将‘幻灯片光标’业务扩展至5000万美元年收入：Gamma创始人Jon Noronha

我们所有人开始新项目时，都会面对这个写着‘输入演示标题’的空白页面。天啊，我该从何入手？我得规划所有幻灯片，构思故事线，还要想好开场如何吸引人。

We all start some new project with this blank page saying enter presentation title. And it's like, my god, where do I start? I have to outline all the slides. I gotta figure out my story. What's the hook?

关键节点是什么？最终呈现效果如何？该用什么字体？什么配色？需要哪些配图？

What are the key moments? What it's gonna look like? What fonts am I gonna use? What colors? What imagery do I need?

然后你就陷入这种反复纠结的漩涡。完全没错。当我们真正开发出解决这个空白页问题的AI时，发现它其实解决了所有人的这个难题。突然间，人们可以从模糊的想法直接得到完整草稿，他们的工作变成了编辑优化，而非从零开始。

And you get stuck in the spiral of, like, thinking through all these things. Totally. And so when we actually built AI that solved our blank page problem, it turned out that it solved everybody else's blank page problem too. Suddenly, they could just go from a vague idea to a fully worked out rough draft, and suddenly their job was editing, not starting from scratch.

今天我们邀请到Gamma创始人John Narona，这款幻灯片制作工具堪称AI应用层成功的典范。图像生成模型是其最初的产品突破点，如今这些AI红利已推动Gamma实现5000万美元以上年度经常性收入，用户超5000万，并实现正向现金流。John将分享关于多模态协调、大规模AB测试、微调与提示工程对比、成本效益等实用经验，并探讨Gamma如何解决冷启动问题，以及未来向更智能编辑体验发展的路线图。

Today we're joined by John Narona, founder of Gamma, our cursor for slides. Gamma is a canonical example of application layer success in AI. Image generation models were the initial product breakthrough, and these AI tailwinds have now carried Gamma to 50,000,000 plus in ARR, 50 plus million users, and cash flow positive. John discusses practical lessons for builders on multimodal orchestration, running extensive AB tests, fine tuning versus prompt engineering, cost efficiency, and more. He also shares thoughts on how Gamma is solving the cold start problem and their future road map towards more agentic editing experiences.

请享受本期节目。John，非常感谢你今天做客《训练数据》。

Enjoy the show. John, thank you so much for joining us on Training Data today.

很荣幸受邀，谢谢邀请。

Great to be here. Thanks for having me.

让我们从基础开始：Gamma是什么？你当初为什么

Let's start with what is Gamma? Why'd you

创立这家公司？我们在2020年创立公司时，最初的愿景至今未变。我们想重新定义演示文档。我的联合创始人Grant在2020年提出这个想法时，描述他整天坐在公园长椅上参加电话会议——因为他妻子占用了家里唯一的办公桌。那时正值疫情最严重时期，大家都被困在家中。

start the company? So we started the company back in 2020, and the original vision has not really changed up till now. We wanted to reinvent presentations. My cofounder Grant came to me with the idea back in 2020, and he was talking about how he was just spending all of his time sitting outside on a park bench trying to call into conference calls because his wife was using the one spot in their house with a desk. This was the depths of COVID when we were all in our houses.

他通过小手机屏幕参加各种会议，艰难地跟着Zoom上播放的PPT。就在他

And he was joining all these conference calls, staring at his little tiny phone, trying to follow along to these PowerPoint decks people were doing on Zoom. The moment he

这个小小的恍惚带来了多少改变？

How brought many this in little dazed?

我知道。我当时就想，这也是我全部的工作内容。我整天就是坐在Zoom会议里看Google幻灯片，偶尔中途可能还要制作一些Google幻灯片。这两件事对我来说都不太有趣。所以他一提到这个，我就觉得，天哪，这真是个重新定义这个事物的绝佳机会——一方面，幻灯片是商业世界的通用语言。

I know. I was like, that's all I do, too. All I do is just sit on Zoom calls all day and look at Google Slides decks, alternating with maybe making Google Slides decks along the way. Neither one's very fun for me. And so as soon as he mentioned it, I was like, man, this is such an exciting opportunity to reinvent this thing, which on the one hand, slides are the language of business.

它们是我们工作中沟通的主要方式，尤其是在讨论重要想法时。比如对你来说，每次有人向你提案肯定都是用幻灯片。但如果你问几乎任何行业的人：你喜欢做幻灯片吗？这是你想要的沟通方式吗？通常都会得到某种厌恶反应，比如'唉，我讨厌做这个'。

They are the way that we all communicate at work, especially around high stakes ideas. Like for you, anytime somebody pitches you, I'm sure, is with a slide deck. And yet, if you ask anybody in almost any walk of life, do you enjoy making slides? Is this how you want to communicate? You generally get some kind of ick reaction, like a, oh, I hate having to do that.

我特别特别讨厌幻灯片让我看起来很糟糕。我讨厌把90%的时间花在排版上，只有10%用在内容上。我讨厌大家都在评判我的幻灯片外观而不是我说的内容。Gamma的愿景正是源于这个几乎普遍存在的问题——人们需要进行视觉化沟通。

I I hate hate how slides make me look bad. I hate how I spend 90% of the time on formatting and just 10% on the content. I hate how everyone's judging me for how they look and not what I'm saying. And so Gamma's vision came out of this almost universal problem. People need to do visual communication.

他们都想呈现得很好，但现有工具无法满足需求。所以我们想重新定义这个领域，不仅仅是做一个新的PPT编辑器，而是从根本上重新思考这种形式本身。

They wanna look great, but current tools don't cut it. And so we wanted to reinvent that, not just by making a new editor for making PowerPoints, but for actually rethinking the format itself.

这太酷了。顺便说一句，我就是那种可耻的、实际上喜欢做幻灯片的人，看来我现在意识到自己是个怪胎了。

That's so cool. By the way, I'm one of those shameful people that actually likes making slides, so I guess I've now realized that I'm a weirdo.

没错，你是那1%的异类。

Yeah, you are the 1%.

