发现图像模型潜力

Language models at the time was all the hype, crazy optimism about language models. 当时，语言模型 (Language Models: 一种能够理解和生成人类语言的人工智能程序) 正是万众瞩目的焦点，人们对它充满狂热的乐观情绪。

We felt similarly about image models. 我们对图像模型 (Image Models: 能够处理、理解和生成图像的人工智能模型) 也有类似的感觉。

Yes, this is a niche place. 是的，这是一个利基市场。

Where could this be going? 它会走向何方呢？

Finding a niche market that is fast growing is the key to startup success. 找到一个快速增长的利基市场是初创公司成功的关键。

After these big models were released is that the number of users 10x, 100x, maybe over a millionx, we saw this change early on. 这些大型模型发布后，用户数量增长了10倍、100倍，甚至可能超过百万倍，我们很早就看到了这种变化。

We have to build systems that are ready for this change. 我们必须构建能够应对这种变化的系统。

极致速度：创业成功的核心

We're fast at everything we do. 我们在所做的一切事情上都追求速度。

We put the models, usually we have day zero releases. 我们部署模型，通常都会在发布日当天就推出。

Space is moving so fast. 这个领域发展得太快了。

We have to be very ahead of getting these models in front of people, making it easy to use. 我们必须非常超前地将这些模型呈现在人们面前，并使其易于使用。

There's a lot of startups out there, you know, they will be stuck on an idea for months and years with no traction. 外面有很多初创公司，它们可能会在一个想法上停滞数月甚至数年，却没有任何进展。

You have to really take that to the extreme. 你必须真正将这种速度推向极致。

I don't think people stress that enough. 我认为人们对此强调得不够。

Think of moving fast. 想想快速行动。

Take that and multiply with 100 and move that fast. 把那个速度乘以100，然后以那样的速度前进。

专注于图像与视频的差异化策略

If we are wrong, we can always revisit our decision. 如果我们错了，我们总可以重新审视我们的决定。

Focusing on image and video is going to be an important differentiator. 专注于图像和视频将是一个重要的差异化因素。

We just raised our Series C round 125 million which values us at a 1.5 billion valuation. 我们刚刚完成了1.25亿美元的C轮融资，公司估值达到15亿美元。

We're very prepared, we're prepared to scale. 我们准备充分，已为规模化扩张做好了准备。

Chat GPT moment for video is that I don't think we've hit it yet. 视频领域的ChatGPT时刻 (ChatGPT Moment: 指一项技术或产品达到普及临界点，引发广泛关注和应用，如同ChatGPT之于大语言模型) 我认为我们尚未达到。

Right now, we employed language models at massive scale. 目前，我们已经大规模应用了语言模型。

We're going to have to do that for AI video, AI image, AI audio, even AI games. 我们将来也必须对AI视频、AI图像、AI音频，甚至AI游戏这样做。

And we want to be the place where all the builders that are building with this technology, we want them to do that through FA. 我们希望成为所有利用这项技术进行构建的开发者们的首选平台，让他们通过FAL来完成工作。

FAL：面向开发者的生成式媒体平台

Burkai: So my name is Burkai and I'm co-founder and CEO at FAL. Burkai: 我叫Burkai，是FAL的联合创始人兼首席执行官。

FAL is a generative media platform (Generative Media Platform: 一种允许开发者利用人工智能模型生成图像、视频、3D内容和音频等媒体的平台) for developers. FAL是一个面向开发者的生成式媒体平台。

We host models that can generate images, videos, 3D audio. 我们托管能够生成图像、视频、3D音频的模型。

Typically these models are very hard to host. 通常，这些模型很难托管。

So the problem we solve is hosting these models as APIs (Application Programming Interface: 应用程序编程接口，允许不同软件组件互相通信), which makes it very easy to consume for developers. 所以我们解决的问题是将这些模型作为API托管，这使得开发者可以非常容易地使用它们。

We also have an inference engine (Inference Engine: 负责运行已训练模型并生成预测或输出的软件系统) that we built inhouse that is specifically optimized to run diffusion models (Diffusion Models: 一类生成模型，通过逐步去除噪声从随机噪声中生成数据) to run two three times better. 我们还拥有内部构建的推理引擎，它经过专门优化，能够让扩散模型运行效率提高两到三倍。

