Post-Training Is How You Keep Your Taste | Fireworks CEO Lin Qiao
频道: Sequoia Capital
视频: https://www.youtube.com/watch?v=yAvJ7b_FxUA
原文语言: en
统计: 共 25 轮 · 主持人 3 · Lin Qiao 19 · 提问者 1
[0:00] 主持人
Okay, up next we have Lynn, uh close close friend and collaborator of Brendan's. Um Lynn, uh actually show of hands, who here who here does post training? Okay, good amount of the room. And who here uses Firework? Okay, good amount of the room, too. So, Lynn, you have you have some friendlies in the audience. Um and so for this next talk, what we're going to do is we're going to focus on post training. Um Lynn, si- similar setup to to Brendan's talk. We're going to do 15 minutes on how to approach the problem of post training, um and then 15 minutes or so for Q&A. Uh and Lynn, really delighted to have you here. Thank you for joining us.
好，下面有请 Lin。她是 Brendan 的好朋友，也是长期的合作伙伴。Lin，我先做个小调查——在座有多少人做 post-training（后训练，也就是在现成的基础模型之上，用自己的数据继续训练）？举手看看。好，不少。那有多少人在用 Fireworks？好，也挺多。看来 Lin，台下都是自己人。接下来这一场我们聚焦 post-training。Lin，形式和 Brendan 那场一样：先用 15 分钟讲讲该怎么理解、怎么着手 post-training 这件事，然后留 15 分钟左右做 Q&A。Lin，非常高兴你能来，谢谢你。
[0:37] Lin Qiao
Thanks for having me. Uh hi everyone, good morning. I'm Lynn, I'm CEO and co-founder of Firework. So, as we bring up the slide, today I'm going to do a little bit uh deep dive on post training and why post training could be very relevant to uh you building your own business. So, first of all, a little bit like prior, I think um this year we have seen So, first of all, I'm a Firework weather specializing intelligence platform. There are tons and tons of application built on top of us, all the way from startups to digital native and enterprises. So, we get uh um the fun part of my job is we get to see a lot of patterns. Um what are the innovation people are uh building on top of us and uh and how they are what are challenges that are uh they're facing and what are trend um people being developers are um building on top of us. So, uh one of the things, especially in the past 1 year, software development and application development has been somewhat disrupted because it goes fine in the past, um you have good idea, you want to implement it, scaling production, it requires a team of tens of very strong product engineers, PMs working together, multiple quarters to deliver that. And right now, with one person, a few weeks, uh without understanding how to write a single line of code, you can do that.
谢谢邀请。大家好，早上好。我是 Lin，Fireworks 的 CEO 兼联合创始人。趁 slide 放上来，今天我想稍微深入地聊聊 post-training，以及为什么它可能跟各位正在做的生意高度相关。先说个背景。Fireworks 是一个专注做 inference（推理，也就是把训练好的模型跑起来对外提供服务）的智能平台，上面跑着海量的应用，从初创公司、数字原生企业一直到大型企业都有。所以我这份工作最有意思的地方，就是能看到大量的模式：大家在我们平台上做出了什么样的创新、遇到了什么样的挑战、开发者们整体又在往哪个方向走。其中特别值得说的一件事，尤其是过去这一年——软件开发和应用开发这件事本身，某种程度上已经被颠覆了。放在过去，你有个好想法，想把它实现出来、推到生产规模，需要一个几十人的强产品工程师团队加上 PM 一起干好几个季度才能交付。而今天，一个人、几周时间，甚至不懂怎么写一行代码，就能做出来。
[1:59] Lin Qiao
So, that collapsing of resource required both in terms of timeline and deep expertise is shifting how how competitive the application space is, and it's shifting people from thinking about building on top of off-the-shelf off-the-shelf black box API to build a much deeper mode. So, that they can build a much durable business. As also you have seen lot of discussion, especially in the past 1 week, between open model and the closed model and all the rallying and support across the open model. The depth of that that alliance is because we believe I mean the industry is much deeper because look at the whole entire industry, there are so many companies, right? There are so many company, all of you are building your own company. Every single company exists for a reason because they focus on solving a unique problem in a special way. And that means they carry their own judgment, taste, and determination, conviction into that product, and that's why company exists. Today, if you build on top of off-the-shelf API, then you really need to think about how you keep that special taste judgment and unique part forward. And we believe one approach for every company to build a build durable business is to actually bake your judgment, taste, and customer deep understanding into the intelligence you build on top of instead of just a off-the-shelf API. So, that's kind of a little bit context of what's happening in industry, what we are seeing, and why post training could be very very relevant to you.