恭喜你们。Gamma如今是AI应用层的一个惊人成功案例。我记得你们宣布ARR已突破5000万美元，现金流为正，团队只有30人。真是了不起的成就。

Congratulations. Gamma today is an amazing success story in the AI application layer. I think you guys announced you're at 50,000,000 plus of ARR, cash flow positive team of 30 people. So congratulations. Thank you.

这一路走来都这么顺利吗？

Did it always feel like this?

不，完全不是。我们花了很长时间才走到今天。没有什么一夜成功的故事。就像我说的，我们在2020年创立公司时，所有这些AI热潮都还没兴起。

No. Absolutely not. It's been it's been a long road to get here. There are no overnight successes. As I said, we started the company in 2020 before all this AI stuff had taken off.

那你们创立公司时就知道AI浪潮会到来吗？

And do you know the AI stuff was going to happen when you started the company?

绝对不是。我其实记得试用过GPT-3的第一个版本，那也是在2020年推出的。我可能是最早获得试用权限的一批人之一。我们测试时，我就想看看它能不能制作幻灯片？

Absolutely not. I actually remember trying the first version of GPT-three, which actually also came out in 2020. And I was one of the first people, I think, to get access. And we tried it out. And I just wanted to see, could this make slides?

因为我们已经在做相关工作了。我想验证：如果给它一份文档，它能总结出关键要点吗？或者反过来，能把要点扩展成段落文字吗？当时的答案是否定的，它还没达到那个水平。

Because we were already working on it. I wanted to see, if I gave it a document, could it summarize the key bullet points? Could it do the opposite, take bullet points and turn it into paragraphs of text? And the answer was no at the time. It just wasn't there yet.

所以我当时觉得这技术挺酷，但没太在意。之后几年我们走了完全不同的路线，专注未来办公理念，研究远程异步协作，试图打造比幻灯片更互动、更像网页、响应更快的替代品。最终我们处于中等吸引力区间——有些产品市场匹配度，但不足以成为风投规模的业务，也没实现指数级增长。创业约两年时，我们陷入了产品市场匹配度不上不下的困境。

And so I kind of thought, wow, this is neat. But I dismissed it. For And several years, we went down a very different path. We were very much focused on this future of work thesis, thinking about remote async work, trying to build an alternative to slides that was more interactive, more web page like, more responsive. And we ended up in the zone of sort of medium traction, some product market fit, but not enough to be probably like a VC scale business, not enough to really take off with exponential growth, and got to a pretty difficult point maybe two years into our journey where we had that middling product market fit.

资金跑道不断缩短，加上当时经济环境动荡——通胀飙升、利率上涨、银行接连暴雷（包括我们的硅谷银行），我们很可能撑不下去。坦白说，拯救我们的是生成式AI的及时成熟。那时我们正面临公司存亡的焦虑：我们真的能活下去吗？

We had a dwindling runway, and we probably wouldn't have made it, especially because the economy around us was just going crazy. It was a period of inflation going up, interest rates going up, banks crashing, including our own Silicon Valley bank. So we had some very rough periods along the way. And, frankly, I would say what saved us was generative AI getting good just in time. So we were already at this point of facing some, I'll say, existential angst of, are we gonna make it as a company?

但这两年间我们持续关注AI进展。真正吸引我眼球的不是语言模型，而是图像模型——那年夏天Stable Diffusion刚问世，我看到推特上所有人都在分享他们用这个工具制作的趣味小图。

But we started checking back in on all this AI stuff over a period of two years. And for me, what really caught my eye was actually not language models. It was image models. Stable diffusion had come out that summer. And I was seeing everybody on Twitter posting these silly little things they were making in stable diffusion.

虽然还不是正经工作场景的应用，但AI这种魔幻般的病毒式传播力让我震惊。DALL·E也同期崛起，我开始意识到这才是能变革演示方式的技术。因为做演示的本质是什么？

It wasn't serious work yet. But there was this magic and this virality to seeing what AI could do. I think Dali was also taking off at the same time. And I started to realize this is the thing that can really change presentations. Because what are we doing when presenting ideas?

我们花费大量时间做视觉设计，常常无意义地装饰内容来填充空间。PPT三十年前的关键创新其实是剪贴画，但至今没人创造出下一代剪贴画。当我看到Stable Diffusion和DALL·E时，立刻意识到这是有史以来最可能实现突破的技术。

We're doing a lot of visual design, and we're often decorating things sort of pointlessly to fill space. Probably PowerPoint's key innovation was actually clip art, maybe thirty years ago, but nobody's really come up with what's the next clip art. And I saw stable diffusion in Dali. And I'm like, this is by far the most compelling thing I've ever seen that could do it.

这个洞察很有意思。所以让你们意识到商业契机的甚至不是语言模型，而是视觉图像模型？

That's a really interesting insight. So it wasn't even the language model side that made you think there's a why now for your business. It was the visual, the image model side.

完全正确。图像模型有种难以抗拒的魔力。但讽刺的是，正是图像模型的爆发让我们重新审视文本模型：之前听说过的GPT-3现在发展到什么程度了？我尝试用两年前相同的提示词测试，突然发现它们奏效了，而且效果相当好。

Absolutely. They're just so magnetic, so magical. But ironically, the image models taking off made us revisit the text models and say, what's been happening with this GPT-three thing that we heard about a while ago? And I remember trying it out and just trying some of the same early prompts I tried two years earlier, and suddenly they worked. And they actually worked quite well.

虽然大众觉得ChatGPT是一夜爆红，但OpenAI其实一直在迭代模型。他们重点突破的是指令微调技术——让人可以用自然语言发出'请写这个'或'格式化那个'的指令，模型就能理解。这个看似简单的创新支撑了ChatGPT的爆发。于是我们突然意识到，现在已具备所有条件，不仅能做演示工具，更能直接替你生成整个演示。

Everyone sees ChatGPT as an overnight success as well, but it turns out OpenAI had been improving the models all that time as well. And in particular, the thing they've been working on was instruction tuning. So making it so that you could actually give a prompt the way a human would, like, please write this thing or format this thing, and it would listen to you. Very simple innovation that I think underlie the whole ChatGPT takeoff. And so suddenly, we realized we had all the ingredients to not just be a tool for making presentations, but actually just make the presentation for you.

