Misha Laskin 谈开放模型：租与买、智能如水电，以及中美开源之争

Misha Laskin

我们预见到这种地缘政治张力：如果只有一个国家在生产开放模型，世界其他地方就会依赖那些模型、依赖那个国家的基础设施。对美国来说，在 AI 上保持竞争力极其重要——不只是有竞争力，而是要拥有最好的开源软件和模型，确保我们的盟友、整个世界以及国内，都建立在我们自己的开放智能之上。

We saw this geopolitical tension coming where if it’s only one country that’s producing open models, then the rest of the world kind of becomes dependent on those models and on the infrastructure of that country. It’s extremely important for America as a country to be competitive, and not just competitive but have the best open-source software and models when it comes to AI, and ensure that our allies, the world, and domestically we all build on our own open intelligence.

Ed Elson

欢迎收听 Prof G Markets 创始人系列，我是 Ed Elson。围绕 AI 建设的担忧并不少：社区在反对数据中心，劳动者担心被取代，还有人担心这项技术会构成生存威胁。但对美国来说，最迫切的挑战之一可能是与中国的竞赛——中国的开放权重模型正在提供一种比最强专有系统更便宜、更快的替代品，并不断扩大地盘。

我们这期的嘉宾在打造一家公司，目标是让美国在这场竞赛中站得更稳。这家初创公司最近发布了首个模型 Beam，称它优于其他西方开放模型，同时运行起来更便宜、更快。随着与中国的竞争加剧，这家公司越来越被视为美国 AI 建设中的战略性玩家，投资人也注意到了：它的投资方包括 Sequoia、Citigroup 和 Nvidia，6 月完成的一轮融资投前估值 250 亿美元。那么，这个新模型能否给美国 AI 带来对华竞争优势？美国要赢下开源 AI 竞赛需要什么？下面是我和 Reflection 联合创始人兼 CEO Misha Laskin 的对话。

Welcome to the Prof G Markets Founder Series. I’m Ed Elson. There’s no shortage of concerns surrounding the AI buildout. Communities are pushing back against data centers. Workers are worried about being replaced. And some fear the technology could pose an existential threat. But one of the most immediate challenges for the US might be the race against China, where open-weight AI models are gaining ground by offering a cheaper, faster alternative to the most powerful proprietary systems.

Our next guest is building a company designed to give the US a stronger foothold in that race. The startup recently released its first model, Beam, which it says outperforms other Western open models while being cheaper and faster to run. As the competition with China intensifies, the company is increasingly seen as a strategically important player in America’s AI buildout, and investors are taking notice. Backed by firms including Sequoia, Citigroup, and Nvidia, the company closed its June funding round at a $25 billion pre-money valuation. So, can this new model give American AI a competitive edge against China? And what would it take for the US to win the race in open-source AI? This is my conversation with Misha Laskin, co-founder and CEO of Reflection.

Ed Elson

Misha，谢谢你来。我最想聊的是开放权重，因为这个话题太重要了，而我觉得很多人并不真正理解。不过在那之前，能先讲讲你的背景吗？你学的是理论物理，在 Google 工作过。讲讲创办 Reflection 之前的故事吧。

Misha, thank you for joining me. We’ll get into open weights, which is really what I want to focus on in this conversation, because it’s such an important topic that I think a lot of people don’t fully understand. But before we get to that, could you tell us a little bit about your background? I know that you have a background in theoretical physics. You worked at Google. Tell us a little bit about the story leading up to starting this company, Reflection.

Misha Laskin

我对物理产生兴趣，是因为小时候搬到美国，住在华盛顿州的乡下，有大把时间。也许部分原因是那一带是曼哈顿计划的场址之一——汉福德，当年在那里生产钚，所以那里有很深的物理历史。但我真正喜欢物理的地方在于它是一门基础科学：基于简单的原理，就能解释我们在世界上观察到的大部分现象，而且是我们每天使用的那些技术的根基。

所以我把物理看作 20 世纪的基础科学。第一个计算机程序是由物理学家带来的——二战时期普林斯顿高等研究院的 ENIAC 项目，那时计算机最早被用来计算弹道导弹的坐标。后来晶体管是贝尔实验室的理论物理学家发明的。GPS 这样的系统之所以能工作，是因为狭义相对论：你必须考虑光速，GPS 才准确。看起来，我们手边所有技术的根基都是物理。

我转向 AI，是因为我注意到 AI 里正在发生类似的事。这个领域原本方法多样，但一直没有一种简单的方式去规模化智能，直到深度神经网络被发现并扩展——从 2012 年的 ImageNet 时刻开始。而当我看到 AlphaGo，一个下围棋的系统——在传统棋类里围棋被认为最复杂，因为它的分支因子最高，每个局面可走的棋步非常多——一个神经网络被训练到了超人水平。而且那个系统从未停止学习：你往里倒更多算力，它就会更强。

我在那里看到的是：一门我认为和物理一样基础的科学，正在被实时发现，它会推动 21 世纪大部分的技术进步。AI 中的一些重大发现，让它看起来像是我们这个时代的物理——21 世纪的物理。这就是我在 2010 年代中期进入 AI 的原因。

The reason I got interested in physics was that I moved to the United States as a kid. We moved to a rural part of Washington State and had a lot of time on my hands, and basically developed an interest in physics. Maybe part of it was that this area was one of the sites for the Manhattan Project, the Hanford site, where the plutonium was produced, so it had a rich history of physics. But what I really liked about physics was that it was a fundamental science, based on simple principles that could explain the majority of things that we observe in the world, and was at the root of the technologies that we are using every day.

So I saw physics as the fundamental science of the 20th century. The first computer program was brought forth by physicists. It was the ENIAC program at the Institute for Advanced Study, during the World War II era, when computers were first being used for calculating ballistic missile coordinates. Then things like the transistor were invented by theoretical physicists at Bell Labs. Systems like GPS only work because of special relativity; you have to account for the speed of light in order to have accurate GPS. So it seemed to me that the fundamental science at the root of all the technologies we had at our fingertips was physics.

And my switch into AI happened when I noticed that something similar was happening in AI, where it went from a field that had a lot of diversity, but there was no real simple way to scale intelligence, until deep neural networks were discovered and scaled, starting in 2012 with the ImageNet moment. But when I saw AlphaGo, a system that played the game of Go, which out of the traditional board games is considered the most complicated one because it has the highest branching factor, meaning the number of different moves each board position affords is very high, a neural network was trained to play it at a superhuman level. And in fact, that system had never stopped learning. You could just pour more compute into it and it would have just gotten better.

What I saw happening there was that a science that I thought was as fundamental as physics was being discovered in real time, one that would power the majority of technological advances coming into the 21st century. Some big discoveries were being made in AI that make it kind of the physics of our time, the physics of the 21st century. And that’s what got me into AI as a field around the mid-2010s.

Ed Elson

你当时在 Google 工作，据我了解你们在 2024 年创立了 Reflection。为什么创办这家公司？你觉得 AI 里正在发生什么，或者缺了什么，是你可以补上的？

So you were working at Google, and my understanding is that you started Reflection in 2024. Why did you start this company? What did you think was happening in AI, or what wasn’t happening in AI, that you could contribute to?