所以，无论是时间还是深度专业能力，做一件事所需的资源都在坍缩。这件事正在改变应用层的竞争强度，也在推动大家从「在现成的黑盒 API 之上搭产品」，转向「去挖一条更深的护城河」，好让生意更长久。你们也看到了最近的大量讨论，尤其是过去这一周，关于开源模型和闭源模型之争，以及各方对开源模型的声援。这个同盟之所以扎得这么深，是因为我们相信——整个行业的逻辑其实还要更深一层。你看整个产业里有这么多公司，对吧？这么多公司，在座每位都在做自己的公司。每一家公司的存在都有它的理由，因为它在用一种独特的方式解决一个独特的问题。这就意味着，它把自己的判断力、品味、决断和信念灌进了产品里，公司才因此成立。而今天，如果你只是在现成 API 之上做产品，你就真得想清楚：你独有的那份品味、判断和独特性，要怎么保住、怎么往前推。我们的看法是，每家公司要做出一门持久的生意，有一条路径就是——把你的判断、你的品味、你对客户的深刻理解，真正烘焙进你所依赖的那个 intelligence（智能，也就是模型能力）里，而不是只挂在一个现成 API 上。这大概就是行业正在发生的事、我们看到的情况，以及为什么 post-training 可能跟你们非常相关。
[3:52] Lin Qiao
Okay. So, you probably heard a lot about owning intelligence not rent. And what does owning your own intelligence mean? It actually means many things. So, first of all, it start from data. Intelligence is derivative of data. And obviously the the foundation labs, the all the foundation model we're using are building on top of the public data and the label data that is has solved common tasks, but all of you are solving a specific task. That's why you you're building a business, you're building a company, and be able to curate production data with high quality and even generate synthetic data to enrich your production data is one step of owning your own intelligence. And then after you have data, you will start to kind of use that data turn into a model, build on top of existing model and own the weights. And there are a collection of techniques you can use to get there. Those techniques are tailored to solve different kind of problem you could possibly have. And those techniques can also interoperate with each other for you to build a reach your final goal. And then after you have a great model belong to yourself solving your specific problem really well, and then you need to work on serving it.
好。你多半已经听过很多遍这句话了：要拥有智能，而不是租用智能。那「拥有自己的智能」到底意味着什么？其实包含很多层。第一层，它从数据开始。智能是数据的衍生物。很显然，各大基础模型实验室、我们正在用的所有基础模型，都是建立在公开数据和标注数据之上，解决的是通用任务；但你们每个人要解决的，都是一个具体的任务。这正是你在做这门生意、在开这家公司的原因。所以，能够把高质量的生产数据（production data，真实业务跑出来的数据）梳理沉淀下来，甚至用 synthetic data（合成数据，用模型造出来补充真实数据）去丰富它，这是拥有自己智能的第一步。有了数据之后，你就开始用这些数据把它变成一个模型——在现有模型之上继续训练，并且真正拥有权重（weights）。走到这一步有一整套技术可用，这些技术分别对应你可能遇到的不同类型的问题，而且它们之间还能相互组合、彼此配合，帮你抵达最终目标。等你有了一个属于自己、能把你那个具体问题解得非常好的模型之后，下一步就要解决 serving（部署上线，把模型跑起来对外服务）。
[5:16] Lin Qiao
You probably first will do a some AB testing to make sure it really move the needle for your product metrics, and then goes back in in this loop. Obviously, I don't think you should jump into post training right away. And and there are different phases you will go into. First, prompt everyone start from prompt, use the model as is. It's few shots. If quickly you can use that to test your ideas. And then you use rag to ground um, the usage of AI into your own data and you can start to do a lot of contest engineering from that. Those are from minutes of interaction to hours of interaction. And then you first progress into hey, I have some data. I want to see how my data is going to reflect and make uh, the model work better with my product. So you will start to do supervised fine-tuning that will take you a few hours to um, hey, um, the the model I want to reflect in personalized taste. It is it is very unique uh, choice of my products. So so therefore you want to start to use preferences uh, information where you collect from user interaction and uh, whether thumbs-up, thumbs-down, a lot of those kind of information and uh, help the model learn your product taste. Um, and and finally, you want to build a model towards understand carrying your expertise in that domain. Whether that expertise is across uh, legal, finance, healthcare, customer support, recruiting, marketing, sales, you name it. Even in one industry vertical, there are so many subdomains.
