Mercor 的 Brendan Foody 谈数据、评估与人类最后的工作

最近论坛一直对数据角度很感兴趣。所以，我们请来了 Merkle 的 CEO Brendan Foody 来和我们聊聊。请大家给他一轮掌声。

For the forum, we've been interested in the data angle lately. And so, we have Brendan Foody, Merkle's CEO, to come chat with us. Please give him a round of applause.

感谢邀请。很高兴来到这里。

Thank you for having me. It's a pleasure to be here.

那么，我们不如先聊聊 Merkle 的起源故事吧？

So, why don't we begin by just talking about like your origin story for Merkle?

当然可以。背景是这样的：我在湾区长大。14 岁时在圣何塞的高中遇到了我的联合创始人 Adarsh 和 Surya。我们 19 岁时创立了 Merkle，最初是在国际上招聘人员，把他们和我们的朋友匹配起来，并自动化所有流程。就像人类会审阅简历、进行面试、决定录用谁一样，我们把这些都自动化了，从印度招聘工程师并把他们和我们的朋友匹配。然后我们从大学辍学，不久后遇到了 OpenAI，当时我们看到了数据市场的巨大转变：它正从 Scale AI 和 Surge 开创的众包模式——即如何让中低技能的人为早期版本的 ChatGPT 写出语法相当正确的句子——转向智能体式数据范式，即如何找到顶尖的软件工程师、机器学习工程师、律师和医生，让他们在团队中协作，在强化学习环境中构建前沿评估。我们从那里极速扩张，成为几乎所有前沿实验室和所有顶级应用层初创公司（从 Harvey、Sierra 到 Cognition、Cursor 等许多公司）的最大数据供应商，在约 24 个月内从 100 万美元的营收运行率增长到超过 20 亿美元，现在则高得多，过去一年增长还在加速。我现在的首要关注点，也是与本次对话相关的，是如何把我们为旧金山的 AI 实验室构建的所有技术，应用到更广泛的企业和初创公司中，这些公司现在正在探索如何后训练自己的模型。我认为这是一个非常迷人的市场，因为它在今天的 token 消费中只占很小一部分，但五年后它将成为 token 消费的绝大部分。所以非常兴奋能讨论这个话题，并与大家分享见解。

Yeah, for sure. So, background is that I grew up in the Bay Area. I've met my co-founders Adarsh and Surya when we were 14 in high school in San Jose. And we started Merkle when we were 19, initially hiring people internationally, matching them with our friends, and automating all the processes of how we did that. So, similar to how a human would review resume, conduct an interview, and decide who to hire, we automated all of those, hiring engineers from India and matching them with our friends. And then, we dropped out of college and met OpenAI shortly thereafter, where we saw this enormous transition in the data market as it was moving away from the crowdsourcing paradigm that Scale AI and Surge pioneered of how do you get low and medium-skilled people writing fairly grammatically correct sentences for early versions of ChatGPT. And it was moving towards the agentic data paradigm of how do you find the top software engineers, ML engineers, lawyers, and doctors that can work collaboratively in teams to build frontier Evals in RL environments. And so, we scaled up extremely quickly from there, becoming the largest data vendor to almost all of the frontier labs and all of the top application layer startups ranging from Harvey, Sierra, to Cognition, Cursor, many others, scaling from a million dollars in revenue run rate to over 2 billion dollars in revenue run rate in about 24 months, now dramatically higher than that with accelerating growth over the past year. And my top focus now, and pertinent to this conversation, is how do we take all of this technology that we've built for the AI labs in San Francisco and apply that to the broader set of enterprises and startups that are now exploring how to post train their own models and I believe that this is such a fascinating market cuz it's a very small part of how tokens get consumed today, but it's going to be the lion's share of how tokens gets consumed in 5 years and so super excited to discuss it and share insights on it with everyone.

太棒了。你提到之前有过一次行业转变时刻，我觉得现在我们几乎又处在另一次转变的边缘。具体来说，比如递归自我改进（RSI）。所以我们有一个问题是：如果现在你可以把一个模型和它的智能体轨迹连接起来，让它自我改进，那么 RSI 对 Merkle 的业务意味着什么？

That's great. So, you mentioned that there was this previous like industry shift moment and I feel like right now we're almost at the cusp of another shift. Um specifically like recursive self-improvement or RSI. And so one question we had is like what does RSI mean for Mercari's business let's say if now you can kind of like hook up a model and it's agent traces and just have it improve itself.

嗯，我想说几点。首先，模型在不可验证的领域确实很挣扎，尤其是自学。这就像学生很难给自己的作业打分一样。比如，数学问题有计算器，是非常干净的模拟，那就容易；但模型在制作幻灯片时，无法识别自己幻灯片中的错误。因此，人类教模型如何改进的市场，受限于人类在经济中能做而模型不能做的事情的数量。当然，我们会在此过程中自动化后训练和爬山过程，但我认为我们会走向一种范式：一旦你有了某个事情的评估，强化学习如此有效，以至于你可以立即饱和那个评估。所以，如何自动化经济中一切事情的瓶颈，就是如何为一切构建评估？而评估在定义上就超出了模型所能做的前沿。

Yeah, well I would say a few things. So, I think first of all is that ultimately models really struggle in non-verifiable domains especially to learn from uh on their own. And that for the same reason a student struggles to grade their own homework. Like if you have uh math problem with a calculator is a super clean simulation, then uh it's easy, but models aren't able to identify when they build a slide deck what are the mistakes in their own slide deck. And so, the market for humans teaching models how to improve is bound by the amount of things that a human can do in the economy that a model can't do. And so of course we'll automate uh the post training and hill climbing process along the way, but I think we'll move towards a paradigm where once you have an eval for something, RL is so effective that you can saturate that eval immediately. And so, the bottleneck for how do we automate everything in the economy is how do you just build evals for everything? And evals are definitionally beyond the frontier of what models are able to do.

你公开说过 80% 的模型差异化来自数据。为什么是这样，另外 20% 来自哪里？

And you spoke in publicly about how 80% of model differentiation comes from data. Like, why is that the case and where does the other 20% come from?

有人问过我这个问题，经过进一步思考，我或多或少会说它是 100%。当然有个前提：会有算法上的转变，比如推理模型，会有巨大突破。但今天大部分情况下，大家都在用相同的算法。只要人们使用相同的算法，模型就只是输入数据的反映。所以，他们可能有不同的技巧来增强专家数据，以及如何最大化利用手头的数据。但最终，大多数都归结为数据整理。因此，这往往是实验室投入大量算力的地方，以及他们如何生成更多数据并最终差异化他们的模型。

Someone asked me this and upon further reflection, I think I think I would more or less say it's 100%. Like, I think with the caveat that there are algorithmic shifts that happen, like reasoning models, where there are like giant breakthroughs. But for the most part today, like everyone's using the same algorithms. And so long as people are using the same algorithms, the model is just a reflection of the data that goes into it. And so, there might be different tricks that they have of how to augment expert data and how to get the most out of the data that they have. But but ultimately, most of those tend tend to come down to just like the data curation. And so, that tends to be where labs invest a lot of their compute and how do they generate more data and how they ultimately differentiate their models.

关于数据这个话题，South Park Commons 充满了非常有野心的人。我感觉几乎每次论坛会议都会遇到类似的问题。比如，如果我们是一个创始人，我们只是能接触到数据，并且认为它很有价值，我们该怎么办？

On the topic of data, South Park Commons is full of very ambitious people. I feel like I kind of get a question along these lines like every forum session. Like, what do we do if we're a founder and we just have access to data and we think it's valuable?

