Ending AI Slop — Thais Castello Branco, Taste Labs

频道: AI Engineer
视频: https://www.youtube.com/watch?v=lCBf9slCanI
原文语言: en
统计: 共 11 轮

[0:13]

Hello everyone. It's great to meet you all. I'm Tais. I'm the founder of Taste Labs. For those of you who don't know us, we came out of Stealth a few weeks ago. Uh, and our whole mission is basically how do we end AI slop? And we believe that to really solve this problem, we have to first decompose and understand subjective domains. Right? I think as probably all of you know, AI has gotten quite good at things like coding and math. Uh but it's still super behind on things like design, creative writing, personality, emotional intelligence. And to understand these domains, I think we have to take a little bit of a different approach than we do with um with objective ones. So our idea is like how do we become this data and infrastructure layer to really uh understand these problems and to become the solution for them across the stack. So from the foundation model layer all the way to how do we build solutions for agents as well. So we work primarily in two ways. We work with the top frontier labs on how do we evaluate benchmark their models understand where they're breaking uh understand how we can fix them and how do we determine also which problem is better fixed through each method. So what things should be turned into RL environments, which things should be turned into post- training data problems. Um, but then we also go and work with a lot of agent and application layer companies on what are things that we actually don't believe should be solved at the foundation model layer and that might be better solved through methods like context uh or understanding user intent. Right? We're basically betting on a world where suddenly you're going to have billions of people creating uh that are not necessarily experts. So this understanding of user intent and contest and context is equally as important as how do we get these models to improve.

大家好，很高兴见到各位。我是 Thais，Taste Labs 的创始人。可能有些人还不认识我们——我们几周前才从 stealth 模式出来。我们的使命说白了就一句话：怎么终结 AI slop（AI 批量生产出来的低质内容）？我们相信，要真正解决这个问题，首先得把主观性的领域拆解开、搞明白。你们大概都知道，AI 现在在写代码、做数学这类事情上已经相当强了，但在设计、创意写作、人格、情商这些事情上还差得远。而要理解这些领域，我觉得我们得换一套跟处理客观领域不太一样的打法。所以我们的想法是：怎么让自己成为那一层数据和基础设施，真正把这些问题搞懂，并且在整个技术栈上都成为它们的解法——从基础模型那一层，一直到怎么为 agent 做出解决方案。我们主要有两种做法。一种是跟最前沿的几家 frontier lab 合作，帮他们对模型做 eval 和 benchmark，搞清楚模型在哪里崩掉、我们能怎么修，以及每一类问题更适合用哪种方法去修：哪些东西该做成 RL 环境（reinforcement learning，强化学习），哪些东西该当成 post-training 的数据问题来处理。另一种是跟很多做 agent、做应用层的公司合作，看哪些问题我们其实并不认为该在基础模型那层解决，而可能更适合用 context（上下文）或者理解用户意图这类办法来解。因为我们赌的是这样一个世界：突然之间会有几十亿人开始做创作，而他们不一定是专家。所以理解用户意图和 context，跟怎么把模型本身做得更好，是同等重要的。

[1:49]

For today I'm going to focus on the model training part. For those that are here tomorrow I'll also be giving a chat on the design track where I'll cover more on what we're doing on the agent side of the house. But there we go. Okay. So most of the world is subjective as I was mentioning. Uh a lot of the world is subjective right? If we talk about these domains of writing even workflows within companies right of sales marketing uh a lot of the times there's this multitude of answers there's not one clear right answer and it's very hard to define what great even means I think at the end of the day like those are why these domains are so difficult um and we oftentimes I would say forget to mention like why uh we treat for example the fact that code is verifiable and measurable as something that is a property about models and models are great at at coding um because we've made them great at coding but realistically it's actually a fact about code. Code is something that decomposes, it verifies, it executes and so it makes it a lot easier for us to be able to train on these domains for something like design or writing like how do you decompose it? How do you verify it? How do you judge if it's actually good? So that's why they become so difficult. Um, so there's two characteristics that I want to touch on today on why fundamentally subjective domains are harder. One is that capability follows measurability. So if we can solve the measurability problem or at least part of it, then we can solve a big portion of these domains. Uh the second which I'll touch on later is basically this like collapse to the mean and why the mean is not necessarily optimal in subjective domains.