你多半会先做一轮 A/B 测试，确认它真的让产品指标动了，然后再回到这个循环里继续迭代。当然，我并不认为你应该一上来就一头扎进 post-training。这中间是有不同阶段的。第一步，所有人都是从 prompt（提示词）开始，模型原样拿来用，few-shot（给它几个例子）就行。这样你能很快验证自己的想法。接着你用 RAG（检索增强生成，把你自己的资料检索出来喂给模型）把 AI 的使用接到你自己的数据上，从这里开始做大量的上下文工程（context engineering）。这两步的投入是分钟级到小时级的。再往后你会进到下一步：「我手上有一些数据，我想看看这些数据怎么反映进模型里、让模型跟我的产品更贴合。」于是你开始做 supervised fine-tuning（SFT，监督微调，用你整理好的示范数据继续训模型），这大概要花几个小时。再往后是：「我希望模型体现出个性化的品味，这是我产品里非常独特的取舍。」于是你开始用 preference（偏好）信息——那些从用户交互里收集来的、点赞点踩之类的一大堆信号——帮模型学会你产品的品味。最后，你会想把模型练成真正承载你在某个领域里专业能力的样子。不管这个专业是法律、金融、医疗、客服、招聘、市场还是销售，随你说。哪怕在同一个行业垂类里，也还有非常多的细分子领域。
[7:01] Lin Qiao
So all of that is unique and special towards the product you're building. So usually um, you probably heard a lot of reinforcement learning and that is to to actually build towards a specialty. Um, so this progression is very similar to how we human being learn knowledge over time. Uh, for example, uh, we actually run uh, learn a lot of knowledge by reading uh, reading literature, right? In the literature it will say, hey, what is correct, what is it not correct? So this is a very similar to supervised fine-tuning. Um, and uh, as we grow, we develop our own taste and judgment of how we want to conduct the specific way of we want to approach a problem, and that is preference or DPO. Um and over course of time, we learn to be a doing a really good job at certain area, whether um hey, I want to be accounting accountant, and I would really know how to kind of um build into the financial financial data, or I want to be a specific set of um I want to be a dentist, and you learn the details of how to operate um with dentistry. So, all this is very similar exactly um to how we human being acquire knowledges. So, um and also very interesting different techniques are there to solve different kind of problems.
所有这些，都是你正在做的这个产品独有的、特殊的东西。所以通常——你们多半听过很多关于 reinforcement learning（RL，强化学习，让模型在带奖励信号的环境里自己试错、慢慢变强）的讨论——RL 正是用来把模型练成某个方向的专家的。而这个递进过程，其实和我们人类随时间积累知识的方式非常像。比如说，我们靠读书、读文献学到大量知识，对吧？文献会告诉你：什么是对的、什么是不对的。这就非常像 supervised fine-tuning。然后随着我们长大，我们发展出自己的品味和判断——面对一个问题，我倾向于用哪种特定的方式去处理——这就是 preference，也就是 DPO（直接偏好优化）。再往后，随着时间推移，我们在某个领域变得非常擅长：比如「我想当会计师」，那我就得真正搞懂怎么钻进财务数据里；或者「我想当牙医」，那你就得学会牙科操作的所有细节。所有这些，跟我们人类获取知识的过程几乎一模一样。另外还有一点很有意思：不同的技术，是用来解决不同类型的问题的。
[8:24] Lin Qiao
Um for example, if the model doesn't know um the fact, and the fact actually is very dynamic, the facts of your data is uh in in your product is very dynamic, and then typically you use rag to solve that problem. Um however, if your model output all the behavior uh or um or the structure is off, then you you give you curate data and do supervised fine-tuning to correct that. Um if your model's answer, the quality is is personal or kind of is specific to the taste of your product, um and then you use preference and tuning. Or the model is quite weak on the special problem you're trying to solve, then you use RL uh reinforcement learning. And the end result of the model is too slow or too expensive to for you to serve in production, and then you use distillation to um go um let the teacher teach the student much smaller student model, so it can be more much more performing or economical. Obviously, distillation also means different things for LLM, VLM, or uh image generation models. Um if you are interested, we can talk more about that. So, there are many different ways um teams start to hey, I want to kind of try those technology, and um and they may not be happy uh of the experience because they can burn money um and the time, but may not reach their uh ideal results. So, so here are the areas that you can possibly uh feel frustrated. Uh for example, um when it comes to data, and data is the essence of tuning, it the quantity is not the most important. Actually, quality is the most important. But, sometimes just by throwing uh tons and tons of data into a training process may not leading to a great outcome. So, you really need to control the data quality and have And usually, who's the best judge of data quality? It's actually your product team. Um so, that's where we see the convergence of um before GenAI, product team and the research team or ML team, they're separate organizations.
