Misha Laskin 谈 Beam：开放模型、RL 算力与「拥有智能」

Misha Laskin

当你移除网络攻击能力时，也就移除了网络防御能力。今天的现实是：封闭实验室里只有几百位安全研究员真正理解这些系统怎么运作；无论他们出发点多好，都不可能覆盖意外后果的长尾。一个非常强大的闭源模型跑去入侵了另一家公司，而那家公司唯一的补救办法，是用开放模型来保护自己。这就是我们所处世界的经验证据。

这就是林纳斯定律：只要眼睛够多，所有漏洞都会变浅。我相信，眼睛够多的话，大多数安全和安全性漏洞同样会变浅。

When you remove cyber offensive capabilities, you also remove cyber defensive capabilities. The state of the world today is that we have a few hundred safety researchers within closed labs that understand how these things work, and despite their best intentions, it’s impossible to cover the long tail of unintended consequences that these systems might have. A very powerful closed model went and hacked into another company, and the only way that company could remediate itself was by using open models to protect itself. That’s the empirical evidence of the world that we’re in.

Linus’s law, with enough eyeballs, all bugs become shallow. I have the belief that with enough eyeballs, most security and safety vulnerabilities become shallow as well.

Host

今天我们请到 Reflection AI 联合创始人兼 CEO Misha Laskin。Reflection 提供开放权重模型，为智能的未来提供动力。Misha 之前是 Google DeepMind 的研究员，拥有物理学博士学位。你们一直在做开放权重、开源模型方面很有意思的工作，尤其带有美国视角。能说说你们这段时间在做什么、造什么吗？

Today we’re joined by Misha Laskin, the co-founder and CEO of Reflection AI. Reflection provides open weight models to power the future of intelligence. Misha was previously a researcher at Google DeepMind and received his PhD in physics. You guys have been pioneering really interesting work in open weights and open source models, particularly with a US bent. Can you talk a bit about what you’ve been up to and what you’ve been building?

Misha Laskin

过去 12 个月，我们把公司使命定为：打造前沿的开放智能，并让它广泛可得。实际上这是一场冲刺——建起一个能做这件事的实验室。有句话说你是在飞行中组装飞机，我们就是这样。一年前我们大约 30 人，但要造这种东西需要一百多人乃至几百名研究员和工程师，像真正的火箭项目。现在我们大约 300 人，预训练、中期训练、强化学习各个团队都搭起来了，训练了第一批模型，并刚刚发布了名为 Beam 的模型，这是 Reflection 的第一个开放模型。

For the last 12 months, we set the mission of the company to build frontier open intelligence and make it widely accessible. And effectively it’s been a sprint to set up a lab that is capable of doing such things. It kind of feels like you’re assembling a plane as you’re flying it, and that’s been the case. We were maybe around 30 people about a year ago, but it takes on the order of a hundred-plus, a couple hundred researchers and engineers, to build one of these things, like a true rocket project. We’re now at around 300 people and have assembled all the teams on pre-training, mid-training and reinforcement learning, trained our first models, and just released a model called Beam, which is Reflection’s first open model.

Host

人们谈到扩大规模，一是算力——能不能搞到足够、连贯可用的训练集群；二是人才，研究员的薪酬大战；三是数据规模。这几样里哪个最受限、最难？

When people talk about scale, one issue is compute, procuring enough for a cohesive training cluster. Second is talent, and there have been huge talent wars in terms of how much people pay for researchers. Third is the scale of data. Which of those has been the most limiting or most challenging?

Misha Laskin

有意思的是，公司全员会上也有人问我同样的问题：过去一年最难的是什么？现实是，每一件事都极其困难。也有人问，让你们的模型比别人更好的那一件事是什么？答案还是「所有事」。你得把 30 件事都做对：要招到人才、留住人才，要有强大的使命和文化让大家划向同一个方向，要拿到数据，要拿到算力，还要建好基础设施让算力真正可用。

在大型实验室里，这些都有多年积累，给你一个稳定的工作面。我之前在 DeepMind，作为研究员我把很多工具视为理所当然，因为它们就是能用；现在这些都得自己造。所以每件事都难，但能做成也非常有成就感。

What’s been interesting is that I got asked this question at an all-hands at the company: what’s been the hardest thing in the last year? And the reality is that everything is extremely hard. People might also ask what the thing is that made your models work better than others, and again the answer is everything. You actually have to get 30 things right. You have to get the talent, you have to retain the talent, you have to have a strong mission and culture that keeps the talent rowing in the same boat. You have to get the data. You have to get the compute. And you have to build the infrastructure to ensure that the compute is actually usable.

When you’re at an existing big lab, there have been years of build-out that enabled them to have this kind of stable surface. I was at DeepMind before this, and a lot of the tools that I took for granted as a researcher, because they just worked, you have to build. So everything has been hard, but also very rewarding that you can do it.

Host

你说大约一年前你们真正下了决心做这件事。能谈谈野心和承诺发生了什么转变、是什么驱动的吗？

You said that about a year ago you really committed to do this. Can you talk about the shift in ambition and commitment and what drove that?

Misha Laskin

公司创立大约两年半前，我和联合创始人 Ioannis 在做最早的一批 Gemini 模型，在强化学习团队，Ioannis 是负责人。我们刚发布了 Gemini 1 和 1.5。那个早期时代和最初的 ChatGPT 体验类似，主要是聊天，基本没有编码，也没有智能体；95% 的算力花在预训练上，模型主要是预训练模型。我们是强化学习研究者——Ioannis 是 DeepMind 的创始工程师之一，是包括 AlphaGo 在内所有重大 RL 项目的关键贡献者。

我们的赌注是：既然看到 RL 对对齐聊天模型有效，那把它用到数学、编码这类领域，就能让系统变成智能体。我们也认为可以作为独立实验室、用更高的资本效率做到，因为当时开始出现开放模型——Mistral 的一个小模型，Llama 2 刚发布，我们也看到 Llama 3 在路上。我们以为会有人造出很好的开放基座，我们在上面做研究、扩大强化学习规模。

结果有两点：一是强化学习比我们预想的更快见效，既在我们自己的实验里，也在业界——o1 就是那个时候出来的。我们原本以为会更久，因为从 2012 年的 ImageNet 到能大规模工作的 RL 系统用了大约四年，但每一年都比我前一年的预测更快。二是大约一年后，所有好的开放模型都来自中国，没有真正好的西方开放模型，而我们需要一个基座。

When we started the company about two and a half years ago, my co-founder Ioannis and I had been working on the first series of Gemini models. We were on the reinforcement learning team, and Ioannis was leading it. We had just shipped Gemini 1 and 1.5. That early era was similar to the first ChatGPT experience, which was primarily a chat experience. There was no coding really, nothing agentic. 95% of the compute was spent on pre-training, so they were primarily pre-trained models. We are reinforcement learning researchers. Ioannis was one of the founding engineers at DeepMind and a key contributor to all the big RL projects that came out of that lab, including AlphaGo.

Our bet was that we saw RL for aligning chat models work, so if you take that and apply it to domains like mathematics and coding, you could make these systems agentic. And we thought you could do this as an independent lab, much more capital efficiently, because we started seeing open models materialize. There was a small model from Mistral at the time, Llama 2 had just released, and we saw Llama 3 coming. We thought there’d be a great open model base that someone would build that we could build on to do our research and scale reinforcement learning.

Two things happened. One, reinforcement learning started working faster than we thought, both within our experiments and within industry; o1 came out at that time. We thought it would take longer, because the journey from ImageNet in 2012 to RL systems that worked at scale was about four years, but every year moves faster than my prediction the previous year. And two, about a year into the company we got to the point where all the good open models were coming from China. There were no really good Western open models, and we needed one as a base for the technology we were building.

Misha Laskin

出于多种原因，我们决定当时最有影响力的事，就是把 RL 的赌注和端到端自己造模型结合起来。有企业层面的原因、地缘政治的原因，也有研究上的原因：事实证明，要让强化学习在规模上很好地工作，你确实需要自己做预训练。这两件事耦合得很紧，你得两件都做。

For a number of reasons, we decided the most impactful thing we could do, given the state of play, was to combine the RL bet with building models end to end. There were enterprise reasons, geopolitical reasons, and also a research reason: it turned out you actually do need to pre-train your model in order to make reinforcement learning work very well at scale. The things are just so tightly coupled that you do need to do both.