今天我主要讲模型训练这一块。明天还在场的朋友，我在 design track 还有一场分享，会更多讲我们在 agent 这一侧在做的事。那好，我们开始。刚才说了，这个世界上大部分东西是主观的。真的，很多东西都是主观的——不管是写作这类领域，还是公司内部的工作流，比如销售、市场，很多时候答案有一大堆，并没有唯一正确的那个，而且「什么算好」本身就很难定义。我觉得归根到底，这就是这些领域这么难的原因。而有一件事我们常常忘了讲清楚：比如代码是可验证、可度量的，我们往往把这当成模型的属性——「模型很擅长写代码，因为我们把它训得很擅长写代码」。但现实是，这其实是代码本身的属性。代码这个东西可以拆解、可以验证、可以执行，所以我们在这些领域上训练模型就容易得多。可换成设计或者写作呢？你怎么拆解它？怎么验证它？怎么判断它到底好不好？这就是它们变得这么难的原因。今天我想讲两个特性，来说明为什么主观性领域从根本上更难。第一，能力跟着可度量性走（capability follows measurability）。所以只要我们能把可度量性这个问题解决掉，哪怕只解决一部分，就等于啃下了这些领域里很大一块。第二个我后面会讲，就是所谓的「向均值坍缩」，以及为什么在主观领域里，均值并不等于最优。

[3:25]

Okay. So to really start solving this problem, we have to turn something that feels fuzzy like if I ask you what is great design into something that is more verifiable. So there's a few questions here, right? Because if I ask you this of what is great design um you could ask yourself, okay, um do you mean great for which type of person, for which type of taste, for which situation? The same slide could be amazing. for example, if you are a startup and completely inappropriate if you are a finance firm. So it's contextual first of all. Second of all, it has this property that it changes over time which is different from other domains. What is considered good today is different than five years ago and different than five years from now. In code that's not necessarily true or in math, right? That's something that is way more consistent over time. So our ability to again decompose it and understand how is this good for a specific audience, how is this good today, how is this good in context um is some of the things that we've been thinking about in terms of how to how to break this down. But I want to give you a very specific example because of course this can mean many things. So let's talk about brand. Um if you're at a company and you've used coding agents, you've probably shipped an internal dashboard. You've probably shipped an internal like landing page. And oftentimes you might wonder, okay, how do I determine if this is slop, if this is actually good? And you have kind of this secret weapon at your disposal, which is really all the work that probably designers at your companies, for example, put into defining your brand. A brand to define takes a lot of effort, takes a lot of care. You're defining all these components about it, when it's good, why you're choosing certain combinations of colors, of typography, of spacing, of texture. Um, but if I just ask you to be like, okay, create something great, that's very

