为何定制 AI 模型将取代基础模型——对话 Fireworks AI 创始人 Lin Qiao

软件开发已经被商品化了。它无法更快了。所以最终结果是应用本身没有很深的护城河。什么无法被复制？代表用户偏好、用户意图的数据。如果我们相信智能源于数据，那么我相信智能的真正前沿尚未被解锁。

Software development has been commoditized. It cannot be faster. So the end result is application itself doesn't have deep moat. What cannot be copied? The data representing user preference, user intention. If we believe that intelligence is derived from data, then I believe the real frontier of the intelligence hasn't been unlocked yet.

欢迎来到 Prompted，一档关于 AI、人物和创意火花的播客。我是 Cameron Adams，Camber 的联合创始人兼首席产品官。你可能最近经常听到“推理”这个词。它是将 AI 带给世界的重要部分，但不是我们很多人都需要思考的。今天的嘉宾经常思考它。我请到了 Lynn Chow。她围绕推理建立了一家价值 40 亿美元的公司。它叫 Fireworks AI，为 Cursor、Uber、DoorDash 和 Shopify 等公司提供推理和基础设施，帮助他们构建、微调和扩展高性能的生成式 AI 应用。在创立 Fireworks 之前，Lynn 是 Meta 的 PyTorch 联合创始人和负责人，她在开发世界上最广泛使用的机器学习框架之一中发挥了关键作用。所以，可以说她近十年来一直在帮助构建 AI 的未来。Lynn，欢迎来到节目。

Welcome to Prompted, a podcast about AI, people, and the creative spark. I'm Cameron Adams, co-founder and chief product officer of Camber. You might have heard the word inference a lot lately. It's an important part of bringing AI to the world, but it's not one that a lot of us have to think about. Today's guest thinks about it a lot. I'm joined by Lynn Chow. She's built a $4 billion company around inference. It's called Fireworks AI and it provides inference and infrastructure to companies like Cursor, Uber, Door Dash, and Shopify in order to help them build, fine-tune, and scale high-performance generative AI applications. Before founding Fireworks, Lynn was the co-creator and head of PyTorch at Meta, where she played a pivotal role in developing one of the world's most widely used machine learning frameworks. So, you could say that she's been helping to build the future of AI for almost a decade. Lynn, welcome to the show.

谢谢邀请。

Thanks for having me.

我非常激动能请到你，聊聊你在 Fireworks 构建的所有东西。但首先，让我们从“提示”环节开始。在这个环节，你要一边画一边告诉我们一个塑造了你的时刻的故事。我们会把那张画交给我们的内部设计师 Harriet 和一点 AI，把它变成神奇的东西。你准备好笔和纸了吗？

I'm so excited to have you here and to talk about all the things that you're building at Fireworks. But first, let's kick off with the prompt. The prompt is where you tell us the story of a moment that shaped you while you draw it. We're going to take that drawing and we'll turn it into something magical with our in-house designer Harriet and a little bit of AI. Do you have a pen and paper there ready?

是的。

Yes.

开始吧。你能同时 multitask 讲故事吗，还是需要一些画画的时间？

Go for it. Can you multitask and tell the story at the same time or do you need some drawing time?

我得说画画很难。我们接受你画的任何东西。

I will say it's hard to draw. We will take whatever you've got.

哦，对此我认为我最好的方式有点抽象。所以，我会解释。就像那样。

Oh, for that I think the best way I have is it's a little bit abstract. So, and I will explain. It's like that.

抽象很好。好的。我看到一个正方形，有一些箭头从里面出来。

Abstract's good. All right. I'm seeing a square with some arrows coming out of it.

是的。

Yes.

那么，告诉我们这幅草图是关于什么的。

So, tell us what this sketch is all about.

所以，对我来说，创办这家公司是自我驱动成长的延续。我从小就一直在寻求这种成长。幸运的是，我在 Facebook 遇到了一群很棒的人，他们把我带到了一个我无法想象的高度，我会说我在 Facebook 的成长超出了我自己的预期。不是头衔什么的，而是接触到最前沿的东西，并能够实现我最初为自己设定的更大目标，这非常令人满足。

So, to me, starting this company is a continuation of self-driven growth. I have been constantly seeking that from very young. Fortunately, I met a great group of people at Facebook and they drive me to a different altitude that I wouldn't imagine myself and I would say my growth at Facebook was exceeding my own expectation. Not in terms of title or anything just in terms of the exposure to the most cutting edge and be able to achieve much bigger goal that I set for myself initially is extremely satisfying.

我听到的是，你的草图是关于推动自己超越边界。这是关于尝试新事物。这是关于成长。我喜欢这种对你个人哲学的洞察以及它如何塑造了你的生活。我们会把那张画发给 Harriet，我们的驻场设计师，她将与 AI 合作把它变成新的东西。我们将在节目结束时揭晓 Lynn、Harriet 和 AI 的创作。所以请继续关注。但在 Harriet 工作的时候，我想和你更深入地聊聊你的 AI 之旅，Lynn。你已经告诉我们一些关于你职业生涯的事情，但我认为关键时刻之一是在 Meta，你共同创造了 PyTorch，我首先很感兴趣的是，对于不熟悉它的人来说，PyTorch 是什么，以及你在开始开发它时试图解决什么问题。

What I'm hearing is that your sketch is all about pushing yourself beyond boundaries. It's about trying new things. It's about growth. I love that insight into your personal philosophy and how it's shaped your life. We're going to send that picture to Harriet, our designer in residence, who's going to team up with AI to whip it into something new. We're going to reveal Lynn, Harriet, and AI's creation at the end of the episode. So stay tuned for that. But while Harriet works, I'd love to go a bit deeper with you in your journey through AI, Lynn. So, you've already told us a bit about your career, but I think one of the pivotal moments came at Meta where you co-created PyTorch and I'm really interested in firstly what PyTorch is for anyone that's not familiar with it and what problem you were trying to solve when you started developing it.