这太酷了。能谈谈你们提到的AI之前遇到的冷启动问题吗？你认为AI是如何改变这一问题的？

That's super cool. Can you say a word on you know, I think you guys mentioned pre AI, you had this cold start problem. How do you think AI is changing that problem for you all?

哦，你这么说很有意思，因为我们确实遇到过冷启动问题——产品测试版发布时反响平平。根本原因在于初始激活率：100人注册后，约95人在使用前几分钟就流失了，因为用户面对的是一个看似创新实则空白的工具页面。我们总说'看这功能多酷'，但...

Oh, it's funny you say that because, yes, we had a cold start problem, which was that we had launched this product into beta with middling traction. And the reason it wasn't taking off was the activation rate at the very start. So 100 people would sign up, and 95 or so would fall off in the first few minutes of using the product because they get dropped into this new innovative tool that was actually just a blank page. And we'd say, look how cool it is. It's got so many features.

用户会反问'我能用它做什么？'，最终除了少数硬核用户能坚持，其他人都流失了。所以我们早期押注AI，就是认为它能解决这个空白页难题——快速向用户展示产品的魔法时刻。

And they'd say, yeah, but what can I make? What can I do with it? And they would pretty much all fall off along the way, except for the few diehards who would manage to power through that blank page. And so the reason we seized on AI early on was we thought it could solve our blank page problem. We thought it could speed us through showing something magical where a user would get it.

直到真正推出后我们才意识到：AI解决的不仅是我们的空白页问题，更是用户的空白页焦虑。回想做演示文档为何令人畏惧？就是因为所有人都要面对那个写着'输入标题'的空白页，瞬间陷入'天啊从哪开始？要列大纲、设计版式...'的恐慌。

I think what we failed to appreciate until we actually launched it was that it wasn't our blank page problem being solved. It was the user's blank page problem. And if you then think back to, like, why are presentations so intimidating and frustrating, it's because we all start some new project with this blank page saying enter presentation title. And it's like, my god, where do I start? I have to outline all the slides.

还得构思故事线：亮点在哪？关键节点是什么？视觉风格怎么定？用什么字体？

I gotta figure out my story. What's the hook? What are the key moments? What it's gonna look like? What fonts am I gonna use?

选什么配色？需要哪些配图？你很快就会陷入这些细节的漩涡里。

What colors? What imagery do I need? And you get stuck in the spiral of, like, thinking through all these things.

完全同意。

Totally.

当我们用AI解决自身空白页问题时，意外发现它同样解决了所有人的这个痛点。用户现在能从模糊想法直接获得完整草稿，他们的工作变成编辑优化，而非从零开始。

And so when we actually built AI that solved our blank page problem, it turned out that it solved everybody else's blank page problem too. Suddenly, they could just go from a vague idea to a fully worked out rough draft. And suddenly, their job was editing, not starting from scratch.

没错。我经常用Gamma做AI主题演讲，直接把文档转成幻灯片的感觉太棒了——再也不用对着空白页苦思冥想演讲内容了。

Yeah. No, I've loved that. And I've actually used Gamma a lot. I get asked to give a lot of presentations about AI. And it's so refreshing to just give Gamma a document and turn this into a slide deck versus having to sit there in front of a blank page and figure out what I'm going to talk about.

感谢你的使用！我们内部也天天用它，这或许正是成功秘诀之一：做出自己都爱不释手的产品。

Well, thank you for trying it. We always appreciate it. And yeah, that's the magic too. We use it every day internally. Think that's part of our success is I love having a dog foodable product, a thing you can use every day.

完全同意。请谈谈你们在基础模型与用户之间所做的工作，这涉及到关于应用层价值多少的整个辩论。

Totally. Tell me about the hard work that goes into and this plays into the whole debate of how much value is in the application layer. Tell me about the stuff that you do in between the foundation models and your users.

是的，这是个非常好的问题。我认为这正是我们过去几年大量工作的重点所在。我想这可以分为几个不同的类别。首先是媒介本身。我提到过，我们最初并不想仅仅做一个演示软件的克隆，比如PowerPoint的复制品。

Yeah, really good question. I think this is where so much of our work the last few years has gone. And I would say there's a couple of different categories of it. The first is actually the medium itself. So I mentioned that where we started was we didn't want to just be a presentation clone, a PowerPoint clone.

我们希望打造自己独特的格式。于是我们创造了一种被用户形容为'Notion与Canva的混血儿'的产品。这意味着什么？Notion的部分在于它基于区块和写作。所以你不需要像在Google幻灯片中那样拖拽和移动方框。

We wanted to be our own unique format. And so we came up with something that our users describe as if Notion and Canva had a baby. So what does that mean? The Notion part is it's block based and writing based. So you're not actually dragging boxes around and moving them the way you are in Google Slides.

你只需要打字。我们试图让你输入的内容变得生动且具有视觉感。而Canva的部分在于我们拥有所有这些视觉元素。我们有图表、图形、视觉效果以及疯狂的主题和模板，这些都让它感觉更神奇。但关键在于，我们为你完成了这部分工作。

You're actually just typing. And we try to bring whatever you type to life and feel visual. But the Canvas side is we have all this visual range. So we have diagrams and charts and visuals and crazy theming and templates that sort of all make it feel a bit more magical. But the key thing is that we do that part for you.

你不需要知道什么看起来是对的。你只需要打字，然后得到一些灰色的内容。所以实际上，我们公司所做的很大一部分工作与提示或微调无关，而是调整那些核心构建块。我们如何拥有最好的图表、最好的布局、最好的视觉主题？

You don't need to actually know what's going to look right. You just type, and you get something grayed out. And so actually, a huge part of what we do as a company is not anything to do with prompts or fine tuning. It's just actually tuning those core building blocks. How do we have the best diagrams, the best layouts, the best visual themes?