好，那要真正开始解决这个问题，我们得把一个感觉很模糊的东西——比如我问你「什么是好设计」——变成更可验证的东西。这里其实藏着好几个追问。因为如果我问你什么是好设计，你可以反过来问：你说的好，是对哪类人好？对哪种 taste（品味）好？在哪种场景下好？同一页 slide，如果你是一家创业公司，可能好得不得了；但如果你是一家金融机构，就完全不合适。所以第一，它是依赖 context 的。第二，它还有一个别的领域没有的性质：它会随时间变化。今天被认为好的东西，跟五年前不一样，也跟五年后不一样。在代码或者数学里就不一定是这样，那些东西在时间上要稳定得多。所以我们能不能把它拆解开，搞清楚「它对某个特定受众为什么好」「它放在今天为什么好」「它在这个 context 下为什么好」——这些就是我们一直在琢磨的拆解思路。不过我想给一个非常具体的例子，因为这么讲太泛了，可以指很多东西。我们来聊聊品牌（brand）。如果你在一家公司里用过 coding agent，你多半已经上线过一个内部 dashboard，或者一个内部的落地页。你多半也会琢磨：我怎么判断这玩意儿是 slop，还是真的做得不错？其实你手上有一件秘密武器，就是你们公司的设计师们为定义品牌所投入的那全部工作。定义一个品牌是要花很多功夫、很多心思的。你要把它的各个组成部分都定义清楚：什么时候算好，为什么选这几种颜色的搭配、这套字体排版、这套间距、这种质感。但如果我只是跟你说「来，做点很棒的东西出来」，那就非常

[5:07]

hard. But suddenly if I'm like okay make something that is on brand that is a much easier problem to define and a brand is something that can become decomposable. Uh so for example if we I'm using the reduct brand as an example here because I I I like their website. Um let's say that we decompose this brand into the colors, the typography, the motion, the animation, the textures. Suddenly you have these very codified things that you can verify against. Verifying in general if something's on brand and you can try this uh by prompting an LLM as a judge to do it is quite hard. But once you start picking apart the exact elements that represent what great is, then it suddenly becomes the shape of something that is codifiable and verifiable. So if you want to turn this into a shape of an RL environment, for example, right, how would you train a model for a capability like brand adurance? Uh LLM LLM as a judge might not necessarily always be the best method. We know that there's a lot of reward hacking. We know that there's uh interesting hallucination patterns there too. And so we oftentimes try to create methods of basically how do we turn a task that feels fuzzy into one where there's a clear ground truth so that it can become the shape of an environment. So in this case the task design itself is really kind of the hardest part of the problem of how do you turn something that appears very fuzzy into something that actually can be arled. Um and so in this case that decomposition that I mentioned becomes the ground truth. So let's say that you start by tasking an agent to create a new page that is going to adhere to the reductal brand but be completely net new and different. Um you would want that output to not only be graded versus the original but to be graded on this ground truth, right?

难。可一旦我说「做点符合品牌调性（on brand）的东西」，这个问题就好定义多了——因为品牌是可以被拆解（decomposition）的东西。举个例子，这里我拿 Reduct 的品牌当例子，因为我挺喜欢他们的网站。假设我们把这个品牌拆成颜色、字体排版、动效、动画、质感。突然之间，你就有了一堆非常明确、成文的东西，可以拿来逐条对照验证。笼统地去验证「这东西是不是 on brand」——你可以自己试试，让一个 LLM 当 judge 去判——是相当难的。但一旦你开始把「好」到底由哪些具体元素构成拆开来看，它突然就变成了一个可以成文、可以验证的形状。所以，如果你想把这件事做成一个 RL 环境的形状，你会怎么为「品牌一致性（brand adherence）」这种能力去训练模型？LLM as a judge 未必总是最好的办法。我们知道那里有大量的 reward hacking（钻奖励函数的空子），也知道那里存在一些挺有意思的幻觉模式。所以我们常常想办法：怎么把一个感觉很模糊的任务，变成一个有明确 ground truth（基准真相）的任务，让它能长成一个环境的形状。在这个例子里，任务设计本身其实才是整个问题里最难的那部分——怎么把一个看上去很模糊的东西，变成真的能拿去做 RL 的东西。而在这里，我刚才说的那套拆解，就成了 ground truth。比方说，你先给一个 agent 派个活：做一个全新的页面，要贴合 Reduct 的品牌，但内容完全是净新的、跟原来不一样。你会希望这个产出不只是拿去跟原件对比打分，而是拿这套 ground truth 来打分。

[6:40]