我想提一个注意事项：PyTorch 是由 FAIR 的一组研究人员创建的。FAIR 是 Facebook AI Research，它最初被设计为一个工具，使研究在构建新的神经网络时更加高效。这是一个出色的设计，那时我从数据领域转向 AI，目标不仅是让 PyTorch 用于研究，而是让 PyTorch 同时用于研究和生产。这是一个全新的设计空间，我创建了一个团队并将其实现，赋能 Meta 的所有内部工作以及外部的研究和生产工作。

One caveat I want to mention is PyTorch was created by a group of researchers at FAIR. FAIR was Facebook AI Research and they were designed originally as a tool to make the research much more productive in making new neural networks. It's a brilliant design that's when I kind of move from data into AI and the goal here is to have not just PyTorch for research but PyTorch for both research and production. That's a completely new design space and I created a team and bring that into fruition empower all Meta's internal work and also external research and production work.

能用于研究的东西和能用于生产的东西有什么区别？

What's the difference between something that can work for research and something that can work in production?

非常不同，对吧？因为研究关乎创造力和创造力的速度，而生产关乎优化。所以这实际上是非常不同的设计空间。我们最初并不清楚这两个设计之间是否有任何重叠。我们可以有一个快速迭代的研究系统过渡到一个高度优化的生产系统，因为这里我们谈论的是生产中的巨大规模。所以但从研究系统到生产的过渡极其残酷。它很慢，把一切砍掉，更重要的是，在纯研究和我们进入生产之前有一个中间阶段，就是实验。实验意味着我们需要在半生产环境中测试那些研究想法，而这个半生产环境实际上是生产，因为它是生产中的一小部分流量来测试这个 AB 测试，所以研究、实验和生产之间的界限非常模糊。然后我们基本上必须解决这个交叉问题。没有别的办法。但这是一个以前没有人解决过的新挑战。我们接受了那个挑战，我说今天的结果是我们工作的验证，PyTorch 现在是整个行业主导的 AI 框架，特别是在 GenAI 方面。大多数模型是在 PyTorch 中创建、在 PyTorch 中训练，并在生产中的 PyTorch 中部署进行推理。

Very different, right? So because research about creativity and the velocity of creativity and production is about optimization. So this is actually very true different design space. And it was not clear to us initially whether there's any overlap between these two designs. We could have a fast iterative research system transition into a hyper optimized production system because here we are talking about huge scale in production here. So but this transition between a research system to production is extremely brutal. It's slow, the hack of everything down and more importantly there is intermediate stage between pure research to and before we hit production is experimentation. Experimentation means we need to test those research idea in a semi-production setting and this semi-production setting is actually production as in it's a small fraction of the traffic in production to test this AB testing so then the line between research experimentation and production is really blurry. And then we basically have to solve this intersection problem. There's no other way around. But that is a new challenge that no one has solved before. We took that challenge and I was like today's result is the validation of our work is PyTorch is now the dominating AI framework across industry especially when it comes to GenAI. Majority of models are created in PyTorch, trained in PyTorch and deployed to inference in production in PyTorch.

所以像 GPT 或 Claude 这样的东西是建立在 PyTorch 上的吗？

So something like GPT or Claude is that built on PyTorch?

我很确定。有趣的是，我相信大约七八年前，我们把 OpenAI 从使用 TensorFlow 转向使用 PyTorch。从那时起，所有主要研究实验室开始全面使用 PyTorch，这成为了漏斗顶部，因为当时仍有许多研究团队将模型开源到社区，这成为了漏斗顶部，让生产团队测试它、进行实验并带到产品表面。所以这就是它自然流动的方式。顺便说一句，我在那段时间遇到了我的其他联合创始人，我们非常紧密地合作，要么在构建支持大规模生产训练和推理的 AI 基础设施方面，要么他们是使用我们构建的 AI 基础设施的客户。所以在那段时间认识联合创始人，我会说是一种极其，我非常幸运。

I'm pretty sure. So fun fact I believe it's probably seven eight years ago we flipped OpenAI from using TensorFlow to use PyTorch. Ever since then all the major research lab start to use all in on PyTorch and that became the top of funnel because there still many research team at that time still open source the model into the community and that becomes the top of funnel to have the production team test it and doing experimentation and bring to product surface areas. So that's how kind of it flow naturally. By the way, I met my other co-founders during that time and we work very closely together either on the building the AI infra side that powers both massive production training and inference or they are customer using the AI infrastructure we build. So like knowing the co-founders during that time is I would say kind of is extremely I'm extremely fortunate.

你开始做 PyTorch 的时候，大语言模型已经出现了吗？还是说它们是稍晚一些才发展起来的？

Were large language models on the scene at the time when you started PyTorch, or were they a development that came a bit later?

两者有不少重叠，但当时还没起飞。我认为 2022 年 11 月是 ChatGPT 时刻，而我们在 2022 年 9 月就创办了公司，在 ChatGPT 之前。所以至今仍有一个巨大的疑问：你为什么会在那时候创业？太早了，时机还没到，你要解决一大堆数据问题。其实也许我该倒回去讲讲我们的历程——它其实非常有意思。我们有七位联合创始人，创始团队非常大。我们大约在同一时期加入 Meta，那正是 Meta 完成移动优先、或者说移动优先逐渐收尾的时候。这是一次巨大的转型，但也是一次精彩的转型，消费者可以在手机上随时随地访问社交网络。产品增长很好，然后从产品表面产生了大量数据，而这些数据成了 AI 的未来。所以移动优先、AI 优先——这个顺序非常重要，因为这个顺序也预示了后来行业转型的顺序。我们刚开始的时候，没有 AI 团队，没有 AI 硬件，一切都在 CPU 上跑着极小极小的机器学习算法。没有 AI 软件——PyTorch 是给研究用的。我们有一些零散的软件，比如移动端的 Caffe2、生产端的 ONNX，但非常碎片化。所以我们从那个点出发，引入了各种不同的 AI 硬件，包括 Meta 自己的 AI 硬件以及 GPU，然后搭建起训练平台，搭建起推理平台，把所有东西统一到 PyTorch 上，让 PyTorch 不仅用于长远的研究，也用于生产训练和生产推理。这不是一段轻松的旅程。我们经历了很多折腾和反复，但今天我们对结果很满意。

