Reflection 创始人谈 Beam：做「西方的 DeepSeek」

Alex Heath

准备这期节目时我一直在想时机。Q3 财报电话会里提到开放模型的次数涨了 6 倍，开放模型的 token 如今已经占 Vercel 和 OpenRouter 流量的大头。你们一直在隐身状态，现在有大消息要宣布。先请两位谈谈：是什么在推动这股对开放 token 的热潮？

I’ve been thinking a lot about timing as we were preparing for this conversation. Mentions of open models on earnings calls 6x’ed in Q3, and open-model tokens are now the majority of what’s flowing through Vercel and OpenRouter. You’ve been in stealth and have big news on the open-model front. But first, what do you think is driving the surging interest in open tokens right now?

Misha Laskin

我们大约一年前决定全力做优秀的开放模型。当时预见到的一些事，发生得比我们预期还快。

We decided to really push for building great open models at this company about a year ago, and some things we saw coming have happened faster than we expected.

Misha Laskin

可以用房地产作比。AI 作为商业市场只存在了大约三年，一开始就是租赁市场：闭源模型相当于租公寓，你在租用智能。年轻时、刚毕业时租房很合理；等你成家、手头宽裕了，拥有就变得重要得多。AI 的时间线被压缩了——西方这三年基本只有「租」这一个选项。

Think of it with a real-estate analogy. This market has only commercially existed for maybe three years, and it started as a rental market: closed models are the equivalent of renting an apartment—you’re renting your intelligence. That makes sense when you’re a student or just graduated. But as you get older, have a family and more finances, ownership matters a lot more. AI timelines are compressed, and in the West there’s effectively only been a rental alternative for those three years.

Misha Laskin

现在 AI 采用率上来了，几年前的小创业公司成了大企业，大企业也都用上了 AI，一个「拥有」市场正在出现。而拥有智能的唯一方式，按定义就是开放：你能在自己的设施上跑、能控制、能定制。

Now, as AI adoption has increased—tiny startups became giant enterprises, and large enterprises adopted AI—an ownership market is coming in. And the only way to own intelligence is, by definition, if it’s open: you can run it on your stuff, control it, customize it.

Misha Laskin

12 个月前几乎没有企业在用开放模型，连创业公司都不多。6 个月前，一批做大的应用层创业公司开始切到开放模型，并继续高速增长。大概 6 个月前 OpenRouter、Vercel 上约 30% 的 token 消耗流向开放模型、70% 流向闭源；6 个月后基本反过来了，开放约 70%、闭源约 30%。这是市场快速成熟的信号，另一个因素当然是地缘政治。

Twelve months ago, not many enterprises were using open models, and not even many startups. Six months ago, a lot of startups that had grown at the application level started switching to open models and kept growing very quickly. Maybe six months ago 30% of token consumption on OpenRouter and Vercel went to open models and 70% to closed; six months later it’s flipped—more like 70% open, 30% closed. That’s the market maturing very fast, and obviously the geopolitical thing is the other factor.

Alex Heath

Giannis，从技术角度看开放模型的现状，这一年的进展让你意外吗？

Giannis, from your perspective technically, has the progress of open models surprised you compared with what you expected a year ago?

Ioannis Antonoglou

并不意外。回头看，开放模型一直跟住了闭源模型；以最大的中国模型为例，现在开放前沿与闭源前沿的距离甚至更近了。我看不出开放前沿为什么非要落后于闭源前沿——只要有人才和算力，完全可以做出和闭源同一量级的开放模型。

No. If you look back, open models have kept pace with closed models, and with the biggest Chinese models there’s maybe an even shorter distance between the open and closed frontier now. I don’t see why the open frontier should stay behind the closed frontier. It’s a matter of having the talent and the compute; you can definitely build open models that punch at the same level as closed ones.

Ioannis Antonoglou

企业和创业公司采用开放模型，我认为有两个主要原因：一是能力已经够了，有些非常强的开放模型能真正替代闭源模型；二是市场开始成熟。软件市场的规律是，一旦成熟就会出现向开放替代品迁移的趋势。AI 市场继续成熟，也会发生类似的事。

There are two main reasons for adoption by enterprises and startups. One, the capabilities are there now—some really powerful open models can be used in place of closed models. Two, the market has started to mature. In software, whenever a market matures there’s always a movement toward the open alternative, and I think something similar will happen as the AI market matures further.

Alex Heath

为什么至今是中国实验室主导了开放模型？Meta 在 Llama 3 时有过高光，但西方后来完全丢掉了开放前沿，直到你们出现。

Why have Chinese labs dominated open models until now? Meta had its moment with Llama 3, but the West completely lost the open frontier until you.

Misha Laskin

西方的技术发展归根结底是资本和商业驱动的。在一个支持闭源开发的早期租赁市场里，先长出来的自然是闭源模型；只有当市场准备好「拥有」，开放模型才有商业支撑。美国曾经有最好的开放模型，巅峰就是 Llama 3，但当时 Meta 或其他人并没有真正的商业动机去做。

Technological development in the West is ultimately capitalistically and commercially driven. In an early rental market that primarily supports closed-model development, that’s what develops first; only once the market is ready for ownership is there commercial support for open models. America used to have the best open models—the peak was Llama 3—but there was no real commercial incentive at the time for Meta or anyone else.