Because it could come up with completely new ways of using these components that are still valid but are different from the original. So you don't necessarily want to just see if it's replicating the original. So this is one example of like how to turn this into a problem of environment shaped um so that we can make it more verifiable. But in a way all of these things are I would say like spectrums right you have uh in a problem like design you have these elements of things like vision uh alignment typography that are closer to objective once you start moving up that scale onto things like style fit creativity how do you judge and measure something like creativity right that's much harder and so you kind of need to think of this as like a routing problem of how do you understand this like vast fuzzy problem break it down into smaller components and what is the best solution for each of these components. So why does something feel like slop? For example, when we're talking about something like creativity, right? I think this is the second reason why um subductive domains are so much harder to solve because if we're talking about the properties of models, they're basically predicting what's the most likely outcome to show up next. And they assume that that outcome is the ideal outcome.

因为它可能会想出全新的方式来使用这些元素，这些用法依然成立，只是跟原件不一样。所以你并不只是想看它有没有在复刻原件。这就是一个例子，说明怎么把这件事变成一个「环境形状」的问题，好让它更可验证。不过说到底，我觉得这些东西都是一条条谱系。像设计这样的问题里，有些元素——比如视觉对齐、字体排版——是靠近客观的那一端；可一旦你沿着这条谱往上走，走到风格契合度、创造力这些东西，你怎么去判断和度量「创造力」？那就难多了。所以你得把它当成一个路由问题来想：怎么理解这个庞大又模糊的问题，把它拆成更小的组件，然后针对每一个组件找出最合适的解法。那么，一个东西为什么会让人觉得是 slop？比如当我们谈创造力的时候。我觉得这是主观性领域这么难解的第二个原因：因为如果我们讲模型的属性，它们本质上是在预测「下一个最可能出现的结果」，并且默认那个最可能的结果就是理想的结果。

[7:48]

And for something like math encoding, that is true, right? You want the answer that your model gives you to be the average answer. If you're asking what 2 plus 2 is, uh, which also happens to be the right answer and the optimal answer. But for something like writing or design, you don't necessarily want the average answer, right? The average meaning the most likely does not necessarily coincide with like the optimal. Uh, a lot of like what I describe it is a lot of greatness and creativity happens actually at the ends of the distribution. It's not the most likely outcome. It's when you actually actively break from rules and actively break from patterns that you can create things that are subjective and and great. Um, and so the reason why this feels like slop and that we have this feeling that we're surrounded by by slop is exactly because of this collapse to the mean and this repetition. And so we have to find ways of okay, how do we break these patterns? How do we break from the mean uh but in a way that's also intentional. So then you kind of shift the problem onto things that are not so easily maybe verifiable. Uh but that are more questions of human preference and judgment and that might be better solved by data for example than by environments. And so again this kind of like mode collapse is is really the thing that we're trying to solve.

对数学和代码来说，这是成立的。你确实希望模型给你的答案就是那个「平均答案」——你问 2 加 2 等于几，那个平均答案同时也是正确答案、最优答案。但对写作或者设计这类事情，你未必想要平均答案。这里的「平均」是指「最可能」，而最可能并不一定等于最优。我常这么形容：很多伟大和创造力其实发生在分布的两端，它不是最可能出现的那个结果。恰恰是当你主动打破规则、主动跳出模式的时候，你才能做出既有主观性又真正出彩的东西。所以，我们之所以觉得这些东西像 slop、觉得自己被 slop 包围，原因正是这种向均值坍缩，以及由此带来的重复感。于是我们得去找办法：怎么打破这些模式？怎么从均值里跳出来，但跳得是有意为之的？这么一来，问题就转移到了那些不那么容易验证的东西上——那些更多是人类偏好和判断的问题，可能用数据来解会比用环境来解更合适。所以说到底，这种 mode collapse（模式坍缩）才是我们真正想解决的东西。

[8:56]