There's quite some overlap, but it didn't take off. I think November 2022 was the ChatGPT moment, and we started a company in September 2022, before ChatGPT. So there's still a huge question: why would you start a company at that time? It's way too early, it hasn't arrived, you're going to solve a lot of data problems. Actually, maybe I'll rewind and talk a little bit about our journey — it's actually very interesting. We have seven co-founders, a very big founding team. We joined Meta around a similar time frame, that's when Meta finished mobile-first or was winding down mobile-first. It's a big transition, but a brilliant transition, where consumers can access social networks anywhere, anytime on a mobile phone. Great product growth, and then from the product surface they have a lot of data generated, and that data became the future of AI. So mobile-first, AI-first — that sequence is very important, because that sequence also foreshadowed the sequence of this transition in the industry later on. And when we started, there was no AI team, there was no AI hardware, everything was on CPU with tiny, tiny machine learning algorithms. There was no AI software — PyTorch was for research. We had some fragmented software, for example Caffe2 for mobile, ONNX for production, but it was very fragmented. So we started from that point and brought in various kinds of AI hardware, including Meta's own AI hardware as well as GPUs, and then brought up the training platform, brought up the inference platform, everything consolidated on PyTorch, making PyTorch not just for far-reaching research but also for production training and production inference. It was not an easy journey. We did a lot of churn and arrow, but we're happy with the result today.

对我们一些不太懂技术的听众，你能不能解释一下训练和推理的区别？训练可能有些人比较熟悉，比如训练 GPT 或 Claude 这样的大模型；而推理正是 Fireworks 主要专注的方向。

For some of our less technical listeners, can you explain the difference between training, which a few people might be familiar with in terms of training something big like GPT or Claude, versus inference, which is a lot of what Fireworks focuses on?

训练是模型学习的一种方式。从根本上说——目前我们主要谈的是神经网络——神经网络学习的方式基本上是：先做一次前向传播，然后产生一个损失，再从损失出发用反向传播往回推，沿途改变权重，然后再前向传播一次，再往回推一次。所以这种来回往复就是训练的本质，它天生就非常依赖数据，因为要从零开始学习就需要海量数据。但训练也有不同的阶段：预训练、中期训练、后训练，使用不同的算法。所以这就是它拿一个大数据集，观察这些数据，找出其中的联系，并从中学习以达成某个目标，对吧？

So training is a way for a model to learn. Fundamentally — currently we're mostly talking about neural networks — the way for neural networks to learn is basically: it has a forward pass and then generates a loss, and then from the loss it uses the backward pass to push back and change the weights along the way, and then go forward, pass again, and then push back again. So this back-and-forth is the essence of training, and it's by nature very data-heavy, because to learn from scratch it needs a huge amount of data. But also training has different phases: pre-training to mid-training to post-training, using different kinds of algorithms. So this is where it takes a big dataset, looks at that data, finds connections between it, and learns from that to achieve something, right?

所以首先，在训练之前你需要设计一个有很多层的模型架构，而学习过程就是把权重填充进这个模型架构。这些权重是某种数值。最终结果是模型架构加上各种不同的权重，然后你把它部署到生产环境。接着你的产品可以输入信息——提示词就是输入——然后前向传播生成预测结果，再返回给产品。这就是推理。所以推理只有前向传播，而训练是多次前向加反向，两者在设计上是根本不同的系统。我会说训练复杂得多，但它有一条共同的路径。其实训练的一部分就是推理——你可以这么理解——正因如此，拥有一个让训练和推理完全对齐的平台极其重要。对齐直接影响质量，而不对齐会导致质量下降和精度损失。所以 Fireworks 把训练和推理作为一个平台提供，来支撑各种应用。这是比较高层面的说法，但我们也经历了算法设计的各个阶段。

So first of all, before you train you need to design a model architecture with many layers, and the learning process is to populate weights into this model architecture. Those weights are kind of numerical values. And the end result is a model architecture combined with various different weights, and then you deploy that into production. And then your product can pump in information — the prompt is the input — and then go forward to generate the projected result and then push back to the product. So that's inference. So inference is forward pass only, and training is forward and backward multiple times, and it's a fundamentally different system to design. Training is much more complicated, I would say, but it has a common path. Actually, part of training is inference — you can think that way — and that's where having a platform that runs both training and inference fully aligned is extremely important. Alignment directly leads into quality, and misalignment leads to lower quality and loss of precision. So that's where Fireworks provides both training and inference as one platform to power applications. That's kind of high level, but we've also gone through phases of algorithm design.

我认为最大的区别在于前 AGI 时代和后 AGI 时代。前 AGI 时代没有基础模型的概念。每个团队都必须自己整理数据、从零开始训练。后 AGI 时代有了基础模型的概念，它从公开互联网、标注公司的标注数据以及大量合成生成的数据中学习，从而创造出通用智能，而这种通用智能——我们听过很多说法，说 AGI 将解决许多问题，并作为黑盒 API 被使用。所以这是一个巨大的变化，因为任何公司要用 AI，都不再需要先建立一个庞大的 AI 团队从零训练；他们可以基于现成的生成模型来构建，或者组建一个更小的团队，在扎实的基础上用自己的数据做后训练。但后训练比从零训练整个领域要容易得多。所以这是对使用 AI 的生产力的根本性解锁。这就是为什么从 2023 年开始，AI 的采用开始指数级增长，尤其是从去年年底开始。我认为我们到了一个拐点，因为模型质量，无论闭源还是开源，都越过了阈值，我们日常中很多原本需要投入人力的工作，都可以由模型自动化。这就是为什么我们看到生成式 AI 的采用现在正指数级增长。

I think the biggest difference is pre-AGI and post-AGI. Pre-AGI there's no foundation model concept. Every team has to curate data and train from scratch. Post-AGI there's a concept of a foundation model that is learned from the public internet and labeling companies' labeled data and a bunch of synthetically generated data to create general intelligence, and that general intelligence — we've heard a lot of terms about AGI going to solve many problems and be used as a black-box API. So that is a massive change, because for any company to use AI they no longer need to first build a huge AI team who can train from scratch; they can build on top of an off-the-shelf gen model, or they can staff up a smaller team to post-train the model based on a solid base but use their own data. But post-training is a much easier problem than training from scratch the whole entire space. So that is a fundamental unlock of productivity using AI. That's why, starting from 2023, the adoption of AI started to go exponential, especially starting from the end of last year. I think we hit an inflection point because model quality, both closed and open, has passed a threshold where a lot of our day-to-day problems, where we were putting in human effort, can be automated by models. And that's where we see the adoption of GenAI going exponentially right now.