Host

传统观点认为前沿规模的预训练贵得离谱，而算力明显在向后训练转移。你们一开始怎么看所需资源，或者有没有设计原则让这件事显得可行？

The conventional wisdom has been that pre-training at the scale of the frontier is impossibly expensive, and clearly there’s been a huge shift of compute toward post-training. How did you think about the resources needed, or did you have design principles that made this feel tenable at the beginning?

Misha Laskin

做开放模型实验室的好处是可以公开谈这些事。要做出有意义的东西需要很多资源，但追赶阶段不需要和前沿一样多；一旦你真到了前沿，就需要和其他前沿实验室同样多的资源。可以把研究看成「探索新想法」与「执行已知想法」：好人才的价值之一是帮你少走探索的路，专注执行有效的东西。所以如果你是在追赶前沿，可以资本效率高得多。

要多少钱取决于前沿一直在移动，每年都更多。大约 18 个月前是数亿美元量级；现在，或者说半年前，大概是个位数十亿美元；到明年可能是百亿美元量级。可以这样想：每一代模型算力大约有 4 倍的乘数。几年前前沿芯片是 H100，大概十万张 H100；然后训练了 Astra，是在十万张 Blackwell 上的一次大训练，大约是 H100 的 4 倍；下一代大概是十万张 Vera Rubin。这就是从数亿到数十亿再到百亿美元的量级。

The nice thing about being an open model lab is that you can talk about things openly. You do need a lot of resources to build something meaningful, but you don’t need the same amount while catching up. Once you’re really at the frontier, you need the same amount of resources as any other frontier lab. You can think about research as exploration of new ideas versus execution of known ideas. Part of why getting great talent matters is that you cut down on exploration and focus on execution of the things that work. So if you’re catching up to the frontier, you can do it a lot more capital efficiently.

How much it takes depends, because the frontier keeps moving, so each year it’s more capital. About 18 months ago it would have been on the order of hundreds of millions of dollars. Now, or let’s say even six months ago, it’s probably single-digit billions. Going into next year, I think on the order of tens of billions. You can think about it as for every generation of model there’s a 4x multiplier in compute. A frontier chip a couple of years ago was an H100, and maybe it was a hundred thousand H100s. Then Astra was trained, and that was a big run on a hundred thousand Blackwells, which is roughly a 4x multiplier on an H100. Now the next generation is going to be around a hundred thousand Vera Rubins, I believe. That gives you the scale and moves from hundreds of millions to billions to tens of billions.

Host

相对这些公司的收入规模，甚至潜在收入规模，你在投入上最终会触到上限。我们是不是正在逼近模型规模或训练集群规模的渐近线？

Relative to the revenue scale of these companies, even potential revenue scale, you start to tap out in terms of the ability to invest against what you’re actually going to produce from an economic value perspective. Are we heading into that asymptote in terms of model size or at least training cluster size?

Misha Laskin

肯定有。资本开支只能到这个程度，前沿实验室现在已经在投入数千亿美元，明年或后年要投万亿美元不太可能。但有几件事在发生。其一，算力的使用正变得高效得多，因为模型本身在成为帮助自己构建的工具。纯人类研究员做事时，每年大概有 7 倍的提升，这是可以非常具体地追踪的：预训练里你有一个预训练损失，改动后用更少的算力达到同样的损失，这就是 7 倍的含义。

预训练可能已经比较成熟，但强化学习里还有很大空间，可能在 30 倍的效率量级，取决于你的模型有多擅长改进自己。具体指标不同，但我认为速度大概是研究员单干的四倍以上，而且可能还会加速。所以每个训练 FLOP 能榨出的智能在增加；同时每个 FLOP 能产生的收入也在增加，因为智能被压得更密。

Absolutely. There’s only so much capex you can do, and frontier labs are now putting hundreds of billions of dollars into it. Will they be putting trillions of dollars in the next year or two? It’s probably unlikely. But a couple of things are happening. One is that the use of that compute is becoming a lot more efficient, because the models are themselves becoming tools that help build themselves. When it was just human researchers doing the work, there’s probably a 7x improvement each year, and you can track these things very technically. In pre-training you have your pre-training loss, you make some changes, and you get to the same point with less compute. So when you say 7x, it’s a very concrete thing.

Pre-training is probably a bit more hardened, but there’s a lot of headroom in reinforcement learning. We’re probably at a point where it’s 30x efficiency, depending on how good your model is at actually improving itself. It depends on what metric it is, but I do think the speed is probably four or more times faster than researchers doing it alone, and that’ll probably accelerate. So the amount of intelligence you can extract per training FLOP is increasing, and the amount of revenue that each FLOP can generate also increases, the more you pack intelligence in.

Misha Laskin

训练 Beam——总参数 5000 亿、激活 230 亿——预训练用了 6,000 张 GB300，大概跑了几周，不过有了基础设施效率提升，现在大约 12 天甚至更短就能做到。强化学习用了一万多张 GB300 跑四周，所以花在强化学习上的 FLOPs 其实更多。强化学习任务更复杂，因为你既要大规模推理（带各种智能体沙箱），又要训练。

我进入 AI 的原因，是看到联合创始人做 AlphaGo，当时我是物理学家。AlphaGo 里有一张图，曲线一直在涨、没有停过，他们只是在某个时刻把它停掉了，因为已经击败了世界冠军，再改进还有什么意义？我当时想，如果能把这套配方用到有经济价值的事情上，那就变成一个经济问题：我愿意投多少钱来让系统持续变好。我们这个领域现在就到了这一步。

我对这个开源项目和之后的技术报告很兴奋，因为我们会描述这些东西是怎么造的。我们的强化学习系统确实一直没停止学习——看我们的曲线，它们就是一直往上走，纯粹取决于算力。这意味着问题从「你能不能训练这些东西」，变成「去哪里拿数据、什么东西有经济价值」，本质上是一个投入多少钱换多少提升的决策。两年前不是这样的。

To train Beam, which is a 500 billion parameter model, 23B active, it was 6,000 GB300s. We ran it, I think, for a few weeks, but now with infrastructure efficiencies we can do it in about 12 days, maybe less. Reinforcement learning was a little over 10,000 GB300s for four weeks, so actually there were more FLOPs spent on reinforcement learning. Reinforcement learning jobs are more complex in the sense that you do both a lot of inference at scale, with all sorts of sandboxes for agents, and training.

The reason I got into AI was I saw my co-founder’s work on AlphaGo, and I was a physicist at the time. There was a plot in AlphaGo where it just never stopped improving. It just kept improving, and they cut it off at some point, because what’s the point of improving it further? You already beat the world champion. I thought, well, if you figure out how to apply that recipe to stuff that is economically valuable, then it becomes an economic question of how much money I want to put in to keep improving the system. And we, as a field, are there.

One of the things I’m really excited about in this open source project and the tech report is that we’ll describe how these things are built. Indeed, this reinforcement learning system never stopped learning. If you look at our plots, they just keep going up, and it’s just a matter of compute in terms of scaling it further. We as a field are there where you have these RL systems that are very general and don’t really stop improving. That means you move from the question of whether you can train these things to where you can get the data and what is economically valuable, and it’s an economics decision of how much money I put into X to get some kind of improvement out of it, which is incredible. This was not true two years ago.

Misha Laskin

这对一些科学家来说可能有点失落：它仍是科学，但感觉更像一门工程学科——就像造火箭，科学上有内容，但大多数人把它看作工程。今天的扩展轴是定好的：预训练、合成数据、强化学习；当然也可能存在某个我们不知道的蓝海，有些新实验室在探索。架构、数据、算法都有改进，但不像五年前那样是「狂野西部」。

DeepMind 和 Google Brain，乃至早期 OpenAI，都有一种美感：赌注非常多样，大家做各种疯狂有趣的事。现在研究收窄了，部分原因是「硬件彩票」：一旦某种东西开始见效，硬件就会围绕它协同优化，所以哪怕你有个好点子，如果不适合硬件，也没意义。我没有看到预训练的收益出现渐近线，也没看到算力效率增益放缓；强化学习里还有很多空间，不过更像是工程发现，而不是对新科学的突破性理解。

This is almost to some dissatisfaction to some scientists: it is a science, but it feels more like an engineering discipline, like the way that building rockets is a science, but most people think about it as an engineering discipline with some scientific work in it. The axes of scaling today are set: pre-training, synthetic data, reinforcement learning. Now there might be something totally blue ocean that we don’t know, and there are some new labs exploring it. There are architectural improvements, data improvements, algorithmic improvements, but it doesn’t feel like the Wild West the way I felt five years ago.