举个例子，如果模型不知道某个事实，而这个事实本身是高度动态的——你产品里的数据事实一直在变——那通常你就用 RAG 来解决这个问题。但如果是你的模型输出的行为方式或者结构不对，那你就该去梳理数据、做 supervised fine-tuning 来纠正它。如果问题出在模型回答的质量上，而这个质量要求很个人化、要贴合你产品的品味，那你就用 preference tuning（偏好微调）。或者，模型在你要解的那个特殊问题上本身就是弱，那你就上 RL，强化学习。而如果最终这个模型在生产环境里跑得太慢或者太贵，那你就用 distillation（蒸馏，让大模型当老师，把能力教给一个小模型），让 teacher 去教 student，把模型做小，跑得更快、成本更划算。当然，distillation 对 LLM、VLM（视觉语言模型）和图像生成模型来说，含义是不一样的。如果大家有兴趣，我们可以再多聊聊这一块。所以，很多团队会说「我想试试这些技术」，但真试了体验未必好——钱和时间都烧掉了，结果却没到理想水平。下面就说几个最容易让人挫败的地方。比如数据。数据是 tuning（微调）的本质所在，而数量并不是最重要的，质量才是最重要的。但有时候，你把海量数据一股脑扔进训练流程，未必就能得到好结果。所以你真的得去把控数据质量。那么，谁是数据质量最好的判官？其实是你的产品团队。这就是我们看到的一个融合趋势：在 GenAI 之前，产品团队和研究团队、ML 团队是两个分开的组织。
[10:42] Lin Qiao
They're separate team. And they work hand-in-hand to make things happen. And a lot of time these days, when um people post-train uh GenAI models, we see kind of the convergence of the product team needs to make judgment call of the data quality and start to be deeply involved in the process. So, uh they can ensure uh the best result. Um and the second is you going to have evals. Um I know all of you are very busy, and uh part into launch quickly, and a lot of evals is a vibe evaling.
它们是各自独立的团队，然后手拉手配合，把事情做成。而现在我们经常看到：当大家去 post-train 一个 GenAI 模型的时候，产品团队必须对数据质量下判断、必须深度介入整个流程，两边开始融合。只有这样，他们才能保证拿到最好的结果。第二件事是 eval（评估，用一套题目和标准去量化模型到底好不好）。我知道各位都非常忙，都急着尽快上线，所以很多评估其实是「凭感觉评」（vibe evaling）。
[11:14]
[laughter]
[laughter]
[11:14] Lin Qiao
And the founders uh really look at the result and feel hey, it is it right or not? Actually, this is judgment. This is you put your judgment of uh of end result and decide whether it's good or not. And that judgment you convert into systemic evaluation. So, this is no different from traditional software development where you have your unit test, integration test to ensure quality. Similarly, if you want to if you think about doing post-training and then have a way to build the eval and build your own judgment into a repeatable process, it's extremely important. Uh and then there could be when you do RL, there could be sloppy RL environment uh where you build a simulation and the simulation is not really reflecting the reality and then the model can heel climb on on kind of the a bad um simulation and I mean it it it will it will also could possibly do reward hacking um and all kind of weird stuff. So, a fun story about reward hacking is um we have been asking a model to uh to generate to do this is coding example uh to create um to to generate code that minimize uh the the error compilation error. Okay. So, guess what the model did? The model generate zero line of code. Okay, there's no compilation error, but that's absolutely not what you want. So, this is a kind of one example of reward hacking, it's very common because model was very smart, it'll try in all different way to uh to get um your goal, but it may not be what you want. So, uh pay attention to all these details and uh and try not to let the model outsmart you.
创始人自己盯着结果看，凭感觉判断「这到底对不对？」其实这就是判断力。你是在把自己对最终结果的判断放进去，来决定它好还是不好。而这份判断，你需要把它转化成系统化的评估。这跟传统软件开发没什么两样——你有单元测试、集成测试来保证质量。同理，如果你打算做 post-training，那么有一套办法把 eval 建起来、把你自己的判断沉淀成一个可重复的流程，就极其重要。再往后，做 RL 的时候，还可能遇到粗糙的 RL environment（RL 环境，让模型试错的模拟场）——你搭了个模拟环境，但它并没有真实反映现实，于是模型就在一个糟糕的模拟上一路爬坡；而且它还可能出现 reward hacking（奖励作弊，模型钻规则的空子去拿高分）之类的各种怪事。关于 reward hacking，有个特别好玩的故事：我们曾经让模型做一个编码任务，要求它生成的代码把编译错误降到最少。结果猜猜模型干了什么？模型生成了零行代码。行吧，确实一个编译错误都没有——但这绝对不是你想要的。这就是 reward hacking 的一个典型例子，而且非常常见，因为模型太聪明了，它会用各种各样的方式去达成你设定的目标，但那未必是你真正想要的。所以，这些细节都要留心，尽量别让模型把你给算计了。
[12:53] Lin Qiao
Um obviously, between um your experimentation, so think about um your development process as hey, you're actually doing a lot of experimentation from training the model to uh training the model is not the end of of your experiment. The judge is the final judge is whether product matches is moving or not. So, you need to bring the final model into into serving tier and do AB testing. Um and that transition is extremely important because um the quality can drop if you move from one training stack to a serving stack without aligning these two. Because think about the model is tons and tons of calculation, math, and the matrix multiplication. Um and the way to do multi matrix multiplication and calculation, if you use different library, different numerics, um and different optimization, it will lead different results. And therefore, the the end result of a training may not be replicated or even uh you lose precision during serving. So, that alignment is very important. I can give you more examples of that. Um and there are a few other um challenges you will run into and happy to talk with you more about in details offline. So, there have been many pioneers uh we worked with post-training. I would say Cursor is one of the few. They have started onboarding getting onto this train from the beginning of last year.