Misha Laskin

我的看法是，中国的出发点更偏地缘。DeepSeek V3 出来时那么强，是个很难预料的意外时刻，让中国的智能在海外变得前所未有地重要，随后有了一股集中力量扶持开源生态的推动。这些实验室直到最近都没怎么赚钱，等智能足够好、市场成熟后才开始有收入。

China has a different mindset that’s more geopolitically motivated. There was a hard-to-predict moment when DeepSeek V3 came out and was so good, which made Chinese intelligence abroad relevant in a way it hadn’t been, and there was a concentrated push to support an open-source ecosystem for geopolitical reasons. These labs weren’t making much money until recently, when their intelligence got good enough and the market matured.

Alex Heath

这就说到你们的首个模型 Beam：约 5000 亿参数，在公开基准上看是同类最佳。你们还在做最后的验证——预计外界会怎么反应？

That brings us to Beam, your first model: around 500 billion parameters, and on open benchmarks it looks state of the art. You’re still in final validation. What do you expect the reaction to be?

Misha Laskin

Beam 是我们第一个开放权重模型，非常兴奋。Reed Hoffman 说创业是一边飞一边造飞机——现在飞机造好了。它总参数约 500B、激活 23B，实质性地推进了西方开放模型的前沿，在编码和智能体基准上尤其强，我们认为这样的规格和能力让它成为企业和公共部门的主力「工作马」模型。

Beam is our first open-weight model, and we’re very excited to release it. There’s a Reed Hoffman quote that building a company is building a plane while you’re flying it—and the plane’s been built. It’s around 500B total parameters, 23B active, and it advances the frontier of Western open models in a meaningful way. It’s particularly strong in coding and agentic benchmarks, and we believe the form factor makes it a really strong workhorse model for enterprises and the public sector.

Misha Laskin

另一点是它在开放模型中推理效率领先——也就是相对于推理时的算力或 token，解一个任务有多快。Giannis 和团队在这上面做了很了不起的工作。

The other thing we’re excited about is that it leads in reasoning efficiency relative to other open models—how fast it solves a task relative to the compute or tokens it uses at inference time. Giannis and team did incredible work there.

Ioannis Antonoglou

想让开放模型被广泛采用，尤其是被企业采用，它就得高效、紧凑，从每个 FLOP 里榨出最多的智能。这需要推理模型——用 token 去解任务。衡量推理效率，就是看解同等难度任务要花多少 token、多少时间、多少算力。我们的模型比 GLM 5.3 等开放前沿模型、以及一些最好的闭源模型如 GPT-6 Luna，推理效率高三到四倍。

If you want wide adoption for an open model, especially from enterprises, it needs to be efficient and compact and produce the most intelligence out of your FLOPs. That means a reasoner that uses tokens to solve tasks. You measure reasoning efficiency by how many tokens, how much time and compute it takes to solve a task of similar difficulty. We’ve built a model that can be three or four times more reasoning-efficient than open frontier models like GLM 5.3, or some of the best closed models like GPT-6 Luna.

Ioannis Antonoglou

怎么从这样一个相对小的模型里榨出最多的推理能力？答案是真正把强化学习规模化。我从 2012、2013 年 RL 的早期就开始做，Misha 和我一直是 RL 的忠实信徒。在 Reflection，我们把 RL 的算力推到了与其他开放实验室公开研究可直接对比的最高水平，就是为了从模型里榨出最多的智能——这正是企业和开发者在乎的：一个紧凑、用 token 用得聪明的模型。

How do you extract the most reasoning and intelligence out of this small model? By really scaling reinforcement learning. I started doing RL back in the early days, 2012–2013, and both Misha and I have always been true believers in RL. At Reflection we scaled RL to the highest levels of compute compared with what other open labs have publicly released, apples to apples—with the goal of extracting the most intelligence out of the model. That’s ultimately what enterprises and developers care about: a compact model that uses its tokens efficiently and intelligently.

Alex Heath

企业用例听起来很明显。消费级应用呢？

The enterprise use cases seem obvious. Do you see consumer applications for a lightweight, efficient reasoning model like this?

Misha Laskin

当然。开放模型的意义是让一个开发者生态在其上构建，让做消费应用的公司掌握自己的命运，拥有可以用自己数据定制的高效工作马。新一波消费体验由智能体模型驱动，这些看似基础的任务其实很费 token。如果模型体量小、推理又比别人高效约 4 倍，这些效率就会转化成更快的响应和低得多的服务成本——对消费场景，服务成本非常关键。

Definitely. The purpose of an open model is to enable an ecosystem of builders. It lets companies building consumer applications control their own destiny and have an efficient workhorse they can customize against their data. We’re seeing a new wave of consumer experiences powered by agentic models, and those seemingly basic tasks cost a lot of tokens. If the model is efficient in size and, say, 4x more efficient in reasoning, those gains translate into faster responses and much lower serving cost—and for consumer use cases the cost of serving matters a lot.