这些成为AI发挥作用的武器库。它们是AI工作的构建块。

And those become our arsenal that AI is let loose on. It's the building blocks that it works on.

这其中涉及什么？是什么让你们能够创造出最好的视觉主题或最好的图表？

And what goes into that? What allows you to create the best visual themes or the best diagrams?

嗯，其中一个因素是品味。这可能与人们想象的在ChatGPT中输入相同内容时的结果有很大不同。我们实际上有一个相当大的设计师团队。事实上，有一段时间，我们公司三分之一的人都是设计师，12人中有4人是设计师。

Well, one thing that goes in is taste. So this is probably a big difference from when, you know, people think about what if I just type the same thing into, say, ChatGPT. We actually have a pretty big team of designers. In fact, for a while, our company was one third designers. It was four out of 12

哦，哇。

Oh, wow.

这在硅谷是非常不寻常的，尤其是在早期阶段。但我们非常注重优秀输出的样子。我们的整个理念是，我们负责优秀的设计，这样你就不必操心。但这也是非常数据驱动的。我们花了很多时间研究用户在我们产品中试图做什么，以及哪些地方会让他们迷失。

Designers, which is very unusual, I think, in Silicon Valley, especially at the early stage. But, we put a ton of care into what will great output look like. Our whole philosophy is that we do the great design so that you don't have to. But it's also quite data driven. We spend a lot of time looking at what users are trying to do in our product and where things get lost.

此外，我们还大量分析了全球各类演示文稿的特点。顺便说一句，AI在这方面非常擅长。它能快速处理成千上万的外部幻灯片，告诉我们其中常见的版式、设计元素等内容。

And also a lot of analysis of what do presentations out there in the world have. AI is great for this, by the way. It's churning through, for example, thousands of external slide decks and telling us what are the common layouts and designs and things that are there.

是啊。从数据中你们获得了哪些关于人们容易在哪些环节卡顿的洞见？又是如何通过产品解决这些问题的？

Yeah. What are some of insights you've learned from the data of where people tend to get lost and where you've solved that by product?

要知道，制作演示文稿的过程其实充满各种障碍。首先是故事构思和结构搭建阶段——我该如何表达这个内容？然后是格式化和设计阶段。我们当然希望在这两方面都做得更好。但特别是在格式化和设计阶段，有些技巧对AI来说轻而易举，对人类却极其繁琐。

You know, if you think about the process of making a presentation, there's actually this whole series of hurdles that come along the way. There's sort of a storytelling and structure phase where it's like, how do I even want to say this? And then there's a formatting and design phase. And we are certainly looking to improve on both. But in particular, in the formatting and design phase, there's actually just certain common tricks that for an AI are so easy, but for a human are so tedious.

一个简单技巧是不要只尝试一种方案。尝试10种不同设计然后选出最佳——这对三秒就能完成的AI来说轻而易举，对人类却极其耗时。另一个技巧是让演示文稿中所有图片保持统一配色方案。

One simple trick is don't just try one thing. Try 10 different designs and pick the best one. Easy to say when you're an AI that can do it all in about three seconds. For a human, that's incredibly time consuming. Another simple trick is, with imagery, making all the images in your presentation the same color palette.

这对能凭空生成图像的AI来说很简单，但对从谷歌图片搜索或个人图库中找素材的人类来说，要完美调校所有图片非常困难。所以我们正试图将这些技巧整合成人类根本无法完成的效果。

Easy to say when you're an AI that can generate the image from scratch. Hard to do when you're a human pulling from Google Image Search or your own library and making everything just perfectly tuned. So I would say that we're trying to stack up a lot of those tricks into something that actually no human could do.

当你们生成10种不同方案时，如何判断哪个是最佳选择？

And when you generate 10 different options, how do you figure out what's the best one?

啊，这问题真好。可能我没提过，在加入Gamma之前，我曾在Optimizely担任产品经理。Optimizely（对不了解的人说明下）是真正实用的A/B测试平台，我们开创了将A/B测试作为营销策略优化和产品发布的行业标准。

Oh, such a good question. So my background, I don't think I shared, is before Gamma, I worked at Optimizely in a product role. Nice. Optimizely, for those who don't know, was AB testing you'll actually use. So we sort of pioneered AB testing as a universal practice in how marketers tune their ideas and then also how products get launched.

这其实贯穿了我的整个产品管理生涯。之前在微软我就做过大量A/B测试，现在Gamma也进行海量测试。每当新模型发布时——比如今天录制时，Sam Altman正在台上宣布GPT-5——我们团队就有人正在更新代码准备接入GPT-5进行测试。

And that's actually been a through line through my whole career in product management. I did a ton of AB testing before that at Microsoft. And now we do a huge amount of it at Gamma. So every time a new model comes out so for example, as we're recording this today, Sam Altman is on stage announcing GPT-five. Literally right now, there's somebody on our team who's actually making a code update on our side so that we can implement GPT-five and run a test on it.

新模型上线我们会进行多阶段测试。虽然会做自动评估，试图量化是否优于前代模型。但创意领域往往没有标准答案，很难直接判定'GPT-5比Claude Sonnet优秀2.7%'——这种指标根本不存在。

And we go through a couple different phases of testing when a new model comes out. We do have evals. So we try to automatically measure, is this thing better than previous models? But it turns out we're in a creative domain where there's often no right answer. And so it's really hard to run evals and just say, oh, GPT five is better than Claude Sonnet by 2.7%.

我们真正拥有的是数百万真实用户。今天下午就会启动GPT-5对比Claude和Gemini的A/B测试，通过用户生成演示文稿时的完整行为数据来进行评估。

There's actually not a metric for that. What we do have are a bunch of real users. And we're lucky to now have millions of users that we can actually run these experiments on. And so sometime this afternoon, we're gonna start an AB test of GPT-five against Claude, against Gemini. And we're gonna actually measure a whole series of user behaviors when we generate a presentation.

我们实际上会通过评分询问用户的看法。我们还会测量他们对最终输出做了多少编辑？导出到其他工具有多少？是否进行了分享？最后，他们是否从免费用户转化为付费用户？

We're gonna actually ask users what they think through ratings. We're also gonna measure how much did they edit the final output? How much did they export it to other tools? Did they share it? And finally, did they convert from a free customer to a paying customer?