这是个有趣的问题。我会说这很大程度上取决于数据种类。有很多方式可以处理，对吧？比如，极端情况下，如果你有一个极有价值的数据集，可以用来训练某个特定能力的前沿模型，你可以建立一个实验室，提供模型服务，从中盈利；或者如果你只有一家倒闭初创公司的数据集，你可以卖掉它。我们购买数百个公司数据集，如果你有强化学习任务，我们也会买。所以，我认为这非常取决于你拥有的数据类型。但你可以利用这些数据来弄清楚如何为你关心的用例训练模型，或者你可以找到一种方式将数据变现，我们很乐意成为买家，我知道许多其他人也会。

It's a interesting question. I would say it depends a lot on what kind of data. And there's, I mean, so many ways you could go about it, right? Like, in the extreme case, if you have a extremely valuable data set where you could train a frontier model for a specific capability, you could build a lab, serve a model, monetize that, all the way to if you just have a company data set from a startup that shut down, you could sell that. And we buy hundreds of company data sets all the way to if you have RL tasks, we would buy that as well. And so, I think it's very contextual to the type of data you have. But you could use those to figure out how to train a model for the use case you care about or you could figure out a way to monetize that data where we'd love to be a buyer of that and I I know many others would as well.

我想追问一下，比如你如何确定一个评估或数据的科学价值，或许还有相关的货币价值？

I guess follow up, like how do you determine the scientific and perhaps maybe correlated like the monetary value of say an eval or that data?

我会说有几个因素很重要。首先是现实性。即它在经济分布中的位置，以及更广泛的 AI 模型用户群体关心什么。第二，特别是对于有任务的评估，通过率或难度是多少？当然，如果一个任务每个模型都能完成，那它根本没有价值。如果一个任务没有模型能完成，那它就变得极其有价值。最后我想说，也许人们对数据价值最大的误解——这与关于轨迹的 RSI 问题有点关联——是模型无法仅仅通过阅读你的 Slack 组织架构和所有公司数据来学习。原因就像我们也无法从这些东西中学习一样。

I would say that there's a couple of things that matter. The first is the realism. Like how much is that in the distribution of the economy and what a broader set of all of the users of AI models care about. The second thing is especially if it's evals where you have tasks, what is the pass rate or difficulty of it? Of course, if you have a task that every model's able to complete, then it's not valuable at all. If you have a task that no model's able to complete, it becomes dramatically more valuable. And the final thing I would say is that maybe the largest misconception in how people think about the value of data and this ties a little bit to the RSI question around traces is that models aren't able to learn from just like reading your like Slack org and like all of the company data. For the same reason that we aren't able to learn from that stuff.

就像如果我只是读一堆幻灯片示例，我并不一定知道如何构建最完美的幻灯片。我需要的是反馈，比如我在制作幻灯片时做得好的地方是什么，做得不好的地方是什么，以及我如何衡量这些，以便我能有更多尝试、迭代、改进等。所以这就是为什么强化学习任务比仅仅拥有一堆数据（比如“这是发生了什么”）要有价值几个数量级。这就是为什么专家在创建这些任务中扮演巨大角色，我们现在每月有大约一百万小时的人在工作来构建这些任务。

Like if I just read a bunch of examples of slide decks, I don't necessarily know how to build the most immaculate slide deck. What I need is feedback of like here is the thing that I did well in producing the slide deck. Here's the thing I didn't do well and how do I measure that in a way that I'm able to have more attempts, iterate, improve, etc. And so that's why RL tasks are orders of magnitude more valuable than strictly just having, you know, a bunch of data of like here's what happened. And that's why experts play a giant role in creating these where we have about a million hours a month now of people that work to build out these tasks.

所以，这是一个很好的过渡，因为我接下来要问你关于专家的问题。你提到创建数据来训练 AI 将是人类的最后一份工作。所以，这暗示了人类数据收集有一个保质期。当模型在某个主题上如此出色，以至于没有或没有足够的人类专家来实际标注或评判它时，会发生什么？

So, this is a great segue because I was going to ask you about experts next. Now you mentioned that creating data to train AI will be the last human job. So, that implies there's like a shelf life for human data collection. What happens when the models are so good at the subject that there aren't like any or enough human experts to actually label it or judge it?

嗯，我认为正如我之前提到的，人类改进模型的市场受限于人类能做什么而模型不能做的事情的数量。所以，我认为最终我们会达到超级智能，那时数字世界或物理世界中没有任何人类能做而模型不能做的事情。但在这条路上，整个道路都铺满了我们如何通过构建评估和训练数据集来将人类智能转化为模型，以便我们可以爬山。嗯，所以我认为要有效做到这一点需要非常长的时间。嗯，但我认为人类将这样贡献于经济的原因是我们将远离人们做重复冗余工作的范式，比如我们如何有一个客户支持代表重复响应 100 张票，或律师重复红lining合同，或顾问遵循他们的麦肯锡模板。他们将转向如何训练一个智能体如何做那件事的范式。这将以各种形式表现出来。有时是构建评估，有时是更新技能文件。但我只是认为人类贡献于经济的整个方式将转向训练智能体如何做事的范式，因为这在结构上是更高杠杆的完成工作的方式。

Well, I think that as I was getting at earlier, the market for humans improving models is bound by the amount of things that a human can do that a model can't do. And so, I think eventually we will get superintelligence where it's like there's nothing in the digital world or the physical world that a human can do that models can't do. But along the way, that entire road is paved with how do we translate the human intelligence into the models through building the evals and the training data sets that we can hill climb. Um, and so I think that's going to take a extremely long time to be able to do effectively. Um, but but I think the reason that that's how humans will contribute to the economy is that we'll move away from a paradigm of people doing repetitive redundant work, of how do we have a customer support representative that is redundantly responding to 100 tickets or lawyer that's repetitively redlining contracts or the consultants that are following their McKinsey template. And they will move towards a paradigm of how do we train an agent how to do that thing. And that will manifest itself in a variety of forms. Sometimes that will be building an eval, sometimes that'll be updating skills files. But I just think like the entire way that humans contribute to the economy is going to move towards this paradigm of training agents how to do things because that is structurally higher leverage uh way to complete work.

好的，那数学呢？有很多这样的东西。

Okay, what about like math, right? There's all this uh stuff like that.

在数学中，有干净的验证器，是的。

In math with clean verifiers, yeah.

好的，那么你会说，也许人类专家在像数学这样的地方正在失去效用？

Okay, so would you say that maybe like the human experts are losing utility perhaps in a place like math?

我想我会这么说。嗯，我会说一个很好的启发式是我们如何学习事物，对吧？就像如果我们想写出更高效的代码，我们可以看看编译后代码的效率。所以我们有一个干净的验证循环，不需要从别人那里获得反馈。所以有些问题像数学、代码效率，甚至后训练，你只需要看评估的分数，感觉这些问题的大部分不需要很多人类专业知识，因为有一个超级定义良好的环境。但如果我想到经济中的绝大多数事情，比如你如何与候选人进行对话让他们兴奋，或者你如何进行销售对话——有太多这些本质上是人类的事情，甚至很多法律问题，比如你如何提出商业策略？对于一个给定主题，你可以有 10 种方式提出世界级的商业策略，100 种方式犯错。这完全相反于有一个干净的验证循环。如果有人只是说“好吧，就在接下来 10 年实施所有这些商业策略并弄清楚”，那真的很难学习。显然，需要一些间歇性奖励，比如我们的导师给我们反馈“嘿，这是进展顺利的事情，这是不顺利的事情”。所以我认为如果我反思我们如何学习，绝大部分是来自他人的反馈，但也有一些元素是我们有超级干净的反馈循环，像计算器一样。