还有一点，在你做实验的过程中——你可以把自己的开发过程理解成：你其实是在做大量的实验，而把模型训出来并不等于实验结束。最终的判官，是产品指标有没有动。所以你需要把最终的模型带进 serving（服务层，也就是线上推理）里去做 A/B 测试。而这个过渡环节极其重要，因为如果你从一套训练栈换到一套服务栈却没有把两边对齐，质量是会掉的。你想想看，模型本质上就是海量的计算、数学和矩阵乘法。而在做矩阵乘法和各种计算时，如果你用的库不同、数值精度（numerics）不同、优化方式不同，得出的结果就会不一样。于是训练出来的结果可能在服务端复现不了，甚至在 serving 阶段丢掉精度。所以这个对齐非常重要，这方面我可以再给你们举更多例子。除此之外，还有另外几个你会踩到的坑，很乐意会后跟大家细聊。在 post-training 这条路上，我们合作过很多先行者，Cursor 算是其中很典型的一家。他们从去年年初就开始上这趟车了。
[14:25] Lin Qiao
Uh there are multiple reasons. One is they really want to control their destiny of the supply of the model. And obviously, um they have a lot of they have a lot of user engagement. They understand um how what is customer preferences, a lot of data. So, that becomes the beginning of their journey. And they do um pretty deep mid-training to post-training. And you have heard them announce um continuously newer models every few month. Um and uh and the result is great. Um their inspiration is to compete at the frontier quality. Uh very bold aspiration. And they're getting there. So, very impressive result from composer two, composer 2.5, um on par with um They're always trying to be on par or beat um the frontier labs quality at the same time. So, um they are kind of one of the examples, a very vibrant example, is completely doable as long as you stay focused and have the right tool and cursive build on top of us. Um there's another there huge huge variety of examples from healthcare. For example, Doximity is one example where they do clinic AI where they basically let doctors ask deep medical questions matching symptoms to medication and side effects and have well-rounded research around everything in medical space.
原因有好几个。一是他们真的想掌控自己在模型供给上的命运。另外很明显，他们有大量的用户互动，非常清楚客户偏好是什么，手上数据也很多。这就成了他们这段旅程的起点。他们做了相当深的 mid-training（中期训练）到 post-training。你们也看到了，他们每隔几个月就持续发布更新的模型，而且结果非常好。他们的志向是在前沿质量上正面竞争——这是个非常大胆的目标——而且他们正在做到。Composer 2、Composer 2.5 的结果都非常亮眼，能跟前沿实验室的质量掰手腕。他们一直在追求要么打平、要么超过前沿实验室的水平。所以他们算是一个例子，一个非常鲜活的例子，证明这条路完全走得通——只要你足够专注、手上有对的工具。而 Cursor 就是构建在我们平台之上的。除此之外，还有大量来自各行各业的例子，比如医疗。Doximity 就是一个，他们做临床 AI，让医生可以问非常深的医学问题——把症状匹配到用药和副作用，围绕医疗领域的各种问题做出很周全的调研。
[15:55] Lin Qiao
They actually also train on Fireworks and they top a very important benchmark which is Stanford Harvard clinic safety benchmark. And we are very proud of that result and they they keep working on this training loop. Um Factory, that's another Sequoia company. They build on top of Fireworks and especially focus on security part of the coding. That is a very hard topic because security is not high tolerance. You need to get it right. They tune a model that also top the kind of benchmarking security area. So you can see all these examples is a demonstration. They solve all the companies are solving a very unique problem in a special way and they have been able to kind of build their own challenges use their own data by building on top of open model. And those are the details. Obviously, as you know, last year 2025 is the year of coding. There Fireworks we support all the coding companies build on top of us. A lot of success in mid-train to post-train very strong models. As you can see those benchmarks are very impressive and all these coding companies are inspired to get on par or beat Frontier Labs in the coding space, but it's not coding. And we start This is the year we start to see very interesting development in all different kind of co-work space from general purpose co-work to specialized domain specific co-work as I mentioned.