我们可能已经进行了数百次这类实验，非常谨慎地了解到哪些模型适合哪些任务。更重要的是，我们掌握了提示词中哪些方面效果良好。可以说我们在提示工程上投入很大，产品中使用相当复杂的提示词和提示编排体系。其中每个环节都经过某种程度的AB测试和实验验证，以找出真正能引起共鸣的内容。

And we have run probably hundreds of these experiments, and we have learned very carefully which models work well for which tasks. And on top of that, what aspects of our prompts work well. So I would say we skew pretty heavily on the idea of prompt engineering. We have quite complex prompts in our products and quite complex orchestration of prompts. And every piece of those is the product of some degree of AB testing and experimentation to find out what actually resonates.

这就是数据飞轮的实际运作。没有人比Optimizely团队更懂如何构建数据飞轮了。

Data flywheel in action. Nobody better to build a data flywheel than the Optimizely guys.

完全正确。

Absolutely.

关于提示词，主要是提示工程而非微调。这是有特定原因吗？

And on the prompt so it's primarily prompt engineering as opposed to fine tuning. Is there a reason for that?

说实话，我们只是发现微调效果并不理想。这个行业对微调炒作过度，坦白说主要来自提供微调服务的公司。但根据我们的经验，微调似乎会削弱模型的实际智能。而且我们发现微调在开源模型中很流行，但这些开源模型实际上远不及闭源的基础模型。

Honestly, we just haven't found that fine tuning works all that well. Fine tuning is something that I've seen so much hype around in the industry, mostly coming, frankly, from companies that do fine tuning for you. But in our own experience, fine tuning seems to hobble the model's actual intelligence. And also, frankly, what we found is that fine tuning seems to be popular among open source models, but we actually have not seen the open source models come nearly close to the closed source foundation models.

有意思。

Interesting.

所以我们日常主要使用Claude、Gemini和GPT，而不是那些被过度炒作的开源模型。

So for us, our daily drivers are like Claude, Gemini, and GPT, rather than some of these open source models that get a lot these.

好吧。那就说说你最喜欢哪个？Claude、Gemini还是GPT？各自最擅长什么？

Okay. So name your favorite child. Claude, Gemini, GPT. Like, what is each best at?

如果要选的话，我是Claude的忠实粉丝。可以说我们是Claude的重度用户。特别是我们发现Claude对'什么看起来好'有种独特的审美，这种创造力很惊人，而且在任何基准测试中都体现不出来。

You know, if I had to choose, I'm a Claude stan. I would say we are very, very heavy Claude users. And in particular, I think what we found about Claude was it just has a certain taste to it about what actually looks good. There's this creativity. And it's remarkable because it doesn't show up in any benchmark.

当这些公司谈论他们的模型时，首先他们总是在讨论软件工程。目前，编码是AI最主要的应用场景，堪称杀手级应用。因此这些公司都在为推理、工具使用和编码进行优化。但我们并非一个编码平台。

When all these companies talk about their models, they're first of all, they're always talking about software engineering. At this point, coding is the dominant use case of AI. It's the killer app. And so all these companies are optimizing for reasoning, tool use, and coding. We are not a coding platform.

我们是视觉表达平台。所以所有基准测试都毫无意义，必须抛弃它们并运行自己的实验。Claude表现最佳——不过有趣的是，早期版本的Claude有时比新版更好用，因为新版似乎走上了编码优化的路线。

We are a visual expression platform. And so all the benchmarks are useless. We have to throw them out and run our own experiments. Claude has been the best. Although funnily enough, earlier versions of Claude sometimes work better than new ones, as they've kind of gone down this coding optimized path.

但我觉得最被低估的可能是Gemini。Gemini是我们使用频率最高的模型，因为他们在成本效益上确实胜出。若论每美元能获得的智能水平，Gemini Flash无人能敌，可以说是我们的日常主力工具。这可能也反映了我们公司的另一个特点：与许多其他AI公司不同，我们非常重视利润率。

I think the one that's most slept on is probably Gemini, though. Gemini is by far our most heavily used model. And that's because they really win on cost efficiency. When it comes to intelligence per dollar, nothing beats Gemini flash, which is probably our daily driver. And maybe this ties into another thing about our company, which is we actually really care about margins, unlike maybe a lot of other AI companies in this space.

自引入AI以来，我们始终努力保持正向现金流，并思考如何可持续地提供这些服务。

We have always tried to operate, since we had AI, at cash flow positive and think about how we can offer these things sustainably.

嗯。其他推理模型呢？它们对你们有帮助吗？

Yeah. What other reasoning models? Have those been an unlock for you all?

出乎意料的是没有。我们实际做过实验，发现模型思考时间越长，创造力反而越差。推理模型在处理有标准答案的问题上表现很好，比如数学证明或编程问题。但在任何创意写作领域似乎都不太适用。

Surprisingly, no. We've actually run these experiments, and what we've found is that the longer a model thinks, the less creative it gets. I think reasoning works really well on questions that have a right answer, like creating a math proof or solving a coding problem. They really don't seem to work well in any kind of creative writing domain.

这真有意思。

That's so interesting.

特别是Deep Sync这个冷门模型，我们发现它完全不适合我们的使用场景，尽管外界都在吹捧'开源模型具备推理能力'。

And in particular, an unfavored child, we just found that deep sync was a total dud for our use case, Despite all the hype of like, Oh my god, these open source models can reason.

明白了。图像生成方面呢？有偏爱的模型吗？

Yeah. What's up on the image generation side? Any favorite models?

图像生成领域我们有很多'宠儿'。产品中支持约20种不同的图像模型，我们采用类似的AB测试方法。在图像模型领域没有绝对的最优解，目前使用最多的是Ideogram和Flux的模型。

We have a lot of favorite children in image generation. I think we have around 20 different image models that we support in our product. And we take a similar approach of AB testing them all. I think in image models, there's not as much of one right answer. I think our most heavily used models come from Ideogram and Flux.