I think I think I would say that. Um and I would say it's like a really good heuristic is how we can learn about things, right? It's like if we want to get better writing really efficient code, we can sort of just like look at like, you know, how efficient was that code after, you know, compiling it. And so we have a clean verification loop where we don't need to get feedback from others. And so there are problems like math, like code efficiency, uh even like uh post training, right? Where you could just look at what is the score of the eval, where it feels like there is uh parts of large parts of those problems that aren't going to require a lot of human expertise because there's a super well-defined environment. But if I think about the vast majority of things in the economy of like, how do you um I don't know, have like a a conversation with a candidate in a way that gets them excited, or how do you have a a sales conver- like there's so many of these things that are intrinsically human, even so many legal questions, right? Of like, how do you uh or how do you come up with a business strategy? Where there's like 10 ways that you could come up with, you know, a world-class business strategy for a given topic, 100 ways that you could make a mistake. That's like the total opposite of having a clean verification loop. Um and it it would be really hard for someone to learn if you just say like, "Okay, just go implement all those business strategies over the next 10 years and like figure it out." Obviously, there needs to be some intermittent reward of like who are our mentors that give us feedback of like, "Hey, here is the thing that's going well. Here's the thing that's not going well." Um and so I think if I if I if I reflect on how we learn, it's like the vast majority of it is feedback from others, but there is um some element of it where we have super clean feedback loops like a calculator.

Mira 是否将其重点放在评估等方面，转向一些更不可验证或不太干净和丰富的领域？

Has Mira kind of been shifting its focus on evals and whatnot towards some of these regimes that have more unverifiable or like less clean and rich

是的，我会说从一开始这就一直是我们的重点。我们确实这样做了，因为任何领域的市场都受限于人类能做什么而模型不能做的事情的数量。所以我们在推理模型的早期开始雇佣像 IMO 奖牌得主这样的人。当然，我们不再真正需要 IMO 奖牌得主了。但与此同时，我们当时雇佣了很多顾问和银行家，我们仍然非常需要那些人。所以只要那些专家中的任何一个能在数字或物理世界中做模型不能做的事情。其中一些可能是感觉本质上是人类的事情，比如如何与某人进行真正引人入胜的对话或表达同理心。我们需要教模型如何做到这一点。

Yeah, I would say from the beginning that's always been our focus. We definitely did cuz the market in any domain is bound by the amount of things that a human can do that a model can't do. So we start out hiring like IMO medalists in the early days of reasoning models. Of course we don't really need the IMO medalists anymore. But at the same time we were hiring a lot of consultants and bankers then and we still really need those people. And so really so long as there is anything that one of those experts can do that the models can't do in the digital or physical world. And some of that might be something that feels intrinsically human like how does it have a really compelling conversation with someone or express empathy. We need to teach the models how to do that.

这是另一个很好的过渡，因为我想和你谈谈比如 Apex，这是你开始做的一些工作。所以一年前在 Apex 上，最好的模型得分在低个位数。现在前沿在哪里，以及那个数字和专业人士之间的差距实际上还剩什么？

This is another great segue because I wanted to talk to you about like Apex for example which is where you're starting to do some of this work. So a year ago on Apex the best model was scoring in like low single digits. Where is the frontier now and what's actually left in the gap between say that number and a professional?

是的，所以我们构建了这个名为 Apex Agents 的基准，我们雇佣了大量专家，涵盖我们在金融、法律和咨询领域看到的顶级专业领域，他们创建了数据室，镜像了他们在经济中完成的工作中的真实数据室，并将所有这些渲染成 Google 工作空间的模拟，然后创建提示和评分标准来验证模型完成响应的能力。有趣的是，01 模型（我想现在大约 18 个月了）在 Apex Agents 上得分不到 1%。当我们今年二月推出 Apex Agents 时，前沿模型大约在 23%，现在它们达到了 65%。所以，这疯狂地展示了这样一个事实：一旦你能为某件事构建评估，实验室就能以荒谬的速度爬山。即使对于跨越非常广泛问题空间的超级多样化评估也是如此。

Yeah, so we built this benchmark called Apex Agents where we hired tons of experts across the top professional domains that we see in finance, law, and consulting and they created data rooms that mirror real data rooms that they had in jobs that they were completing in the economy and rendered those all into mocks of Google workspace and then created prompts and rubrics to verify how well models can complete responses. And the fascinating thing is that 01 which I guess is about 18 months old at this point scored less than 1% on Apex Agents. When we launched Apex Agents in February of this year frontier models were at about 23% and now they're at 65%. And so, it's been just insane demonstration of the fact that once you can build an eval for something, the labs can hill climb it ridiculously quickly. Even for super diverse evals across a really broad problem space.

你提到了 Apex 智能体，但我觉得你还有更多。

So, you mentioned Apex agents, but I think you have more.

是的。所以，我们有与 Ramp 合作的 Apex 会计。

Yeah. So, you have Apex accounting with Ramp.

你们还与 Harvey 在另一个评估上有合作，对吧？

You have another partnership with Harvey on another eval, right?

我们与 Harvey 构建了法律智能体基准。我们与 Sierra 构建了表格基准。我们还与 Cognition 构建了许多 Apex 套件。

We built legal agent bench with Harvey. We built tab bench with Sierra. We built—yeah, lots of these Apex suite with Cognition as well.

那么，这种事情是如何整合的？这些公司给了你什么，可能是你无法在内部完成的？

So, like how does that sort of thing come together? Like what do those companies give you that maybe you can't just like do in-house?

他们理解的是其用户群中存在的超级现实且具有经济价值的使用模式。他们贡献这一点的动机是，应用层公司的存在理由——或者说他们几乎要对前沿模型拥有一定防御力的方式——是由他们能在多大程度上商品化前沿模型所决定的，对吧？举个例子，如果 Claude 在编码方面比其他任何模型都好得多——达到极端程度——那么与 Claude Code 竞争的应用层公司就没有市场了。所以，如果他们发布一个公开评估，那么所有中国实验室都会针对它进行蒸馏。其他每个实验室都会对它进行爬山优化。而差距——你知道，来自前沿模型的杠杆作用开始减少，这样他们就能在他们关心的分布上改进整个生态系统，这对他们的业务至关重要。所以我认为我们将看到这一趋势在每个初创和企业中以急剧增加的程度出现——那些公司想要构建公开评估——因为，你可以花一百万美元构建一个公开评估，而实验室们已经花了数亿美元对它进行爬山优化。因此，这是一种非常高杠杆的方式来改进你关心的分布中的模型能力。

So, the thing that they understand is the super realistic and economically valuable usage pattern that exists in their user base. And their motivation to contribute that is that an application layer company's reason for existence—or like almost the way that they would have some defensibility against the frontier models—is framed by how much they can commoditize the frontier models, right? Like if you take an example, if Claude were way, way better than any of the other models at coding—like to an extreme extent—there'd be no market for an application layer company that's competing against Claude Code. And so, like if they put a public eval out in the world, then all the Chinese labs will distill against it. Every other lab will hill climb it. And the gap between—you know, there starts to be less leverage from just the frontier models so that they're able to improve the entire ecosystem at the distribution they care about, which is incredibly important for their businesses. And so I think we will see this trend to a dramatically increasing degree across every startup and enterprise—those companies wanting to build public evals—because, like, you can build a public eval for a million dollars and the labs have spent hundreds of millions hill climbing it. And so it's a very high leverage way to improve model capabilities in the distribution that you care about.

你能否论证说，像这样的公司可以利用他们拥有的独特见解和优势，直接在内部对模型进行爬山优化和后训练？

Could you make the argument that like maybe a company like that could use that unique insight and edge they have to just like hill climb a model and post train internally?