所以对世界上大多数人来说，如果你不是研究者，你与 AI 交互的方式大概就是通过推理层。如果你在用 AI 助手，你的提示词被处理、然后结果返回给你的方式，就是通过推理。对吗？

So for most of the world, if you're not a researcher, the way that you interact with AI is probably through the inference layer. If you're using an AI assistant, the way that your prompt gets processed and then a result returned to you is through inference. Correct?

没错。

That's right.

那你认为推理改进最重要的方面会出现在哪里？如果推理变得更好，它究竟会怎样影响我这样一个 AI 助手或模型的终端用户？

So where do you see the most important aspects of inference improvement happening? How will it actually affect me as an end user of an AI assistant or a model if inference gets better?

我真的认为推理最重要的要素是前沿质量。这最终是开发者构建产品时最大的价值所在。任何产品市场契合度都取决于产品带给消费者或专业消费者的质量，而这种质量决定了他们使用的模型必须是前沿的。所以有意思的部分是如何获得前沿质量的模型，有两条路径。一条成熟的路径是使用 AGI 模型——来自前沿实验室的模型——你专注于上下文工程、提示词工程，从这些模型中得到结果，然后你的产品在此基础上构建。

So I really think the most important element of inference is front quality. That's ultimately the biggest value to a developer building a product. Any product-market fit is hinged on the quality that product brings to a consumer or prosumer, and that quality dictates that the model they're using needs to be frontier. So the interesting part is how to get a frontier-quality model, and there are two paths. One established path is to use the AGI model — the model coming out of the frontier labs — and you focus on context engineering, prompt engineering, and get a result out of those models, and then your product builds on top of that.

不过，过去一年里，行业在宏观层面发生了一件很有意思的事：软件开发已经被商品化了，把一个应用想法转化为实现、再扩展到生产环境，速度不可能更快了。过去这需要几十名优秀工程师组成的团队、花上好几个季度才能做到，而现在只需要一名工程师，或者一个人——不一定是工程师——一个人，几周甚至几天就能交付。所以最终结果是，应用本身并没有很深的护城河。因为很多人只要有想法就能做出应用，他们甚至可能一行代码都不会写。而且应用非常容易被复制，你只要截几张图，让 Claude Code 生成一个类似的就行。那么真正的护城河是什么？什么无法被复制？是代表用户偏好、用户意图、用户参与度的数据，以及最重要的——激活这些数据的算力。

However, in the past one year, an interesting thing happened at the macro level in the industry: software development has been commoditized to convert an application idea into implementation, to scale in production. It cannot be faster. In the past it required a team of tens of strong engineers and multiple quarters to do that, and now it's one engineer, or one person — not necessarily an engineer — one person, multiple weeks or even days, and they can deliver that. So the end result is the application itself doesn't have a deep moat. Because a lot of people can create applications as long as they have ideas. They may not even know how to write a single line of code. And an application is very easy to copy, because you can just take screenshots and ask Claude Code to generate a similar one. So then what is the real moat? What cannot be copied? The data that represents user preference, user intention, user engagement, and the most important compute to activate that data. And if you look at the world's data, actually 90 to 95% of data is private data like this, locked inside applications, locked inside enterprises.

而且如果你看全球的数据，实际上 90% 到 95% 都是这样的私有数据，锁在应用里，锁在企业里。

And if you look at the world's data, actually 90 to 95% of data is private data like this, locked inside applications, locked inside enterprises.

是的。

Yeah.

而这些数据并没有被激活，没有被充分激活。如果我们相信智能源自数据，那么我认为智能真正的前沿还没有被解锁，因为前沿实验室拿不到这些数据，而数据所有者又没有技术在这些数据上构建模型。所以这就是 Fireworks 的切入点。我们的使命是充分激活私有数据，并构建面向具体应用的前沿模型。所以这非常简单：私有数据加开放模型，等于你自己的前沿模型。我们希望每个人都能成为自己的前沿实验室。路径就是使用 Fireworks 的训练和推理平台，结合你的数据，然后你就能从中得到最好的模型。

And it's not being activated. It's not being activated fully. If we believe that intelligence is derived from data, then I believe the real frontier of intelligence hasn't been unlocked yet, because frontier labs don't have access to that data. And the data owner doesn't have the technology to build a model on this data. So that's where Fireworks comes in. Our mission is to fully activate private data and build application-specific frontier models. So it's very, very easy: private data plus open model equals your own frontier model. So we want everyone to be their own frontier lab. And the path is to use the Fireworks training and inference platform combined with your data, and then you will get the best model out of it.

更重要的是，这其实不是一次性的事情。你提到了一堆客户。并不是说 Uber 训练一次模型、训练出一个 Uber 模型，然后就永远好用了；也不是说 Shopify 训练一个 Shopify 模型就永远好用了。这必须持续进行，因为应用在不断演进，基础模型也在不断进步——每两周就有新模型冲上排行榜榜首——这就像软件开发一样，需要持续 rebase。所以这种训练会一直进行下去。但同时，我们生活在智能体的时代，这个过程应该尽可能简单、尽可能自动化。

More importantly, it's actually not a one-time thing. You mentioned a bunch of customers. It's not like Uber would train this model once and train an Uber model and then they will be good forever, or Shopify will train one Shopify model and it would be good forever. This has to continuously happen, because applications keep evolving, base models keep improving — every two weeks there's a new model popping up on top of the leaderboard — and it's like software development, continuous rebase. And that's why this training will keep going. But at the same time, as we live in the age of agents, this process should be as easy to do as possible, as automated as possible.

所以我们看到，要达到前沿质量有两条策略，而前沿质量是打造产品最重要的因素，我相信这两条策略会共存。但我不认为世界会被少数几个基础模型主导。我相信未来会有数百万个定制模型，每个应用、每个用例一个。

So then we see, hey, there are two strategies to get to frontier quality, which is the most important factor to build a product, and I believe these two strategies will coexist. But I do not believe the world will be dominated by a few foundation models. I believe the future will be millions of customized models, one per application, per use case.