Misha Laskin

有意思的是，2024 年其实只是两年多前，但按 AI 的时间线感觉要久得多，那时这个领域看起来很不一样。2024 年初是什么样？GPT-4 已经发布，ChatGPT 已经起飞，第一批 Gemini 模型发布了，Gemini 和 1.5——我和联合创始人 Ioannis 都参与了这些模型，尤其是强化学习部分。但那时的智能只擅长聊天。编程、智能体、推理这些东西都不存在，也做不成。这些模型只是好用的聊天机器人，并不擅长任何需要自主完成的认知工作。

我们当时的赌注其实有两层。第一，我们看到开源生态开始萌芽，一些相当不错的开放模型陆续出来，比如 Meta 等公司的模型。我们觉得这些模型可以作为很好的基础，不必自己砸大笔研发经费去做预训练。第二，也是公司最初的大赌注：我们认为强化学习这个范式——正是造出 AlphaGo 的范式，我的联合创始人是 AlphaGo 的关键贡献者之一——也就是让 AI 系统通过试错来学习，可以在语言模型之上规模化，造出不只会聊天、而是擅长认知型智能体工作的系统。我们以编程为切入点：擅长写代码、能构建软件、能在电脑上完成其他智能体任务的模型，以及法律、金融等领域。

结果在公司第一年里很快就发生了这样的事：我们在小模型上做出了强化学习的小规模概念验证，但与此同时，整个领域也发现了如何规模化强化学习。OpenAI 的第一批 o 系列模型出来了，把推理带进了公开市场。所以强化学习作为一个问题，被解决得比我们预想的快——我们以为要好几年，结果不到一年，无论在外部实验室还是在我们自己的实验里都是如此。

然后我们看到另一件事：西方开放模型开始明显落后于中国。很强的中国开放模型陆续出来，大家应该都记得 DeepSeek V3 的时刻，差不多和推理模型同一时期。之后一整个中国模型生态都非常强。大约一年多前，我们做出决定：公司能做的最有影响力的事，是承担起开放智能的使命，因为我们看到有几件事会到来——原本以为需要更长时间，结果来得快得多。

其一是地缘政治张力。开放模型之所以重要，是因为开源是世界会广泛采用的东西。想想最普及的操作系统：电脑上是 Linux，手机上是 Android，都是开源的。互联网协议是开放的。这些都起源于美国，然后被世界其他地方采用。开放技术会成为所有人采用的基础，同时也会带上发源国的基础设施和软件。所以我们预见到：如果只有一个国家在生产开放模型，其他国家就会依赖它的模型和基础设施。对美国来说，拥有最好的开源软件和模型、确保盟友和国内都建立在我们自己的开放智能之上，极其重要。我们看到了这一点。

It’s funny, 2024 is only a bit more than two years ago, but it seems a lot longer on AI timelines, and the field looked very different then. So what did the field look like in early 2024? GPT-4 had shipped, ChatGPT had taken off, the first Gemini models were shipped, Gemini and 1.5. These are models that my co-founder Ioannis and I contributed to, especially on the reinforcement learning side. But at the time, intelligence was really only good at chat. Coding, agents, reasoning, none of that existed and it didn’t work. These models were just good as chatbots. They were not really good at any sort of cognitive work that needed to be done autonomously.

Our bet at the time was actually twofold. We saw the beginnings of an open-source ecosystem. Some open models had started coming out that were pretty good, the models from Meta and others. So we thought those models would give us a good base to build on top of, instead of having to sink our own large R&D money into pre-training these models. But the big bet that we made as a company initially was that the paradigm of reinforcement learning, the same paradigm that created AlphaGo, and my co-founder was one of the key contributors to AlphaGo, the paradigm of teaching AI systems to learn through trial and error, could be scaled on top of language models and create systems that were not just good at chat but good at cognitive, agentic work. We went in with coding as the entry point: models that are good at coding, that can build software, that can do other agentic tasks on a computer, legal, finance and so forth.

Now what happened very quickly, in the first year of the company, is that we were able to get proof of concept of reinforcement learning working at small scale on small models, but at the same time the field had actually discovered how to scale up reinforcement learning. OpenAI’s first models started coming out, the o-series of models that introduced reasoning into the public market. And so reinforcement learning as a problem was solved faster than we anticipated. We thought it would take years. It took less than a year, both within external labs and even within our own experiments.

And then we saw another thing happening, which is that Western open models started significantly lagging behind China. Chinese open models started coming out that were very strong. I think we all remember the DeepSeek V3 moment that happened roughly around the same time as the reasoning models were coming out. And then a whole ecosystem of Chinese models came out that was very strong. And a bit over a year ago, we made the decision that the most impactful thing we could do as a company was to take on the open intelligence mission, because we saw a couple of things coming that, again, we thought would take longer to play out, but they played out a lot faster.

One was that we saw the geopolitical tensions coming. The reason open models are so important is because open source is the thing that the world adopts broadly. When you think about the most widely adopted operating system, it’s Linux for computers and Android for mobile. These are open-source operating systems. The internet protocols are open. These originated in America and were adopted by the rest of the world. So open technology becomes the foundational thing that everyone adopts, and with it, it brings in the infrastructure and software of whatever nation originated it. So we saw this geopolitical tension coming, where if it’s only one country producing open models, the rest of the world becomes dependent on those models and on the infrastructure of that country. And it’s extremely important for America to be competitive, and not just competitive but have the best open-source software and models when it comes to AI, and ensure that our allies, the world, and domestically we all build on our own open intelligence. So we saw that coming.

Ed Elson

打断一下。开源到底是什么意思？一个模型是「开放」的意味着什么？你拿 Linux 举例，开源和闭源的区别是什么？

Can I just interrupt for one moment? What does open source actually mean? What does it mean to be an open model? When you describe Linux, which is open source, what is the difference between open source and closed source?

Misha Laskin

高层次的区别是：封闭的东西你无法拥有，只能租；开放的东西你可以完全拥有。这有点像房地产里租房和买房的区别。租公寓时你并不拥有它，受制于房东的心情，可能涨租、可能被赶走，各种事都可能发生；但它很方便，你付房租，别人帮你打理物业，水管漏了打电话给房东，他来解决。另一边是买房市场：你买下房子，也继承了拥有它带来的所有麻烦；但出于财务、控制等各种原因，买房是合理的。软件和 AI 里的开源也是一回事：你要么从别人那里租软件或 AI，要么自己拥有它。

所以说到 AI 里的开源，我们想的是开放模型：要么从别人那里租，要么我可以把它装在一个 U 盘里给你，你拥有它，跑在自己的基础设施和电脑上。这是根本区别。

The high-level distinction: closed, you cannot own, you can only rent. Open, you can own entirely. It’s kind of like the difference between the rental and the ownership real estate market. When you are renting an apartment, you don’t actually own it. You are subject to the whims of your landlord, which could mean price hikes, it could mean an eviction, all sorts of things. But it’s very convenient. You just pay rent and someone else manages your property for you. You don’t have to worry about fixing the pipes when they’re leaking; you ask your landlord and they figure it out. On the other hand, in real estate there’s the ownership market: you buy a house, but then you also inherit all the problems that come with ownership, and for various reasons it makes sense to buy homes, for financial reasons, for control reasons, and so forth. You can think about open source in software and AI as really the same thing. You are either renting software or AI from someone, or you’re owning it yourself.

So when we think about open source in AI, we think about open models. You’re either renting it from someone, or I can give it to you on a flash drive, you can own it, and you can run it on your own infrastructure and your computers. That is the fundamental difference.

Ed Elson

我的理解跟这个比喻不完全对得上：开源和开放模型的一个要素是免费。买房要花钱，但开源模型据我理解任何人都能用、都能在上面构建和定制，这是很关键的一点。我理解错了吗？

My understanding, which I don’t think fully aligns with the analogy, is that an element of open source and open models is that they’re free. When you buy a home, you’ve got to pay money to buy the home. But with open-source models, my understanding is anyone can use it, anyone can build on it, customize it, and that is a crucial element. Do I have that wrong?