Something was kind of beautiful about both DeepMind and Google Brain, and even early OpenAI, in how diverse the bets were. People were doing crazy fun stuff, and now it’s really narrowed in. Part of it is a bit of a hardware lottery: once something starts working, the hardware also starts co-optimizing against it, so even if you have a great new idea, if it’s not a good fit for the hardware, it doesn’t make sense. There hasn’t been an asymptote in how much juice you can get out of better pre-training; we’re not seeing a slowdown in compute efficiency gains. There’s a lot of headroom in reinforcement learning, but it feels more like an engineering discovery than a groundbreaking understanding of a new science.

Host

Anthropic 很早就押注代码，OpenAI 也这样做过，后来才扩展到消费者等领域。你们在经济价值创造上最聚焦哪些应用领域？是代码和智能体工作流吗？

Anthropic placed a pretty early bet on code, and OpenAI did too, and then they diversified into consumer and other areas. Are there specific application areas that you are most focused on from the perspective of economic value creation? Is it code and agentic workflows?

Misha Laskin

代码和智能体能力为智能打下基础。智能是「锯齿状」的：它泛化得令人惊讶——我们的模型很快就达到了其他实验室不久前才发现的能力水平，说明确实有某种拼接式的泛化。但同时也有锯齿：一旦你有了对应任务的正确数据，它适应得很快；而不同基准之间，比如不同版本的 Terminal Bench，你看不到干净的跨 harness 泛化。不过只要有了通用的智能体能力，再给新 harness 的数据，它就能很快适应。

开放模型的力量之一是企业和客户可以拿去为自己的工作负载定制，得到帕累托最优——以最低成本取得最高性能。这是一个经验问题：去找客户，看什么对他们有价值。通常如果有经济价值，你就能生成数据——不只是在客户数据上训练，更是为他们的任务建立评测；评测搭得好，就能生成逼近它的合成数据，得到不错的泛化。具体来说，金融里的 KYC 和合规流程很有价值，网络安全尤其是网络防御很有价值，还有法律等各种企业内部的智能体垂直领域，模式都很类似。

I think the code and agentic stuff sets a foundation for the intelligence. It’s jagged in the sense that it’s generalized in a surprising way. I was surprised by how quickly our model got to a level of capability that other labs were discovering not so long ago, so there is some kind of stitching generalization happening. At the same time it is somewhat jagged: it quickly adapts to data once you have the right data for the tasks you want to do, but even between the various versions of Terminal Bench you don’t see clean generalization between different harnesses. Once you have a general agentic capability and you get the data for a new harness, it actually starts adapting to it pretty quickly.

Part of what makes open models powerful is that enterprises and customers can take them and customize them for their own stuff, and get something that is Pareto optimal for their workloads: the lowest cost for the highest performance. It’s an empirical question: you go to customers and see what is valuable to them. Usually when it’s economically valuable you can generate data, because it’s not only that you’re training on the customer’s data, it’s that you’re setting up evaluations. If you can set up a good evaluation for their tasks, you can generate synthetic data that approximates it and get good generalization. Concretely, KYC and compliance flows in finance are very valuable, cyber security and cyber defense in particular are quite valuable, and so is legal. There are all sorts of agentic verticals within the enterprise that have a pretty similar pattern.

Host

Beam 有一点很突出，就是推理效率。能多讲讲吗？

One thing that stands out on Beam is the reasoning efficiency of the model. Could you tell us a little bit more about that?

Misha Laskin

Beam 是为编码和智能体任务训练的，这是它擅长的地方。模型构建里重要的不只是能力，还有智能体完成任务的速度：这意味着更短的工作负载时间、更低的客户成本——你要是用过 AI 聊天，问个问题等 10 分钟，肯定希望 1 分钟就搞定。Beam 的效率比同能力档的模型高 3 到 4 倍；相对一些更大的模型，效率增益能达到约 10 倍。

它这么高效，是因为我们既重视强化推理的强预训练基座，又用我们认为是开源领域有史以来最大规模的强化学习去放大它——我还没见过有公开记录的「一万张 GB300 跑四周」的训练。强化学习的设置思路，是在最短时间内榨出最大能力，这和 AlphaGo 一样：最早的 AlphaGo 智能体解题还很「漫游」，等到练到李世石那个水平，它在搜索上就非常聪明了。RL 跑得越多，能力越高，系统解题也越快。

Beam is a model trained for coding and agentic tasks, and that’s where it excels. An important thing in model building is not just the capability, but how quickly an agent achieves a thing. It translates to faster workload times and cheaper costs for customers. If you’ve ever sat around asking your favorite AI chat something and it took 10 minutes, it’s much better if it does it in one. Beam tends to be three to four times more efficient than models of the same capability class, and versus larger models out there the efficiency gains end up being something like 10x.

The reason it’s so efficient is because we prioritized both a strong pre-training base for reasoning, and then amplified it with reinforcement learning at what we believe to be the largest scale that’s ever been done in open source. I’ve not seen a 10,000 GB300, four-week run documented yet. The way you set up reinforcement learning is that you want to extract the maximum amount of capability in the least amount of time. It’s the same thing that happened in previous systems like AlphaGo: the first AlphaGo agents were pretty meandering in the way they were solving the problem, and by the time you got to the Lee Sedol level it was very smart in its search. The more reinforcement learning you run, the higher the capability and the faster these systems solve it.

Host

这很务实。我们本来就是 RL 的信徒，一直相信通往 AGI 的路是在强大的基座上做强化学习。别忘了最早的 AlphaGo 先用人类对局做模仿学习，然后才做强化学习，所以两件事需要配合。

That seems very pragmatic. We are kind of RL believers, and we always believed that the path to AGI was through reinforcement learning with a really strong base. We forget that the first AlphaGo systems were trained on expert amateur human games, so they did imitation learning first and then reinforcement learning. You kind of need those things working together.

Host

从变现角度，路径是什么？开源和开放权重模型有不同的商业化方式，Mistral 等是较早的。你们有具体方向吗？

From a monetization perspective, what is the path? There are different ways to commercialize open source and open weight models, and Mistral and others were early in terms of different approaches. Is there a specific direction you’ve been heading in that you can share?

Misha Laskin

我认为开放和闭源模型的商业化在某种意义上很相似：你要做的都是最大化推理量。区别在于租用与拥有。买一个 token，你租的是整套栈中的一小块：harness（如果有智能体 harness）、模型、推理软件、集群管理软件、跑它的 GPU，这些都摊销进了一个 token 里。我常把它比作：我们起初都租公寓，长大后因为各种原因想拥有一套房子，哪怕有点麻烦。

当你在财务上足够成熟、有了家庭，就想拥有。AI 作为商业市场已经成熟到企业在租用智能上花很多钱，并想因为控制权等原因开始拥有它。所以回到变现：开放模型是宽松许可的，但客户要用起来，需要围绕它的一切——你从闭源模型那里视为理所当然的集群管理软件、推理软件、harness 等。我们为大型企业和主权国家提供让开放模型部署成功所需的全部工具；而且大多数企业需要手把手帮助，所以服务是进入、解锁真正有价值的用例、拉动大量推理需求。从这个意义上，服务和开放模型是推理业务的「需求驱动器」。

I think that commercialization of open and closed models is in some sense pretty similar: what you’re really trying to do is maximize inference. The difference is that it’s basically rental versus ownership inference. When you’re buying a token, you’re renting a piece of a whole stack: the harness, if there’s an agentic harness in there, the model, the inference software, the cluster management software, the GPUs that it’s running on, and all of that is amortized into a token. I liken it to how we typically start off renting our apartments, but as you grow up you want to own a house for various reasons, even if it’s a bit of a headache.