那真是个很有意思的未来。你提到应用已经商品化了，这在很大程度上是因为现在编码智能体的能力，它们能把相当宽泛的指令变成任何人的现实。你在更深的层面也看到同样的情况吗？你看到编码智能体影响你正在做的工作，让它更快、更迅速，甚至让 AI 系统几乎能自我进化吗？

That's a really interesting future. You mentioned that applications have become commoditized, and that's largely because of the power of coding agents now and the ability for them to take pretty broad instructions and turn that into anyone's reality. Are you seeing the same thing at a deeper layer? Are you seeing coding agents affect the work that you're doing and make that faster and rapid, and enabling AI systems to evolve themselves almost?

绝对有。我觉得有意思的地方在于——我坚信 AI 的速度——我开始看到不只是软件工程师在用 AI 工具，这理所当然，而是我的会计主管，她也在用 AI 工具，她还用 Fireworks 来管理账目、管理电子表格。你定义各种技能，让她的团队可以重复使用，这省下了好几个 Python 工程师，她的生产力大大提升。而且不只是财务，它横跨不同领域。我们构建了自己的招聘智能体，因为我们需要招很多优秀的人，而收到的申请太多了，所以必须自动化。我认为人才质量对我们非常重要，而筛选所有这些申请工作量很大。我们还有数据团队、RevOps，他们在用 AI 工具构建各种仪表盘，这样我们能更快获得可见性，而不用等很久让流水线跑完。我开始看到它进入我们职业生活的许多不同方面，这强烈表明模型质量已经跨过了一个门槛。当模型质量跨过门槛，定制、使用数据、把模型质量拉到新高度就变得容易得多。所以我们坚信方向就是定制模型这个正确方向，如今 Fireworks 上 95% 的流量都来自定制模型。

Absolutely. I think what makes this interesting, and I'm a firm believer in the velocity of AI, is I start to see not just software engineers adopting AI tools — which is a no-brainer — but my head of accounting, she used AI tools, she also used Fireworks to manage your books, right, to manage spreadsheets. You define all kinds of skills that her team can repeatably use, and that saved multiple Python engineers, and then she just can be so much more productive. And it's not just finance. It cuts across different areas. We build our own recruiting agent because we need to hire a lot of great people and there's a lot of inbound applications — it's just too many — so we have to automate this. I think talent quality is very important for us. It's a lot of work to scan through all these applications. We also have our data team, RevOps, they are using AI tools to build various kinds of dashboards, just so we have more visibility quickly instead of waiting for a long time for the pipeline to finish. I start to see it start to go into many different aspects of our professional life, and that is a strong indication that the model quality has passed a threshold. When the model quality passes a threshold, it becomes much easier to customize, use the data, and pull the model quality to a new altitude. So that's where we firmly believe the direction is the right direction of a customized model, and today 95% of traffic on Fireworks is through customized models.

你之前提到 Fireworks 有七位联合创始人，我觉得这在 AI 创业公司里似乎是一种模式，它们往往比以前的创业公司有更多联合创始人。你看到这种模式在发生吗？背后有原因吗？

You mentioned earlier that there are seven co-founders at Fireworks, and this seems to be a pattern I think in AI startups, that they tend to have a few more co-founders than previous startups might have. Is that a pattern that you see happening, and is there a reason for that?

我不确定这是不是一种模式，我们只是恰好如此。这肯定有很多好处。我觉得如果团队动态真的很好，拥有一个更大的创始团队——我们在讨论中非常坦诚——能帮助我们获得非常全面的视角，去思考产品、思考业务、在哪里下注、在哪里聚焦、在哪里减少投入。我认为这些视角极其重要。但同时，我们从不陷入分析瘫痪；决策速度与创始团队多大无关。你知道，公司变大后总会有更多人。决策速度、判断力、判断的立场和判断的质量都非常非常重要。所以当我们扩展到创始团队之外时，我们非常注重人才密度和人员质量。我们也很幸运，有许多优秀的人加入创始团队之外，把公司建设成今天的样子，我们也会继续招聘优秀人才。

I'm not sure if it's a pattern. We just happen to be that way. And there's definitely a lot of benefits. I feel if the team dynamics is really good, having a larger founding team — we're very candid in discussions — and it helps us get a very comprehensive perspective in thinking about product, thinking about business, where to make bets, where to focus, and where to drop the weights. And I think those perspectives are extremely important. But at the same time, we never get analysis paralysis; decision-making velocity doesn't matter how big the founding team is. You know, when the company grows bigger we'll always have more people. Decision-making velocity, the judgment, the position of the judgment and quality of judgment is very, very important. And that's where, when we grow beyond the founding team, we pay a lot of attention to talent density and the quality of the people. And we're also fortunate that we have many great people joining us beyond the founding team and building the company towards today's shape, and we continue to hire great people into the company.

你开始看到 AI 在决策本身中扮演更主动的角色，而不只是当分析师给你一些数据吗？它真的开始改变你的决策了吗？

Are you starting to see AI play a more active role in decision-making itself rather than just being an analyst and giving you some data? Is it actually starting to change your decisions?

还没到那个程度。我认为很多决策判断仍然由团队里的每一位成员来做。但我确实认为 AI 正在接管大量琐碎的工作，因为它非常擅长遵循指令——只要指令足够清晰，它就会去做。

Not to that point yet. I think a lot of decision judgment is still being done by every single team member in the team. But I do think AI is taking away a lot of mundane work, because it's very good at following instructions — if the instruction is clear enough, it just goes off and does the work.

这就把我们解放出来，去专注于更重要的战略思考、创造力和实验。我认为这正是人之所以为人的地方：探索、创造、人际互动。所以我觉得它给了我们更多时间去做那些让我们成为人的事。我认为这很重要，因为否则我们一直忙于日复一日的重复性工作，如果这些能被 AI 取代，那我们专注于真正重要的事，严格来说只会更好。

And that frees us up to focus on more important strategic thinking, creativity, and experimentation. I think that's what makes us human: exploration, creativity, interpersonal interaction. So I think it gives us more time to do what makes us human. And I think that's important, because otherwise we've been very busy day-to-day doing repeatable work, and if those can be replaced by AI, it's just strictly better for us to focus on the right things.

这正是这档播客的主题。谢谢你这段精彩的话。当我们谈到优化推理、让它尽可能做到最好时，数据中心在其中扮演什么角色？你们是直接和数据中心合作，把推理优化到极致吗？

That is exactly what this podcast is about. So thank you for that amazing quote. When we're talking about optimizing inference, making it the best it can be, how do data centers come into the picture? Are you working directly with data centers to make inference super optimized?