在某些情况下，这是有道理的。所以他们每个人也有其他私有的评估。但他们也有一组公开的评估。我认为这非常取决于你想在业务中建立竞争优势的地方，以及你想商品化的领域。当然，这是每个玩家公司都在努力解决的有趣战略问题。

In some cases that makes sense. And so each of them also have other evals that are private. But then they have a set of evals that are public. And I think it's very contextual to where do you want to build competitive advantages in your business versus what are the areas that you want to commoditize? And that's of course like an interesting strategic question that every player company is working through.

正如你所说，我有时也进行一些后训练。

I have had some sometimes post training as you're getting at.

是的。

Yeah.

所以是的。

So yeah.

那么现在让我们多谈谈训练。

So now let's talk a little bit more about training.

好的。

Yeah.

如果明天发布了一种超级样本高效的学习或建模技术并席卷全球，Mirko 会怎样？

What happens to Mirko tomorrow if a super sample efficient way of learning or modeling technique releases and takes the world by storm?

我认为这最可能表现为拥有更好的流程来创建合成训练数据以对评估进行爬山优化。然后我认为最终所有数据生产都将纯粹收敛于如何构建评估？因为即使你可以自动化整个训练过程——就像一旦你有一个评估，你可以在一夜之间饱和它——你仍然需要某种感觉，比如什么是好的？超越前沿的东西是什么，你如何衡量它？所以我认为很多人没有意识到我们业务的一点是，我们做的大部分实际上是构建评估，而不是交易数据。我认为随着后训练在经济每个领域变得自动化，这将越来越成为趋势。

I think that the thing this would most likely look like is having much better processes for creating synthetic training data to hill climb an eval. And then I think that ultimately all data production will just purely converge on how do you build evals? Because it's like even if you can automate the entire training process—like once you have an eval you can saturate it overnight—you still need some sense of like what does good look like? What is the thing beyond the frontier and how do you measure that? And so I think one thing a lot of people don't realize about our business is majority of what we do is actually building evaluations, not trading data. And I think that will increasingly be the trend as post-training gets automated throughout every domain in the economy.

你那里有一个大概的百分比吗？比如 80% 的工作是评估创建之类的？

Is there like a vibe percentage you have there? Like 80% of work is like eval creation or something?

总是有一点模糊的界限，因为这是相同的任务形状，所以并不总是完全清楚，但我会说略微多数是评估。而且通常是为了那个用例的评估：如何创建合成变体，然后，你知道，如何在这些评估上进行超级样本高效的爬山优化。

There's always a little bit of a blurry line because it's the same task shape and so it's not always perfectly clear, but I would say it's a slight majority is eval. And it would often be eval for that use case of how do you create synthetic variants and then, you know, how super sample efficient hill climbing at those evals.

好的，现在我们之前稍微谈过这个，但我想问你另一条推文。你预测一家前沿实验室将在未来 3 个月内推出或重新推出一个微调产品，以与 RL 即服务公司竞争。但我觉得——至少据我观察——趋势似乎正朝着相反的方向发展，因为 OpenAI 曾经有一个微调产品，但后来他们弃用了它。那么，你怎么看？

Okay, now we were talking about this a little bit earlier, but I want to ask you about another tweet that you have. You made a prediction that a frontier lab will launch or relaunch a fine-tuning product in the next 3 months to compete with the RL-as-a-service companies. But I think I—at least as I observe it—it seems like the trend is going in the opposite direction because OpenAI had, let's say, a fine-tuning product, but then they deprecated it. So, what do you think?

所以，我认为他们确实有一个微调产品并弃用了它，其中 SFT 没有真正起作用，RFT 没有真正起作用，原因是——

So, I think that it's certainly right that they had a fine-tuning product and they deprecated it, where SFT didn't really work, RFT didn't really work, and the reason—

起作用？

Work?

嗯，他们没有获得采用。我认为尝试采用它的人无法使用它，因为基础设施运行不佳，而实验室没有投资修复基础设施，因为没有——模型改进得太快，以至于没有市场需求。我认为实验室将开始拥有更多后训练产品的理由是，对于经济中的某些工作流程，我们正在达到一个充分性阈值，你不需要为每个用例都使用超前沿模型。从结构上讲，更有效的方法是只使用一个较小的模型，该模型使用他们超大型前沿模型的蒸馏版本针对特定工作流程进行定制。当然，已经开始出现这个初创生态系统，许多公司正在进行后训练和 RL 即服务以及所有相关产品。所以也许 3 个月有点夸张，但感觉如果这继续成为一个非常大的市场，在未来 3 个月以看起来的趋势发展，那么实验室不可避免地处于最佳位置来蒸馏和训练较小的模型，特别是在他们的客户已经达到充分性阈值的情况下。

Well, they didn't get adoption. I think people that tried to adopt it couldn't use it because the infra didn't work well and the labs didn't invest in fixing the infra because there wasn't—the models were improving too fast for there to be market demand. The reason that I think the labs will start having more post-training products is that it feels like for a certain set of workflows in the economy we are reaching a sufficiency threshold where you don't need the ultra frontier model for every use case. And it's structurally more efficient to just have a smaller model that is customized using a distilled version of their super large frontier model for a specific workflow. And there's starting to be this startup ecosystem, of course, with a lot of companies that are post training and doing RL as a service and all the products around that. And so maybe 3 months is a little bit of an exaggeration, but it feels like if this continues to become a very large market over the next 3 months in the way that it looks like things are trending, then it's inevitable that the labs are the best position to distill and train smaller models, especially in cases where their customers have hit the sufficiency threshold.

这难道不是有点搬起石头砸自己的脚吗？

Is that not kind of like shooting themselves in the foot a bit?

嗯，不一定。就像我认为在某些方面它会降低他们的总收入，因为他们可能有一个较小的模型或每 token 收费更低，但这比仅仅有一个竞争对手，你知道，在成本上具有竞争优势，从而失去客户关系和用例要好。

Well, not necessarily. Like I think it in some ways it drives down their total revenue because they might have a smaller model or charge less per token, but that's better than just like having a competitor that, you know, has a competitive advantage on cost where they lose that customer relationship and use case all together.

那么关于数据，你如何区分适合后训练的数据和不适合的数据？

Now on data, how do you differentiate between data that is suitable for post training and data that is not?

最重要的是，除了我之前提到的真实性和难度之外，就是验证器的质量。如果我们以我之前提到的构建优秀幻灯片的问题为例，构建评分标准是一个极具挑战性的问题，因为创建优秀幻灯片有十几种方法，而犯错的方式有上百种。因此，最好的评分标准需要捕捉所有这些，问题空间的全部潜力。你可以走的所有路径是什么？你可能犯的所有错误是什么，并避免任何假阳性和假阴性，当你在训练中展开数百条轨迹来给所有这些打分时。所以最有价值的数据是这些强化学习任务，你可以构建出色的验证器，我们很多技术都投入到如何以最有效的方式做到这一点？如何优化这个问题，将——如果我们认同人类贡献经济的方式是通过训练智能体而不是做重复性工作——我们如何优化将一小时人类时间转化为尽可能多的模型改进的比率，针对整个经济分布？

The most important thing is outside of like realism and difficulty that I talked about earlier is the quality of the verifiers, where if we take this problem I was mentioning of building a great slide deck, it is an incredibly challenging problem to build a rubric because there's a dozen ways you could create a great slide deck and a hundred ways you can make a mistake. And so the best rubric needs to capture all of that, that full potential of the problem space. What are all the paths you could go down? What are all the mistakes you could make and avoid any false positives and false negatives when you roll out the hundreds of trajectories during training to score all of those. And so the most valuable data is these RL tasks where you can build spectacular verifiers and a lot of our technology goes into like how do you do that in the most efficient way possible? How do you optimize this problem of converting, if we subscribe to the fact that the way humans are going to contribute to the economy is by training agents rather than doing repetitive work. How do we optimize the ratio of converting an hour of human time into as much model improvement as possible against the entire distribution of the economy?