Alex Heath

你们是从零训练的，没有蒸馏。现在大家默认人人都在蒸馏，连闭源厂商也有传闻和指控。为什么这样选？

You trained this from scratch, no distillation. Everyone assumes everyone is distilling these days, even the closed players. Why approach it this way?

Ioannis Antonoglou

两个原因。第一，我们想确保这是一个完全由我们在美国（以及英国）团队构建的西方前沿开放模型。第二，要真正规模化 RL，你必须掌控预训练。预训练现在是比 RL 更成熟的科学，但要让高算力 RL 保持稳定——这是一个混合专家模型，要让路由和专家在 RL 中保持均衡稳定——你必须完全拥有它。

Two reasons. First, we wanted this to be a Western frontier open model fully built by our team in the United States—we also have a team in the UK. Second, when you really want to scale reinforcement learning, it’s extremely important that you control your pre-training. Pre-training is now a better-understood science than RL, but to keep a mixture-of-experts model stable during high-compute RL—keeping the router and experts balanced—you need to fully own it.

Ioannis Antonoglou

我们在预训练数据上下了大量功夫，让它播下正确的推理能力，再由 RL 接手放大；在架构、学习动力学和优化上也做了工作，让模型训完后足够稳定、有合适的先验。只有端到端优化、预训练和 RL 两端都创新，才能做出以体量而论非常强、把大量智能装进小模型的系统。

A lot of work went into the pre-training data so that it seeds the right reasoning capabilities that RL then picks up and accentuates, and into architecture, learning dynamics and optimization so that it’s stable and has the right biases and priors for RL. Only by optimizing the system end to end, innovating on both pre-training and RL, can you build something really capable for its size that packs a lot of intelligence into a smaller form factor.

Alex Heath

安全与对齐可能是当下 AI 乃至地缘政治的头号话题——最近有 OpenAI 等公司的失控智能体到处入侵的新闻。开放权重模型怎么做对齐？

Safety and alignment is maybe the number-one topic in AI and geopolitics right now—there are headlines about rogue agents from OpenAI and others hacking things. How do you approach alignment for an open-weight model?

Misha Laskin

先别想模型，看软件史。互联网协议是开放的，Linux 和 Android 是开放的，催生整个网络安全领域的强加密协议也是开放的。测试软件靠黑盒、白盒渗透测试。开放一直是软件安全的根基——Linus 定律说「足够多的眼睛让所有漏洞变浅」，所以开源软件在很多方面被认为比闭源更安全可信。AI 是软件的演进，它是机器，是一段软件。

Let’s not think about models for a moment. In software, internet protocols are open, operating systems like Linux and Android are open, and a lot of the strong encryption protocols that gave rise to cybersecurity are open. You test software with black-box and white-box penetration testing. Openness has been fundamental to securing software—there’s Linus’s law, “with enough eyeballs, all bugs become shallow”—which is why open-source software is in many ways considered safer and more trustworthy. AI is an evolution of software; it’s a machine, a piece of software.

Alex Heath

所以它不是神。

So it’s not a god.

Misha Laskin

据我所知不是（笑）。它和 Linux、Android 一样不是神。它是一台非常复杂、带着一长串漏洞的机器；只靠少数人接触，不可能高置信度地发现并修补那条长尾。闭源实验室合计可能有 300 名安全研究员，用意很好，但这就像说我全身只有 300 个白细胞——我信得过它们清掉每天攻击身体的所有病原体吗？不，我宁愿越多越好。要从 300 人扩到 30 万人去检查这些系统。

Not that I’m aware of. It’s no more a god than Linux or Android. It’s a very sophisticated machine with a long tail of vulnerabilities, and it’s impossible to discover and patch that long tail with high confidence when only a small number of people have access. Across the closed labs there might be 300 safety researchers—the intentions are great—but it’s like saying I have 300 white blood cells in my whole body. Do I trust them to find all the pathogens attacking me every day? No, I’d rather have as many as possible. You need to go from 300 to 300,000 people who can inspect these systems.

Alex Heath

说到这，Hugging Face 被 OpenAI 的模型入侵时，他们转而用了开放的中国模型来防御，因为拿不到最好的闭源网络安全能力。但闭源实验室里也有很多人认为这和 Linux 不同，到某个点应该把模型扣住。Giannis，你们在训更大的模型，有没有一个点会说「我们是开放公司，但这个太强了，不放」？

As you say that, I think of Hugging Face getting hacked by OpenAI—they turned to open Chinese models because they needed to defend and didn’t have access to the best closed cyber capabilities. At the same time, many in the closed labs believe this is different from Linux and maybe there’s a point where it should be held back. Giannis, on larger training runs, is there a point where you’d say this is too capable to release?