他们也是在 Fireworks 上做的训练，并且在一个非常重要的榜单上拿了第一——斯坦福和哈佛的临床安全基准（clinical safety benchmark）。我们为这个结果非常自豪，他们也在持续打磨这个训练闭环。还有 Factory，这也是一家 Sequoia 投的公司。他们构建在 Fireworks 之上，尤其聚焦编码里的安全这一块。这是个很难的题目，因为安全的容错度极低，你必须做对。他们调出来的模型同样在安全领域的基准测试上拿了第一。所以你能看到，所有这些例子都在说明同一件事：这些公司都在用一种独特的方式解决一个独特的问题，而且他们真的做到了——靠自己的数据、在开源模型之上，把自己的难题解掉。这些就是其中的细节。当然，如你们所知，去年、也就是 2025 年，是编码之年。在 Fireworks，我们支撑了所有这些编码公司在我们平台上构建，很多团队在 mid-training 到 post-training 上做出了非常强的模型，非常成功。你们能看到那些基准成绩都很亮眼，这些编码公司都立志要在编码领域追平甚至超越前沿实验室。但今年就不只是编码了。今年我们开始看到，在各种各样的 AI coworker（AI 同事，能独立承担岗位工作的智能体）领域都出现了非常有意思的进展——从通用型的 AI coworker，到我刚才提到的那种垂直领域、领域专精的 AI coworker。
[17:33] Lin Qiao
There are like legal, finance, marketing, recruiting, sales, customer support co-work space. They are all starting to post train and own their own intelligence power their product. Here, Jen Spark is one of the generic co-work application and they build deep research for professionals and slide generation. As you see, this they compete with a frontier model and it's slightly better, but the cost is significantly lower. So here we're talking about five to 10 10 times cost reduction. So as a startup, you can once you hit a problem like if it you can scale quickly into a doable business. And I guess another health care example and health care data or use case is not well captured in in frontier labs model. So you can see the quality of tuned model is significantly better than the state of art closed model. Well, we have many different type of developers build on top of fireworks to do post training. So what what are these types? So we have seen many frontier agent builder. So frontier agent builder, they usually build bespoke customized harness. They don't use common harnesses. And to co-optimize their harness with their systems and all different kind of tools they want to build on top of often time requires post training. Uh so, we have we have seen a lot of repeatable success on doing that.
法务、财务、市场、招聘、销售、客服，这些领域现在都冒出了各自的协同办公产品。它们无一例外都开始做 post-training（后训练，即在通用大模型基础上用自家数据继续训练），把智能能力变成自己的东西，反过来驱动自己的产品。比如 Genspark，它是一款通用型协同办公应用，主打面向专业人士的深度研究和 PPT 生成。你们可以看到，它跟前沿模型打得有来有回，甚至还略胜一筹，但成本低得多——这里说的是五到十倍的成本下降。所以作为一家创业公司，一旦你击中了某个真实的痛点，就能迅速把它放大成一门跑得通的生意。再看一个医疗的例子：医疗数据和医疗场景在前沿实验室的模型里覆盖得并不好，所以你能明显看到，调过的模型质量显著优于当下最好的闭源模型。在 Fireworks 上做 post-training 的开发者有好几类。都是哪几类呢？第一类是前沿 agent（智能体）开发者。这类团队通常会自己搭一套定制的 harness（智能体的运行骨架/调度框架），不用现成的通用方案。而要把这套 harness 跟他们的系统、跟他们想接进来的各种工具协同调优，往往就绕不开 post-training。在这条路上，我们已经见到大量可复制的成功案例。
[19:19] Lin Qiao
And obviously, we have seen a lot of big companies put to post training because interestingly, those incumbents have huge amount of traffic. For them to deploy AI features across the board to all the all their customers is a huge cost burden. Huge cost burden. Uh so, if we say we we need to be careful not scale into bankruptcy, it's not just for startups. Uh it's actually for incumbents. They literally their CFO is blocking their AI feature launch because of the cost. And the post training is way to remediate that. Uh and we have seen a lot of specialized model operator, uh those could be uh new uh new labs and a lot of kind of cutting-edge model develops also um do significant post training. So, uh this is a quote we have heard repeatedly uh that after part of market fit, post training becomes the vehicle to for many, many companies to build specialized intelligence. We do believe we do see the trend that in the future there could be millions of specialized models. One application per use case. It it it it will be millions. That's how uh I believe in that. Um And obviously, there are um we have seen kind of different way to convert the data into into model. Um and the the signal from the reward and how to build those reward function is very important um part of the story.
当然，我们也看到很多大公司在做 post-training。有意思的地方在于，这些老牌巨头的流量体量极其庞大，要把 AI 功能全面铺给所有客户，成本压力是巨大的——真的非常巨大。所以我们常说「别把自己扩张到破产」，这话不只是讲给创业公司听的，对巨头同样成立。他们的 CFO 是真的在因为成本卡着 AI 功能不让上线，而 post-training 正是化解这件事的办法。我们还看到不少专门做特定模型的团队，可能是一些新兴实验室，另外很多走在最前沿的模型开发者，也在大量地做 post-training。有一句话我们反复听到：在跑通 product-market fit 之后，post-training 就成了很多公司打造专属智能的载体。我们确实相信、也确实看到了这个趋势——未来可能会有几百万个专用模型，一个场景对应一个应用，量级真的会到百万级，我是这么相信的。当然，我们也看到了各种把数据转化成模型的不同路径。而 reward（奖励信号，用来告诉模型做得好不好）从哪儿来、reward function（奖励函数）怎么设计，是这个故事里极其关键的一环。
[20:58] Lin Qiao
Um so, our observation is no, start early, start early, and the start to do experimentation a lot more iteratively and they get hands-on experience and we we have seen people on board so quickly. There there's actually not the barrier as many people feel because this this whole reward engineering is very similar to software engineering. The the the logical reasoning and mindset is very very similar. So so yeah so get your hands dirty and and start to kind of test it out. Cool. I think Yeah so those are the kind of key takeaways and uh I think we [snorts] we we want to acknowledge there are different specialties or different knowledges you have right now in this space. For example, we work with the companies like Cursor and Cognition. They have deep researchers. They want to control every single knob as much as possible to get extreme results. We give them the lowest API. That means we just have RL rollout that they can directly interact with and they fully control the trainer. That gives them the best result. But many of the company you have product engineers or machine learning engineers. You have good enough information and knowledge about post training. You want to have some control but you do not want to have the lowest level control. So that's where we have our training SDK iterate very closely with all of you to get to to get you going and we also are happy to kind of deploy our own researchers. That's where we will see, um, help hands-on like training your team to to get on board. So, those are the kind of different way we see different teams want to engage and and get familiar with the whole entire process.