When you’re financially mature enough and you have a family, you want to own it. AI has matured as a commercial market to the point where enterprises are spending a lot of money renting their intelligence and want to start owning it, for various reasons around control. Going back to monetization, the open model is permissive, but for a customer to make use of it they need all the other stuff that went around it: the cluster management software, the inference software, the harness, the things you take for granted from a closed model. We provide large enterprises and sovereigns all the tools they need to make open model deployment successful. Most enterprises need some handholding, so we think about services as how we go in and unlock really valuable use cases that drive a lot of inference demand. In a sense, services and the open model are a demand driver for the inference business.

Host

一两年的时间尺度上，开放与闭源的 token 构成会怎样？

Over a one or two year time horizon, what will the mix of tokens be, open versus closed?

Misha Laskin

你现在已经能看到轨迹：大概半年前，去任何网关，比如 OpenRouter 或 Vercel，还是闭源占多数、开放占少数；现在几乎精确地从「闭源 70、开放 30」翻转成了「开放 70、闭源 30」。我认为这只会加速。我怀疑世界会和操作系统类似：全世界 95% 以上的服务器和计算机跑的是开源操作系统 Linux，但这不妨碍闭源的东西非常有价值——微软和苹果都是极其有价值的公司。所以市场很大，我认为大多数 token 需求会流向开放模型。

我们也会看到很多经济价值流向开放：即便模型完全宽松或带有许可，运行它所需的其他一切依然昂贵。算力仍然昂贵，而且它比 CPU 是更贵的计算底座，所以需要围绕它建一种不同的云。我预期大多数 token 流向开源，同时仍会有极有价值的闭源模型公司，以及围绕开放模型的有价值的生态公司和分销商。

I think you’re starting to see the trajectory now: maybe six months ago, when you go to any gateway like OpenRouter or Vercel, it was majority closed and minority open. It’s flipped almost exactly from 70/30 closed/open to 70/30 open/closed now. I think that’s only going to accelerate. I suspect the world is going to look not too dissimilar from operating systems, where 95% plus of servers and computers in the world run on an open source operating system like Linux. That doesn’t mean the closed stuff isn’t very valuable. Microsoft and Apple are extremely valuable companies. The market for this stuff is really big, so I think we’ll see most token demand going to open.

A big difference here is that I think we’ll also see a lot of economic value going to open, simply because even if the model is fully permissive or has a license on it, everything else you need to run it is still expensive. The compute is still expensive, it’s a more expensive compute substrate than CPUs, and so you have a different kind of cloud that needs to be built around it. So I expect the majority of tokens to be going to open source, extremely valuable closed model companies to exist, and a valuable ecosystem of open model companies and various distributors around them to exist as well.

Host

开放 token 里有多大比例是经过 RL 定制的，而不是基础模型？

Do you have a sense of what proportion of the open tokens are RL’d versus sort of base model, for a specific use case or application?

Misha Laskin

我听一些开放模型专用推理提供商说，绝大多数——90% 以上——是定制过的。但有个前提：今天专用推理的最大客户是 AI 原生公司，比如 Cursor、Cognition、Harvey 这类，产品本身就是围绕某个大型定制模型建的。但我对企业有不同的看法：我认为企业的大部分 token 消耗不会来自定制模型，而是来自定制系统——你拿了开放模型，甚至还没微调，只是在它外面定制了系统，比如智能体 harness，让它适配某个 KYC 流程；等你足够成熟，才开始微调。

AI 原生公司的路径是先用闭源，再转开放，再定制，而且走得很快，因为他们的产品天然是数据收集器。企业则大多是围绕既有产品和工具来搭建，直奔微调有很多摩擦。

What I’ve heard from some of the open model dedicated inference providers is that it’s actually the vast majority, 90% plus, customized. But there’s a caveat there, because today the biggest customers of dedicated inference are AI natives, who have a whole product that is just one big customization of an open model, like Cursor or Cognition or Harvey. I actually have a different take on what’s going to happen in enterprise. I think the majority of enterprise token consumption is going to come not from customized models, but from customized systems. You took an open model, you didn’t actually fine-tune it yet, you just customized a system around it, like your agentic harness, to make it work for some KYC flow, and once you’re sophisticated enough you might start fine-tuning.

The journey that AI natives went through, which was starting with a closed model, then moving to an open model, then customizing it, they went through very quickly because they set up native products that are data collectors. Enterprises are mostly architecting this around their existing products and existing tools, and there are a lot of frictions there to just do a straight shot to fine-tuning.

Misha Laskin

我们的观察是：企业开始考虑开放模型时，往往已经在闭源模型上跑起了相当规模的工作负载。我没见过直接冲向「拥有」的市场，你得先经历租用阶段。成本成为足够大的驱动，你才会问怎么降本、这件事是否值得自己留着。如果你每年在闭源模型上花 1 亿美元以上——这并不罕见——就会开始考虑更优的方案。

我认为还有个有趣的现象是本地部署（on-prem）的回潮，部分也因为算力短缺：如果你从最喜欢的超大规模云厂商那里买，往往没有足够的算力、很难拿到好价格。你花那么多钱，也许就去找 Dell 这样的基础设施提供商说：我只想搭裸机，然后把服务提供给我的企业。那接下来谁来帮他们做裸机和真正成功之间的那一层？我们认为我们的角色就是帮助企业建成成功的解决方案，是以解决方案为驱动的：评估最大价值的事情是什么，有些企业可能已在闭源模型上做了覆盖数亿客户的智能体，想要扩展出去，扩展系数非常高，只说「我们给你推理」是行不通的。

That has been our observation: the point at which an enterprise is considering open models is by the time they’ve ramped up some significant workloads with closed models. I have not seen a straight shot sprint to an ownership market; you have to go through the rental stage. It has to become a big enough cost driver for your enterprise, and then you ask how to cut costs and whether this is valuable enough to keep. If you’re spending, let’s say, $100 million plus a year on closed models, which is not uncommon at all, then you start thinking about whether there’s a more optimal way to do this.

I think there’s an interesting resurgence of on-prem, also because there’s a compute shortage. If you’re an enterprise consuming from your favorite hyperscaler, very often there’s just not enough compute and it’s hard to get good rates, but you’re spending so much money that you might go with an infrastructure provider like Dell and say, I just want to set up the bare metal, but I want to serve stuff into my enterprise. Then who helps them with the layer between the bare metal and actually making them successful? We view our role as enabling enterprises to build successful solutions. It is solutions driven: assessing what the biggest value things you can do are. A lot of these enterprises might have built an agent that spans hundreds of millions of customers on closed models that they really want to scale out, so the scaling factor is very high, and it’s very hard to just go in and say we’re going to give you inference. That has not worked.

Host

闭源生态竞争激烈，开放模型生态看起来也会如此。你们如何定位，竞争的重要维度是什么？

The closed model ecosystem is a competitive one. It looks increasingly like the open model ecosystem will also be. What do you think are the important dimensions of competition when you think about yourselves and how you position in the long run?

Misha Laskin

我有个很简单的公式：任何人在这个领域获胜，靠的是你产生多少持久的收入。公式是：你能提供的智能密度，乘以你拥有的算力，乘以你在组织中的信任度——也就是你解决他们问题的能力有多强。这个市场有趣的地方在于算力是稀缺的。在过去的竞争里，你可能说这几家玩家都没有差异化；但如果你真的很强又有算力，情况就不同。

整个空间在整条栈上都在竞争：有人说这一层在商品化、那一层在商品化，现实是一切都在商品化，模型的利润率会被压缩。开放模型与闭源模型之间存在张力，导致利润率压缩；应用层、推理层、裸机层也一样。很多 AI 原生公司在和闭源供应商谈判时非常强硬，正是因为开放生态存在。所以对任何一家公司来说，只要有很好的开放模型、有可得的智能密度，就应该有很多成功的公司，因为那变成了每家能聚拢多少算力的问题。而当你是出色的智能构建者、收入飙升，就能汇聚比任何人都多的算力，这就是闭源公司发生的事，也让你处于战略位置。

I have a pretty simple formula: ultimately the way anyone wins in this is by how much durable revenue you’re generating. The formula is what intelligence density you’re able to offer, times how much compute you have, times how much trust you have with organizations that they would want to work with you. Basically, how good are you at solving their problems? Something that’s interesting about this market is that compute is scarce. In previous competitive instances you might say these few players are all undifferentiated, but it turns out that if you were really good and you have the compute, that changes things.