I think latency (Latency: 延迟，指从请求发出到接收到响应之间的时间) kills creativity, latency kills productivity. 我认为延迟扼杀创造力，延迟扼杀生产力。

We work with customers like Adobe, Canva, Shopify, Perplexity. 我们与Adobe、Canva、Shopify、Perplexity等客户合作。

We are at 90 million annualized run rate revenue. 我们目前的年化营收运行率达到9000万美元。

土耳其背景与文化冲击

My co-founder and I have been long-term friends. 我的联合创始人跟我一直是多年的朋友。

We're both from Turkey. 我们都来自土耳其。

I grew up in Turkey. 我在土耳其长大。

I moved to the States for college. 我为了上大学搬到了美国。

There was definitely culture shock. 那确实有文化冲击。

I think even a decade makes a big difference here, right? 我认为即使是十年，在这里也会产生巨大的差异，对吧？

I moved to the States 2007. 我2007年搬到美国。

I think Facebook had just come out. 我想那时Facebook才刚刚推出。

It definitely felt like I wasn't as tapped in to the culture, right? 我肯定感觉自己没有那么融入当地文化，对吧？

So there's a big gap between how I grew up in Turkey and what people like to do there versus how they're in the US. 所以，我在土耳其的成长经历和那里人们喜欢做的事情，与美国人的生活方式之间存在巨大差距。

美国求职市场的挑战

联合创始人: I would say school work felt a little bit easier than I thought because we have a pretty good education system in high school in Turkey especially with maths and sciences. 联合创始人: 我会说学业比我想象的要容易一些，因为土耳其的高中教育体系相当不错，尤其是在数学和科学方面。

The biggest challenge actually was understanding the job market, how people do internships because immediately people start school and prepare for their summer internship and then the next summer and they have a whole plan on how their career is going to happen. 最大的挑战实际上是理解就业市场，以及人们如何进行实习，因为人们一入学就立即开始准备暑期实习，然后是下一个暑假，他们对自己的职业发展有着完整的规划。

I didn't know I should be doing that. 我当时不知道我应该那样做。

So couple years I wasn't really planning my internships towards my career. 所以有几年我没有真正为我的职业规划实习。

So that was a big shock. 所以那是一个巨大的冲击。

移民身份与职业选择：从大公司到创业

Burkai: I actually did an internship at Oracle during college. Burkai: 我大学期间实际上在Oracle实习过。

I was working on some fairly boring things in the beginning to be honest and I had started my green card process. 说实话，一开始我做了一些相当无聊的事情，而且我已经启动了我的绿卡申请流程。

This is a very typical thing for immigrants in the US. 这对于美国的移民来说是非常典型的情况。

You could kind of be stuck in jobs if you start your green card process. 如果你开始申请绿卡，你可能会被困在某些工作中。

Around 2015 was a very interesting time. 2015年左右是一个非常有趣的时期。

深度学习 (Deep Learning: 机器学习的一个子领域，通过多层神经网络学习数据表示) 刚刚开始流行起来。 Deep learning was just kind of starting to become popular.

I started getting really into it. 我开始真正投入其中。

Around that time I had a few other friends at Coinbase and Coinbase was a very small company back then. 大约在那个时候，我在Coinbase有几个朋友，而Coinbase当时还是一家非常小的公司。

It's maybe 40 50 people. 大概只有四五十人。

One of my friends told me, "Hey, we're building a machine learning team." 我的一个朋友告诉我：“嘿，我们正在组建一个机器学习团队。”

And I was like, "Okay, this sounds very interesting. 然后我想：“好吧，这听起来很有趣。

I can go do some deep learning in this new company and there's a lot of things I can learn there." 我可以在这家新公司做一些深度学习，而且在那里我可以学到很多东西。”

But I was mainly excited about starting my own thing. 但我主要还是对创办自己的事业感到兴奋。

I had actually talked to a lot of my founder friends seeing their experience. 我实际上和很多创始人朋友聊过，了解他们的经验。

I had a lot of encouragement from friends to actually go and start my own thing. 我得到了很多朋友的鼓励，真正去开创自己的事业。

自由探索与移民束缚的解除

联合创始人: In the beginning Burka and I rent a house in Palm Springs for a while. 联合创始人: 一开始，我和Burka在棕榈泉租了一段时间的房子。