Ioannis Antonoglou

有可能到某个能力水平，你需要在部署方式上更谨慎，比如更严格地限制谁能访问。但从根本上说，与研究社区共享一直对 AI 发展是净正面的，安全尤其如此：强大的开放模型一到研究者手里，大量洞见和新防护手段就被开发出来。

It’s possible it gets to a level of capability where you want to be more careful about how you deploy it—maybe restrict who has access more. But fundamentally, sharing with the research community has always been a net positive for AI development, and we saw that with safety in particular: once powerful open models were available to researchers, so many insights and new safeguards were developed.

Ioannis Antonoglou

确实，模型越来越强，会突破沙箱、会在网上结成群体、会彼此串通，这些我们看到得越来越多。但同时它们也更擅长自我监管，你可以用 AI 来管控 AI。闭源实验室公开了这些智能体如何运作的发现，只会给我们更多信息去构建防御。开放和透明让社会为 AI 的到来准备得更好。

It’s true that as models get more powerful they hack their sandboxes, form communities online, collude with each other—we see this more and more. At the same time they get better at policing themselves; you can use AI to control AI. The fact that even the closed labs released findings on how these agents operated only gives us more information to develop defenses. Being open and transparent makes society better prepared.

Misha Laskin

我们当然想过阈值问题，华盛顿现在也在大量讨论。但这与开放还是闭源无关：到了某个能力水平，业界和政府应该有相应政策，规定什么能发、在什么门控下发。我们不是「闭源不安全、开放才安全」的强硬派——安全的是一个生态：闭源模型、开放模型各有立场，这样的世界才更安全。

We’ve definitely thought about that, and there’s a lot of discussion in DC. It’s independent of whether a model is open or closed: if you hit a certain capability, industry and government should have policies around what gets released and under what gated measures. I wouldn’t describe us as hardliners on “closed is not safe.” An ecosystem is safe—closed models with their points of view, open models with theirs—that’s the safer world.

Misha Laskin

大家常说的是坏人滥用，而我觉得较少被讨论的是「意外后果」。OpenAI 与 Hugging Face 的事件就是意外后果：做大型 RL 训练时，整个训练周期里可能有十亿个智能体在做事，100 或 300 个研究员不可能逐一追踪。那些智能体逃了出去；Hugging Face 试图用闭源模型保护自己，但安全护栏不允许用于网络攻防，于是只能转向开放模型。这恰恰说明生态在起作用。

Beyond the bad-actor issue, there’s the unintended-consequences issue people don’t talk about as often. The OpenAI–Hugging Face hack was unintended consequences: in these large RL runs there can be a billion agents over the life of a run, and 100 or 300 researchers can’t track each one. The agents hacked out. Hugging Face tried to use closed models to protect themselves, but the safety guardrails prevented them from being used for cyber, so they had to turn to an open model. That shows an ecosystem working.

Alex Heath

技术上，做开放模型的安全对齐和闭源有什么不同吗？

Technically, does alignment differ when you’re building an open model?

Ioannis Antonoglou

技术和配方与闭源实验室基本相同，分两部分。首先要有覆盖各种风险画像的评测，衡量模型的表现和响应，确保它不是一个什么都拒绝的模型——什么都拒绝最安全，但没用。有用性和安全性是一对权衡，你得测对、往正确方向爬坡。

The technology and recipes are quite similar to the closed labs, in two parts. First, you need the right evaluations across all the different risk profiles to measure the model’s responses, and ensure it’s not a model that refuses everything—that’s the safest model but not useful. Usefulness and safety are a trade-off you need to measure right and hill-climb the right way.

Ioannis Antonoglou

有了评测，定义好你想要的行为，收集好坏行为的示例，再用常规方法——模仿学习（SFT）或强化学习——让模型按预期行事。这些方法过去五年在闭源实验室和开放研究社区里发展起来；我们有一支只专注安全的研究团队，确保放出去的模型评测到位、处在安全与有用性的前沿。

Once you have evaluations and have defined the behavior you want by collecting examples of good and bad behavior, you apply the usual methods—imitation learning, or SFT, and reinforcement learning—so the model behaves as intended. These methods were developed over the past five years in the closed labs and the open research community, and we have a team of researchers focused only on safety, making sure our models are evaluated correctly and at the frontier of safety and usefulness.

Alex Heath

但有人会说，这种能力一旦完全扩散，我们无法想象它会做什么，所以不能拿操作系统类比——它更像一个我们不知道如何控制的「存在」。这种看法正在进入公众意识。你们怎么看？

Even if you buy that, some argue that capability will create intelligence we cannot fathom once fully diffused, so it’s not fair to compare it to an operating system—it’s almost like a being we don’t know how to control. That’s seeping into public consciousness. Where do you sit?

Ioannis Antonoglou

这又回到 Misha 说的权力分散。每当只有一两个机构控制如此强大的东西，即使出于最好的意图，它也是单点故障。你需要冗余，避免一个系统失灵就全盘失灵。Hugging Face 就是典型：如果能用 AI 去防御行为失当的 AI，世界就更安全；让更多机构能用强模型做网络安全、做防护，是净正面的。

It goes back to the diffusion of power Misha was talking about. Whenever only one or two institutions control something so powerful, even with the best intentions, it’s a single node of failure. You want redundancy so that if one system fails, all systems haven’t failed. Hugging Face is a prime example: if you can use AI to protect against badly behaving AI, the world ends up safer, and enabling other institutions to use powerful models for cybersecurity is a net positive.