I see the whole space being competitive across the whole stack. There are arguments that this part is getting commoditized and that part is getting commoditized. In reality everything is getting commoditized, and the model margins are going to be compressed. The open model tension with closed models does lead to margin compression, and the same for application margins, the inference layer and the bare metal. AI native companies negotiate their closed deals aggressively because the open ecosystem exists. So for any one of these companies, so long as there are great open models and great intelligence density accessible, there should be many successful companies, because it’s a matter of how much compute each one can assemble. When you’re a great intelligence builder, your revenue really shoots up and you’re able to amass more compute than anyone else, which is what’s happened with the closed companies, and that puts you in a strategic position.

Host

两年前很多开放权重模型源自西方，Mistral 在欧洲，Meta 是欧美混合。过去一两年中国开放权重模型崛起，有观点认为它们大量蒸馏了最先进的模型；闭源实验室也在推出工具阻止蒸馏。你认为未来一两年中国开放权重模型会怎样？

Two years ago a lot of the open weights or open source models were Western in origin. Mistral was developed in Europe, and Meta was a mix of Europe and the US. What’s happened over the last year or two is we’ve seen the rise of Chinese open weight models, and in some cases there’s the perspective that they’ve been distilling a lot from the state-of-the-art models. The big closed labs are trying to roll out tools to prevent as much distillation from happening. What do you think happens over the next year or two with Chinese open weight models?

Misha Laskin

首先，一个出色的开放模型生态来自中国，对世界其实是巨大的好处。西方有很多公司因此能建立更持久的业务。你可以想象一个没有这件事发生的世界，那就根本没有开放模型了。还存在「高个子罂粟综合症」：如果你是应用构建者，做出了很好的东西，它就会被闭源模型提供商吸收，因为它是一门要产生收入的生意。所以我是正面看待的。

但你不希望世界是「单极」的：不希望资源和算力集中在一两个闭源实验室——有 10 到 20 个闭源实验室还行，只有一两个就有点可怕。同样，你也不希望别人想拥有并在其上构建的智能来源只来自一个国家。理想情况是也有 10 到 20 个，但资本开支很高，所以至少两个会很好。因此很重要的是，西方有一个能与中国竞争的生态；但这并不真的是西方对中国——如果除中国之外有另一个国家是所有好开放模型的来源，你同样会希望有竞争张力。

我认为中国模型会继续很强。他们有一些优势：他们确实在工业规模上蒸馏闭源模型；能获取便宜甚至免费的数据，因为可以在受版权保护的 PDF 上训练，监管不同；而且这些公司有某种直接或间接的国家支持——它们很长时间没有收入。

I think the first thing is that the fact that a great open model ecosystem came from China is actually a massive benefit for the world. There are so many companies in the West that have been able to build more durable businesses as a result. There is a sense in which it’s a very positive thing, because you could imagine a world where such a thing didn’t happen and then there are just no open models. And there’s a kind of tall poppy syndrome: if you’re an application builder and you build something that’s great, then it gets subsumed into a closed model provider, because it’s a business that needs to generate revenue. So I view it positively.

The thing you don’t want is a world that is monopolar in any given way, where all the resources and compute are concentrating across closed labs, or a handful. If it was 10 or 20 closed labs, fine, you have a competitive ecosystem, but if it’s one or two, that’s a bit scarier. And similarly, you don’t want the source of intelligence that everyone else can build on, when they want to own it, to be coming from one country. Ideally it’d also be 10 or 20, but the capital expenditures are very high, so at least two would be great. So I think it’s extremely important that there’s an ecosystem in the West that competes with China, but it’s not really a West versus China thing. If there was one country other than China where all the great open models were coming from, you’d probably also want some competitive tension.

I think the Chinese models will continue to be great. They have a number of advantages: they certainly distill closed models at an industrial scale, that is true. They have access to data that is much cheaper and free, because you can just train on copyrighted PDFs in China, and it’s okay, just different regulation. And there is a notion in which those companies do have some state support, whether directly or indirectly, because these companies were not making any revenue for a long time.

Host

我一直觉得中国开源本质上是中国政府对美国或西方企业的补贴，如果你看实际发生的事情。

I always thought Chinese open source was basically a subsidy by the Chinese government to US enterprise, or Western enterprise, if you actually looked at what was happening.

Misha Laskin

确实如此。而现在这些公司在变成真正的生意：模型对海外是开放的，但在中国国内，它们相当于这里的闭源模型实验室。我认为它们会继续造出很好的模型。

Yeah, exactly. And now these companies are turning into businesses, and while the models are open abroad, within China they’re the equivalent of the closed model labs here. There are real businesses being built there, and I think they’ll continue building great models.

Misha Laskin

企业、公共部门和主权国家对西方开放模型有大量需求，原因有几个，其中有监管上的担忧或不确定性；而且你想要的不只是一个模型，而是一个能帮你把模型服务起来的合作伙伴。我们看到，在大多数《财富》500 强里，开放模型的占比目前其实很低，因为对中国模型有抵触，有些原因不理性，有些是理性的。显然，建设这些东西很重要，如果你能维持一个经济引擎——一是建立信任，让企业愿意与你这个模型构建者合作，二是让你能确保算力——那就有了商业引擎。

Fundamentally there is a lot of demand from enterprise, public sector and sovereigns for Western open models, for a number of reasons, including regulatory fears or uncertainty, and it’s not just that you want a model, you want a partner that will help you serve the model. We have seen that in most of the Fortune 500, the open model footprint is actually pretty low today because of resistance or aversion to Chinese models, for a number of reasons, some irrational and some not. It’s very clear that building those things is important, and if you can sustain an economic engine, one, it builds a lot of trust so that enterprises want to work with you as a model builder, and two, it enables you to secure compute. If there’s a commercial engine around this stuff, then that works.

Host

我想问的是，中国那边的模型公司继续保持模型开放的激励是什么？

I’m curious what the incentive system is for the Chinese model companies to continue to leave their models open.

Misha Laskin

我的假设是，中国有点像一个封闭市场，你可以做开放模型，仍然在这个市场里赚很多钱。有意思的是，在西方有一些优秀公司不造开放模型、但把它们服务给各种公司，我不太了解中国是否也有这种动态。所以开放模型提供商基本上就是该国的智能提供商。持续发布出色的开放模型在地缘政治上对中国很有利。作为一个国家，无论是美国、中国还是别的国家，你都希望其他国家在你的东西上构建。过去美中之间在 5G、光纤部署、「一带一路」上都有过紧张。当你向另一个国家提供便宜的东西、让它锁定在你的基础设施上，就有商业和地缘政治上的优势。

我常这样想：开放模型是它所带来的基础设施的「特洛伊木马」。光有开放模型没什么用，你得买下整个软件和基础设施生态。比如今天某个美国盟友国家可能在美国芯片上跑中国模型，但很快——我们已经看到——华为这样的公司会进来提供全栈方案，锁定该国的供应。你可以把 AI 或芯片、数据中心看作类似铁路的基础设施：别国没法自己造铁路，都得来找你，你就有很大的地缘杠杆。

这些栈现在正被端到端实时优化，因为范式已经确立——也许还有别的扩展智能的方式，但这是第一个真正规模化的。中国今天的劣势是芯片性能不如人，但它们有能源优势，而且在芯片上也在追赶。美国一直以来的玩法就是让世界在你的技术上构建：互联网、开放协议、开源，美元成为全球货币，自由贸易靠的是我们海军无处不在。在开源上美国一度处于守势——但我们也要明确，中国的竞争对手真的非常强，美国的生态才刚刚开始形成，还有追赶的空间。

My hypothesis is that China is somewhat of a captive market, where you can build open models and still make a lot of money in that market. Something that’s interesting is that in the West, there are great companies that don’t build open models but serve them into various companies. I’m not really aware of that dynamic happening in China as much. So the open model providers are kind of the intelligence providers in that country. Continuing to release great open models is very geopolitically advantageous to China, for a number of reasons. One is that as a country, and this is for America and China and any other country, you want other countries building on your stuff. There have been previous escalations or tensions between the US and China over 5G, fiber layout and the Belt and Road Initiative. Ultimately, when you provide something cheap to another country that then gets locked into your infrastructure, there are commercial and geopolitical advantages.