Misha Laskin

没错，你说得对。模型可以是免费的，不过也有些模型选择带许可证。你可以有一个别人能拥有、但附带许可证的开放模型——比如 Meta 最新的模型 Muse，对分发方有商业许可；Kimi K3 对分发方也有许可。所以开放也可以附带可以变现的许可证。顺带一提，Linux 也有许可证，如果你做了某些修改，那些修改也必须开放。开源软件和开放模型周围有一整套许可证生态，所以开放并不必然等于免费。

当然，今天开放模型上的大多数许可证是免费的，我们发布的模型 Beam 用的是 Apache 2 许可，也就是免费。但模型免费不代表周围的一切都免费。如果你是客户或企业，想把开放模型跑好，就得从某处拿到算力，要为算力付费；要有管理算力、提供模型服务的软件——光是模型服务的技术本身就很复杂、很难做，你得为此付费；它通常还嵌在某种智能体或定制工作流里，你一般会找人合作并为此付费。你可以把开放模型看作发动机，但没有车，发动机没用。所以即便模型本身因为许可宽松而免费，单靠模型并没有那么有用。我可以把我们的开放模型装在 U 盘里给你，但要让它有用，你需要它周围的一切。

No, I think you’re right. The model may be free, even though some models choose to license. You can have an open model that someone can own that has a license attached to it. Meta’s latest model, Muse, has a license attached to it, a commercial license if you’re a distributor. Kimi K3 has a license attached to it if you’re a distributor. So you can have open where there’s a license attached that can be monetized. And by the way, even Linux has a license attached to it, where if you make certain modifications to it, those have to be open as well. So there’s a whole ecosystem of licenses around open-source software and open models, and it doesn’t necessarily mean that open means free.

Now, most licenses on open models today are free, and even our model, Beam, that we released, is an Apache 2 license, meaning that it is free. But just because the model is free doesn’t mean that everything around it is free. If you are a customer or an enterprise and you want to run open models well, you need to get the compute from somewhere, so you need to pay for the compute. You need to get the software for managing that compute and for serving the models. Just the technology for serving the models is complex and hard to do, so you have to pay for that. It typically is embedded in some kind of agentic workflow or customization workflow, and typically you partner with someone and pay for that. So you can think about the open model as the engine, but the engine is useless without a car. You have to have the car around it. So even while the model itself could be free because the license is permissive, the model in its own right is not that useful. I can give you our open model on a flash drive, but in order to make it useful you need to have all the stuff around it.

Ed Elson

那是否可以这么说：对开放模型而言，训练是免费送出去的，但推理不是？另外也想请你澄清一下「开放权重」和「开源」的区别——开放权重是我们在 AI 圈新听到的说法，并不完全清楚有什么不同。也就是说：这是底层模型，就像你说的发动机，免费给你，随便你折腾；但你要真正运行它、围绕它做生意，就得付钱给别人来跑，很可能是付给你们。这样理解对吗？

Would it make sense then to say that for open models, and one thing that would also be helpful as a clarification is the difference between open weight and open source, open weight is the new terminology that we are hearing in the AI world and it’s not totally clear what the difference is, would it be correct to say that for open models you’re giving away the training for free, but you’re not giving away the inference for free? You’re saying, here’s the underlying model, as you say the engine, that’s free, you can have it and mess around with it however you want. But if you want to run that model, and if you actually want to build a business around it, you’re going to have to pay someone to run the model, probably us. Is that the right way to frame this?

Misha Laskin

对。开源和开放权重的区别在于：开源是更高层的类别，开放权重是它的一个具体形态，因为以前的软件根本没有「权重」这种东西，权重是 AI、尤其是深度学习特有的。也有一些可以说是开源纯粹主义者的人认为，如果你不开放数据和训练配方，那就不算开源。我们的立场是：公司的使命是打造开放智能并让它广泛可得，所以我们的目标是尽可能多地开源那些让模型易于使用、广泛可得的东西。

所以我们发布 Beam 时，有 40 家合作方一起上线，我们把它集成进了多个开源微调框架，还发布了自己的开源智能体 harness。我们也即将发布一份技术报告，相当于把科学部分开源出来，较详细地写清楚这些系统是怎么造出来的，因为我们认为让研究者生态理解这些东西如何构建很重要。这是科学的基本信条：你希望科学家能在你的工作上继续前进，而唯一的方式就是让他们理解你做了什么。

但这并不意味着要开放一切。比如要做出好模型，你得购买大量专有数据，而合同里不允许你把数据发给任何人，因为卖数据就是有些人的生意。所以开源不是把绝对所有东西都开源，而是尽可能多地开源，让模型对科学界、研究界、开发者社区都有用。至于在企业或公共部门里落地，商业模式确实是对让模型变得有用的那套系统收费：为企业提供推理服务，为它们打包智能体，提供微调服务，等等。

That’s correct. The difference between open source and open weight: I think open source is the high-level category around open-source software, and open weight is a specific instantiation of it, because in previous iterations of software there was no such thing as weights. Weights are a thing specific to AI, and in particular to deep learning. So it’s a more specific instantiation, and there are some people, open-source purists I would say, who would argue that it’s not open source if you don’t open source the data and training recipes and so forth. Our position is that the mission of the company is to build open intelligence and make it widely accessible, and so our goal is to open source as much as possible that makes these models easy to consume and widely accessible.

So when we launched Beam, we launched it with 40 different partners. We integrated it into a number of different open-source fine-tuning frameworks. We released our own open-source agentic harness. We’re also about to release a tech report that is effectively open-sourcing the science, writing up in some detail how these systems were built, because we think it’s important for an ecosystem of researchers to understand how these things are built. This is a fundamental tenet of science: you want scientists to build on your work, and the only way to do that is for them to understand what it is that you did.

But it doesn’t mean that, for example, in order to get good models you have to buy a lot of proprietary data, and in the contracts you’re not allowed to release that data to anyone, because that’s some people’s business, selling data. So open source doesn’t mean open-sourcing absolutely everything, but we try to open source as much as possible to make these models useful to the scientific community, the research community, the developer community. But when it comes to adoption within an enterprise or public sector context, then indeed the business is charging for the systems around the model that make it useful, and these systems are effectively serving inference to these enterprises, packaging up agents for those enterprises, fine-tuning services for these enterprises, and so forth.

Ed Elson

欢迎回来，继续和 Misha Laskin 对话。我之所以觉得搞清这个区别很重要，是因为这种区别似乎正在颠覆我们所知的整个 AI 生态。就像你说的，我们经历了 DeepSeek 时刻：一方面是「这是一个非常强的中国模型」，另一方面是「这是一个非常强的开放模型」。之后 token 价格大幅下降，OpenAI 降价最高 80%。我们熟悉的那些 AI 公司，特别是闭源的 OpenAI 和 Anthropic，基本被迫重新定价，因为开放模型把成本拉了下来。所以我才对「免费」这一点感兴趣：并不是整件事免费，而是其中一个环节免费，就导致了价格显著下降。

关于开源和闭源最后再问一个：开源和开放权重是否也意味着任何人都可以看懂模型的底层架构，并自行修改？而 Anthropic 和 OpenAI 的模型你看不到架构，也不知道如何在上面构建。这样区分对吗？

We’re back with Misha Laskin. The reason that I think it’s so important for us to understand the difference here, and why I’m so interested in hearing from you, is that this difference seems to be upending the entire AI ecosystem as we know it. As you pointed out, we had the DeepSeek moment, and part of that was, here is a Chinese model that is very powerful, but also part of it was, here is an open model that is very powerful. And since then we have seen token prices come way down. We’ve seen OpenAI reduce its prices by up to 80%. The AI names that we are so familiar with, specifically OpenAI and Anthropic, whose models are closed, have been forced to basically reprice everything because all of these open models are bringing the costs down. And that’s why I’m interested in this idea of it being free, because it seems as though it’s not that the entire thing is free, but this one element of it is free, which is resulting in significantly lower prices.

I guess just one final question on the difference between open source and closed. Does open source and open weight also mean that anyone can go in and understand the underlying architecture of the model and change it on their own time, whereas with a model from Anthropic and OpenAI you can’t see the architecture behind it and you can’t know how to build off of it? Is that a correct distinction or not?