制作验证器的难点是什么？

What is hard about making a verifier?

当问题空间如此之大时，所有的假阳性和假阴性都是难点。我认为这会更复杂——在数学中很容易，你只需检查正确答案，但大多数领域不是这样，对吧？即使在代码中，通过单元测试也相对容易。我们仍然看到很多奖励黑客行为，比如模型会删除整个代码库然后重新生成并通过所有单元测试，或者添加大量注释。但在其他没有单元测试、没有定量答案的领域，要真正构建避免假阳性和假阴性的验证器是极其困难的。

It's all the false positives and false negatives when the problem space is so large. And I think it gets compounded—it's so easy in math where you just check the right answer and most domains aren't like that, right? Even in code it's somewhat easy with unit tests. We still see a lot of reward hacking where a model will like delete the entire code base and then reproduce it and pass all the unit tests or proliferate tons of comments. Um but in other domains where you don't have unit tests and you don't have like quantitative answers, it is unbelievably challenging to like actually build verifiers that avoid false positives and false negatives.

当你拥有用户信号时会发生什么？让我们以你的 PowerPoint 例子为例，对吧？我是一家像 Gamma 这样的公司，对吧？我正在生成 PowerPoint，也许用户接受了这些，不接受那些。这能作为训练信号或评估信号吗？

What happens when you have like user signals? So let's take your PowerPoint example, right? I'm a company like Gamma, right? I'm generating PowerPoints and maybe the user like accepts these ones and doesn't accept these ones. Like can that be like the training signal or the evaluation signal?

这是一个有用的标志，可以告诉你应该为哪些提示创建评分标准，但它不是有用的训练信号，原因是它没有回答这个问题：所有出错的方式是什么，所有重要标准是什么，因为你需要的是当你从正在训练的模型中展开一百条轨迹时，如何给所有这些轨迹打分，以对应其好坏程度？所以你不能仅仅通过简单的点赞和点踩来推导出幻灯片中所有好和所有坏的完整问题空间。但至少你可以理解模型基于点踩会犯错的失败模式，这样你就可以针对性地进行评分标准或任务生产，在最高杠杆的地方。

It's a helpful flag for what are the prompts that you should create rubrics over, but the reason it's not a helpful training signal is it doesn't answer the problem what is—what are all of the ways that there were mistakes and what are all of the like important criteria cuz that's what you need is when you roll out a hundred trajectories from the model that you're training, how do you get a score on all those trajectories that corresponds how good or bad it is? And so you can't really derive the full problem space of all the good and all of the bad in that slide deck just from a simple thumbs up and thumbs down. Um but you are able to understand at least what are the places—what are the failure patterns where a model would make mistakes based on the thumbs down so that you can target uh rubric or task production uh the highest leverage places.

你认为可以自动化吗？比如查看错误并找出共性和弱点？

Do you think you could automate that? Like looking through the errors and finding like the commonalities and weaknesses?

我认为在那些非常干净的环境中，你可以自动化。比如我回到数学和代码效率。但对于幻灯片的美观程度，我认为你不能。原因是这一定义上超出了模型能力的前沿。模型如何能可靠地衡量它不知道如何做的事情？而且很多部分涉及理解人类境况，比如人类觉得什么样的幻灯片美观，对吧？所以即使你有超级智能，也有这个问题：如何制作人类喜欢看的电影，如何制作人类喜欢读的幻灯片？如何让对话具有说服力？我认为这些东西往往是最难构建验证器的。

I think that you can automate it in these domains with like very clean environments. Like I go back to math and code efficiency. But for like how aesthetic is a slide deck? I don't think you can. And the reason is that it is definitionally beyond the frontier of model capabilities. Um like the reason that the—like how would the model, you know, reliably measure the thing that it doesn't know how to do? Um And even many parts of these are sort of like understanding the human condition of just like what slide deck do humans find aesthetic, right? And so even if you have superintelligence, like there's this problem of like how do you build a movie that humans love to watch and how do you build a slide deck that humans love to read? How do you have a conversation that humans find persuasive? Um and I think that stuff is often some of the hardest to build verifiers for.

既然这么复杂，那你如何为这类事情构建验证器？因为 Mercury，就像你刚才谈到的为什么难。但 Mercury 正在承担并解决这个问题，对吧？

Given it's so complex then, like how do you build verifiers for that kind of thing? Because Mercury like you just talked about why it's hard. But Mercury is like taking that on and like solving that, right?

是的，没错。所以方法是，你需要展开大量轨迹，给所有这些轨迹打分，验证打分方式以查找评分中的任何假阳性和假阴性，并通过人类和智能体的某种组合来做到这一点。这通常是事物尝试的趋势。而且通常你甚至无法捕捉一切，因为比如在构建优秀验证器的过程中，你展开十几条轨迹。可能只有当你开始展开数百条轨迹时才会出现奖励黑客行为。所以这几乎是强化学习训练中最具挑战性的问题。

Yeah, exactly. So the way you do it is you would need to roll out a ton of trajectories, score all of those trajectories, validate the way that they're scored to uh look for any false positives and false negatives in the scoring, uh and do that through some combination of humans and agents. Um and that's generally what things have tried trended towards. Uh and oftentimes you don't even catch everything cuz like say you roll out a dozen trajectories in the process of like how to build a great verifier. There might be reward hacking only when you start rolling out hundreds of trajectories. Um and so it's a it's almost certainly the most challenging problem in RL training.

奖励黑客？

Reward hacking?

是的。

Yeah.

你多久遇到一次？

How often do you run into that?

我是说一直都有，对吧？就像我们每月在数百万小时内生产数十万个任务。所以我们考虑的最大事情是这些验证器的噪声是什么，以及我们如何构建一个非常好的任务形状和评估信号来爬山。

I mean all the time, right? It's like we're producing hundreds of thousands of tasks a month across those millions of hours. And so uh the largest thing we think about is like what is the noise of those verifiers and how we um structure uh yeah, a really good task uh shape and eval signal to hill climb.

现在，我们在论坛上有一个持续的辩论。

Now, we have like a ongoing debate in the forum.

嗯。

Mhm.

那就是强化学习环境是否有保质期？你提到过创建数据可能是人类的最后一份工作，所以你可能对这类事情也有一些看法。你的想法是什么？

Which is do RL environments like have a shelf life? Um you mentioned I guess that like creating data is like the last human job, so you probably have some opinions on this kind of thing as well. What are your thoughts?

嗯，我会——我知道我有点重复，但我会回到这一点：这取决于具体市场以及人类能做什么而模型不能做的事情有多少。当然，如果你在编码领域构建强化学习环境，而在 12 个月内我们达到一个范式，人类软件工程师能做的而模型在编码中不能做的事情很少，那么感觉继续在那里构建更多环境的市场就不大了。或者至少那个市场会被实验室内部化，因为对于纯合成的东西，实验室通常希望内部完成。

Well, I would—I know I'm being a little bit repetitive, but I would tie back to this thing of like it depends on the specific market and how many things a human can do that a model can't do. Cuz like of course, if you build RL environments in coding and in, you know, 12 months we get to a paradigm where there's nothing that a human or there's very few things that a human software engineer can do that a model can't do in coding, then it feels like there's not a super large market for how do we continue building more environments there. Um or the very least like that that market will be in-housed by the labs cuz for things that are purely synthetic labs would generally want to do that in house.