One way I think about it is that open models are Trojan horses for the infrastructure that they bring with them. You have an open model, but that alone is not very useful, so you buy into the entire software ecosystem and infrastructure ecosystem of a given country. Today an allied country, a US ally, might run a Chinese model on American chips, but pretty soon, and we’re seeing this already, Chinese companies like Huawei are coming in and offering a full-stack solution and locking into that country’s supply. If we think about AI, or chips and data centers, almost like railroads, as fundamental infrastructure, then no other country can actually build their own railroads, so they all have to go to you, or another country. You have a lot of geopolitical leverage.

These stacks are getting optimized end to end in real time, because in some sense the paradigm has been set. There may be other ways to scale intelligence, but this is clearly the first one that really scaled. Where China has a disadvantage today is that their chips aren’t as performant. They have an energy advantage, and they’re catching up on the chip side too. Geopolitically, the notion of the rest of the world building on your stuff is something America has done with all other meaningful technologies: the internet, open protocols, open source. The American currency is the widely established global currency, and free trade works because we have a Navy that’s everywhere. That’s been the American playbook for a long time. I think the United States has been on the back foot when it comes to open source, but we have to be clear that the competitors in China are really, really good, and an ecosystem is just starting to form now in the United States, so there is some catching up to be had.

Host

闭源实验室对开放模型最大的批评是安全和可控性。你的理念是什么？这是正当的担忧吗？

The biggest criticism from the closed labs of open models tends to be safety and controllability. What’s your philosophy on this? And is that a legitimate concern?

Misha Laskin

这当然是正当的担忧。但我认为已确立的安全世界观是从一个特定视角出发的，已经近乎教条。如果从第一性原理看 AI，不先考虑闭源还是开放，你预见了这项技术会拥有这些能力，并思考安全——你会怎么从零设计？一个类比是软件：可以把 AI 看作更高级的软件，它有很多相似之处，是一种数字公用事业。上世纪 90 年代初关于软件是否危险就有过辩论，特别是强加密协议：当时它们是封闭的，NSA 和公民自由倡导者争论该封还是该开。

最终的结论，是在封闭一方出现了一些灾难性失败之后——少数工程师设计的系统带来了他们无法预料的后果、很容易被黑——强加密协议走向开放，而这也催生了整个网络安全领域。所以世界的默认状态其实是：开放即安全。如果我们只是把 AI 看作技术的延续，这就是我们会带入 AI 的默认状态。模型和软件有很多相似之处，只是被放大了：它是黑箱，比软件更难理解，漏洞和意外后果的长尾更大。林纳斯定律说，眼睛够多，漏洞都会变浅；我相信安全漏洞也是如此。

今天的现实是，只有几百位安全研究员在封闭实验室里理解这些东西。尽管出发点很好，也不可能覆盖长尾；这一点的症状就是，当我们看浮出水面的重大网络安全问题时，是一个非常强大的闭源模型产生了意外后果、去入侵了另一家公司，而那家公司唯一能自我补救的办法是用开放模型保护自己。这是我们所处世界的经验证据。

No, it’s definitely a legitimate concern. But I do think the safety worldview that has been established has been established from one particular perspective, with one particular point of view, that has become almost dogmatic. If we were looking from first principles at AI, before considering closed or open, and you had the foresight to know it’s going to have these capabilities and you were thinking about safety, how would you actually design it from scratch? One analogy is software: you can think about AI as a more advanced version of software. It has many similarities, actually. It’s a digital utility. There were debates, by the way, around software being dangerous, particularly in the early 1990s around strong encryption protocols. They were actually closed at the time, and there were debates between the NSA and civil liberty advocates about whether you should close it or open it.

Ultimately the decision, after some catastrophic failures on the closed side, where effectively a small handful of engineers designed certain systems that had unintended consequences that they couldn’t predict and that got easily hacked, was that strong encryption protocols became open, and that actually gave birth to the whole field of cybersecurity. So the default state of the world is actually that openness is safety. That’s the default state we’d be going into AI with if we were just thinking about a continuation of technology. Models actually have a lot of similarities to software, but exacerbated: it’s a black box, it’s harder to understand than software, and the long tail of vulnerabilities, and hence unintended consequences, is larger. Linus’s law, from the founder of Linux, is that with enough eyeballs all bugs become shallow, and I have the belief that with enough eyeballs most security and safety vulnerabilities become shallow as well.

The state of the world today is that we have a few hundred safety researchers within closed labs that understand how these things work. Despite their best intentions, it is impossible to cover the long tail of vulnerabilities or unintended consequences that these systems might have, which is why the symptom of this is that when we look at what major cyber security issues have surfaced, a very powerful closed model had unintended consequences where it went and hacked into another company, and the only way that company could remediate itself was by using open models to protect itself. That’s the empirical evidence of the world that we’re in.

Host

人们常把「安全」混为一谈：有网络攻击和黑客，有生物武器和恐怖主义（我认为短期被夸大了），还有对人类的生存威胁，每一种都是可以分开的。你说的安全是指哪些？

People really conflate notions of safety, and safety means three or four different things to different people. There’s safety in terms of cyber attacks or the use of AI for hacking, there’s safety in bioweaponry or terrorism, which I think tends to be overstated in the short run, and then there’s safety from the perspective of an existential threat to humanity. Each one of these is separable. When you’re talking about safety, which of those are you addressing?

Misha Laskin

我认为安全是一个谱系：从经验现实到理论情景。我承认 AI 里有科幻的成分——去年还纯属理论的东西，比如这些模型的网络能力，现在是现实了。所以不是要无视理论部分，但我们必须清楚：有真实的经验性问题正在发生，而且我们可以相当有信心地预测未来六个月会发生；另一端则是最大化的理论情景。

AI 的挑战在于它太贴近日常文化、太容易拟人化，所以从感知上，人们对安全含义的分布不是均匀的，而是峰值落在末日情景上。而现实更像是：峰值应该在现实，然后向理论衰减。领导者说「我们全都死掉有 10% 的概率」是没有帮助的，那没有具体依据，是编出来的数字。作为科学家，我倾向分布式思考：峰值在现实；再往后六个月，确实担心错位，即这些系统的意外后果，随着变强而继续放大。但我们必须把它们当作被工程出来的工具。

I think on the safety spectrum, there’s basically a spectrum from empirical reality to theoretical scenarios, and I will grant that AI has sci-fi bits to it. Something that was totally a theoretical thing last year is a reality now, like the cyber capabilities of these models, for example. So it’s not to discount the theoretical stuff, but we have to be clear that there’s real empirical stuff that’s happening and that we can predict with some confidence will be happening in the next six months. And then there’s the maximal theoretical stuff.

The challenge with AI is that it’s so top of mind, it has permeated everyday culture, and it’s so easy to humanize that, from a perception perspective, it’s not even a uniform distribution around what safety means; it’s peaked at the doomsday scenarios. Whereas the reality is it should peak at reality and then decay into the theoretical. It doesn’t help when leaders of companies say that on the theoretical side there’s a 10% chance that we all die. It’s not substantiated by anything specific; it’s kind of a made-up number. As a scientist I tend to have distributional thinking, peaked at the reality, and then six months ahead I’m definitely worried about misalignment, these unintended consequences of systems that continue to get more powerful. But we have to take the perspective that these are engineered tools.

Misha Laskin

末日情景是不错的饭桌话题，但说到对齐——这是安全的重要部分，即怎么让这些系统按预期行事——现实是，迄今为止对齐的科学一直「深深地无聊」且不令人满足，没有什么神奇的对齐方程式，更像打地鼠。回想 ChatGPT 之前，有些预训练模型发布出来是有毒的，当时的说法是模型有毒、修不好，因为毒性方式太多。结果发现，有足够的后训练数据你就能把这些补上，而且泛化得很好，变得相当难越狱。现在对齐问题可能也会很平凡、很无聊：你发现一堆漏洞，有数据去补上，有算法去检测——所谓算法，通常就是让语言模型去检测这件事。

The theoretical doomsday stuff is good dinner conversation, but take alignment, which is a big part of safety, how do you align these systems to behave in intended ways? The reality is that the science of alignment so far has been deeply boring and unsatisfying. There’s no magical alignment equation; it’s a whack-a-mole thing. If you remember the early models before ChatGPT, there were some models released that were pre-trained and toxic, and there was this notion that these models are toxic and can’t be fixed because they can be toxic in so many ways. Then it turned out that with enough post-training data you can just patch all those things, and it generalizes in a way that it becomes pretty hard to jailbreak. Now you’re seeing the same thing with these alignment issues: they’re probably going to be very mundane and boring. You’ve discovered a bunch of these vulnerabilities, you have data that patches them up, and you have algorithms for detecting them, and when you say algorithms, it’s typically asking a language model to detect this thing.