他们确实有一些非常出色的模型和惊人的进展。但我们也对OpenAI的GPT图像能力感到惊叹。它在文本和信息图表类内容上表现极为强大。因此我们在适当的地方混合使用，同时也在探索如何经济高效地提供服务。

They have some really great models and amazing progress. But we've also been really amazed by what OpenAI's GPT image can do. It is so powerful at text and infographic type content. And so we're mixing in where we can, but also trying to figure out how to serve it cost effectively.

完全同意。我想回到数据飞轮和提示调优这个概念。你熟悉DSPY吗？你们是否使用这个或其他方法来自动优化提示？

Totally. I want to go back to this data flywheel and kind of prompt tuning concept. Are you familiar with DSPY? Yeah. Are you guys using that or any other methods to kind of automatically engineer your prompts?

实际上我们没有用。我们尝试过一点DSPY，但发现对我们帮助不大。我认为有几个原因：我们非常重视自身提示的可读性和清晰度。

We're actually not. No. We've tried DSPY a little bit, but we haven't found it helpful for us. I think there's a couple reasons for that. We've actually really valued sort of the legibility and readability of our own prompts.

我们几乎将提示视为用户体验，是设计师和产品经理可以参与优化的部分，而不仅仅是工程师编写和优化的内容。另一个原因是，我们非常关注跨模型的可复用性。比如我提到过，Claude、Gemini和GPT各有优势，我们会协同使用这些模型，并持续进行对比测试。

We almost view prompts as user experience, as things that a designer can work on, things that a PM can work on, not just things that engineers write and optimize. I think another thing is that we really care about cross model reusability. So I mentioned that, for example, Claude and Gemini and GPT each have their own strengths. We use all those models in concert. We test them against each other constantly.

此外，这些模型随时可能出现故障，这时就需要切换使用其他模型。因此我们非常重视这种可移植性。我们发现像DSPY这样的工具擅长微观优化特定敏感组合，但我们更关注通用性。

Also, at any given time, one of these models is broken, and you have to use a different one. And so we care a lot about that portability. And we found that tools like DSPY are really good at micro optimizing one very specific sensitive combination, but we care a lot about generality.

所以你们是在手工打造提示词吗？

And so are you handcrafting your prompts then?

应该说Cloud在帮我们手工打造。我们大量使用Cloud代码来迭代提示，但也会进行大量人工审查和迭代，并对内容方向给予指导。

Well, Cloud is handcrafting them for us, I would say. We use a lot of Cloud code to iterate on our prompts, but we do a lot of human review and iteration and Yeah. Guidance on what to

非常有意思。现在该谈谈房间里的大象了——我认为是竞争问题。确实，我同意你的观点，视觉传播是商业世界中最普遍的沟通形式之一。

Super interesting. Well, that's about the elephant in the room, which I think is competition. Yeah. I mean, slides, I agree with you. Visual communication, it's one of the most pervasive forms of communication, especially in the business world.

因此这个领域存在着许多重量级竞争对手。绝对如此。有哪些竞争对手让你们感到担忧？

And therefore, have a lot of 800 pound gorillas in the room Absolutely. That are going be competing with you. Who scares you?

最让我们担忧的仍然是PowerPoint，其次是Google Slides，因为它们仍是市场主导者。我们很自豪拥有数百万月活用户，但据我所知，PowerPoint和Google Slides各有5亿月活用户。全球约有10亿幻灯片用户，我们仍只占其中极小部分。因此我们仍视自己为挑战这些巨头的革新者。

The one that scares us most is still PowerPoint and, to a lesser extent, Google Slides because they are still the dominant incumbents. We're proud to have millions of monthly active users. But the stat that I heard is that PowerPoint and Google Slides each have 500,000,000 monthly active users. So there's like a billion or so Slides users out there, and we are still only the tiniest fraction of that. So we still view ourselves very much as an insurgent trying to take on these very big incumbents.

幸运的是，那些公司在AI领域的动作相当迟缓。我认为我们其实很幸运，因为自主编码吸引了所有人的注意力，导致这些生产力应用场景反而被一些大公司忽视了。我认为另一个主要竞争来源是基础模型本身。几周前我们刚看到ChatGPT助手问世——它能制作PPT。

Luckily for us, those companies have been fairly slow moving with AI. I think we're actually lucky that agentic coding has taken up so much of everyone's attention because these productivity use cases have actually been neglected by some of the big players. I think the other big source of competition for us is the foundation models themselves. So we saw ChatGPT agent come out just a few weeks ago. It makes PowerPoints.

值得庆幸的是（对我们而言），目前做得还不够好，但我们知道这终将改变，会越来越完善。所以我们现在仍在努力强化自身差异化优势：如何与PowerPoint、ChatGPT形成鲜明区别？如何在模型基础功能之上注入我们独特的审美品位？但这需要微妙平衡——我们既高度重视独特性...

Thankfully for us, not very good ones, but we know that'll change, and we know these will get better. And so I think for us, we're still trying to lean into our differentiation. How can we be as distinct as possible from PowerPoint, from ChatGPT? How can we bring a huge amount of our own taste on top of just what the models can do? But it's this delicate dance because on the one hand, we care a lot about being different.

我们不想最终沦为PPT制作工具，因为PowerPoint终将进化成出色的PPT制作工具。我们要成为卓越的Gamma构建平台，打造属于自己的视觉叙事媒介。但同时必须保持易用性——我们正跨越鸿沟：前5000万用户是热衷尝试新事物的早期采用者，而下个1亿用户可能是需要便捷性、熟悉感、并能融入现有工作流程的早期大众群体，尤其在我们从纯B2C向B2B拓展的转型期。

We don't wanna just be a PowerPoint builder at the end of the day because PowerPoint will eventually become a great PowerPoint builder. We wanna become a great Gamma builder. We wanna be our own visual storytelling platform and really our own medium. And yet, we also have to be familiar because we're kind of crossing this chasm where our first 50,000,000 users were early adopters who wanna try the newest thing. I think our next 100,000,000 or so users are gonna be maybe this early majority that need things to be convenient, need things to be familiar, need things to fit into existing workflows, especially as we also make this transition from pure b to c to start to do b to b as well.