我们直接与具体的硬件打交道。今天我们支持八种不同的英伟达 SKU 和三种不同的 AMD SKU，而且随着硬件不断向前发展，我们计划增加硬件的多样性。作为这波 AI 浪潮的一部分，硬件开发的速度也是前所未有的。以前每三年才有一代新硬件，现在一个厂商每年就有一到三代硬件。

We directly interact with specific hardware. Today we support eight different Nvidia SKUs and three different AMD SKUs, and we plan to increase the diversity of the hardware as the hardware keeps moving forward. Hardware development is at an unprecedented pace as part of this AI wave as well. Before, every three years there was a new hardware generation, and now every year there are one to three hardware generations from one vendor.

太不可思议了，对吧？所以这就形成了一种非常有趣的动态：模型的贬值速度不可能更快了，因为每两周就有新模型冒出来。而硬件是每四个月、六个月就有新硬件冒出来。有趣的是，更新的模型在新硬件上总能跑出更好的性价比。

It's incredible, right? So then it's a very interesting dynamic that model depreciation cannot be faster, because every two weeks there's a new model popping up. And the hardware is every four months, six months there's new hardware popping up. And the interesting thing is the newer model can always run more price-performance better on newer hardware.

我们始终想把最好的带给客户。所以我们的策略是速度优先。速度优先。而这个速度是客户的速度。我们希望他们能用上最好的模型，拥有最好的质量、最快的速度和最高的效率。这就是我们一直全神贯注的最大价值所在。

We always want to bring the best to our customer. So our strategy is velocity first. Velocity first. And that velocity is our customer's velocity. We want them to get onto the best model at the best quality, best speed, and the best efficiency. So that is the biggest value-add we have been laser focused on.

抱歉。随着越来越多公司部署 AI，成本、性能、效率正成为真正重大的关切，尤其是成本。你今天看到这些权衡是如何显现的？人们又是如何应对的？

Sorry. As more companies deploy AI, cost, performance, efficiency, they're becoming really major concerns. Costs in particular. How are you seeing these trade-offs showing up today? And how are people managing them?

我认为成本确实已经成了问题。有意思的是，在这个时代，产品市场契合和可持续的生意是两个不同的概念，对吧？你可以有很好的产品市场契合，也就是客户喜欢你的产品、愿意付费，但你无法扩张，你卡着闸门，因为一旦扩张你就会把钱烧光。根本原因在于，建立在 GPU 上的 AI 基础设施，其单位经济模型与 CPU 有着根本不同。而效率是个真问题。在我们今天所处的供给受限的世界里，效率是更大的问题，因为那是机会成本。如果你能用同样的硬件支撑 10 倍的 token，就意味着你能获取 10 倍的客户，覆盖 10 倍大的市场，而所有人都在供给受限，这对很多公司就成了真问题。如果你效率不高，你就会把市场份额输给竞争对手。这不只是从营收角度看，而是市场份额，对吧？而且它会复利式累积。

I think cost has really become a problem. It's interesting that in this time, product-market fit and a durable business are two separate concepts, right? You can have a great product-market fit, as in the customer loves your product and they want to pay for it, but you cannot scale, and you're holding the gate because once you scale you're going to run out of money. The fundamental reason is that the AI infrastructure built on GPU has unit economics fundamentally different from CPU. And efficiency is a real problem. Efficiency is an even bigger problem in a supply-constrained world, which we live in today, because it's opportunity cost. If you can power 10 times more tokens with the same hardware, that means you can acquire 10 times more customers, you can cover a 10 times bigger market, while everyone is supply-constrained, and that becomes a real problem for many companies. If you're not efficient, you're losing market share to your competitors. It's not just from a revenue point of view, it's market share, right? And that compounds.

所以我们在效率上创造了很多价值，但更重要的是，我们和这么多客户合作，尽管他们首要关心的是效率和成本，质量也必须达标。质量的门槛是前沿实验室级别的质量，对吧？最终一切都会归结到这一点。所以我不认为以牺牲质量为代价去解决效率问题是个办法。我们想在不降质量的前提下解决效率问题，同时把控制权交给应用开发者，让他们不把自己的命运押在一个黑盒 API 上，他们需要在自己正在成长的市场中建立竞争优势和差异化。而最好的方式就是拥有自己的定制模型。那个定制模型将拥有前沿的质量、前沿的速度，以及最高效的运行时。这就是我们为客户创造的价值。

So we add a lot of value in efficiency, but more importantly we work with so many customers even though their first concern is efficiency and cost, quality has to meet the bar. The quality bar is frontier-lab quality, right? At the end it always gets distilled down to that. So I don't think solving the efficiency problem at the cost of quality is a solution. We want to solve the efficiency problem by holding the same quality while in addition giving control to application developers, so that they do not put their destiny in a black-box API, and they need to create a competitive edge and differentiation in the market they're growing into. And the best way is to own their own customized model. That customized model will have frontier quality, frontier speed, and the most efficient runtime. So that's what we add value to for our customers.

我们目前看到的模式是，公司发布高质量的模型，给人们更好的体验，但随后他们直接把成本转嫁给客户。所以如果你想用这个高质量模型，你就要消耗更多 token，我们就要收你更多钱，你需要订阅更大的套餐。你认为这会永远是 AI 的经济学吗，还是你能看到这种关系正在发生变化？

The current pattern that we're seeing is that companies are releasing high-quality models and giving people a better experience, but they then just directly pass the cost on to the customer. So if you want to use this high-quality model, you're going to eat more tokens, we're going to charge you more, you need to subscribe to this bigger package. Do you think that will be the economics of AI forever more, or can you see a change happening in that relationship?

是的，我认为转嫁成本是一回事，但没有建立护城河是更大的问题，因为如果你建立在一个你很容易被替换的基础上，那不是个好位置，不是一门可持续的生意。打造产品最重要的作用是建立一门可持续的生意。而可持续的生意往往意味着你必须完全掌控你依赖关系中的关键部分。我确实认为，从模型中衍生出的智能是产品依赖关系中的关键部分，它应该回到产品团队手中。而把代表产品的数据转化为产品专属、应用专属的智能，正是正确的方向。

Yeah, I think passing on cost is one thing, but also not building a moat is a bigger problem, because if you're building on top of a foundation that you're just easily replaceable on, it's not a good position to be in. It's not a durable business to be in. The most important role of building a product is to build a durable business. And that durable business often means you have to have full control of the critical part of your dependency. And I do think the intelligence derived from the model is a critical part of a product dependency that it should be returned to the hands of the product team. And that's where turning the data that is representative of a product into product-specific, application-specific intelligence is the way to go.