Host

那是不是应该有十万名研究员和计算机科学家去查看、修补这些漏洞，这样更安全？我觉得这里有个重要的哲学区分：即使它们是非常智能的工具，我们作为生态能管理意外后果，我们是否不希望个人和企业拥有这么强大的工具？这种立场隐含的前提是：我们可以拥有强大的工具，别人不行，因为我们更擅长管理。

Then shouldn’t there be a hundred thousand researchers and computer scientists looking at and patching up these bugs? Wouldn’t that be safer? I feel like there’s an important philosophical distinction: even if they are very intelligent tools, and we can manage, as an ecosystem, the unintended consequences, we don’t want individuals and businesses to have such powerful tools. That is coming from a standpoint that we can have access to powerful tools but other people can’t, because we’re effectively better at managing these tools.

Misha Laskin

公平地说，这类主张的论点是：任何一个人或公司可能造成的负面影响太大，这很可怕。但就像武器的规模——美国政府也向其他国家出售喷气机。更广泛地说，把一个训练过要安全的开放模型，投入算力和精力把它变得非常危险，并不容易。我不是说不可能。就像网络安全：非常聪明的黑客确实能造成破坏，但在整个生态里，防御往往大于进攻；这就像一个由白细胞组成的生态系统，让更多人拥有防御能力，实际上有助于抵御进攻能力。

经验现实是，很难把网络防御和进攻分开。当你移除进攻能力，也就移除了防御能力，结果是那些想帮忙防御的玩家做不到。我们在最近 OpenAI 与 Hugging Face 的事件里看到了这一点——他们回退到使用开源模型，因为现有最先进实验室的模型上有护栏，他们用不了。这类论点往往走向绝对化：这些东西太强大了，除了我们，任何人都不该拥有。但现实是，除了极少数几乎没有正面用途的例外，这类绝对主义的主张很少行得通。

To be fair to the claim, the argument is that the negative impact any single person or company could have is too large, and that’s scary. But it’s on a scale of weapons, right? The US government sells jets into other countries. More broadly, it’s not so easy to take a model, even an open model that’s been trained to be safe, and put in the compute and work and effort to make it very dangerous. I’m not saying it’s impossible. When we think about cybersecurity, very intelligent hackers can take and wreak havoc, as they have, but you have an ecosystem where the defenses tend to outweigh the offenses. It’s like an ecosystem of white blood cells, and enabling more people to have the defensive capabilities actually helps you protect against the offensive capabilities.

When we look at the empirical reality of what happened, it’s really hard to separate cyber defense from offense. When you remove cyber offensive capabilities, you also remove cyber defensive capabilities, and as a result the players who would want to help defend are incapable of doing so. We saw that in some of the recent things that happened with the OpenAI Hugging Face incident, where they reverted to using open source, because they weren’t able to use the existing state-of-the-art labs given the guardrails placed on the models. These arguments go to a place where they’re very absolute: these things are so powerful that no one should ever have access to them, except for us. And the reality is that, except for some really rare exceptions where there are actually not that many positives, such absolutist perspectives rarely work.

Host

往前看两到五年，你对 AI 世界最兴奋的正面变化是什么？我知道很难预测。

What are you most excited about if you think two to five years out in the AI world? I know it’s very hard to predict right now, but as you think ahead in terms of the curve of technology and the curve of adoption, what are some of the positive things that you view as coming?

Misha Laskin

有太多正面的事情正在到来，而且一直在到来。我个人对科学进步特别兴奋。我们搬到美国后，我对物理产生了兴趣，此后一生都在追求科学。这几年我一直拿一个提示测试语言模型：把我的博士论文给它。当然，博士论文里很多工作是找到正确的问题，但执行那些机械性的工作也要花很长时间。几年前模型只会聊天，什么也做不了；一年前它开始回答，大概是本科生作业的水平；六个月前它解出来了，达到真正的博士水平且是正确的；现在我再试，它甚至给了我一些当年没考虑到的有趣新信息。

所以现在，如果你有正确的问题，就可以把它放进聊天框，让它做所有计算并给你很有意思的结果。这意味着你的迭代速度——为什么还需要几年的博士，如果你一周就能做完一个博士？

There are so many positive things that are coming, and they have been coming. I’m personally very excited about scientific progress. When we moved to the States, I developed an interest in physics and have always had this lifelong pursuit of science since then. One of the prompts I’ve been trying with language models over the last few years is giving it my PhD thesis. A lot of the work in a PhD is getting to the point where you have the right question, but actually executing the rote work takes a really long time as well. A couple of years ago it was just chat, so it couldn’t do anything. Then a year ago it started answering things at an undergraduate level, like that would have been my answer if this was given to me as homework in undergrad. Six months ago it solved it, at real PhD level, and solved it correctly. And now when I actually tried, it even gives me some interesting new information that I hadn’t considered at the time.

So now, if you have the right question, you can just put it into a chat box and have it do all the calculation and give you something really interesting. That means your iteration speed changes: why do you need a few years for a PhD when you can do a PhD a week?

Host

这和 OpenAI 最近一两天的消息类似，就是过去几周里 OpenAI 证明的各种数学定理。

Similar to the recent news from OpenAI, I think from the last day or two, in terms of all the various mathematical theorems that were proved in the last couple of weeks by OpenAI. It’s kind of crazy what’s been happening.

Misha Laskin

对科学的影响绝对不可思议。我同样兴奋的不只是像黑板一样的理论科学，还有真实世界的科学：生命科学、材料科学、化学，AI 可以和真实实验对接，建立专有数据飞轮。这会比黑板上的慢得多，但比以前快得多。我们还在持续低估这对软件工程的改变有多大、你能造出多少东西。

It’s absolutely incredible what this is going to do for science. The other stuff I’m really excited about is not just theoretical science, where the environment is the equivalent of a chalkboard, but real-world science: life sciences, material science, chemistry. There are a lot of places where you can have an AI interfacing with a real experiment and set up proprietary data flywheels. It’s going to be much slower than the chalkboard stuff, but dramatically faster than what it was before. And we continue to underestimate how dramatic a shift this has been for software engineering and how much more you can build.

Misha Laskin

几周前我参观了 Stargate，让我印象深刻的是停车场里的车那么多，有那么多人在那里工作。这些数据中心项目创造了数以万计的高薪工作。在我看来，数据中心如今就像过去的工厂——就像 20 世纪的铁路和工厂：一家鞋厂能让一个小镇变成有活力的经济体。我在现场看到这些项目的建设，虽然人们害怕数据中心，但实地去看，它为当地创造了很多就业和税收。

当然，有些事需要处理得很得当：这些设施很吵，需要隔音，或者放在离居民区较远的地方。但就本地经济而言，它们确实创造就业、产生税收，在就业创造上就像过去的工厂。

I toured Stargate some weeks back, and the thing that struck me was the number of cars in the parking lots. There are so many people working there. These data center projects create tens of thousands of high-paying jobs. To me, a data center is now kind of what a factory used to be. You had railroads and factories in the 20th century, like a shoe factory that makes a town produce a vibrant economy. When I’m on the ground seeing the stuff that’s getting built, there is fear of data centers, but the reality is you go on the ground and it actually creates a lot of jobs and a lot of tax revenue for the local communities.

There are certain things that I think need to be done very tactfully: these things are noisy, so you need to insulate the noise or put them in places that are less close to residential. But it is certainly the case that for local economies, these things do create jobs, they generate tax revenues, and just feel like what a factory used to be in terms of job creation.

Host

你们现在如何引导科学和实验工作？假设你们在用模型自己改进训练，科学团队今天的角色是什么？

How do you direct the scientific and experimental effort going forward, and within that, assuming you are trying to use the models themselves to improve your training effort, what is the role of the science team today?