Misha Laskin

对。实际上对闭源模型，我们不只是不知道架构，连大小都不知道，基本一无所知。还是用车来比喻：一切都封装好了，引擎盖是焊死的，你没法往里看，没法折腾，没法真正理解或定制它。开放模型你就能看到架构，可以把它接进你喜欢的开源微调框架，加载到你的 GPU 上，按需定制。你可以把它拆开、理解它。

从根本上说，这意味着你从这样一个世界——所有封闭实验室加起来也许有 5,000 名研究员理解这些系统并在上面工作，其中只有少数做安全，比如 5,000 人里大概 500 人在做这些模型的安全，这是个很小的群体——走向另一个世界。这是如此强大的基础技术，这么小的研究群体不可能真正理解关于这些模型需要理解的一切，以及如何让它们安全、如何对齐。而开放模型生态邀请了大得多的研究员、工程师、开发者群体，任何有兴趣的人都能来看、来贡献。

这从根本上就是科学的做法。科学不是炼金术，科学本质上是开放的。网络安全也是这样运作的：整个渗透测试领域的前提，就是你能接触到可以看见、可以理解的软件。实际上世界上最坚固、最安全的软件往往是开源软件，因为有句话叫「只要眼睛够多，所有漏洞都会变浅」。开源软件之所以如此普及、如此稳健，是因为有成千上万开发者的眼睛在看各种漏洞；他们未必能修补所有漏洞，但开源软件通常比闭源软件坚固、稳健得多。

That’s correct. In practice, it’s not even that we don’t know the architecture of a closed model; you don’t even know the size. We sort of know nothing about it. Again, with the car analogy, it’s all packaged up, the hood is sealed effectively, so you can’t peek under the hood, you can’t tinker with it, and you can’t really understand it, customize it and so forth. With an open model, you can see the architecture. You can plug it into your favorite open-source fine-tuning framework, load it up on your GPUs, and actually customize it for your needs. You can take it apart, you can understand it.

And fundamentally what that means is that you move from a world where across the closed labs there are maybe 5,000 researchers who understand these systems and are working on them, and a minority of them are safety researchers, so let’s say out of the 5,000 maybe 500 are working on the safety of these models, which is a very small community. And this is such foundational technology, so powerful, that it’s impossible for such a small community of researchers to really understand everything there is to understand about these models and how to make them safe, how to align them and so forth. With an open model ecosystem, you’re inviting a much larger group of researchers, engineers, developers. Anyone who takes an interest can take a look and contribute to these things.

And that is fundamentally how science is done. It’s not alchemy. Science is fundamentally open. That’s the way cybersecurity works. The whole field of penetration testing is based on the notion that you have access to software that you can actually see and understand. And actually the most hardened software in the world, the safest software, tends to be open-source software, because there’s this saying that with enough eyeballs, all bugs become shallow. That’s why open-source software is so ubiquitous and robust, because it’s had hundreds of thousands of developer eyeballs on it, looking at various sorts of vulnerabilities. And while they might not patch all of them, open-source software tends to be much more hardened and robust than closed software.

Ed Elson

但这和传统的知识产权生意很不一样：如果你创造了极有价值的东西，通常公司的第一要务是申请专利，确保它高度封闭，围起高墙然后变现。所以开放模型的生意让我很着迷，我很好奇它怎么运转。看看 Anthropic 和 OpenAI，它们在大模型领域排名第一第二，但每年仍在亏损数十亿美元。对它们来说，商业模式的一大问题是怎么提价、怎么把价格维持住，而开放模型生态在把价格往下压。对你们这样的公司——我对 DeepSeek、Qwen 以及其他中国开放模型也有同样的疑问——商业模式究竟怎么成立？因为据我理解，构建和训练模型是整件事里最贵的部分，而这部分在技术上是送出去的。长期来看商业模式怎么运作？

The thing that it’s quite different from, though, is the traditional business of intellectual property, where if you create something that is extremely valuable, usually the number one priority of the business is to put patents on it and make sure that it is extremely closed, so they can build gates around it and then monetize that thing. And so the open model business fascinates me; I wonder how it’s going to work. I look at companies like Anthropic and OpenAI, who are number one and number two in the world of LLMs, but they are still losing billions of dollars a year. For them, a big question around their business model is how can they raise prices, how can they keep prices up, while the open model ecosystem is bringing prices down. But for a company like yours, and something I wonder about with DeepSeek, with Qwen, with all of the other Chinese open models, how is that business model actually going to work? Because my understanding is that building and training a model is the most expensive part of the whole thing, and that’s the part that is technically being given away. So how will the business model work over the long term?

Misha Laskin

我们所知道的大多数基础技术，即便在软件时代，都是从封闭开始的；但如果这项技术足够重要、足够基础，随后就会出现一个开放版本，并占据大部分使用量。操作系统一开始是封闭的，Linux 是后来才出现的。封闭的比如 MS-DOS 和 Windows，微软的闭源操作系统，围绕它建立了很好的生意；开源版本 Linux 后来出现，现在全球超过 90% 的服务器——我觉得数字更高，超过 95%——运行着某种开源操作系统的变体，大多是 Linux。Android 也是：第一个真正的移动操作系统来自苹果，是封闭的，随后出现了开源的 Android，如今世界上大多数手机都运行 Android。互联网是开放协议，但一开始也是封闭的，基本是在政府和大学内部建起来的。所以现在发生的事，是一件非常自然的事的延续。

我觉得我们对 AI 的思考方式可能需要调整：模型是一种商品，就像电和水。说到电和水，我们不会想「我的能源公司怎么维持生意，它需要 90% 的毛利」。我们把能源和水看作商品，这些公司有一定利润，实际上不高，但这是为整个国家、整个世界提供的大规模基础设施服务。所以模型这一侧正在发生的利润压缩是健康的，因为智能应该被视为像电和水一样的基础品，它是商品。

不过推理的利润率，即便在开放模型上，实际上也相当高，原因是算力稀缺——GPU，也包括 CPU。算力之所以稀缺，是因为每瓦能装下的智能不断增长：模型变聪明的速度快于我们建设算力的速度。而当你身处智能市场，它有点像人工劳动力的市场，是一个非常正和的市场。你愿意把多少人工劳动力投向真正困难的问题，比如药物发现或网络防御？你大概会尽可能多地投入，去获得最好的网络防御系统、发现新药。

我认为在可预见的未来——我不知道会持续多久——智能会持续跑在我们建设算力的能力前面，因为算力是在物理世界里建起来的。所以每瓦得到的智能会持续增加。于是你希望尽可能压低智能的成本，这样就能用得更多，这是经典的杰文斯悖论。说到底，智能不应该被看作苹果 iPhone 那样的高端产品，它更像电和水，应该是商品。围绕智能的整套系统是昂贵的，它们会有一些利润；而今天无论闭源还是开放模型，推理的利润率其实都很健康。

For most foundational technologies that we’re aware of, even in the software era, they started off as closed, but if the technology is important and foundational enough, it was followed by an open version of it that then took up the majority of consumption. Operating systems started closed, and Linux came after. Closed could be MS-DOS and Windows, Microsoft’s closed operating system, which built a great business around it. The open-source version, Linux, came after, and now over 90% of the world’s servers, I think the number is even higher, over 95%, run on some variant of an open-source operating system, most often Linux. The same thing is true for Android. The first real mobile OS came from Apple and was closed, followed by an open-source version, Android, and now the majority of the world’s phones run on Android. The internet is an open protocol, but it started off closed; it was built basically within the government and universities. So what’s happening here is a continuation of a very natural thing.

And I think the way we probably need to reposition our thinking about AI is that the models are a commodity, like electricity or water. When we think about electricity and water, we don’t think, how does my energy company sustain its business, it needs to be making 90% margin. We think about energy and water as commodities. There is some margin that these companies sustain, which is actually low, but it’s a massive service that’s provided as fundamental infrastructure to the entire country and the entire world. So I think the margin compression on the model side is a healthy thing, because intelligence should be viewed as being as foundational as electricity and water. It’s a commodity.

Now, inference margins, even on open models, are actually quite high, and the reason for that is because compute is scarce, GPUs effectively, but CPUs as well. Compute is scarce because the amount of intelligence you can pack per watt keeps growing. These models get smarter faster than our ability to stand up compute. And when you’re in the intelligence market, that’s kind of like the market for artificial labor, and it’s a very positive-sum market. How much artificial labor would I throw at really difficult problems like drug discovery or cyber defense? You’d probably throw as much as possible at it, in order to get the best cyber-defensive systems, in order to discover new drugs.