And how do we force that distribution back? Uh, oftentimes, by the way, we we work, for example, with a community of designers, um, over like a thousand experts that are experts in different types of medium, different styles, and we purposely want to force that distribution when we're breaking down the problem. Exactly. So, we don't end up in this mode collapse, but there's of course a lot of other pieces of that puzzle. Uh, but I think this is an interesting framework is like the closer you are to something that it becomes verifiable, especially programmatically, the better for something like RL, right? And I think the challenge is how do we turn things that feel fuzzy into things that become more verifiable by establishing this ground truth and designing tasks in a way that allow for that. But then the more that it does shift to things that are contextual or that depend on that time element that we talked about or this like distinction in preference, the more this shifts towards something that requires human judgment. So uh I I I like this analogy of basically kind of pulling things toward verification. So brand adherance on its own would be something that's very hard and that is probably better judged by a human than for example by an LLM as a judge or something deterministic. But by codifying it by understanding which pieces matter and how do I turn that into something that is observable and measurable we can kind of pull it into this realm of verification. uh so this routing logic is I would say if you have one takeaway uh take this away it's like how do we break down this problem into something that you understand what's actually the best method to solve it and so when we are talking about these things that are more subjective uh we do require human judgment I think this is something that we um we believe human judgment is still at a much higher level than any LLM as a judge and this like human taste is

那我们怎么把那个分布重新撑开？顺便说一句，我们比如会跟一个设计师社区合作，一千多位专家，他们各自擅长不同的媒介、不同的风格；我们在拆解问题的时候，会刻意去把分布撑开，就是为了不掉进 mode collapse。当然，这个拼图还有很多别的部分。但我觉得有个框架挺有意思：一件事越接近「可验证」，尤其是可以程序化验证，它就越适合用 RL 来做。而挑战在于，怎么把那些感觉模糊的东西变得更可验证——办法就是建立起 ground truth，并且把任务设计成能支撑这种验证的样子。可反过来，一件事越是依赖 context、越是取决于我们刚才说的那个时间因素、或者越是取决于偏好上的差异，它就越往「需要人类判断」那一端偏。所以我挺喜欢这个比喻：把事情往验证那一侧拽。品牌一致性这件事本身是很难判的，交给人来判大概比交给 LLM as a judge、或者某种确定性方法要靠谱得多。但通过把它成文化、搞清楚哪些部分真正重要、再把这些部分变成可观察、可度量的东西，我们就能把它往验证这个领域里拽一截。所以这套路由逻辑——如果今天你只带走一条，就带走这一条——就是：怎么把问题拆开，拆到你能看清每一块究竟该用什么方法去解。而当我们谈那些更主观的东西时，我们确实需要人类判断。我们相信，人类判断的水准仍然远高于任何 LLM as a judge，而这种人类的 taste（品味），

[10:37]

really how do we encapsulate this in a way that uh can be turned into high quality data so that we can train these models better. And I think one of the tricky things here is often times in the past you had kind of this uh collection of preference data that would collapse again to the mean because you would collect it from a bunch of different people without necessarily understanding who they are, what they like, why they like it, when they like it. And if you don't break up that problem accordingly, you then end up again with preferences that kind of don't agree with each other. Because that naturally happens in the world, right? I bet that some of you might like one style better than another. And that doesn't mean either of those things are wrong or it doesn't mean that the best answer is the average of what two people might like. It means that we need to fundamentally understand that the world is multi-preference and how do we do that matching accordingly. So this almost like understanding of how do we create like a preference vector let's say for someone and attach that to even something like preference data can help us to train in a way that allows for that more pluralism of preferences intentionally instead of that data being turned into something that's noisy.