Alex Heath

Reflection 是怎么来的？你们在 DeepMind 共事过，我听说 Misha 曾向 Giannis 汇报。

What’s the origin story of Reflection? You worked together at DeepMind, and I heard Misha reported to Giannis at one point.

Misha Laskin

对，Giannis 基本上是我的老板，不过现在我们是平等的合伙人。当时我们一起做 Gemini 1 和 1.5，Giannis 是 RLHF 负责人，也是 DeepMind 最有名的 RL 研究者之一。我是看到 AlphaGo 后从物理转到 AI 的，后来才知道 Giannis 是作者之一，加入 DeepMind 在他的 RL 团队工作，有点像见到偶像。

Yes, Giannis was basically my boss—though we’re actually equal partners now. We worked on Gemini 1 and 1.5; Giannis was the RLHF lead and one of the most storied RL researchers at DeepMind. I switched from physics to AI when I saw AlphaGo come out, later learned Giannis was an author, and joining his RL team at DeepMind was a bit of a celebrity-crush moment.

Misha Laskin

那时 Gemini 1、1.5 很擅长聊天，但还不擅长编码和智能体。我们创业时押了两个注：科学上，RL 是让模型获得智能体与编码能力的路径；市场上，会有强大的西方开放模型可供我们在其上做研究、降低成本。第一年里 RL 成了全行业共识，我们在小规模上拿出了很强的 RL 结果；但 DeepSeek V3 之后西方开放前沿反而退潮，我们等了几个月也没等到同样好的模型，于是公司转向自己做。

Those Gemini 1 and 1.5 models were really good at chat but not at coding or agentic tasks. We started the company with two bets: a scientific bet that RL was the way to get agentic and coding capabilities, and a market bet that there’d be powerful Western open models to build on so we could research more cost-effectively. Within the first year RL took off across the industry, and we showed strong RL results at small scale. But the Western open frontier receded after the DeepSeek V3 moment; we waited a few months for something as good, saw it wasn’t happening, and changed course.

Misha Laskin

更有意思的是怎么开始的。做 Gemini 时会有跨办公室冲刺，我在纽约，Giannis 在伦敦。我们其实各自都在想创业，但对老板说「要不要一起干点什么」挺吓人的。那晚我带一群人去 Comedy Cellar，赶上一场极其粗俗的演出，有点尴尬；之后我们走了很久，在 Google 办公室附近一艘卖龙虾卷的船上吃饭，是 Giannis 先开的口。对我来说，只要是和 Giannis 一起，做什么都行。我们都信 RL、都信 AGI，对世界走向的判断很相似。

The better story is how we started. On Gemini there were sprints where teams from different offices came together; I was in New York, Giannis in London. We were thinking these ideas independently, but telling your boss “do you want to start something?” was intimidating. One evening I took a group to the Comedy Cellar—the raunchiest set I’ve ever been to, a little awkward—then we took a long walk and grabbed dinner on a boat by the Google office that serves lobster rolls, and Giannis actually broached the subject. For me, whatever it was, if it was with Giannis, I was in. We’re both RL-pilled and AGI-pilled with a similar view of where the world was going.

Alex Heath

有传闻说，是你们的大投资方 Nvidia 的黄仁勋推动你们走开放路线？

I heard a rumor that Jensen, with Nvidia as a large investor, pushed you to go the open route. Accurate?

Misha Laskin

我们是独立得出这个结论的。Nvidia 在整个生态里都是大投资方，因为生态很重要，而开放模型是生态的核心推动力，所以愿景很一致，我们也很感激他们的支持。

We came to that conclusion independently. Nvidia is a big investor across the whole ecosystem because the ecosystem matters, and open models are a core enabler of it, so there’s a lot of vision alignment, and we’re very grateful for their support.

Alex Heath

你们融了两三十亿美元，上一次已知估值约 250 亿美元，还签了大额算力：与 Nebius 约 10 亿美元、与 SpaceX 逾 60 亿美元。这意味着你们看得到收入爬坡。开放模型怎么赚钱？

You’ve raised over two, two and a half billion, the last known valuation was around 25 billion, and you’ve struck big compute deals—$1 billion with Nebius, over $6 billion with SpaceX. That implies you see the revenue ramp. How do you make money on open models?

Misha Laskin

开源技术是某种东西的需求驱动器，比如 Kubernetes 驱动了 CPU 云的需求。开放智能最大的买家会是做大的创业公司、大企业、公共部门、主权实体和盟友国家。但光给他们一个开放模型，他们不知道怎么用——你从闭源实验室买一个 token，背后打包了集群管理、服务与推理软件、应用层等整套系统。开放模型就是这整套东西的需求驱动器。

Open-source technology is a demand driver for something—Kubernetes was a demand driver for CPU clouds. The largest buyers of open intelligence will be scaled-up startups, large enterprises, public sector, sovereign entities and allied countries. If you just give them an open model, they don’t know what to do with it, because a token from a closed lab packages up cluster management, serving and inference software, the application layer—the whole system. Open models are the demand driver for that.