And what I think will continue happening is that, for the foreseeable future, and I don’t know how long that lasts, intelligence is going to outpace our ability to stand up compute, because with compute you’re standing up stuff in the physical world. And so the amount of intelligence you’ll get per watt is going to keep increasing. And so you want to drive the cost of intelligence down as much as possible so you can use it more. There’s this classic Jevons paradox kind of thing. So all this to say that intelligence should not be seen as a premium product like an Apple iPhone. Intelligence is more like electricity and water, and it should be a commodity. The whole system around intelligence is expensive; they’ll have some margin, but the margins today on inference are actually very healthy for both closed and open model inference.

Ed Elson

公用事业这个观点很有意思。我记得 Sam Altman 说过，他认为 OpenAI 有可能走到那一步，虽然从他们实际追求的业务方向看并不像。我也同意推理利润很健康。但这么说时，我们漏掉了训练这一块，那似乎是一笔非常重要、非常大的开支。而公用事业一般会以某种方式得到联邦、州或地方政府的补贴，好让人人都用得起，比如电和水。你觉得 AI 经济会不会因为训练非常昂贵，最终出现类似公用事业那样的补贴？

On the inference, I think the utility point is really interesting. I think Sam Altman has said that he thinks there’s a world where that might happen with OpenAI. It doesn’t seem to be the direction that they’re actually headed in terms of what they’re pursuing as a business, but the idea of AI as a utility is really interesting, and agreed, inference margins are very healthy. But when we say that, we’re also leaving out the question of training, which seems to be a very important and significant expense. On utilities, generally speaking, they are in some way subsidized by federal, state and local governments in order to make them accessible for everyone, electricity, water, etc. Do you see a world where, perhaps because training is very expensive, there will be some level of subsidization similar to what we see with utilities?

Misha Laskin

也许在未来某个时候。我觉得现在市场还太早，说不准。我描述的其实是 AI 理想的终局：它是一种公用事业，就像互联网——互联网实际上就是公用事业，是大多数人负担得起的商品，还有很多项目在研究怎么让偏远地区也用上 Wi-Fi。所以我不确定 AI 会走电力的路还是互联网的路。但无论是否有政府补贴，两者都被视为大规模的商品化服务，而在它们之上建起了很多非常成功的高利润公司。很多高利润公司建在互联网之上，尽管接入互联网本身相当便宜。

Perhaps in some future. I think right now the market is too early to tell. What I’m describing is kind of the desirable end state for AI, that it’s a utility, like the internet. The internet is effectively a utility. It’s a commodity that is affordable to most people, and there are a lot of projects around how you get Wi-Fi to work in rural areas as well. So it’s unclear to me whether AI goes down the path of electricity or the path of the internet. But nonetheless, whether one is government-subsidized or not, both are seen as large-scale commodity services, and a lot of very successful high-margin companies are built on top of them. A lot of companies with high margins are built on top of the internet, even though access to the internet in its own right is fairly affordable.

Ed Elson

中国开放模型很有意思。DeepSeek 出来，我们都记得市场发生了什么：大家突然恐慌，美国 AI 股票纷纷下跌。而在中国发生的事——这也是你们成为美国战略性公司的原因，说错请纠正——似乎是中国用大量政府补贴来做这些模型，然后免费发给全世界，改变了整个 AI 模型市场的格局，大幅压低了价格。但最终据我理解，是政府、是中共在买单，至少 DeepSeek 看起来是这样。这么说错了吗？中国政府在多大程度上参与了资助和推广这些开放模型？

The Chinese open models are so interesting, because DeepSeek comes along and we all remember what happened in the market. Suddenly everyone panics and you see all of the US AI stocks going down. But what is happening in China, which is why you have emerged as this strategically important company for the US, and correct me if I’m wrong, is that it seems China is using vast amounts of government subsidies to create these models, which they then distribute to the world for free, changing the entire complexion of the AI model market and bringing down prices significantly. But ultimately, from my understanding, the CCP seems to be paying for it, at least when it comes to DeepSeek. Is that wrong? To what extent is the Chinese government involved in funding and promoting these open models?

Misha Laskin

关于中共的参与，我知道的可能并不比你多。但我同意，如果没有政府某种形式的深度支持，很难解释中国的开放模型生态。就是很难解释。这些公司其实已经存在一段时间了，花了好几年才走到真正收取可观费用的地步，即便如此，你说得对，需要投入巨额资金。

我认为中国这些公司的商业模式和西方实验室、分发商之间有很大差别。从外面看，中国有点像一个封闭市场：虽然有开放模型，但我在中国没看到多少除了模型开发方之外的分发商。DeepSeek 有自己的 API，Kimi 有自己的 API，GLM、智谱也有自己的 API。从外面看，它们的商业模式和西方的闭源模型并没有太大不同——只是在中国国内市场。

I probably don’t know much more than you do on the subject of the CCP’s involvement, but I would agree with you that it’s very hard to explain the Chinese open model ecosystem without deep support from the government in some way. It’s just hard to explain. These companies were around for some time. It took a few years for them to get to the point where they actually started charging meaningful money for their products, and even so, you’re right, there’s a large investment that needs to be put into it.

I think there’s a big difference between the business models that these companies have in China versus Western labs and distributors. From the outside looking in, it looks in a sense like a captive market, in the sense that while there are open models in China, I don’t really see too many distributors of those open models in China outside of the companies that are building them. DeepSeek has their API, and Kimi has their API, and GLM, or Zhipu AI, has their API. And it seems to me, again from the outside looking in, that their business models are not dissimilar from closed models in the West, but in the domestic market in China.

Ed Elson

澄清一下：你是说它们在中国国内按闭源模型运作，在国外则是开放的？

To clarify, you’re saying that they operate as closed models within China, but outside of China they’re open?

Misha Laskin

至少从外面看我是这么感觉的，因为我没看到中国冒出一批分发商，比如在 DeepSeek 之外提供 DeepSeek 服务的。所以那里的市场结构有点意思。我觉得在西方，人们不太信任使用中国的 API；而中国是个大市场，所以我猜它们在国内的运作方式和闭源 API 差不多，然后在国外把模型免费送出去——反正在国外也没法真正变现，因为国外的人并不想用中国的 API。我确实认为这有地缘政治上的好处，也可以看作一种经济上的动机，去压低国外竞争对手的利润。

At least that is my perception from the outside looking in, because I’m not seeing distributors popping up in China that are serving DeepSeek outside of DeepSeek, for example. So something is interesting in the market structure there. I think there isn’t much trust in consuming from Chinese APIs in the West, and China is a big market, so I suspect the way their businesses operate within China is not too dissimilar from a closed API, and then you give away the model for free abroad, because you can’t really monetize it anyway, since people abroad don’t really want to use the Chinese API. And I do think it has geopolitical advantages, and one could think of it as an economic incentive to drive down the margins of your competitors abroad.

Misha Laskin

西方的情况相似但又不同：这是一个竞争性市场，你发布一个开放模型，会有一整个分发商生态去提供这个模型的服务。但开放模型开发方确实会获得一些我们正看到兑现的优势。第一，在造模型的过程中，你也造出了前沿的模型服务栈、智能体构建、微调等等，你在技术开发上大约领先竞争对手六个月，因为等你的模型发布时，支撑它的技术早已建好。所以你有技术领先。

另一点是，因为你是对国家、对西方具有战略意义的公司，你和企业、公共部门、主权实体建立的信任也非常高。从根本上说，我们看到开放模型最大的消费市场会在企业——比如大银行——以及公共部门和主权实体。卖给初创公司可能是很技术化的销售：「来试试我的产品」。卖给这些机构，你当然得在技术上赢、展示很强的技术差异，但它们也在看：你会不会是那个帮我用开放模型改造组织的战略伙伴？从这个角度看，开放模型开发方作为合适的战略伙伴，拥有很多信任。

In the West, it’s similar but different, in the sense that it’s a competitive market. When you release an open model, there’s an ecosystem of distributors that can basically go and serve that model. But you do get a few advantages as an open model builder that we’re seeing come to fruition. One is that by building the model, you’ve also built a frontier, state-of-the-art stack for serving the model, for building agents around it, fine-tuning and so forth. And you are about six months ahead of your competition in terms of your technology development, because by the time your model is released, the technology for supporting it was built out well in advance. So you have a technological lead.