但对其他非常依赖品味的事情，或者另一个极端的例子，是物理世界里的所有事情——我们怎么得到能跑科学实验的模型，以及机器人领域所有这些令人兴奋的问题。感觉要等很长时间，我们才会有在物理世界所有事情上都比人类更强的模型，强到人类在任何任务上都没有比较优势。所以我认为那是人类的最后一份工作，但我觉得那是在一个长达几十年的时间尺度上。AGI 很远、ASI 很近，也许可以这么概括。也许未来 3 到 5 年我们能自动化经济中 50% 的任务，但我认为存在一条长尾。很多事情的“人味”是经济结构本身内生的。然后我认为物理世界还要再花几十年。

But I think for other things that are super taste-based, or even another extreme example is all the things in the physical world — how do we get models that can run scientific experiments, and all these exciting problems in robotics. It feels like it's going to be a really long time before we have models that are better at absolutely everything in the physical world that humans can do, where there's no task that humans have a comparative advantage at. So I think it's the last human job, but I think that's on a many, many decade time horizon. AGI long, ASI short is maybe one way of putting it. Maybe we'll be able to automate 50% of tasks in the economy in the next 3 to 5 years, but I think there's a long tail. I think there are a lot of things that are intrinsically human just in how the economy is wired. And then I think the physical world is going to take another few decades.

我想说，那会是数据收集的下一个前沿，因为归根结底，我们需要能够衡量那个前沿：模型在物理世界里所有成功的事情和所有失败的事情——不只是人们常说的低技能领域，也包括电工、机械师、科学家等真正高技能的领域。我们非常期待能利用我们的产品和能力去改进那里的模型。

I would say that I think that will be the next frontier of data collection, because ultimately we'll need to be able to measure that frontier of all the things that models succeed at and all the things they fail at in the physical world — not just in the low-skilled domains that people talk about, but also in the really high-skilled domains of electricians, mechanics, scientists, etc. And we're very excited about how we can leverage our products and capabilities to improve models there.

比如说，为“当电工”准备数据或评测，有什么独特之处？

What is unique about, let's say, preparing data or evals for being an electrician?

这个我不是专家。我猜——

I'm not an expert in this one. And I guess that's —

那就随便挑一个物理世界的任务吧。

Pick a physical world task, let's say.

好，那我给你一个我们一直在反复思考的：湿实验室。客户用我们的部分原因是，我并不真的知道怎么跑湿实验室实验，AI 研究人员也不知道。所以我们需要的是在化学、生物或任何必要领域有专业知识的人来跑这些实验。我们需要实验室里与实验相关的设备。我们需要确保这些人在所有合适的时间都能到位来跑实验。所以，如何协调这一切的后勤，是一个非常难的问题。有意思的是，我和一些在物理世界做自动化的领先公司聊过，比如 Neil Labs，它们中很多已经融了九到十位数的资金，并计划把其中一半以上花在数据上。因为数据正是它们提升模型能力的主要瓶颈。

Yeah, well, I'll give you one that we've been thinking about a lot: wet labs. Part of the reason our customers use us is that I don't really know how to run wet lab experiments, and neither do the AI researchers. And so what we need is people that have the expertise in chemistry, biology, or whatever is necessary to run those experiments. We need the equipment relevant to that experimentation in the lab. We need to make sure that those people are available at all the right times to run experiments. So it's a really hard problem to have the logistics of how do you coordinate this. And it's fascinating that I've talked with some of the leading companies, like Neil Labs, that are automating things in the physical world, and many of them have raised, you know, nine or 10 figures and plan to spend more than half of that on data. Because that is the primary bottleneck to how they can improve model capabilities.

那迭代速度呢？因为跑代码、测代码可以很快，对吧？但你不能那么快地“修补”一个电路。

What about iteration speed? Because you can run and test code pretty quickly, right? But you can't patch a circuit as quickly.

确实如此。所以你没法轻易对它做强化学习，对吧？就是那种把尝试和评分过程自动化的做法。这让它难得多。这是物理世界的一个结构性特点。物理世界的另一个结构性特点是，失败容忍度往往低得多。写代码的好处是，你可以随手 vibe code 一个东西，如果不太行，直接扔掉重做，或者对实验效果做个 A/B 测试。而无人驾驶之所以花了那么长时间、那么多聪明人投入其中，是因为 99.99% 的可靠性阈值都不够。再想想电工、机械师，甚至洗碗，对吧？你不能有 40% 的时候搞砸，因为你可能把所有盘子都打碎，或者把厨房弄得一团糟。所以我认为物理世界所需的可靠性阈值，绝对是我怀疑物理世界的超级智能还要几十年才能实现的原因之一。

That's certainly the case. And so you can't do RL with it easily, right? Where you just automate the process of attempts and grading. So that makes it a lot harder. And that's something structural about the physical world. The other structural thing about the physical world is that oftentimes the failure tolerance is a lot lower. With coding it's nice because you can vibe code something and if it doesn't really work, you can just toss it out and redo it, or you can have some sort of A/B test for how well your experiment works. Whereas the reason driverless cars took so long, with so many brilliant people focused on it, is that a 99.99% reliability threshold wasn't enough. And if you think about something like electricians and mechanics, or even doing dishes, right? You can't mess up 40% of the time, because you might break all the dishes or mess up the kitchen and all these things. So I think the reliability threshold needed for the physical world is definitely one of the reasons I suspect it'll take decades before we actually have superintelligence in the physical world.

那你会说，在这种场景里，合成数据和合成环境要重要得多吗？

So would you say synthetic data and environments are much more important in this kind of a gym?

不完全是。我会说真正重要的是模拟。我们做了很多这方面的工作。不知道你们熟不熟悉一个叫 SciCode 的基准。SciCode 是顶级的科学基准，由 Min Yang Tian 构建，他当时是 UIUC 的物理学博士生。然后他去找了他所有读博的朋友，把他们最难的科研问题都收集来。后来我们当然就招了 Min Yang，并构建了 SciCode 的未来版本。其中一个前提是：我们如何构建模拟，来预测那个湿实验室实验会跑得怎么样，从而能把它用于强化学习？因为如果我们能在代码里模拟它，哪怕只是粗略近似，那就能让这个问题对实验室来说变得可攻克。所以在数字世界和物理世界如何融合方面，有很多很酷的事情可以做。

Not exactly. Simulations is the thing I would say is really important. We've done a lot of this. I'm not sure if you all are familiar with this benchmark called SciCode. SciCode is the top science benchmark, built by this guy Min Yang Tian, who was a PhD at UIUC in physics. And then he went to all of his friends who are PhDs and got all of their hardest science problems. So then we of course hired Min Yang and built out the future versions of SciCode. And one of the premises is: how do we build simulations that can predict how well that wet lab experiment is going to run, so that we can use that for RL? Because if we can simulate it in code, or even a rough approximation of it, then that makes it a tractable problem for labs to climb. So there's lots of cool stuff you can do in how the digital and physical world blend together.

我知道 SBC 有一群人在做这类事情，比如 sim-to-real 差距等等。我猜，在你们处理这些问题时，这个差距对你们来说是多大的问题？

I know there are a bunch of people at SBC working on this kind of thing, like the sim-to-real gap and so on and so forth. How much of a problem is that for you, I guess, as you approach these problems?