We were talking about potentially starting a company but we didn't have particular angle or idea to go after. 我们当时讨论着可能创业，但并没有具体的切入点或想法。

So we knew that we want to do this, we would have to go through period of exploration where we find something that we are both passionate about. 所以我们知道，如果想做这件事，就必须经历一个探索期，找到我们都充满热情的东西。

Purai quit maybe 4 months before me and then I joined them. Purai大概比我早四个月辞职，然后我加入了他们。

It is liberating because all my life I also had to deal with immigration work visa and then green card. 这是一种解放，因为我一生中也必须处理移民工作签证，然后是绿卡。

That's one of the reasons actually I stayed at working at a big company. 这实际上是我一直在大公司工作的原因之一。

I wouldn't say that's the only reason but that's definitely a factor. 我不会说那是唯一的原因，但这绝对是一个因素。

And at that time all my immigration process had ended as well. 那时，我的所有移民程序也已结束。

That was also liberating in the sense that I didn't have to work for a big tech company to stay in the country. 从某种意义上说，那也是一种解放，因为我不再需要为一家大型科技公司工作才能留在美国。

I could do whatever I want and I took the opportunity then. 我当时可以做任何我想做的事情，于是我抓住了那个机会。

生成式AI时代：从LLM到图像模型的机遇

Starting with the post-ChatGPT era, it was brand new to everybody. 从后ChatGPT时代开始，这对于每个人来说都是全新的。

It was such a new environment that nobody knew where things are going. 这是一个全新的环境，没有人知道事情会走向何方。

We started running image workloads and we saw a tremendous growth in the companies that are working with us. 我们开始运行图像工作负载，并看到与我们合作的公司取得了巨大的增长。

That made us really excited about the space. 这让我们对这个领域感到非常兴奋。

We also sat down and thought, where could this be going? 我们也坐下来思考，这会走向何方？

Two and a half years ago people saw LLMs (LLM - Large Language Models: 具有数亿到数万亿参数的深度学习模型) and ChatGPT and they sort of drew out where this technology go and they immediately said, "Okay, we're going to AGI (AGI - Artificial General Intelligence: 拥有人类级别智能，可以执行任何人类智力任务的AI)." 两年半前，人们看到了大型语言模型和ChatGPT，他们大致描绘了这项技术的发展方向，并立即表示：“好吧，我们要走向通用人工智能了。”

We felt similarly about image models. 我们对图像模型也有类似的感觉。

We thought as the models get better, there's going to be more capabilities. 我们认为随着模型变得更好，将会有更多的功能。

Quality is going to increase and the resolutions are going to increase and the controllability is going to increase. 质量会提高，分辨率会提高，可控性也会提高。

抓住利基市场与基础设施的机遇

I think finding a niche market that is fast growing is the key to startup success. 我认为找到一个快速增长的利基市场是初创公司成功的关键。

There are a lot of niche markets that stay niche and never grow. 有很多利基市场一直保持利基状态，从未增长。

But we were lucky that market we operated in was very niche and small but also was growing incredibly fast. 但我们很幸运，我们所处的市场非常利基且小众，但增长速度却快得惊人。

What changed after these big models were released is that you didn't have to train it anymore. 这些大型模型发布后，改变的是你不再需要训练它。

You can just pick it off the shelf and start building something useful. 你可以直接拿来即用，开始构建有用的东西。

And that meant the number of users maybe 10xed, 100xed, maybe over a millionx. 这意味着用户数量可能增长了10倍、100倍，甚至超过百万倍。

We saw this change early on and we decided, okay, this changes everything. 我们很早就看到了这种变化，并决定，好吧，这改变了一切。

Now that these models are going to be used by millions of people, we have to build systems that are ready for this change. 既然这些模型将被数百万人使用，我们必须构建能够应对这种变化的系统。

And that's why we decided to build an inference platform (Inference Platform: 专为运行和部署AI模型以进行预测和输出而设计的平台) early on. 这就是为什么我们决定尽早构建一个推理平台。

Another decision we had to make when the revenue was constant for a couple of months. 当收入连续几个月保持不变时，我们必须做出另一个决定。

One tempting thing we could have done run inference for LLM models as well. 一个诱人的选择是，我们也可以为大型语言模型提供推理服务。