And the other thing that you have, because you are a strategic company to the nation and to the West, is that the trust it builds with enterprises, public sector, sovereign entities is very high as well. And fundamentally, where we see the markets for consumption of open models, the biggest markets are going to be in enterprise, like the big banks, the public sector and sovereign entities. Whereas selling to a startup might be a very technical sale, like, hey, just go pick up my product and try it, selling to those entities, yes, you have to win on the technical stuff and show really strong technical differentiation, but they’re also looking at you as, are you going to be the strategic partner that helps me transform my organization with open models? And the open model builders have a lot of trust from that perspective, as being the right kind of strategic partner to transform the enterprise.

Ed Elson

欢迎回来，继续和 Misha Laskin 对话。你提到中国开放模型背后可能有某种地缘政治战略。如果情况是这个生态在中国国内有点封闭、在国外不封闭，那就完全符合我们的共识：中国开放权重模型的扩散——他们真的加倍下注、投入了大量资源——部分，我甚至觉得几乎全部，是为了像你说的那样侵蚀美国现有 AI 公司的利润、颠覆 AI 市场。我在 Prof G Markets 的搭档 Scott Galloway 把它叫作「AI 倾销」：中国进来，往经济里注入海量供给，搅乱那些在美国真正领先的公司的生意。至少我认为是这样。

所以问题是：你们 Reflection 被视为对此的回应。你们获得了世界上一些最重要机构的投资，包括 Nvidia，黄仁勋明确表示支持开放模型，认为这非常重要。你们还把模型卖给了政府基础设施相关的机构，我相信五角大楼也在用 Reflection 的模型。有一种说法是，Reflection 会成为美国对中国的回应。我的问题是：这会怎样反制中国？怎样改变市场、让美国重新领先？如果一切顺利，Reflection 在战略上的重要性在哪里？

We’re back with Misha Laskin. You mentioned that there might be some sort of geopolitical strategy in the Chinese open models. If it’s the case that the ecosystem is somewhat closed in China and not closed outside of China, I think that fully aligns with what we are in agreement on, which is that the proliferation of the Chinese open-weight models, which they have really doubled down on and invested significant resources into, was partly, actually I think almost entirely, a mission to, as you say, erode the margins of the existing AI businesses in America and upend the AI market. My co-host on Prof G Markets, Scott Galloway, calls it AI dumping: China goes in and says we’re going to inject a huge amount of supply into the economy, and that’s going to screw up the businesses that have really gotten ahead in America. At least that’s what I think is going on.

So I guess the question is, your company, Reflection, is seen as the response to that. You’ve had significant investment from some of the most important firms in the world, from Nvidia as well, and Jensen Huang has specifically stated his support for open models; he thinks it’s really important. You’ve also sold your model to government infrastructure groups. I believe the Pentagon is using Reflection’s models. There is this idea that Reflection is going to be America’s response to China. My question: how will this get back at China? How will this change the market and put America in front again? What is strategically important about Reflection, given things go right?

Misha Laskin

首先，我其实不同意「AI 倾销」这个前提，不认为它对西方市场是负面的。

First, I actually disagree with the premise of AI dumping and it somehow being negative for the Western markets.

Ed Elson

嗯，我其实也有点同意你，我只是这么设定问题。

Yeah, I kind of agree as well. I’m sort of framing it.

Misha Laskin

世界上存在优秀的开放模型，不管来自哪里，对我们所有人都是礼物。这件事碰巧发生在中国，而世界上有了大家可以在其上构建的优秀开放模型，这就是礼物。不管他们的意图是什么，他们给了我们一份礼物。看看我们最好的初创公司：它们开始被一些封闭实验室「削高罂粟」。这并不是说封闭实验室有什么不好，它们有自己的生意要做。如果编程非常赚钱，它们当然会围绕它做垂直产品，为什么不呢？但你需要一个有竞争的生态，让同样在做编程智能体的初创公司也能做成生意。那些初创公司做了什么？它们开始转向中国开放模型，在其上构建、为自己的工作负载定制，自立门户。所以这是中国实际上送给美国初创公司的礼物。如果因为某些地缘政治担忧把开源模型统统拿走，反而会伤害美国的创新生态。

所以第一点要想清楚：从构建者、开发者、创新的角度看，这些东西是礼物，这很好。我们完全可能身处一个没有好的开放模型的世界，那会是一个更令人担忧、更黑暗的世界。另一点是，AI 不是一个固定大小的零和蛋糕。当你压低某样东西的成本，需求会上升，又是杰文斯悖论。到目前为止，每次压低成本，这些智能系统的需求都会上升。现在企业显然变得很在意成本，并不是因为它们不愿投入——只要能换回数倍的价值，企业很乐意往 AI 里投几十亿美元——只是价值显现需要时间，而迄今为止 AI 一直相当昂贵。所以第一点，优秀开放模型的存在是净正面的。

那么从美国的地缘政治和战略角度看，为什么美国成为开放模型的领导者、并有一个遍布全球的生态如此重要？国内国外都是如此。国内，我们的财富 500 强企业出于风险认知、监管等各种原因，不想用中国模型。

The fact that there are great open models out there, period, doesn’t matter where they come from, is a gift to us all. It happened in China, and there are great open models in the world that people can build on. That is a gift. So whatever their intentions were, they gave us a gift. And how do we see that? Our best startups started getting tall-poppied by some of the closed labs. Again, it’s not anything bad or negative about the closed labs; they have businesses to build. So of course, if coding is very lucrative, they’re going to go build a vertical product around it. Why not? They have to build businesses. But you want a competitive ecosystem where the startups that are also building coding agents can build businesses too. And what did those startups do? They actually started moving to Chinese open models, building on top of them, customizing them for their workloads and standing up on their own two feet. So this is a gift that China gave to American startups. And if you were to take away open-source models altogether because of some geopolitical worries, it would actually hurt the American innovation ecosystem.

So I think the first thing we have to think about is that, from a builder, developer, innovation perspective, these things are gifts, and it’s great. We could be in a world where there were no good open models, and that would be a much more worrisome, darker world to be in. And the other thing is that AI is not a fixed-pie, zero-sum thing. When you drive down the cost of something, in this instance the demand goes up. It’s a Jevons paradox sort of thing. So far, every time you’ve driven down the costs, the demand has gone up for these intelligence systems. Right now enterprises are obviously getting very cost-conscious, not because they wouldn’t invest; an enterprise would happily put billions of dollars into AI if it yielded some multiple of that in value back to them, but it just takes some time for that to manifest, and to date AI has been quite expensive. So on the first point, the existence of great open models is a net positive.

Now, when we think about it from a geopolitical and strategic perspective for America, it’s very important that America is the leader in open model building and that there’s an ecosystem that’s distributed across the world, both domestically and abroad. Domestically, our Fortune 500 enterprises do not want to use Chinese models, for various reasons around risk perception and regulation.

Ed Elson

不过有些企业甚至在用，这让我挺惊讶。我很确定微软曾经测试过 DeepSeek 一阵子，我挺意外的。希望我没记错。

But some of them even have, which I’ve been surprised by. I’m pretty sure Microsoft was testing out DeepSeek for a moment, which I was really surprised by. I might be getting that wrong, hopefully not.