你处在驱动 AI 的技术的核心。所以我确信你对如何使用它、如何让每一天更高效有独到的见解。我很想知道 Lin 在日常生活中是怎么用 AI 的。你用它做什么，它真正改变了你工作方式的哪些方面？

You're at the heart of the technology that drives AI. So I'm sure you have insights into how to use it and how to make your day better. I would love to know how Lin uses AI in her day-to-day. What do you use it for, and what has it really changed in the way that you work?

我的日常包含很多事情，但我经常做大量的市场分析。所以我确实觉得 AI 工具很擅长做深度研究。

So my day-to-day consists of many things, but oftentimes I do a lot of market analysis. So I do feel like AI tools are pretty good at doing deep research.

是的。

Yeah.

所以它让我的工作轻松多了，我专注于分析我从各种 AI 工具收集回来的数据。我们建立的公司价值观或文化之一就是全面拥抱 AI。对每一个人来说，无论他们的主要工作是什么，拥抱 AI 工具都非常重要。而这显著地——这不是关于刷 token，不是，这真的是关于速度。真的是关于在这个动态环境中的决策速度和产品开发速度。速度压倒一切。我们是工具领域的一部分，提供一个平台，让应用拥有自己的定制模型和智能。我们自己也在深入使用 AI 工具，让自己高速、高效。

So it makes my work much easier and I focus on analyzing the data I collect back from various AI tools. One of the company values or culture we establish is fully embrace AI. It's very important for every single person, no matter what their main job is, to embrace AI tools. And that significantly — it's not about token maxing, no, it's really about the velocity. It's really about decision-making and product development velocity in this dynamic environment. Velocity trumps everything. We're part of a tool space providing a platform for applications to own their own customized model and intelligence. We ourselves also deeply work on using AI tools to make ourselves high-velocity and highly productive.

你提到了深度市场研究，它非常擅长广泛地铺开，然后再收回来为你综合整理。

You mentioned deep market research, and it is fantastic at going broadly out and then bringing it back and synthesizing it for you.

但你怎么验证它带回来的信息是正确的，或者它从大型数据集中得出的结论是正确的？

But how do you verify that what it's bringing back is correct or that the conclusions that it reaches from a large data set are correct?

是的。验证的方法有很多种，通常我会用多个来源。我会从不同角度做研究，然后尝试匹配。如果不匹配，那就不是一致的数据。看待同一件事有不同的方式。从视角上，你需要匹配。另一种方法是向多个 AI 工具问同一个问题，看看它们是否给出一致的答案。但通常我会从不同角度提问，而这正是做研究的一部分，对吧？就是从一个问题的不同视角去看，看看整体图景是否一致。

Yeah. So there are various different ways to do verification, and usually I have multiple sources. I will do research from different angles and then try to match it. If it doesn't match, then it's not consistent data. There are different ways to look at the same thing. And in perspective, you need to match. And the other way is to ask multiple AI tools the same question and see if they give consistent answers. But usually I will ask from a different angle, and that's part of doing research, right? It's to look at different perspectives of a question and see if the picture is consistent or not.

最后一个问题。如果我们从现在回望 5 年后，你希望 Fireworks AI 在塑造 AI 的未来中扮演了什么角色？

Very last question. And if we look back 5 years from now, what role do you hope that Fireworks AI has played in shaping the future of AI?

我认为我们很早就押注在推理上。我觉得这是对的。我不会改变我们的定位。我认为我们对前沿推理、前沿质量推理的关注是正确的焦点。所以实际上在这段旅程中我不会改变任何东西。我们的使命、我们的梦想是——这里稍微讲讲 Fireworks 这个名字的由来，对吧？这个名字从何而来，因为它是一个有趣、多彩、好玩的名字。

I think we bet on inference early on. I think that's right. I wouldn't change our positioning. I think our focus on frontier inference, frontier quality inference is the right focus. So actually I would not change anything in this journey. Our mission, our dream is—here's a little bit about the Fireworks name, right? Where the name comes from, because it's a funny, colorful, playful name.

是的，我喜欢这个名字。

Yeah, I love the name.

它来自我们 PyTorch 的起源。我们的工作——PyTorch——而火炬承载着火。我们希望 AI 之火无处不在。所以我们称自己为 Fireworks。我认为我们非常忠于我们一开始设定的使命，我们正在构建前沿质量的推理来为每一个应用提供动力。所以一路上我们取得了很多进展，我对我们下一阶段要构建的东西非常兴奋。

It came from our origin of PyTorch. Our working—PyTorch—and torch carries fire. We want the fire of AI to be everywhere. So that's why we call ourselves Fireworks. And I think we are very faithful to the mission that we set ourselves on at the beginning, and we are building frontier quality inference to power every single application. So we have made a lot of progress along the way, and I'm very excited about what we build into the next phase.

我现在想换个话题，很想了解一下你的全明星创意团队。这是一个由四人组成的团队，由你来选择。任何时代的人都可以。他们可以是已故的，可以是活着的，也可以是来自 500 年后的未来，但我很想知道你的全明星创意团队会有谁。

I'd love to change gears now and I would love to get some insight into your all-star creative team. This is a team of four people that you get to choose. Anyone from any era. They can be dead, they can be alive, they can be someone from 500 years in the future, but I'd love to know who you would have on your all-star creative team.

实际上我会选我的联合创始人。

So actually I will pick my co-founders.

哦，选四个。你有七个。你打算选哪四个？

Oh, do four of them. You got seven of them. Which four are you going to choose?