所以我们就像在走钢丝，既要保持亲切感又要坚持创新。

So And we're kind of on this tightrope, trying to figure out how to be familiar and yet innovative.

完全同意。Canva呢？我认为他们是以设计审美著称的标杆，显然其用户主要用幻灯片功能。

Totally. What about Canva? I view them as somebody that's been beloved for their taste in design. Obviously, of their usage is slide driven.

怎么

How do

看待他们？

you think about them?

首先必须说我们视其为灵感源泉。他们取得的成就令人振奋，为我们正在探索的道路树立了标杆。他们验证了几个关于企业建设的深刻洞见：产品驱动增长(PLG)模式可以做到巨大规模...

First and foremost, would say we view them as an inspiration. I think they're a very inspiring company in terms of what they've achieved. They've really paved the path that we're trying to follow. I think they've proven a couple really interesting things about how you can build a company. They've proven that PLG can work at enormous scale.

我听说Canva有个惊人数据——在年经常性收入达到5亿美元之前，他们没雇佣过任何销售人员。这很疯狂，因为多数公司穷尽任何策略都难以企及这个数字。他们通过聚焦中小企业和海量模板库实现了这点。我们想追随他们开辟的路径，但选择略有不同——Canva本质是设计工具，依赖拖拽操作和模板；而我们自疫情初期就植入了写作工具的基因。

So they have The stat that I've heard about Canva is that they didn't hire a single salesperson until they had $500,000,000 of ARR, which is crazy because I think most companies aspire to ever reach 500,000,000 using any strategy whatsoever. Certainly, we do. And yet they've managed it with this sort of SMB focused strategy, and a huge library of sort of templates. And so we want to follow that trail that they blazed. I think we're taking a somewhat different path than they are, though, because Canva is first and foremost a design tool, where you drag and drop things into place and use templates to do it.

这也是我们常对标Notion的原因：无需设计技能就能创造精美作品。Notion在AI平台整合方面也是典范。是的，我们与Canva的正面竞争会加剧，但我希望我们能走出一条原创道路，而非简单模仿。

We've always had this different DNA from where we started in the pandemic of actually being a writing based tool. And that's why we kind of also compare ourselves to Notion as maybe another index point of this company where you didn't need to have any design skill whatsoever to make something really beautiful and magical. And I think Notion's also an inspiration at integrating AI into their platform as well. And so, yes, I think we'll compete with Canva more and more head on over time. But I'm hoping we can forge our own path versus just trying to be a copycat.

确实。我记得几年前曾有三四款这类AI幻灯片制作工具——暂且这么称呼。而你们确实脱颖而出，其他几款要么转型要么发展不顺。你认为原因是什么？

Totally. I remember at some point a couple of years ago, were three or four of these AI slide creation tools, for lack of a better word. And you guys really you really won out, and a couple of others have pivoted or or gone sideways. Why do you think that is?

没错。这个领域过去竞争非常激烈，至少目前我们成功让竞争缓和了些。我认为有几个关键差异点：首先是始终保持精简和危机意识。你提到的某些公司融资很多，组建了庞大团队。

That's true. It's been a very competitive space in the past, and I think we've managed to make it a little bit less competitive for now at least. I would point to a couple differentiators. The first one is that we've just always had a pretty lean paranoid mindset. So I think some of companies you're referring to raised quite a lot of money and built pretty big teams.

我们始终认为AI世界变化极快。决定成败的最重要特质是快速转向和适应的能力。因此我们团队规模相对业务量一直很小，非常注重保持这种精简机制。这也影响了我们的支出策略——比如在推理成本上极其谨慎。

And we always had the mindset that AI world is changing really, really fast. And the most important trait that will determine our success is our ability to pivot and adapt. And so that's led us to have quite a small team, I would say, relative to our traction. We've cared a lot about keeping that sort of lean mechanic. It's also impacted our spending, just in a very tactical way in terms of how many dollars we're willing to spend on inference.

现在很多AI公司都在亏损运营，包括那些使用GPT-4和DALL·E等昂贵工具的竞争对手。这就像用0.75美元的成本卖出1美元——虽然能快速扩张但不可持续。很多公司最终都陷入这种不可持续模式的困境。因此成本效率已成为我们的DNA核心。

So many AI companies now are operating at a loss. I think that includes many of our competitors who were running these tools like GPT-four and DALL E that were just tremendously expensive when they came out. And I think it's this classic mechanic of selling dollars for $0.75 You can get a lot of growth, but it's not sustainable. And I think a lot of companies really hit the crunch of doing something unsustainable. And so that's also become a part of our DNA is we really care a lot about cost efficiency in everything that we do.

我们极其重视保持正向现金流，因为这是让我们持续领先的基础动力。

We really care about maintaining that cash flow positivity because we think it's the foundation that will propel us to keep moving faster than everyone

对于其他在应用层创业的创始人，你有什么建议？

What advice do you have for other founders building in the application layer?

首先要思考你对目标市场的独特见解。Gamma的独特之处在于革新媒介本身。我们非常明确：不只是做PPT生成器，而是要创造取代幻灯片的新格式。

First and foremost, I would think about what is your unique perspective on the market that you're tackling. For us at Gamma, our unique perspective is all about differentiating the medium itself. So we've always been very clear to ourselves. We are not just trying to make a PowerPoint builder. We're trying to create a new format to replace the slide deck.

这个理念指导了我们无数产品决策，虽然有时会让用户不满，但帮助我们在红海市场中找到了方向。我强烈建议创业者思考自己的独特视角。

And that's guided many, many of our product choices. It's also angered our users sometimes. It's not a sort of a panacea. But I think it's helped us to navigate what is a very competitive market. And so I would definitely encourage other founders in this space to think about their unique lens.

