落后一点，就像 OpenAI 现在发布旧模型的方式一样。」
Chris Dixon · Chris Dixon on How to Build Networks, Movements, and AI-Native Products

分歧五：谁替开源付前沿训练的账

这是全语料里唯一一条各方都承认没解决的分歧。Lin Qiao 把账摆得最直白：

"once the model is there, whoever is using those models, there's literally no cost. But there's fundamental cost for the Frontier Labs to invest in those models and recoup the R&D cost back."
「一旦模型存在，使用这些模型的任何人实际上没有成本。但前沿实验室投资于这些模型并收回研发成本则是根本成本。」
Lin Qiao · 20VC: Are OpenAI and Anthropic Overvalued?

Simon Mo 观察到的解法正在成形，而且形状很值得注意——许可证，不是捐赠：

"especially now the labs are trying to figure out a way to economically fund it, especially when they're with open-source model. Everybody can just take it and run it themselves, whereas nobody will use their API anymore in many cases"
「尤其是现在实验室正在想办法从经济上给它筹资，特别是当他们做开源模型时。所有人都可以直接拿走自己跑，很多情况下就没人再用他们的 API 了。」
Simon Mo · The Engine Powering Open-Source AI

他给的类比是制药，而制药恰恰是一个靠专利租金养研发的行业：

"it's really about sustainability in the end. It's about how do you make sure that all this initial CapEx almost to train the model fail again and again and train the model again. … how do you make sure that the R&D process of new drugs are properly funded and is proper sustainable method to making sure that people are willing to take big risk, big bet to go to do research for new drugs"
「归根结底这是可持续性的问题。……你怎么保证新药的研发过程得到恰当的资助，有一套可持续的方法让人们愿意冒大风险、下大赌注去做新药研究。」
Simon Mo · The Engine Powering Open-Source AI

Babushkin 想做的也是同一件事，只是他站在生产方：

"I'd love to train the best open model, but right now we're very interested in how to build a business on these open weights so that the company can be self-sustaining."
「我很想训练最好的开放模型，但现在我们非常感兴趣的是如何基于这些开放权重建立一个可自我维持的公司。」
Igor Babushkin · Ep 92: xAI Co-Founder Unpacks the Future of Model Development

都没说透的

1. "开源"这个词在整场讨论里从没被界定过。 唯一点破的是 Matt Bornstein 一句带过的 "open-weights, which is a little bit different than true open-source"，之后所有人——包括他自己——继续把两者混用。而这恰恰是抗租金论证的要害：权重可下载但训练数据、配方、许可都不公开的东西，抗掉的是使用租金，抗不掉复制租金。

2. 许可证正在把开源重新分层，但没人算过那条线画在哪。 Simon Mo 提到 Minimax、Kimi 都在加"按收入/衍生品"的条款，Babushkin 提到中国实验室可以随时改许可。也就是说：一个按你的收入向你收费的"开源"模型，和一个 API 的差别在哪？这条界线全场没有一个人试图定义。

3. 抗租金叙事的发言人几乎全是卖方。 这批访谈里主张开源抗租金的，是推理引擎（vLLM）、推理云（Fireworks、Baseten）、应用层（Decagon、Glean）、以及投了它们的 VC。没有一个纯买方——某家企业的 CIO——出来讲他们迁移之后账单实际降了多少、迁移成本是多少。Decagon 提到"研究团队很贵"，但没人把这笔钱和省下的 token 钱放在一张表上比。

4. "控制 guardrail"的另一面没人碰。 Simon Mo 和 Lin Qiao 都主张 guardrail 应该由使用方自己设。但语料里没有一个人问：当每家公司都能自定义边界时，被误报挡住的那类研究和被有意放开的那类滥用，是不是同一个开关。这在 agent-security 那条线里被讨论过，但两条线在语料里从未接上。

我的看法

这是判断，不是事实。 我认为语料在 2026 年发生的真实变化，不是"开源追上来了"（那是 2025 年就有的说法），而是抗租金的主战场从价格挪到了控制权——而且这个位移让抗租金论证变得更稳固，不是更脆弱。理由是：价格差可以被前沿实验室降价抹平（Arvind Jain 就听到了 OpenAI 要大幅降价的传闻），而 Simon Mo 描述的那种"两小时任务被误报清零"的痛点，是闭源商业模式的结构性副产品——只要有中心化的 moderation 责任，就一定有误报，降价解决不了它。

我对这一点把握中等偏上。把握不高的部分在于：这批访谈里没有任何一个大规模买方给出迁移的完整成本核算，而 Decagon 那句"研究团队很贵"暗示控制权是要用工程编制换的。如果换算下来是"用 20 个研究员换掉 guardrail 误报"，那对绝大多数企业来说，这场抗租金运动只会发生在有能力自己养模型团队的那一小撮公司里——那就不是抗租金，是租金的重新分配。

Clay Bavor 那句"基本输入是 GPU 和电力"我认为是这批访谈里最容易被忽略、也最难反驳的一句：抗租金抗掉的是实验室的毛利，抗不掉 Nvidia 和电网的毛利。

还想知道什么

1. 一家非 AI 原生的大企业，完整的迁移账：从前沿 API 迁到自托管开源模型，token 成本降了多少、工程编制增加了多少、迁移周期多长。这一条能直接证伪或坐实"抗租金只属于养得起研究团队的公司"。 2. 一位前沿实验室内部的定价负责人的访谈。目前所有关于"开源压低前沿定价"的证据都来自被压的对手方或旁观者，没有一句来自定价一侧。 3. 许可证条款的实际执行案例：有没有公司真的因为收入越线而被要求签商业协议、金额是多少。这决定了"开源"和"折扣 API"之间还剩多少实质差别。 4. 如果能力真的触顶（Ev Randle 的分岔点），前沿实验室的应对是降价、转产品、还是收紧权重发布。这个分岔一旦落地，本文大半判断需要重写。

取材

继续读
线索
阅读线索：agent 不到 9 nines 可靠，那应该卖给企业还是消费者
线索
阅读线索：如果 coding 是 commodity，AI startup 的护城河长在哪儿
线索
阅读线索：推理成本一年跌 10×，AI app 该按什么定价
线索
摘报
线索
摘报
线索
摘报