关键就在于怎么把它封装起来，变成高质量的数据，好让我们把模型训得更好。我觉得这里有个很棘手的点：过去很多时候，你收集来的那批偏好数据（preference data）最后又会坍缩到均值上，因为你是从一大堆人那里收上来的，却没搞清楚他们是谁、喜欢什么、为什么喜欢、什么时候喜欢。如果你不把这个问题相应地拆开，最后拿到的就是一堆互相打架的偏好。而这在现实世界里本来就会发生——我敢打赌，在座各位里就有人更喜欢这种风格而不是那种。这不代表其中哪一种是错的，也不代表最好的答案就是这两个人喜好的平均值。它意味着我们必须从根子上认识到：这个世界是多偏好的，问题在于我们怎么把匹配做对。所以，如果能为一个人建立起类似「偏好向量」这样的东西，并且把它挂到偏好数据上，就能帮我们在训练时有意识地容纳这种偏好的多元性，而不是让这些数据变成噪音。

[11:40]

when we talk about data quality as well for these domains, I think it becomes very uh I don't know if any of you have bought data for example for these domains or or have tried to curate data yourselves. Um but there's it's very hard to define okay like now I'm realming to the side of data. What is actually good? What is actually going to be helpful to my model? Um and there's obviously a few things that are harder to define but a few that I think become more controllable and how do we actually understand patterns of quality so that we can measure them. So one I would say is that problem decomposition. How do we force like true distributions of what you see in the world? How do you force true like expert selection across these u different buckets? And these are things that you can totally control, right? You can see who those experts are. You can do a selection pattern that is very strict um and you can decompose that problem and this is something that is completely in your control and that totally helps with the um results let's say of the experiment being being good. Um the second is I would say that like flow of like how do you determine then the right problem routeed to the right solution. Um we also I would say do a lot of what we call essentially like QA on this data and I think there's two ways to do that is understanding what are properties about that data point that correlate with being it being rich and high signal. So for example, specificity when you're trying to ask an expert to define is this good, is this bad or put reasoning behind it or create a whole observation system around uh how they would judge an asset. The specificity of their language and of how precise they're being able to be with how they're doing that description is what will determine that data quality. Or for example, if you can tie their commentary with actually which piece in the code does this relate to.

再说说这些领域里的数据质量。我不知道在座有没有人给这类领域买过数据，或者自己动手做过数据整理——这事真的很难，很难说清到底什么才算「好」。我顺着数据这条线往下讲：什么样的数据才是真正对我的模型有帮助的？有些东西确实很难定义，但有一些我觉得是可控的，关键是怎么把质量的规律摸清楚、让它变得可衡量。第一点，我会说是 problem decomposition，也就是问题拆解。怎么让数据真实反映现实世界里的分布？怎么保证在不同的类别上都能选到真正的专家？这些完全在你的掌控之内——你能看到这些专家是谁，你可以把筛选标准定得非常严，你可以把问题拆开来，这些都是你说了算的，而且对最后实验结果好不好帮助极大。第二点是整个流程：你怎么判断该把哪个问题路由到哪个解法上。我们还会对这些数据做大量我们叫做 QA 的工作。我觉得有两条路：一是搞清楚一个数据点身上，哪些属性跟「信息量大、信号强」是相关的。比如说具体程度——当你让一个专家判断这个东西好还是不好、让他给出背后的推理，或者围绕他怎么评判一件作品建立一整套观察体系，他语言的具体程度、描述得有多精确，直接决定了这份数据的质量。再比如，你能不能把他的点评跟代码里具体的那一块对应起来。

[13:22]