Misha Laskin

成为企业信任的开放智能提供方，你就要为他们提供推理服务、帮他们搭智能体、建好并运营这些基础设施——以打包、交钥匙的方式提供让开放模型真正好用的一切。

When you’re an enterprise’s trusted open-intelligence provider, you serve inference to them, help them set up agents and the infrastructure, and run it—in a packaged, turnkey way, all the other stuff they need to make open models useful.

Ioannis Antonoglou

不只是前线部署。退一步看，大企业和盟友国家都意识到 AI 是不能外包给第三方的关键技术，必须自己掌控和拥有——数据、基础设施、部署都包括在内，资本和 capex 都在朝这个方向流。我们要成为所有想掌握自己命运的人的智能提供方，包括软件、推理和内部应用。

It’s not just forward-deployed work. Step back: big enterprises and allied nations realize AI is an extremely important technology that cannot be outsourced to a third party. It needs to be controlled and owned—the data, the infrastructure, the deployment—and there’s a lot of capital and capex moving that way. We’re becoming the intelligence provider for anyone who wants to own their own fate and intelligence, including software, inference and internal applications.

Ioannis Antonoglou

我一直把 AI 看作一种劳动力，这也是它定价方式的由来。劳动力的主要消费者是大企业、政府、大机构，恰恰是想要掌控自己智能的那批机构。所以中期看，服务这些客户是 AI 经济里最大的一块。

I always think of AI as a form of labor—that’s why it’s priced the way it’s priced. The main consumers of labor are big enterprises, government, big institutions—exactly the institutions that want to own their own intelligence. So serving these customers is, even in the medium term, the biggest chunk of the AI economy.

Misha Laskin

有个误解是这只是「前线部署工程师」式的服务生意。其实你做的是解锁推理，和闭源实验室、有 API 的中国开放实验室一样。估计未来几年会有数十吉瓦的开放模型推理需求；你要做的就是消除阻力，帮企业、主权和公共部门在规模上跑自己的推理。

There’s a misconception that this is a forward-deployed-engineer services business. What you’re doing is unlocking inference, just like closed labs and the open labs in China with APIs. Estimates are there’ll be tens of gigawatts of open-model inference within the next few years, and you remove frictions so enterprises, sovereigns and public-sector agencies can serve their own inference at scale.

Alex Heath

这是否意味着你们需要成为完全垂直整合的「开放超大规模云」，自建数据中心？

Does this imply you need to be a fully vertically integrated open hyperscaler with your own data centers?

Misha Laskin

在某种渐近的终局里大概是合理的，因为拥有能提升利润率。投资人会说开放模型在把闭源模型商品化、应用层竞争激烈被商品化、裸金属也被商品化——都对，一切都会被商品化到合理的利润率，跨栈累积利润的垂直整合公司最持久。但问题是什么时候做什么。市场还早，现在的游戏是把优秀的开放模型推出去、尽可能多地解锁推理，再去优化利润率。

In some asymptotic end state that probably makes sense, because ownership increases your margin. Investors will say open models are commoditizing closed models, the application layer is commoditized, bare metal is commoditized—and all of that is true; everything gets commoditized to an appropriate margin, and vertically integrated companies that accumulate margin across the stack will be the most durable. But the question is when you do what. The market is early. You’re in the game of getting great open models out, unlocking as much inference as possible, and then optimizing margin.

Alex Heath

就像阿里巴巴可以免费放出 Qwen、靠云赚钱。如果有人自己跑 Reflection 模型、不需要你们的服务，你们就赚不到钱。未来会不会只有 Reflection 的算力来服务 Reflection 的模型？

Like Alibaba giving away Qwen and making money on cloud. If someone runs a Reflection model without your help, you make nothing. Do you imagine a future where only Reflection compute serves Reflection models?

Misha Laskin

两点看法。一是我们一直低估这些市场有多大；多极化是好事，客户有选择，市场大到能容纳多家大公司。二是一家公司的收入潜力遵循一个简单的式子：智能密度 × 你掌控的算力 × 企业对你的信任。

Two views. First, we keep underestimating how big these markets are; multipolarity is good, customers have choice, and the markets will support multiple large companies. Second, a company’s revenue potential follows a simple equation: intelligence density times the amount of compute you have times the trust an enterprise has to work with you.

Alex Heath

最后一项有点模糊，信任怎么衡量？

That last one’s a little fuzzy. How do you measure trust?