我会选联合创始人，还有 George，刚加入我们担任总裁的 George。他们是我共事过的最聪明的人，思维的范围非常广，从详细的实现，无论是运营，到高层次的战略。而这个战略不仅仅是关于软件、软件、硬件，各种不同层次的应用。能够和这些人一起工作，这就是我的梦想，我说我的梦想在这段旅程中成真了，能够和他们一起工作。我不会改变。是的，我会改变。如果有的话，我可能会——如果我从这段旅程一开始就有 George，那会更有趣。

I will pick the co-founders and also George, George who just joined us as our president. They are the most intelligent people I've ever worked with and the range of thinking is so wide, all the way from detailed implementation, whether operation, to high altitude strategy. And this strategy is not just about software, software, hardware, various different layers of application. And to be able to work with those people together is just—it was my dream and I say my dream has come true in this journey to be able to work with them. I wouldn't change. Yeah, I will change it. If anything, I'll probably have—if I have, you know, George from the beginning of this journey, that would be even more fun.

好吧，我希望你的团队在听，因为他们应该为你和他们一起取得的成就感到无比自豪。我以为你会在那里挑选一些你最喜欢的联合创始人，那会让你陷入非常棘手的境地。

Well, I hope your team was listening because they should be incredibly proud of what you've done with them. I thought you're going to have to pick some of your favorite co-founders there, which would have put you in a really sticky position.

我们别这样。我喜欢他们每一个人。

Let's don't do it. I like every single one of them.

好的，我想 Harriet 已经有足够的时间把你的原始草图输入 AI，并与之合作，想出一种新的创意格式。所以，让我们把她请回来，看看这个人-AI-人创意三角的最终结果。Harriet，你在吗？

Okay, I think Harriet has had enough time to feed your original sketch into AI and work with it a bit to come up with a new creative format. So, let's bring her back to see the final results of this human-AI-human creative triangle. Harriet, you there?

你好。你好。是的，我在。我一直在忙，积分用完了，但我还是设法拼凑出了一些东西。你实际上必须升级才能获得更多积分吗？

Hello. Hello. Yes, I am. And I've been rushing around running out of credits, but I've managed to pull something together. Did you actually have to upgrade to get more credits?

不，我只是换了工具。嗯，但我这里有你原始草图的全部笔记。我试图处理的东西相当多，但我已经有了一些简单和象征性风格的灵感。所以，我就先做了第一个。我觉得那一刻对我来说真的很有启发。我觉得很酷。然后我试着从你的笔记中做出一个故事弧线，把它变成几个不同的幕，我在几个不同的场景中发展了这个图像，我会给你看最终作品，然后边讲边解释。它有点科幻、有点怪异的氛围。所以，我尝试对故事进行抽象的重述，从你年轻时经常搬家的最初部分开始。改变你的视角，连接所有的点，转向 STEM，结合你的学术背景。拥抱所有这些，搬到美国，进入研究，从 LinkedIn 到 Facebook，谈论你倾听客户。嗯，然后我只是没时间把它拉成一个漂亮的循环，但最后这种完整的循环图像，有点像在说你现在已经到达了现在的位置。

No, I just changed tool. Um, but I've got my notes all here from your original sketch. Quite a broad amount of things I was trying to deal with, but I already had some inspiration of simplistic and symbolic type of style. So, I just went for my first one. I thought that's really inspiring for me at the moment. I thought it was cool. And then I have tried to make a bit of a story arc out of your notes and turning it into a few different acts and I have evolved that image across a few different scenes and I will show you the final piece and then kind of explain as I go. It's a little bit sci-fi ticky kind of eerie vibes. So, I've tried to do an abstract retelling of the story from the first parts of moving around a lot when you were younger. Changing your perspectives, connecting all the dots, turning towards STEM with your background with academia. Kind of embracing all of that, moving to the US and moving into research and moving from LinkedIn to Facebook, talking about you listening to customers. Um and then I've just kind of ran out of time to pull it into a nice loop, but kind of this full circle image at the end kind of saying you where you you've reached where you are now.

嗯，即使你没时间了，它看起来仍然很棒，Harriet。我喜欢你如何把最初的概念草图变成了它自己独特的风格，然后把它做成了一段令人难以置信的视频。你觉得呢，Lin？

Well, it still looks amazing even though you ran out of time, Harriet. And I love how you took that initial concept sketch, turned it into its own unique style, and then made it into an incredible video. What do you think, Lin?

不，这是我们每天越来越多看到的创造力。我喜欢你在这么短的时间内整合出的叙事，以及通过这部电影表现出来的——代表出来，显然我们甚至没有提前想过。所以，这就是我们未来会看到的随着 AI 工具不断增强的创造力。我超级——我坚信这一点。我喜欢你刚刚给我看的视频。如果你能发给我，那就太好了。

No, that's a creativity we're seeing more and more every day. And I love the storytelling you pulled together in such a short time and the representation—represent that through this movie and obviously we didn't even think ahead of time. So, that's the increasing amount of creativity we're going to see into the future with AI power tool. And I'm super—I'm a firm believer there. I love the video you just showed me. If you can send to me, that'd be great.

当然。谢谢。

Sure. Thank you.

音频听众，如果你想看 Harriet 创作的内容，请前往节目笔记，我们会有一个视频链接供你观看。谢谢你，Harriet。谢谢你，Lin。这是一次关于 Fireworks AI 推理以及当今将 AI 带给人们的所有深度技术的迷人对话。感谢你深入探讨并给了我们一些非常有趣的答案。

Audio listeners, if you want to see what Harriet created, just head over to the show notes and we'll have a link to the video which you can watch. Thank you, Harriet. Thank you, Lin. That was a fascinating conversation about Fireworks AI inference and all the deep tech that goes into bringing AI to people nowadays. Thanks for diving deep and for giving us some really interesting answers.

谢谢。谢谢大家。

Thank you. Thanks everyone.

感谢听众们加入我们这一期的 Prompted，一个关于 AI、人和创意火花的播客。要跟上未来的节目，请在你喜欢收听播客的任何地方订阅，或者前往 YouTube 观看所有创意荣耀。如果你喜欢这个节目，请给我们留下评分或评论。我们会阅读每一条，我们总是喜欢听到你的想法。下次见，继续追逐那创意的火花。

And thank you listeners for joining us on this episode of Prompted, a podcast about AI, people and the creative spark. To keep up with future episodes, please subscribe wherever you love to listen to podcasts or head on over to YouTube to see all the creative glory. And if you're enjoying the show, please leave us a rating or review. We read every single one and we always love hearing your thoughts. Until next time, keep chasing that creative spark.