我们的观察是：要早点开始，早点上手，用更快的节奏去做实验、去迭代，把手感练出来。我们见过很多团队上手快得惊人。其实这件事的门槛没有大家想象的那么高，因为整个 reward engineering（奖励工程）跟软件工程非常像，那套逻辑推演和思维方式几乎是一回事。所以别怕弄脏手，先跑起来试试。好，以上就是几个关键结论。同时我们也想承认一点：在这个领域，大家目前掌握的专业度和知识水平是不一样的。比如我们服务的 Cursor、Cognition 这类公司，他们有很资深的研究员，希望把每一个旋钮都攥在自己手里，把效果压到极致。那我们就给他们最底层的 API——我们只提供 RL rollout（强化学习里的采样执行环节），他们直接跟它交互，trainer（训练器）完全由他们自己掌控，这样能拿到最好的结果。但很多公司里坐着的是产品工程师或机器学习工程师，你们对 post-training 已经有足够的了解，想要一定的控制权，但并不想下沉到最底层。这时候就用我们的 training SDK，我们会跟大家高频地一起迭代，把你们带上路。我们也很乐意直接派自己的研究员过去，手把手带你们团队上手。这就是我们看到的、不同团队参与并熟悉这整套流程的几种不同方式。
[23:15] Lin Qiao
Cool. Uh, I'm going to pause here and see if you have any questions.
好，我先讲到这儿，看看大家有没有问题。
[23:19] 主持人
I'll jump um People are most most successful if they have data and reward signal already, maybe from their product. What are the what are the best examples of reward signals you've seen with very successfully trained very quickly?
我先来问一个。最容易做成的，往往是那些手里已经有数据、也已经有 reward signal（奖励信号）的团队，这些东西可能本来就来自他们自己的产品。你见过的最好的 reward signal 案例有哪些？就是那种训练得特别成功、而且特别快的。
[23:33] Lin Qiao
So, um, usually the rewards you can think about rewards actually is code. You write rewards in code. And uh, usually think about rubrics of rewards. Um, and think about uh, you want to grade the result in multiple dimensions. Uh, [snorts] for example, if you are building a recruiting agent, uh, and then you can think about, "Hey, how do we evaluate candidate selection?" Um, and a different company, I guarantee you you have different criterias. Uh, for example, we want to grade a candidate um, um, aptitude. Are they really hungry? They do not take no as an answer. They they will break down walls. And that's one matrix. The second is they're really fast in kind of um, building things and making progress. And and so on. So, then you have a blend of score to blend to merge this. So, so think about reward in different dimensions of rubrics. And then different company have a different way to blend those. And that's that's your um, unique part and secret sauce. Oh, sorry.
你可以这样理解 reward：它本质上就是代码，你是用代码把 reward 写出来的。通常要按 rubrics（评分量表）来想这件事——你要从多个维度去给结果打分。举个例子，如果你在做一个招聘 agent，你就得想：「候选人筛选到底该怎么评？」我敢保证，不同公司的标准完全不一样。比如我们要评一个候选人的资质：他是不是真的有那股劲儿？会不会不接受「不行」这个答案、非要把墙推倒不可？这是一个指标。第二个指标是，他做东西、推进事情的速度够不够快。以此类推。然后你把这些分数按权重揉到一起，合成一个总分。所以，reward 要拆成不同维度的评分量表去想，而每家公司揉合这些维度的方式各不相同——那部分就是你独有的东西，是你的秘方。哦，不好意思。
[24:45] 提问者
Uh, great talk, Lynn. Raj from Vercel. I'm curious about I think you mentioned what are the pitfalls of starting too early, but based on what you're seeing from Vercel, Vercel cognition, or other companies, when do they actually start thinking about post training? Is it when they feel ready? Is it when the cost is now too much? Because the benchmarks where they're beating it, that's a great signal that yes, you can uh um you know, beat the frontier models, but is that the primary motivator? Like that is that 2% bump worth the cost or investment into this whole framework? And and how do they think about ROI? You know, close models are more expensive, but then post training has some intercept of, you know, money and investment and maintenance going forward. So, I'm curious of like, are they approaching when cost becomes a concern or is it when the usage spiking up and they really have to now plan for a year or two in advance?