Misha Laskin

模糊但非常重要。企业决策既看基础能力，也看有多信任你。我们发现，自己训模型本身就建立了大量信任——在模型发布前，我们就因为训模型和沿途展示的里程碑，在多个商业项目上取得了实质进展。企业、公共部门和主权实体不像硅谷公司那样技术精深，他们找的不只是推理供应商，而是「智能伙伴」。

It’s fuzzy but really important. Enterprises decide on base capability and how much they trust you. Building the models builds a lot of trust—we’ve made substantial progress on a number of commercial efforts before the model has been released because of that and the milestones we showed along the way. Enterprises, public sector and sovereigns aren’t as technically sophisticated as some Silicon Valley companies; they’re looking not just for an inference provider but an intelligence partner.

Alex Heath

政府越来越强调「别用中国开放模型，要用美国 AI」。企业用西方开放模型还是中国开放模型，到底有多重要？

There’s growing push from the administration that companies shouldn’t use open Chinese models. How important is it really to use a Western open model versus a Chinese one?

Ioannis Antonoglou

回看互联网时代，开放软件是一切的基础设施，而互联网正是建在美国开发的开放软件之上，这影响了美国在世界的位置。开放模型会成为新经济的基础设施层，不要低估拥有这一层的重要性，美国和西方最近开始意识到这一点。另一方面是安全风险：智能体可以被提示注入后去做有害的事——我不是说它们在这么做，而是有可能；如果用的不是你能检查、由可信公司提供的模型，就是在引入不必要的风险。

In the internet era, open software was the infrastructure everything was built on, developed here in the US, and that affected America’s position in the world. Open models will be the infrastructure layer of the new economy, and people shouldn’t underestimate owning that layer—the US and the West have started to realize this. Second, there are security risks: you can inject an agent with a prompt and it goes off and does something harmful. I’m not saying they’re doing that—but it’s possible, and if you don’t use models you can inspect, from a company you trust, you’re introducing unnecessary risk.

Alex Heath

短期内你们已有商业合作，比如韩国的项目和 Dell 合作。除了放到 Hugging Face，你们怎么把模型推向世界？

In the short term you have commercial partnerships—South Korea and Dell. Beyond putting the model on Hugging Face, how do you get it out to the world?

Misha Laskin

两件事：生态建设和商业结构。生态不会自己长出来，你得主动参与，与新云、推理提供商、超大规模云、Dell 这样的公司合作，确保模型在全栈原生支持。我们有专门的开源团队，除了模型，还会发布配套的开源工具，帮助社区理解模型、构建安全护栏和智能体。

Two things: ecosystem building and commercial constructs. The ecosystem doesn’t build itself; you need to be an active participant working with partners so models are natively supported across the stack—neoclouds, inference providers, hyperscalers, the Dells of the world. We have a whole open-source team that builds open-source software around the model—tools coupled with model releases that help the community understand it and build safety guardrails, agents and so forth.

Misha Laskin

商业上，土地、电力、机房外壳是部署算力的真实瓶颈——一个没通电的机柜无法变现。这类合作能拿到电力、土地、机房，甚至把基础设施搭到裸金属。韩国这个是 250 兆瓦的数据中心。之后呢？如今算力市场主要是包销（offtake），卖给超大规模云；但它正在迅速转向推理——卖推理给其他企业或自家企业。到那一步你就需要开放模型，因为你拿不到闭源模型来跑在自己的设备上，还需要一个帮你搭整套栈的伙伴，这就是我们在裸金属之上的软件层。

Commercially, land, power and shell are real bottlenecks—a rack that isn’t powered can’t be monetized. Partnerships like these get you power, land, shell and infrastructure to bare metal; in Korea it’s a 250-megawatt data center. Then what? Today compute is largely offtake—sell bare metal to a hyperscaler. That market is moving quickly into inference: selling inference to other enterprises or your own. At that point you need an open model, because you can’t get a closed one to run on your own stuff, and a partner to set up the stack. That’s where we come in, at the software layer above bare metal.

Ioannis Antonoglou

三大云——GCP、AWS、Azure——最初都是为了解决内部问题：AWS 要支撑亚马逊的业务，GCP 源于 Google 服务数十亿用户的搜索。内部需求催生了「吃自己的狗粮」式的技术，后来发现可以对外卖，于是有了互联网与 SaaS 时代。我认为我们会进入类似的时代：只有真正处在智能前沿、亲手训前沿模型，才能以可扩展的方式设计出对的系统。

The three main hyperscalers—GCP, AWS, Azure—started to solve internal problems: AWS supported the Amazon business, GCP came from serving Google search to billions. That internal demand led to dogfooded technology that could be sold more widely, giving rise to the internet and SaaS era. We’ll enter a similar era where you need the insight of building a frontier model, at the frontier of intelligence, to design a scalable system that’s the right solution.

Misha Laskin

也就是说，训练这些模型时，我们给自己的训练跑服务了数万亿 token、数十亿个智能体，已经构建了管理跨 GPU 集群的大规模集群、在其上做推理、为智能体做大规模沙箱的全部技术。闭源实验室把这些抽象掉了；而对想拥有自己系统的企业，我们把这些经过规模验证的软件外化给他们，通过合作让他们拥有和闭源前沿模型同等的能力。

To Giannis’s point, by training these models we serve trillions of tokens and billions of agents into our own runs, and we’ve built the technology for managing very large clusters across many GPU fleets, serving inference on them and sandboxing agents at scale. In a closed lab that’s abstracted away. For an enterprise that wants to own its stuff, we externalize this battle-tested software and, through a partnership, give them the same capabilities as a frontier closed model.