Like let's say you have an expert that's judging a landing page. Uh they might be able to just like write a paragraph describing this. But we know that models have a tricky time kind of actually connecting the piece of the code to the visual. And so if you can find for example a method to tie that exact code component to the commentary of the expert, suddenly you have data that is way less noisy and way more clear. So these are some things that we can do both in terms of the flow of how you connect the data uh of how you collect the data but also in these like QA checks of how do you determine characteristics about it that will uh correlate highly let's say with with valuable data and same with human QA. I think by the way human QA is tricky here because there's two sides to it. There is a side of for example let's say you're collecting uh a bunch of preference data about slide design. Um you could have human QA be like okay is this again high quality data is the following the specs which most people would agree on but suddenly if you ask for expert consensus and you try to have another expert see if they agree with the initial designers like votes you might start seeing some disagreement there and you I think the key is understanding is that disagreement something that's actually a flaw in the data meaning are they for example disagreeing on something that they should be agreeing on such as alignment like alignment is something that's pretty objective So it would be kind of odd to see experts disagreeing on that front. But suddenly if they're disagreeing on things like sty style or um aesthetics that is not necessarily bad data that's actually good data. It shows you that there is a distinction for what people like. And so uh this almost like analysis of how do you run human QA in a way that both screens for kind of the fundamentals but then when you use consensus I think in an intentional way is another big big piece

比如说，有位专家在评判一个 landing page，他可能就写一段话把它描述一下。但我们知道，模型其实很难把代码的某一块和视觉上的效果真正对应起来。所以如果你能找到一种办法，把专家的这段点评精确绑定到那个代码组件上，你的数据一下子就噪声小得多、也清晰得多了。所以这些是我们能做的事——既包括数据怎么串联、怎么采集的流程，也包括 QA 环节里怎么判断哪些特征跟「有价值的数据」高度相关。人工 QA 也是同理。顺便说一句，人工 QA 在这里挺棘手的，因为它有两面。一面是，比如你在收集一堆关于 slide 设计的偏好数据，人工 QA 可以判断这算不算高质量数据、符不符合规范，这些多数人都能达成一致。但如果你要的是专家共识，你让另一位专家去看他同不同意最初那位设计师的投票，你可能就会开始看到分歧。我觉得关键在于分辨：这个分歧到底是不是数据本身的缺陷？也就是说，他们产生分歧的那个点，是不是本来就该有共识的？比如对齐——对齐这件事是相当客观的，专家在这上面还有分歧就有点奇怪了。但如果他们分歧的是风格、是审美，那这就未必是坏数据，反而是好数据，它恰恰说明人的偏好本来就有分野。所以，怎么设计人工 QA，既能在基本功层面把关，又能有意识地用好共识这件事——这是另一个非常非常重要的部分。

[15:05]

of this. Um and then obviously kind of seeing actually how this data is is interacting with models. Uh we run a lot of research on our side. Obviously when you're interacting with labs as we do um it there's a lot more involved and we oftentimes don't get that feedback loop of exactly what affected this cause which is why we're so adamant on focusing on these things we can control in terms of data quality because these are things that uh we can completely measure on on our side. Um so I think one other uh piece of advice or message of the day is I think especially when it comes to subjective domains I would advocate for a quality over quantity approach. I think creating high quality data is expensive. It's difficult. It takes a lot of understanding and depth around a specific domain. and having that be incredibly high quality done by people that also are incredibly high taste or whatever you want to call it in that domain uh yield far better results than getting a bunch of noisy data or a bunch of messy data uh that was not necessarily intentional or didn't have all those things we talked about of like the problem breakdown. So that is my my message of of the day. Thank you.

然后当然，还要看这些数据实际上跟模型是怎么互动的。我们自己做了大量研究。当然，像我们这样跟各家 lab 打交道的时候，情况就复杂多了，我们往往拿不到那个反馈闭环，不知道到底是什么造成了什么结果——这也正是为什么我们这么执着于盯住数据质量里那些我们能控制的部分，因为这些是我们在自己这一侧就能完整衡量的。所以我今天想给的另一条建议、或者说今天的核心信息是：尤其是在这些主观领域，我主张质量优先于数量。做高质量数据很贵，也很难，需要你对某个具体领域有非常深的理解。但把数据做到极高质量，而且是由那些在这个领域里 taste 极高——你想怎么叫都行——的人来做，效果远远好过堆一大堆嘈杂的、混乱的数据，那些数据往往做得并不用心，也没有我们刚才讲的那些问题拆解。这就是我今天想说的。谢谢大家。