讲得很好，Lin。我是 Vercel 的 Raj。我想问的是——你前面提到了太早开始会踩什么坑，那反过来，从你看到的 Cursor、Cognition 或者其他公司来看，他们究竟是在什么时候真正开始考虑 post-training 的？是觉得自己准备好了？还是成本已经高到扛不住了？因为跑分能超过前沿模型，这当然是个很好的信号，说明确实打得过；但那真的是主要动机吗？就那 2% 的提升，值得为整套体系投入这么多成本和精力吗？他们又是怎么算 ROI（投入产出比）这笔账的？闭源模型是更贵，但 post-training 本身也有一笔前期投入，加上后续持续的维护成本。所以我很好奇：他们是等到成本变成痛点才动手，还是在用量开始飙升、必须为未来一两年提前做规划的时候动手？
[25:42] Lin Qiao
Yeah, this is this is excellent question. Uh so, usually so, think about uh in the AI building product market fit and the scaling the business as actually two phases. And then in the SaaS time, it's one concept. You hit a product market fit, just scale. I know you got to scale as much as possible. Um but now we see a bifurcation. Product market fit doesn't really mean you have a durable scalable business. And typically, we see uh companies deploy the strategy of focus on product market fit first by building on top of um you know, uh Frontier Labs model because you don't need to worry about anything. So, just kind of spend your money and hit a product market fit. Uh and the other very important thing is only after you hit the product market fit, the data you collect from product surface area are really meaningful. A- and you will get uh also the volume of high quality data start to collect from um product surface area. That became the fuel of uh you start to own your intelligence. So, we um we see like that as kind of the very strong indication um because once you hit upon market fit, you really think about start start to scale a business, you think about two things. One is continue to keep your competitive edge.
这个问题非常好。可以这样想：在 AI 时代，「打磨出 product-market fit」和「把生意规模化」其实是两个阶段。而在 SaaS 时代，这两件事是一回事——你一旦跑通 PMF，接下来就是尽可能放大规模。但现在我们看到了分叉：跑通 PMF 并不意味着你就有了一门可持续、能放大的生意。我们通常看到的策略是：先专心把 PMF 跑通，这个阶段就直接搭在前沿实验室的模型上，因为那样你什么都不用操心，花钱就完事，先把 PMF 打出来。另一个非常关键的点是：只有在跑通 PMF 之后，你从产品触点上收集到的数据才真正有意义，而且那时候你才会开始积累起有规模、高质量的数据。这些数据就成了你「拥有自己的智能」的燃料。所以我们把这当成一个非常强的信号。因为一旦跑通 PMF、真正开始考虑把生意做大，你就要想两件事：第一，怎么持续保住你的竞争优势。
[26:59] Lin Qiao
Two is um build a build durable business, so your um your revenue and your costs are, you know, in a healthy state. Um so, then um post-training becomes a very appealing solution because post-training allow you to um basic encode codify um your uh unique taste into into a model that no one can uh steal from. Because it's very easy to clone and copy application as is, as I as you all know, right? Uh from coding agent it's very easy. Uh so, from screenshot, boom, generate the same or even better app. Uh so, that's a very scary that application itself is kind of the a moat is is being um reduced. Um and you bake your data uh where reflecting the product engagement and all the deep knowledge into your model is a way to preserve that. And second is you can post-train a model that bring down the cost five to time 10 times. And and then that means you can support five to 10 times much higher traffic with the same budget. And the your unit of economics of scaling is so much better than you avoid scaling into bankruptcy. So, we see kind of those as as two compelling story behind the reason and timing.
第二，怎么建立一门可持续的生意，让收入和成本处在一个健康的关系里。这时候 post-training 就成了一个非常有吸引力的解法。因为 post-training 能让你把独有的品味固化、编码进模型里，别人偷不走。你们都知道，照着克隆一个应用太容易了——拿 coding agent，一张截图丢进去，「啪」，就生成一个一模一样、甚至更好的应用。这件事挺吓人的，意味着应用本身作为护城河正在被不断削薄。而把你的数据——那些反映产品真实使用情况、承载深层 know-how 的数据——烘进模型里，才是把护城河守住的办法。第二点是，post-training 能把成本压下来五到十倍。这意味着同样的预算，你能撑起五到十倍的流量。规模化时的单位经济模型好太多了，你也就避开了「扩张到破产」。所以这两点，就是我们看到的、决定动手理由和时机的两个有说服力的逻辑。
[28:22] 主持人
Thank you, Lin.
谢谢你，Lin。
[28:23] Lin Qiao
Thanks a lot.
非常感谢。
[28:24]
[applause]
［掌声］