Misha Laskin

有些企业在试，因为需求实在太大；但很少看到中国开放模型真正进入财富 500 强的生产环境。这意味着我们的国内企业不具备同样的竞争优势，比不上那些能承担风险、不在乎这些的初创公司。所以我们需要提供好的方案。你可以把企业想成：它们不想租用智能，而想拥有，或者两者兼顾——你得给它们好方案。而我们现在的处境是，我们最大的企业和公共部门机构，相对中国公司、也相对我们生态里愿意承担那种风险的公司，处于竞争劣势。所以拥有能和中国模型竞争的东西、支持我们自己的公司和机构，极其重要。

So some enterprises are testing it out because, again, the demand for it is so high, but it’s very rare to see Chinese open models go into production in the Fortune 500. And what that means is that our domestic enterprises do not have the same competitive edge as the startups that are able to assume the risk and don’t care about it. So we need to provide good solutions. You can think about enterprises who, instead of renting their intelligence, want to own it, or want to do a bit of both. You have to provide them a good solution. And where we are in the market is that our largest enterprises and our public sector agencies are operating at a competitive disadvantage to Chinese companies, and also to other companies within our own ecosystem who are willing to take that sort of risk. So it’s extremely important to have something that’s competitive with Chinese models to support our own companies and agencies.

Misha Laskin

而在国外，让我们的盟友建立在美国技术之上非常重要，这就是地缘政治的部分。海外世界会采用最好的技术栈，不管是什么，只要最好就采用。Linux 和 Android 来自美国而不是中国，是有原因的：美国历来非常擅长打造世界上最好的开源技术。这是第一次，在如此重大的事情上，起源地不是美国。这意味着海外世界正在中国技术之上构建。

这会导致什么？开放模型可以看作它们所带来的基础设施的特洛伊木马。因为它只是发动机，你需要车，车是什么？是芯片。所以当中国公司把开放模型卖进中东、其他亚洲国家、巴西等地时，它们也会开始卖芯片和数据中心。这意味着其他国家、盟国政府会开始被锁定在中国的基础设施上，这才是美国国家安全层面的真正风险。所以无论国内还是国际，为了国家安全，我们都需要最好的技术栈——凭实力，我们的开放模型最好，我们的芯片最好，世界选择在我们的技术上构建。

所以这件事有微妙之处：中国模型很了不起，它们的存在是给世界的礼物；但为了我们的国家安全，我们需要一个盟友可以采用的优秀产品。

And then abroad, it’s really important that our allies build on American technology, and this is where the geopolitical part comes in. The world abroad will adopt whatever the best stack is. Whatever the best technology stack, they’ll adopt it. There’s a reason why Linux and Android came from America and not from China. America has traditionally been really good at building the best open-source technology in the world. This is the first time where, for something so major, it is not originating in America. And what that means is that the world abroad is building on Chinese technology.

What that will lead to is that open models, you can think about them as Trojan horses for the infrastructure that they bring. Because, again, it’s just the engine; you need the car around it, and what’s the car? Well, it’s the chips. So now, when Chinese companies go and sell open models into the Middle East or other Asian countries or Brazil or what have you, they will also start selling chips and their data centers. And what that means is that other governments, allied nations, will start getting locked into Chinese infrastructure, and that is really the risk to the United States from a national security perspective. So it’s important, both domestically and internationally, for our national security as a country, to have the best stack on merit: our open models are the best, our chips are the best, and the world chooses to build on our technology.

So it’s nuanced. Chinese models are incredible; it’s a gift to the world that they exist. But for our national security, we need to have a great offering that our allies can adopt.

Ed Elson

你说中国给了我们一份礼物时，我心想，他们真的会给美国送礼吗？我觉得你指的是，关于他们为什么这么做有两种理论。一种是这根本不是礼物，他们想侵蚀美国 AI 市场的经济性——听起来你并不完全同意。在你看来，更贴切的类比可能是「一带一路」：他们出去建设，让其他社会高度依赖他们的技术；一旦有了这种上风——就像我们高度依赖他们的关键矿产，这显然已成为大问题——就能在之后用作很大的筹码。

When you say the Chinese gave us a gift, I’m like, would they ever really give America a gift? And I think what you’re pointing to is that there are two theories behind why they would do this. One is that it’s not a gift at all and they’re trying to erode the economics of the AI market in the US. It sounds like you don’t fully agree with that. And in your view, maybe the analogy would be something like the Belt and Road Initiative, where they’re going out building, making other societies incredibly dependent on their own technology. And as soon as you have the upper hand in that way, maybe similar to the way we have a huge dependence on their critical minerals, which has obviously become a huge problem, that’s a lot of leverage that they can use later down the line.

Misha Laskin

没错。

That’s correct.

Ed Elson

那我就假设正在发生的是这个。下一个问题是：美国为什么不直接禁用中国模型？我一直在等这件事发生。为什么不干脆禁了 DeepSeek？为什么让它们用开放模型充斥我们的市场？我想是因为有用吧，但考虑到本届政府对中国的立场，我有点意外我们一直对它们这么开放。

So I’m going to assume that’s what is happening here. I guess the next question is, why wouldn’t the US just ban Chinese models? I’ve been sort of waiting for that to happen. Why not just ban DeepSeek? Why are we letting them flood our market with their open models? I guess it’s useful, but I’m a little surprised, given this administration’s positioning on China, why we have been kind of open to using their models.

Misha Laskin

我认为没有开源——不管来自哪里——对我们的经济才是真正毁灭性的。突然之间，我们的初创公司没法打造自己的智能，没法和大型在位者竞争；我们的研究员被拒之门外，除非你在某家大实验室，否则没法真正理解这些系统是怎么造的。我认为美国作为一个国家之所以成功，在于它能在全国范围内培育非常多元的竞争和创新，而开放让这成为可能。

实际发生的情况是：DeepSeek V3 出来时出现了市场恐慌，一些股票下跌，但那其实是非理性的反应。那些股票反弹了，而且表现很好，因为对算力等各种东西的需求上升了。现在有了一个开放模型，就有了更多算力买家。这是拥有型市场，拥有意味着我拿到模型、放在自己的算力上。所以只要有好的开放模型，对算力的需求就会更多、更多元。初创公司，以及那些做更多科学探索的团队，能够探索的思路范围会比任何一家公司都广得多。创新经济就是这样运作的，保护好这一点极其重要。

有一个简单的答案，我们会很乐意相信：中国不好，所以中国的开放模型不好。但实际情况更复杂：中国的开放模型很好，对生态很好，但我们也需要有自己的优秀开放模型。这是世界上的两个超级大国，美国是首要的超级大国。你在构建一项基础技术，就好比电只存在于一个地方而不在另一个地方：你必须有竞争力，必须确保世界建立在你的东西之上，这关乎你的国家安全。两件事看起来矛盾，但可以同时成立：中国开放模型的出现是好事，而这同时也是这个国家的国家安全问题。

I think the lack of open source, period, wherever it comes from, will be the thing that’s devastating for our economy, because all of a sudden our startups can’t build their own intelligence and can’t compete with large incumbents. Our researchers get locked out and can’t really understand how these systems are built unless you’re in one of the big labs. I think the whole notion of America’s success as a country is in its ability to foster very diverse competition and innovation across this country, and openness enables you to do that.

What actually happened is that when DeepSeek V3 came out, there was a market scare and some stocks went down, but the reality is that it was an irrational reaction. Those stocks rebounded and started doing very well, because the demand for things like compute went up. Now you have an open model out there, so you have more buyers for compute. It’s an ownership market, and ownership means that I get the model and I put it on my own compute. So by nature of having a good open model, there’s just going to be more, and more diverse, demand for compute. The startups, and the ones doing more scientific pursuits, will be able to explore a much wider set of ideas than any one company could. This is just the way the innovation economy works, and it’s extremely important that we preserve that.

So there’s a simple answer that would be nice for us to believe, that China is bad, therefore open models from China are bad. But it’s actually more complicated. Open models from China are very good, really good for the ecosystem, but we also need to have our own great open models. These are the two superpowers of the world, and America is the predominant superpower. You’re building a foundational technology; it’s like if electricity existed in one place but not in another. You have to be competitive, you have to ensure that the world builds on top of your stuff, for your national security. And while they seem at odds, both things can be true: it’s great that Chinese open models are out, but it’s also a matter of national security for this country.