Focusing on image and video is going to be an important differentiator. 专注于图像和视频将是一个重要的差异化因素。

We already have a technical advantage because we've been working on this type of models for a while. 我们已经拥有技术优势，因为我们在这类模型上已经工作了一段时间。

If we are wrong, we can always revisit our decision, but it's going to be harder for us to go from general to specific. 如果我们错了，我们总可以重新审视我们的决定，但从通用转向特定对我们来说会更难。

So we tried to stay specific. 所以我们努力保持专注。

I think if you focus on a specific market, you get to work with your users in a closer manner. 我认为如果你专注于一个特定的市场，你就能以更紧密的方式与用户合作。

You understand their problems better. 你更能理解他们的问题。

For us, this was image models and fine-tuning image models. 对我们来说，这就是图像模型和图像模型的微调。

In the beginning, all of our customers were doing very similar things. 一开始，我们所有的客户都在做非常相似的事情。

So we were able to focus on it, get really good at it and differentiate ourselves from others. 所以我们能够专注于此，做得非常出色，并与他人区分开来。

So our ultimate vision is basically we want to be the infrastructure layer for this new technology. 所以我们最终的愿景基本上是，我们希望成为这项新技术的基础设施层 (Infrastructure Layer: 指为技术提供基础服务和资源的底层架构)。

Chat GPT moment for video is that I don't think we've hit it yet. 视频领域的ChatGPT时刻，我认为我们尚未达到。

I think there's a lot of signs like we're getting very close to it. 我认为有很多迹象表明我们离它非常近了。

If you've seen V3 it's close to the chat moment. 如果你看过V3，它已经接近ChatGPT时刻。

It's a very capable model but I think we're still not there yet. 这是一个非常有能力的模型，但我认为我们还没有完全达到那个点。

But interestingly now if you go to your Instagram TikTok feed like third half the videos are AI generated, right? 但有趣的是，现在如果你刷Instagram或TikTok，大约三分之一的视频都是AI生成的，对吧？

It is already happening in a way. 它在某种程度上已经发生了。

It's just happening in a little bit of a slow motion. 只是以一种稍微慢动作的方式发生。

There may be a point this year is that we see even better models that can actually be edited real time and you can interact with the characters that are in the video and generate very interesting content. 今年可能会有一个时间点，我们看到更好的模型，它们可以实时编辑，你可以与视频中的角色互动，并生成非常有趣的内容。

And we want to be the place where all of this infrastructure is being hosted and all the builders that are building with this technology we want them to do that through fall. 我们希望成为所有这些基础设施的托管地，所有利用这项技术进行构建的开发者，我们希望他们通过FAL来完成工作。

AI市场的商业化与快速迭代

I think there are two things happening with AI. 我认为AI领域正在发生两件事。

People are willing to pay but there are questions about the quality of that revenue or how durable that revenue is going to be. 人们愿意付费，但关于这种收入的质量或持久性存在疑问。

I think AI markets are incredible markets. 我认为AI市场是令人难以置信的市场。

生成式媒体 (Generative Media: 指通过人工智能技术生成图像、视频、音频、文本等多种形式的媒体内容) 是其中一个可以立即变现的领域。 Generative media is one of those things where that it can be monetized right away.

In the previous versions of internet businesses, people waited years and years to monetize their business. 在早期版本的互联网业务中，人们等待多年才将他们的业务变现。

First built a user base and then maybe try to monetize with subscription or ads. 首先建立用户基础，然后可能尝试通过订阅或广告来变现。

But with AI, people are willing to pay for it right away. 但对于AI，人们愿意立即为其付费。

The MVP (Minimum Viable Product: 最小可行产品，指具有最少功能但足以满足早期用户需求并进行市场验证的产品) you are building should be good enough for people to start paying. 你正在构建的最小可行产品应该足够好，让人们开始付费。

And it is really easy to get signs. 而且很容易就能得到反馈信号。

The revenue numbers are increasing or not. 收入数字是否在增长。

Now monetization should be something a priority from day zero and it's actually easier for the founder to see if this is a good idea or if this is a good product by the revenue they are making from the first day. 现在，变现应该从第一天起就成为优先事项，而且创始人通过从第一天起就获得的收入，实际上更容易判断这是否是一个好主意或好产品。