这极其关键。所以当你问关于把人们的数据变现的问题时，我第一句话就是真实性。数据只有在能预测经济中某个比例、以及人们愿意为它赋予的价值时，才有意义。所以 sim-to-real 差距极其关键。我们做的很多事情，就是去弄清楚人们在 sim-to-real 差距上可能忽略的那些反共识想法。一个很有意思的例子是，我经常喜欢问别人：他们工作中多大比例的任务需要和别人互动？仔细想想，对大多数工作来说其实超过 50%。然而，如果看这在基准里的体现，是 0%。Tau-bench 里有一些多轮对话，但并不是和不同同事互动。所以我们在研究各种有创意的东西，关于如何在你能想象到的每一个维度上衡量物理世界和数字世界里的 sim-to-real 差距，从而把它弥合起来。

It's super critical. That's why when you asked the question around monetizing people's data, the first thing I said is realism. The data only matters insofar as how it is predictive of a certain percentage of the economy and the value that people would place in it. And so the sim-to-real gap is super critical. And a lot of what we do is we try to figure out what are these contrarian ideas for the sim-to-real gap that people might be missing. One fascinating example is I often like asking people what percent of the tasks that they do in their job require interacting with other people. And if you think through it, for most jobs it's actually over 50%. Yet, if we think about what is the representation of that in benchmarks, it's 0%. There's some multi-turn with Tau-bench, but there's not really interacting with different co-workers. And so there's all sorts of creative things that we do research on with respect to how do we measure the sim-to-real gap across every dimension you can imagine, in the physical world and the digital world, so that we can bridge that.

那说到合成数据，前沿实验室正在大力投资合成数据流水线，并且全部留在内部。你觉得有没有可能，最终大部分数据都会是合成的？

So on the topic of just synthetic data, frontier labs are investing significantly into synthetic pipelines and keeping them all in-house. Do you think there's maybe a world where much of the data ends up being synthetic?

我认为很多训练数据会是合成的。而评测数据绝对需要是人类的。我们一直以来的判断是，实验室会留在内部的是合成数据，或者几乎完全合成的流水线，因为那是他们建立竞争优势的方式。

I think a lot of the data for training will be synthetic. I think that the data for eval definitely needs to be human. And our thesis always has been that the thing that labs will in-house is synthetic or almost entirely synthetic pipelines, because that's how they build a competitive advantage.

如果模型 100% 只是数据的体现，那么他们当然需要在数据管道的管理上有自己的优势。但所有实验室都不想内部做、有些尝试后失败的事情，是建立一个 500 万人的庞大人才网络，我们每天支付 600 万美元，已经完成了 700 万次面试，以动员经济中各个类别的人。所以，我认为人类在评估中会扮演重要角色。我认为实验室会将合成数据管道内部化。这大概就是世界的趋势。

Like if 100% of the model is just a manifestation of the data, then of course they need to have alpha in how they curate their data pipelines. But the thing that none of the labs want to do in-house, and some of them tried and failed, is building a giant talent network of 5 million people where we're paying out $6 million a day and have done 7 million interviews to mobilize people across every category in the economy. And so, I think humans will play a giant role in eval. And I think that labs will in-house their synthetic pipelines. And that's sort of how the world will trend.

有些公司押注基础模型包含了人们所有的偏好和品味，比如说 Simile。你觉得这是否有点与这里的叙事相悖？

There are some companies that are taking a bet on the fact that the foundation models contain all of the preferences and tastes, let's say, of people like Simile, for example. Do you think this goes against a little bit of the narrative here?

我认为它们包含了一部分。我的意思是，我喜欢 Simile。我很早就投资了他们。我觉得 Jun 和 Percy 很棒。我认为预训练中可能嵌入了一些能力，这些能力可能有价值，尤其是对于他们为所有这些消费品进行市场发现的用例，但似乎总是存在结构性问题：模型无法可靠地理解它在幻灯片中哪里犯了错误，而这些评估需要由人类创建，因为它们超出了前沿。

I think they contain part of it. I mean, I love Simile. I invested early on in them. I think Jun and Percy are awesome. I think there are certainly some capabilities that might be embedded in pre-training that could be valuable, especially for their use case in doing market discovery for all these consumer products, but it just feels like there's always going to be the structural problem of a model can't reliably understand where it made a mistake in the slide deck, and those evals need to be created by humans because they're beyond the frontier.

现在，前沿是一个移动的目标。

Now, the frontiers are a moving target.

嗯。

Mhm.

你如何确保今天收集的数据明天不会过时？我觉得甚至 NeurIPS，对吧？人们谈论向 NeurIPS 或只是会议提交，对吧？也许他们会针对某些模型进行交叉切片，但等到会议真正开始时，也许它已经过时了。你处理这个问题吗？

How do you ensure that data you collect today isn't obsolete by tomorrow? I feel like even NeurIPS, right? People talk about submitting to NeurIPS or just conferences, right? And maybe they'll take a cross-slice against some set of models, but by the time the conference is actually live, maybe it's obsolete. Do you deal with this?

是的，我们做了很多。我想说可能两点。一是如果你把问题弄得非常非常难，它们的保质期会更长。举个例子，假设你创建一个 RL 任务，通过率大约 60%。也许你发布它，然后在 2 个月或更短的时间内，模型会达到 80% 或 90%，那没有多少保质期。但如果你找到真正困难的问题和模型在结构上糟糕的事情，那么也许发布时通过率低于 10% 或低于 25%，它可能会有更长的保质期。然后第二点是前沿数据按定义是定制的，对吧？就像我们总是在推动前沿，而那个前沿按定义就是新东西。所以我们经常思考的一件事是如何复合生产前沿数据的引擎，对吧？因为我们知道我们将总是需要更大的才人网络。我们知道我们将总是需要所有这些基础设施，围绕为工具和数据集上的议程质量控制启动模板。我们知道我们将总是需要伟大的研究文化来发布这些新的评估。所以我们可以复合所有这些，拥有这个真正壮观的生产新前沿数据的引擎，即使它不断进化。

Yeah, we do a lot. I would say that maybe two things. One is if you make the problems really, really hard, they'll have a longer shelf life. And so, an example is like say you create an RL task where the pass rate is like 60%. Maybe you launch it and then in 2 months or less, the models are going to be at 80% or 90%, and that didn't have much shelf life. But if you find really hard problems and things that models are structurally terrible at, then maybe with under 10% pass rates or under 25% when you launch it, it might have a longer shelf life. And then the second thing is that frontier data is definitionally bespoke, right? Like we are always pushing the frontier and that frontier is definitionally something new. And so, one thing we think about a lot is how do we compound the engine for producing frontier data, right? Because we know that we will always need to have a larger talent network. We know that we will always need to have all of this infrastructure around spinning up templates for tooling and agenda quality control on the data sets. And we know that we'll always need to have great research culture to put out these new evals. And so we can compound all of those things to have this really spectacular engine for producing new frontier data even if that is constantly evolving.

所以你在构建这个引擎。你拥有所有数据。你有专家网络。Merkle 会直接进入建模领域吗？你说开放，就像你没说 OpenAI。你说前沿实验室尝试建立人才网络但失败了。但你有那个。所以你不能直接构建一个模型吗？

So you're building this engine. You have all of the data. You have like the network of experts. Would Merkle ever move into the modeling space directly? You said open like you didn't say open AI. You said frontier labs try to set up talent networks and they failed. But you have that. So like can't you just build a model?

我们做了很多后训练。我的意思是，如你所知，我们聘请了 Edward Hu，LoRA 的第一作者。所以他发明了最常见的微调技术，然后成为 o1 的核心贡献者。他在 o1 和 OpenAI 工作，现在领导我们的后训练团队，我们做大量的定制后训练。我们肯定与所有后训练模型的 AI 初创公司和企业密切合作。但我们希望在与谁竞争和不与谁长期竞争方面保持策略性。是的。

We do a lot of post training. I mean as you know we hired Edward Hu the first author on LoRA. So he invented sort of the most common fine tuning technique and then was a core contributor to o1. And he on o1 and OpenAI and he now leads our post training team where we do an enormous amount of custom post training. We definitely partner very closely with all of the AI startups and enterprises that post train models. But we want to be tactful about who we choose to compete against and who we don't choose to compete against long term. Yeah.