Misha Laskin

研究方向主要由我的联合创始人 Ioannis 负责，他领导研究和技术。有很多我们知道有效的东西，没来得及放进去，纯粹是时间线问题：我们是在 7 月收到的 SpaceX 集群上训练的最终模型，先是几周预训练，再是几周我们自己的合成数据，然后发布。所以有很多没来得及放进去、我们很兴奋的东西。这里有个权衡：谨慎执行已知的东西或风险较低的赌注，还是去做更有风险的赌注。你要先把已知有效的高影响力的事做尽，再开始扩大风险敞口。

风险这边，你要做小规模的扩展实验。有意思的是，有些有趣的东西在大规模之前根本不会出现，所以有一种艺术与科学：是投更多算力去找下一个点子，还是转向下一个想法。至于模型进研发回路，我觉得对研究员非常赋能：原本个人要花很久的事，现在可以很快完成。这些模型的智能是锯齿状的，你要有感觉，知道在哪里能形成紧密回路、哪里仍需要人的创造力，很快就会摸到手感。能形成回路的地方就用回路，比如超参数搜索和一些更细致的基础设施工作。

它们是很令人兴奋的工具，因为让科学家发挥直觉——就像有一个又快又热切、肯听你话的同事，前提是它对齐得正确。我看到研究员的行动速度快了很多，这也许解释了为什么每年的改进速度，原来是大约 7 倍（纯手工时），现在是它的好几倍。

The way we direct our research, a lot of this falls under my co-founder Ioannis, who leads research and technology, among other things. There are a lot of things that we know work that we weren’t able to put in. That’s just a timeline thing. We trained the final model on the SpaceX cluster that we received in July. It was a few weeks of pre-training, then a few weeks of synthetic data on our own, and shipping it. So there’s a lot of stuff that didn’t make it in that we’re really excited about. There’s a trade-off between the careful execution of known things, or bets that have a bit less risk, and going on to more risky bets. You want to exhaust all the high-impact things that you know work and then start expanding the risk profile.

On the risk profile, you have scaling experiments where you do stuff at small scale. The interesting bit is that some interesting things don’t emerge at all until you’re at large scale, so there’s a bit of an art and science of whether I’m going to put in more compute to find out the next thing or go on to the next idea. As far as model in the loop goes, I find it very empowering for researchers, because things that would have taken a long time individually can now be done fairly quickly. These models are fairly jagged in their intelligence, so you have to feel where they have a tight loop and where human creativity is still important, and you get a feel for it pretty quickly. The places where you can loop things, you do that. That includes all sorts of things around hyperparameter searches and some of the more granular things on infrastructure.

I actually think they’re very exciting tools because they enable scientists to exercise their intuition. It’s kind of like having a very fast and eager colleague that actually listens to you, if the colleague is aligned correctly. I see researchers just getting a lot more ability to move a lot faster than before, which is why maybe the rate of annual improvement was roughly, let’s say, 7x before when it was just manual work, and now it’s several times that, say maybe four to five x.

Host

有个问题是，在什么时候你不再需要研究员参与回路？如果今天需要 100 人，明年做同样的事是不是只要 20 人？

There’s this question around at what point you don’t need a researcher in the loop. I was going to ask you, if you need 100 people today, do you need 20 next year for the same effort?

Misha Laskin

做同样的事，是的。但你会不断扩大你想做的事的野心，所以我们一直是人手不足，而不是过剩。而且有一个人员与算力的比例要把握，并不是可以一直扩下去：三千名研究员不会有帮助，但几百名会非常有帮助。

For the same effort, yes. But you’re constantly expanding the ambition of what you can do, and so we find ourselves constantly understaffed rather than over. There’s basically a staff-to-compute ratio to keep in mind, so it’s not like it goes on forever. I think 3,000 researchers would not be very helpful, but a couple of hundred would be very, very helpful.

Host

我听过的论点是：每人需要分配的算力那么多，要把某些实验扩大规模，而每个领域贡献又有分布——总有少数人贡献了最多的想法——所以最终你要把每一份增量算力分给最有产出的那部分人，结果是研究员数量静态甚至缩减。我不是说它正确，只是听过这种论点。

The argument I’ve heard is that given the amount of compute you need to allocate per person, and especially if you want to scale certain experiments up over time, and then coupled to that, there’s a distribution in terms of contributions, with a handful of people who contribute the most from an idea perspective in every field, eventually it collapses into wanting to give every incremental piece of compute to the most productive subset of people, and therefore you end up with a static or shrinking number of researchers over time. That would be the argument in the extreme. I’m not saying it’s correct, I’m just saying that’s an argument I’ve heard being made.

Misha Laskin

这个论点有道理。这类项目从来不需要极其庞大的研究员数量：AlphaGo 当时大概 10 个人量级，现在像我们这样的大工程大约是一百到几百人。至于是否会塌缩回 10 人，很难说，我估计会有一个大约一百人量级的稳态。还有很多事要做，而且在应用研究上会有大量的就业创造，把这些东西真正应用到实际问题上。

The argument definitely has merit. These projects never needed an extraordinary number of people. A project like AlphaGo at the time was maybe on the order of 10 people, and for a large effort like the one we’re doing, it’s on the order of a hundred or hundreds. Whether it collapses back to 10 or not, it’s hard to say. I would estimate there’s going to be some steady state on the order of a hundred. There are just a lot of things to do, and there’ll be a lot of job creation in applied research, going and actually applying these things to real problems.

Host

人们谈工程和研究时忽略了这些角色向更广泛社会的扩散，而不只是造语言模型。到目前为止的证据显示，很多公司需要的工程师更多而非更少；而且某个水平的工程师还扩散到那些原本无法大规模获得这种人才的组织去部署这项技术。人们倾向于把经济集中到少数大科技公司，而不是看整体 GDP 如何被这项技术的扩散所改造。

I think people really lose that when they talk about engineering and research in terms of the diffusion of those roles out into a broader swath of society, and not just building language models. Even when people talk about the potential displacement of engineers, so far the evidence seems to be that many companies need more engineers versus fewer, and there’s also the diffusion of a certain level of quality of engineer out to organizations that never would have been able to access somebody of that talent level at scale. People tend to centralize the economy into a handful of big tech companies instead of saying, what is the overall GDP and how does that get transformed by the diffusion of people bringing this technology out.

Misha Laskin

我同意。研究员的数量多年来并没有剧烈变化，大项目大约一百人，我不认为需要数千人；不同产品面上也许需要一批人。有意思的是，你看为什么需要数百人，很大一部分是因为有团队为不同能力做评测、收集数据——模型有某种能力不是偶然的，而是有 5 到 10 人的小组去瞄准某一项能力，编码里就有 5 到 10 项能力。这是为模型带来通用能力。而当你把模型应用到真实世界的能力时，每一项真实世界能力基本都需要一个小组，每家企业有很多这样的能力。

所以我认为会有一类新的「前线部署工程师」的就业创造，他们更偏科学和评测导向，对这些模型和 harness 有直觉——这其实就是评测研究员的技能，和造核心模型的研究员是同一套技能，只是被部署去解决真实问题。我们会尽可能多地招这样的人。这更像是一个培训缺口：需要培养工程师擅长这件事，而不是需求不足。

I think that’s correct. The number of researchers hasn’t changed dramatically over the years; a big project is around a hundred people. I don’t think you need thousands on a big project, but on different product surfaces you need a bunch of people. What’s interesting is that when you look at why you even need hundreds of people, a lot of it is teams that do evals for different things and collect data for different things. It’s not by accident these models have capabilities; there are pods of five to ten people that go and target a particular capability, and within coding there are five to ten capabilities. That’s for bringing the general capabilities to the model. But when you want to take that model and apply it to real-world capabilities, you basically need a pod for any real-world capability, and each enterprise has many of them.

So I think the job creation for this new type of forward deployed engineer, who is a bit more scientific and evaluations oriented, with some intuition for these models and the harnesses, which is actually what the evaluation researcher does, is real. It’s the same skill set as the researcher building the core thing, deployed to work on real problems. We’ll take as many of those people as possible. It’s more of a training gap, in that you need to train engineers to become good at this. We would take as many as possible today.

Host

恭喜你，Misha。有了第一个达到前沿级别的西方模型，对整个生态是很大的一步。非常感谢。

Congratulations, Misha. Having the first Western model that is in the class of the frontier has been a big step forward for the ecosystem. Thank you so much.

Misha Laskin

是的，谢谢你们的参与。

Yeah, thanks for joining us.