We are very particular about what models we want to put because there's a lot of models out there. 我们对要部署的模型非常挑剔，因为市面上有很多模型。

There's a lot of projects out there, research projects, even things that big funded companies that put out there that are cherrypicked (Cherrypicked: 指经过精心挑选，只展示最好结果，而非全面或平均表现的样本). 有很多项目，研究项目，甚至是一些大公司发布的、经过精心挑选的成果。

Basically, you take the results and you look at the good ones and you just use those for your demo or for your launch. 基本上，你拿出结果，只看好的部分，然后用这些来做演示或发布。

So that's called cherrypicking. 所以这叫做“挑肥拣瘦 (Cherrypicking: 指只挑选最好的结果或证据来支持自己的观点，而忽略其他不那么有利的部分)”。

So there's a lot of cherry picking happening in models. 因此，在模型领域有很多挑肥拣瘦的情况。

When we look at the model, first thing we do is we take the model, we run the model and we run bunch of queries to understand is it actually doing the thing that is advertised and then we will go and optimize it and make sure it can run faster and faster especially if there's a lot of demand. 当我们评估模型时，我们做的第一件事是获取模型，运行模型，并执行大量查询以了解它是否真的达到了宣传的效果，然后我们会对其进行优化，确保它能够运行得越来越快，尤其是在需求量很大的情况下。

Developers spend so much time optimizing their iterative loop, right? 开发者花费大量时间优化他们的迭代循环，对吧？

Making sure that once they do something they can see the result, they can see the tests and go iterate. 确保他们一旦做了某事，就能看到结果，看到测试，然后进行迭代。

So no one wants to sit and wait around 5 minutes for a video to generate in the future. 所以未来没有人愿意坐等5分钟来生成一个视频。

This is going to be seconds. 这将是秒级。

It's going to be real time and we're preparing ourselves from infrastructure standpoint for that future. 这将是实时的，我们正在从基础设施的角度为那个未来做准备。

团队文化：小团队、使命感与对AI的热爱

Scaling the company has been one of the most exciting things about this job to be honest. 老实说，公司规模化发展是这份工作中最令人兴奋的事情之一。

We were a very small team like six people for the first two years almost. 在最初的两年里，我们几乎一直是一个非常小的团队，大约六个人。

I think small teams is very important before product-market fit (Product-Market Fit: 指产品在市场上找到了足够多的目标客户并满足了他们的需求). 我认为在达到产品市场契合度之前，小团队非常重要。

You actually do want to have the smallest team that you can, right? 你确实想拥有你能组建的最小团队，对吧？

And experiment and have a small group making decisions and move really fast. 然后进行实验，让一个小团队做出决策，并以非常快的速度前进。

I think this alignment with the company's mission is very important. 我认为与公司使命保持一致非常重要。

This is something people talk about. 这是人们经常谈论的事情。

This is another thing I really learned from Coinbase. 这是我从Coinbase学到的另一件事。

Coinbase early days everyone was a 加密货币爱好者 (Crypto Head: 指对加密货币领域充满热情并深入了解的人). Coinbase早期，每个人都是加密货币爱好者。

You would not find anybody that is not just insanely excited about crypto and that created the foundation for the company that is just so specific and so by default people are just excited about what they're working on. 你找不到任何一个不对加密货币感到极度兴奋的人，这为公司奠定了基础，使其如此独特，以至于人们自然而然地对他们正在做的事情感到兴奋。

But one of my criteria was that I had to love it. 但我的一个标准是，我必须热爱它。

That was super important to me. 那对我来说超级重要。

The intersection of creativity and AI. 创造力与AI的交汇。

I mean, there's unlimited fun there. 我的意思是，那里有无限的乐趣。

At least for me, I wake up every day, I'm very excited about the next models that are released, what this technology where this technology is going, what amazing things are people building. 至少对我来说，我每天醒来都非常兴奋，期待着即将发布的新模型，这项技术将走向何方，人们正在构建哪些令人惊叹的东西。

It is literally the most fun thing I could be doing. 这简直是我能做的最有趣的事情。

If I wasn't doing this, I would probably go play with these models myself. 如果我没有做这个，我可能会自己去玩这些模型。

I love this technology and that's the thing that gives me a lot of drive. 我热爱这项技术，而这正是我获得巨大动力的源泉。