Ed Elson

你觉得 Anthropic 和 OpenAI 会不会在某个时候转向开放？如果我们正朝这个方向走——基本上免费送出发动机，理念是这样会激发更多活动、更多增长、更多创新，带来更繁荣、更高效的生态，回报会在未来体现——而且看起来我们可能真在这样走，尤其中国已经决定这么做。为什么不用更便宜的开放权重模型？这难道不意味着 Anthropic 和 OpenAI 也得跟上吗？它们会怎么做？

Do you think Anthropic and OpenAI will at any point switch to being open? If this is the direction we’re headed, where we’re basically giving the engines out for free, and the idea is that if you give the engines out for free, that’s going to spur a lot more activity, growth and innovation, resulting in a more prosperous and productive ecosystem where the returns will eventually be seen in the future, and it seems like maybe we are, especially if China’s just deciding to do that, why would you not use cheaper open-weight models? Does that not mean that Anthropic and OpenAI are going to have to get with the program? What are they going to do?

Misha Laskin

我觉得它们会做得很好。这些是了不起的公司，以前所未有的速度扩张，不断在 AI 上发现新突破。这是一个正和的世界，是「这个和那个」并存：会有成功的开放模型公司、在开放模型之上构建的公司和分发商，也会有成功的闭源模型公司。有租赁市场，也有拥有市场。拥有的需求开始增长，并不意味着每个人都会买房，仍然会有人租、有人买。

OpenAI 过去开源过模型。我觉得当某件事不是公司的核心战略时，你往往不会那么上心，往往会开源一个比你实际在变现的东西弱一些的版本。我们在开源发布上看到的正是这样：Google 通过 Gemma 为开源社区做了很多，但它和 Gemini 是很不一样的模型档次；OpenAI 发布 gpt-oss 时也类似，和它们手里真正的内部模型完全不是一个档次。所以我不认为它们的策略一定会改变，因为光是深挖现有业务，它们面前的机会就已经太多了。它们也许会，但我觉得它们都会做得很好。

Yeah, I think they’re going to do just fine. These are incredible companies that have scaled at an unprecedented rate and keep discovering new breakthroughs in AI. I think you can be in a world where it’s positive-sum, this and that. There will be open model companies and builders on top of open models and distributors that are successful, and there will be closed model companies that are successful. You have a rental and an ownership market. And just because ownership is starting to grow in its demand doesn’t mean that everyone will then own a house. There will still be people who are renting and people who are owning.

OpenAI has open-sourced models in the past. I think that when something is not the core strategy of the company, you tend to not pay as much attention to it, and you tend to open source maybe a lesser version of the thing that you’re actually monetizing. That’s what we’ve seen from open-source launches. Google has done a lot for the open-source community with the Gemma launches, but it’s a very different model class than the Gemini models. Similarly when OpenAI launched gpt-oss, it was very different from the class of actual internal models they were sitting on. So I don’t see that strategy necessarily changing for them, because they have so much opportunity ahead of them just chipping away at their current business. They might, but I think they’re also going to do just fine.

Ed Elson

Reflection 最近上了新闻，据报道 Nvidia 曾在洽谈收购你们，或至少加深合作、追加投资。Nvidia 已经是 Reflection 最大的股东之一。如果你能谈谈这件事最好。更宽泛地说，Reflection 接下来几年的计划是什么？

Reflection has been in the news recently. Reportedly Nvidia was in talks to acquire your company, or at least deepen the relationship, deepen their investment. Nvidia is already one of Reflection’s largest shareholders. If you can speak to that, that would be great. But more generally, what’s next for Reflection? What is the plan over the next few years?

Misha Laskin

公司的目标是继续打造优秀的开放模型，部署到企业、公共部门、主权这些客户群，围绕它建立一家具有世代意义的公司。我们也希望自己不是唯一一家：希望今年、明年在美国出现一个开放模型开发者的生态，在质量上能与中国开放模型生态相媲美。我认为中国开放模型建设之所以强大，在于它不是一家公司，而是好几家公司、一整个生态，你追我赶，确保整体上总有很棒的模型出来。所以我们希望 Reflection 成为这个生态里独立的领导者，但也希望在西方看到一整个这样的公司生态涌现。至于媒体的猜测，我们不予置评；公司的计划就是继续打造优秀模型，并把它们分发到企业、公共部门和主权客户。

The goal for the company is to continue building great open models, deploy them into customer segments, into enterprise, public sector, sovereign, and build a generational company around this. And we hope that we’re not the only ones. We hope to see an ecosystem of open model builders come out of the States this year and next year that is comparable in quality to the Chinese open model ecosystem. I think the thing that makes China so powerful in its open model buildout is that it’s not just one company, it’s a handful of companies and a whole ecosystem. And the fact that one keeps leapfrogging the other ensures that, as a composite, there are always great models coming out from there. So we want Reflection to be a standalone leader in this ecosystem, but we want to see a whole ecosystem of such companies emerge in the West. We don’t really comment on media speculation, but the plan for the company is to continue building great models and distribute them into enterprise, public sector and sovereign accounts.

Ed Elson

对于正在考虑创业、考虑 AI 的听众：你打造了一家极有价值的公司，最近一轮估值 250 亿美元，在 Google 工作过，在这个领域有很多经验。对正在考虑自己创业的年轻人，你有什么建议？

For anyone listening to this who is thinking about entrepreneurship, thinking about AI: you’ve built an incredibly valuable company, $25 billion in the most recent round, you’ve worked at Google, you have a lot of expertise in this area. Any advice for younger people listening to this thinking about entrepreneurship and maybe starting a company themselves?

Misha Laskin

我最大的体会很直白：要深刻地知道自己对什么感兴趣。现在还不知道也没关系，不用急着去创业。为了创业而创业不是好理由；为了钱创业也不是，有很多好得多、风险低得多的赚钱方式。它必须来自你深深感兴趣、愿意投入大量时间、即便艰难也会坚持到底的东西。

对我来说，走到这家公司之前，是多年学习物理、在那里培养出对科学的兴趣；然后在 2010 年代中期深度神经网络开始奏效时发现了对 AI 的兴趣；再进入前沿、真正理解某个领域的前沿，理解问题空间是什么；然后找到一个我完全信任、很兴奋要一起做事的合伙人，因为我们之前已经在项目上非常紧密地合作过。所以如果有人在考虑创业，我的建议是别急，不要觉得「我必须为了创业而创业」。它应该来自深深的兴趣和投入，时机对了你自己会感觉到。我们都有自己的「胡扯探测器」，费曼有句话：最容易骗的人是你自己。所以也许我的建议就是：别骗自己去创业，只有当它来自你对那个领域和想法的深切兴趣时才去做。

The biggest learning that I’ve had, and it’s pretty straightforward, is to have a deep sense of what it is that you’re interested in. And it’s okay if you don’t know what that thing is now. There’s no rush to jump into building companies. Building a company for the sake of building a company is not a good reason. Building a company for financial reasons isn’t either; there are plenty of way better, less risky ways to make money. So it has to come from something that you’re deeply interested in, that you’re going to commit a lot of time to and see through even when things are tough.

For me, the journey of getting to this company was preceded by studying physics for years, developing an interest in science there, discovering this interest in AI in the mid-2010s as deep neural networks were starting to work, going in and working at the frontier, really understanding the frontier of something, understanding what the problem set is, and finding a business partner that I trust fully and was very excited to build something with, because we had already worked together very closely on projects before that. So if someone is considering starting a company, I’d advise it not to be rushed, in the sense of, oh, I need to start a company for the sake of starting it. It should just come from a place of deep interest and commitment, and people will feel when it’s right. We all have our own BS detectors, and there’s a quote by Feynman that the easiest person to fool is yourself. So maybe the advice is not to fool yourself into building a company, but only do it if it’s coming from a place of deep interest in that space and idea.

Ed Elson

Misha Laskin 是 Reflection 的联合创始人兼 CEO。Misha，感谢你的时间。

Misha Laskin is the co-founder and CEO of Reflection. Misha, we appreciate your time.

Misha Laskin

Ed，谢谢邀请。

Ed, thanks for having me.