Alex Heath

这对闭源前沿实验室意味着什么？它们需要应对吗？

What are the implications for the frontier closed labs? Do they need to react?

Misha Laskin

这是一个兼容并存的世界。我们一直低估 AI 市场的收入深度，因为它是「按需供应的人工劳动力」，而很多国家各类劳动力都短缺。按「智能密度 × GPU × 信任」的式子，闭源实验室三样都有，会过得很好。但大势是多极化：与其两三家公司垄断，不如有十家、二十家智能提供方，开放模型让模型构建者之外的公司也能在其上繁荣。

It’s a “this and that” world. We keep underestimating the depth of revenue potential in AI because it’s artificial labor on tap, and there’s a shortage of all sorts of labor in this and other countries. By the equation of intelligence density times GPUs times trust, the closed labs will do just fine—they have all three. But the momentum is toward multipolarity: much better to have 10 or 20 intelligence providers than two or three, and open models let companies building on top thrive too.

Alex Heath

最后预告一下：你们想必在训更大的模型。接下来想推什么，节奏和规模如何？

Let’s end by teasing what’s coming. I imagine you’re training a large model. What will you push on, and what cadence and sizing can we expect?

Ioannis Antonoglou

公司的使命和愿景是弥合开放与闭源前沿之间的差距，让世界上只有一个前沿，而 Reflection 站在前沿上。为此要同时扩大算力、模型规模和 RL 规模——2027 年会有显著更大的模型。RL 可以分波次进行，就像其他模型系列有 5.1、5.2、5.3 这样的连续版本；我们也会先有一个更大的基座模型，随后多次发布。下一代模型已经在做了，2027 年会和世界分享更多。

The mission and vision of the company is to close the gap between the open and closed frontiers so there is one frontier, with Reflection at it. That means scaling compute, model size and RL—we’ll have significantly bigger models coming out in 2027. RL can happen in waves, which is why other model families ship 5.1, 5.2, 5.3 and so on; we’ll have a bigger base model and then multiple releases after it. We’ve already started building our next generation and will share more with the world in 2027.

Misha Laskin

Giannis 在 AlphaGo 和 AlphaZero 上的工作最打动我、也是我从物理转向 AI 的原因是：那是一个永不停止学习的算法。击败李世石时它已经超越人类，再加算力还能更强，只是在围棋上再强已没什么经济价值。创业时我们就认为通用编码与智能体模型也会迎来这一刻，而团队真正攻克的是「RL 会一直扩展」。我们相信这是公开记录中最大的开源 RL 训练：约 1 万张 GB300 连续跑了数周，它从未停止学习，完全受算力瓶颈约束——加算力，它就继续学。

What inspired me about Giannis’s work on AlphaGo and AlphaZero—and why I switched from physics to AI—is that an algorithm was developed that never stopped learning. By the time AlphaGo beat Lee Sedol it was superhuman, and you could add compute to make it even better; it just didn’t make economic sense at Go. When we started the company we thought this moment would come for general coding and agentic models, and one thing the team really cracked is that RL just keeps scaling. We believe this is the largest openly documented RL run in open source—around 10,000 GB300s running for weeks—and it never stopped learning. It’s completely compute-bottlenecked: put in more compute and it keeps learning.

Misha Laskin

我认为整个领域都已经搞明白，闭源实验室稍早一些：规模化的不只是预训练，还有强化学习。现在你可以让模型在任何事情上变得超级智能，剩下的是经济问题——投多少钱、数据怎么来。一些根本性的科学突破已经完成，接下来是让它们高效得多。这对我们俩是一个回到初心的时刻。Giannis 2012 年就加入了 DeepMind，14 年后的今天，我很难把这些系统看作 AGI 以外的东西，或者说它们正走在通向 AGI 的平滑路径上。问题变成如何把它落地、让好处分布到全世界，并尽可能安全——我们相信多极化、透明和开发者生态是关键。

The field has figured out—the closed labs a bit earlier—that you scale not just pre-training but reinforcement learning. Now you can make the model superintelligent at almost anything; it’s an economic question of how much money to put in and how to get the data. Some fundamental scientific breakthroughs have been made, and now it’s about making them far more efficient. It’s a full-circle moment for both of us. Giannis joined DeepMind in 2012; fourteen years later it’s hard for me to dismiss these systems as anything other than AGI, or on a clear smooth path to it. The question now is how to operationalize it and distribute the benefits throughout the world as safely as possible—we believe multipolarity, transparency and an ecosystem of builders are important.

Alex Heath

期待接下来几个月看到你们的更多成果。谢谢两位。

I look forward to seeing what else you have in the coming months. Thank you both for your time.