好的。所以你有 Edward Hu，你正在关注后训练。这对 Merkle 的路线图意味着什么？

Okay. So you have Edward Hu and you're looking at post training. Like what does that mean for Merkle's road map?

我的意思是，我认为为了让我们知道数据集有用，最终我们需要能够在其上训练。所以后训练是一个巨大的肌肉，我们将继续投资。我们也是所有 RLHF 服务公司的最大客户。所以我们中的大多数是第一个客户。我认为当然会有客户希望我们直接为他们做模型的端到端后训练的情况，我们有非常强大的能力来做到这一点。我们还有所有 RL 即服务供应商的市场，所以我们能够外包所有工作流程。所以我们会以某种形式为客户后训练模型，但我认为最终这是从如何最好地让他们利用数据和我们与他们一起构建的所有知识产权的竞争优势的角度来看待的。

I mean I think that for us to know that a data set is useful, ultimately we need to be able to train on it. So post training is a giant muscle that we're going to continue investing in. We're also the largest customer of all of the RLHF service companies. So like we are the first customer of most of these. And I think that of course there are going to be cases where our customers want us to just do the end-to-end post training of the model for them, and we have really powerful capabilities to do that. And we also have this marketplace of all the RL as a service vendors, so that we're able to outsource all of the workflows. And so what we will post train models for customers in some form or another, but I think that ultimately that is viewed through the lens of how do we best allow them to leverage the data and the competitive advantage of all the IP that we build with them.

在过去一两年里，你的心态有没有改变？比如你最近聘请了 Edward，谁？在 Edward 之前的日子发生了什么？

Has your mindset on that shifted at all in the last like year or two? Like you recently hired Edward who? Like what happened before in the pre-Edward days?

嗯，之前——聘请顶尖研究人员一直非常困难，所以这是不变的事情。我认为公司现在有更多的资本，终于可以投资于前沿后训练工作。你知道，现在有数十亿现金和数亿月收入。所以这是一个变化。我想说的另一个变化是，定制后训练现在开始成为一个市场，这在 12 个月前是没有的，因为 A 我们谈到的充分性阈值，B 推理支出的规模，以及 C 开源模型正在接近前沿模型。所以这些都为一个市场创造了非常好的条件，感觉每个初创公司和企业都需要后训练模型，以在未来 5 年保持相关性。

Well, the pre- It's always been very hard to hire top researchers, so that's been a non-changing thing. I think the company now has a lot more capital to finally invest in having frontier post training efforts. You know, now with billions in cash and hundreds of millions in monthly revenue. And so that is one thing that's changed. I would say that the other thing that's changed is that it feels like custom post training is now starting to emerge as a market in a way that it wasn't 12 months ago because of A the sufficiency threshold that we talked about, B the scale of inference spend, and C the fact that open source models are approaching frontier models. And so those all create really good conditions for a market where it feels like every startup and enterprise needs to post train models to stay relevant over the next 5 years.

你是否需要以任何方式降低预训练数据的风险，或者你会说你最近更关注后训练？

Do you have to de-risk like pre-training data in any way, or would you say that you have more of a focus on post training these days?

更多是后训练。我们在预训练上做了一点。而且总是有这种中间训练的混合。但我想说绝大多数焦点是在 RL 数据上。

Much more on post training. We do a little bit on pre-training. And there's always this blend of also mid-training. But I would say the vast majority of the focus is on data for RL.

随着时间的推移，这有什么变化？

How's that changed over time?

变化不大。没有。我想说也许随着时间的推移，它已经从纯粹的 RL 数据扩展了。比如，我们进行了大量的公司收购，我们拥有财富——像前财富 500 强公司这些品牌名称的公司数据集的独家许可，并已经收购了数百个公司数据集。

Not much. No. I would say maybe it's expanded away from purely data for RL over time. Like, we've done an enormous amount of company acquisition where we have exclusive licenses to company data sets on Fortune like former Fortune 500 companies that are brand names and have acquired hundreds of company data sets.

你觉得放到物理世界里，这个故事会有什么变化吗？因为据我所知，我们还没到所有人都在为机器人或物理世界做后训练的阶段。

Do you think the story changes at all for the physical world? Because we aren't necessarily, at least to my knowledge, in a place where everybody's trying to post-train for robotics or just the physical world.

是的。我得说我对机器人不算专家。我特别喜欢 Jim Fan 的那场演讲——不知道你看过没有——他把 LLM 后训练的发展路径和机器人做了很好的类比。这一点一直让我很有共鸣：最初在 LLM 领域收集的行为克隆数据，加上 SFT 和 RLHF；现在机器人领域也在大量做类似的事，用的是第一人称视角数据。然后我们最终走到了强化学习范式——怎么做大量 rollout 并给它们打分。他说，这最终也是机器人领域会抵达的终局。

Yeah, I would say I'm less of an expert on robotics. I really love this lecture that Jim Fan gave — I'm not sure if you saw this — it drew a great parallel between the development of post-training for LLMs and how that compares to robotics. And that's always resonated a lot with me: how we basically had all of the behavior cloning data that was initially collected in the LLM world, with SFT and RLHF. And now there's a lot of that happening with egocentric data and robotics. And then eventually we got to the RL paradigm — how do we do a bunch of rollouts and score them. And that's eventually what he calls the end game of what we'll get to in robotics as well.

那世界模型呢？

What about world models?

对，我们在世界模型上做了很多工作。因为数字世界和物理世界本来就高度交融。我们为评估强化学习构建的很多数据集，往往也非常适用于世界模型，所以我们也和很多这类公司合作。

Yeah, I mean, we do a lot of work on world models. Because naturally the digital world and physical world blend together so much. And a lot of the data sets that we build for eval RL tend to also be super applicable to world models, and so we work with many of those companies, too.

好，我准备的问题都问完了，不过还有一个。你们在招人吗？

Now, that's all my prepared questions, but there's one more. Are you hiring?

我们一直在招人，而且招得非常快。说来有点讽刺，两年前大家还在说“哦，经济里不再需要更多人了”。但我觉得，随着我们团队的生产力越来越高，我们招的人反而比以往任何时候都多。从和 Edward、Min Yang 以及许多其他优秀人才一起工作的 AI 研究团队，到我们的工程组织，再到大量负责管理这些人类数据项目的运营人员。我们非常愿意和任何人合作。

We're always hiring. We're hiring incredibly quickly. And it's ironic, because I feel like two years ago everyone was saying, "Oh, we're not going to need anyone else in the economy." But I feel like as we make our team more and more productive, we are hiring more than ever before. And so, ranging from our AI research team working with Edward and Min Yang and many other exceptional people, all the way to our engineering org, all the way to lots of operators to help manage these human data projects. We would love to work with anyone.

人们为什么要来 Mercor 工作？

Why should people work for Mercor?

我觉得我们正处在一个时刻：我认为最重要的问题是，未来十年人类在经济中扮演什么角色。这个问题比以往任何时候都更不确定。而我相信人类会继续扮演非常重要的角色。Mercor 正是未来十年定义这一角色的最重要、最有影响力的公司。这一点的体现之一，是我们帮助构建了数十亿人使用的所有前沿模型；另一部分体现，则是我们是历史上增长最快的就业创造者。

I think that we're at a moment in time where I believe the most important problem is what role do humans play in the economy over the next decade. And that's more uncertain than ever. And I believe humans will continue to play a really important role. And Mercor is the most important and impactful company in defining that over the next decade. And part of that manifests itself in how we help to build all of the frontier models that billions of people use. And part of that manifests itself in how we're the fastest growing creator of jobs in history.

时间差不多了。感谢大家到场，让我们再给 Brendan 一轮掌声。

I think we're at time. So I wanted to thank everybody for coming and one more round of applause for Brendan.

谢谢你们的邀请。

Thank you for having me.