比如我会警惕再做一个跟风的编程工具——这个领域同质化太严重。应该寻找被忽视的、AI尚未应用的领域，甚至反基础模型训练方向而行。那些编程工具火爆正是因为基础模型本身擅长编码。

So for example, I'd be wary of making yet another vibe coding startup. There are so many companies that are doing pretty much the same thing in this space. I would think about what is the neglected area where people are not applying AI. And even potentially, you want to go against the grain of what the foundation models are training for. So the reason these Vibe coding tools are taking off is that the foundation models themselves are so good at coding.

如果你选择基础模型不太专注但仍相关的领域，就更有可能在巨头碾压下突围。另外要建立实验机制，尝试多种模型。人们容易锁定某个供应商就认定最好，但技术创新既快速又不均衡，必须做好随时切换最佳模型的准备。

And it's so clear that the foundation models are optimizing for that. I think if you work on something they're not optimizing for as much but is still adjacent, you'll have a little bit better luck at making progress without being smooshed by the giant dinosaurs out there. I would also really try to incorporate some form of experimentation and really encourage trying multiple different models. I think people tend to lock into one model provider and just assume that it's the best out there. But the innovation out there is so rapid and uneven and unpredictable that I think you have to plan for a world where on any given day of the week, there's a different best model out there.

是的。我们这里已经掌握了Optimizely的核心技术。最后我想谈谈您对Gamma未来发展的看法。目前，您们的平台已经生成了约2.5亿个Gamma作品。就目前的使用情况而言，您观察到了哪些趋势？

Yeah. We've got the Optimizely DNA coming out here. I'd love to close by talking a bit about the future of how you see Gamma. Maybe so right now, you all have something like two fifty million Gammas have been creating created on your platform. What are you seeing in terms of the usage so far?

人们主要用Gamma来做什么？随着底层模型的改进和平台功能的增强，这种使用方式是如何演变的？

What are folks using you for? And how has that evolved over time as the underlying models have gotten better and as your platform has gotten better?

是的。我们最初是通过演示文稿这个应用场景实现产品市场匹配的。而演示文稿本身其实是个非常广阔的领域——从视觉化的TED演讲型演示，到文字密集的咨询报告或投行幻灯片。所以我们发现即使在演示文稿领域，仍有巨大的改进空间。

Yeah. So, you know, we started and really reached product market fit with this use case of presentations. And even presentations turn out to be a very wide world. Everything from a very visual TED Talk type presentation to a very wordy consulting or iBanking slide. And so we actually see quite a lot of room within presentations to just keep improving and doing better.

但我们也发现了很多意料之外的相邻领域。第一个就是文档——这里我说的不是纯文本的Google文档，而是像PDF提案、宣传册、给客户的精美报告和白皮书这类视觉化文档。

But we also discovered all these interesting adjacencies that we hadn't necessarily planned on from the start. The first one was simply documents. And when I say documents, I don't mean like plain text Google Docs. I mean things like, PDF proposals and brochures and shiny reports that you give to a customer and white papers. There's actually all of these things.

如果说PowerPoint的用户规模约5亿人，那么PDF用户可能接近10亿。人们创作的大量视觉化输出其实并不完全符合幻灯片的标准分类。我们在这个领域看到了巨大需求，因此开发了许多相关产品。另一个惊喜是网站建设——很多人开始用Gamma制作个人网站或小型公司官网，就像Notion文档被当作迷你简历使用那样。

You know, if the market share of PowerPoint is like 500,000,000 people, I think for PDF, it's like a billion people. I think there's just an enormous number of these visual outputs people are creating that don't quite fit the classification of a slide deck. And we've actually seen tremendous pull there and built out a lot of product for that. The next one that really surprised us was websites. So a lot of people started making these things in Gamma and realizing that the Gamma they made worked well as a personal website or their small company or agency website, similar to how Notion docs actually kind of get used as these little mini resumes.

我正想作同样的类比。

Was going make the exact same comparison.

没错。Notion确实给了我们很大启发。网站这个市场很有吸引力，因为它是人们在网络世界的门面。所以我们去年也推出了网站产品。这些相邻领域的探索让我们意识到，Gamma远不止是个AI演示工具。

Yeah, yeah. Again, they're a big inspiration to us. And we realized this is a great market to compete in because websites are so sticky, and they're so meaningful to people in terms of being their public face to the world. And so we launched a Sites product last year as well. And so we keep finding these adjacencies and pulling, and it's making us realize that we are much more than an AI presentation tool.

我们仍在思考如何准确定位，但目前认为Gamma是个视觉叙事平台。当您需要传递重要内容，特别是当'以貌取人'效应显著时——比如人们会特别在意呈现效果时，Gamma就是理想选择。我们希望通过多种使用场景来实现这个愿景。

We're still thinking about how to formulate it. But I think our our latest mindset is that we're a visual storytelling platform. And what that means is that any time you have something high stakes and important you want to convey, and especially when you're kind of like a book being judged for your cover where people really care how it looks, Gamma is a really great place to go. And we want to serve that through many different use cases.

这个定位很棒。能透露下近期产品会如何演进吗？用户在未来几个月可以期待哪些新功能？

I love that. Say a word on what should users expect from you over the near to medium term? How how will the products evolve?

下个月（九月）我们会有重大更新。先稍微剧透下：正在大幅改进核心编辑体验，扩展视觉表现力。我们的目标是让每个Gamma作品都独具特色，用户只需简单提示就能实现任何创意构想。

We have some really big things coming out just next month, so in September. Maybe just to give a bit of a tease of what's coming. We are doing a lot to improve our core editing experience, really expanding our visual range and variety. So we have this goal of making every gamma feel wildly different. And anything that you sort of imagine that you wanna achieve with it, you can just get, through easy prompting.

这对我们来说是个重大进展。我们正在构建更多代理式编辑功能，这样你只需委托AI，比如告诉它：这份五页的演示文稿其实应该扩充到二十页，能否详细展开第三部分的内容，并重新设计其他部分的视觉效果？只需通过对话就能实现。另一个让我非常兴奋的重磅消息是，我们将推出API接口。实际上，我们已经收到了超乎预期的热烈反响。

So that's a big one for us. We're building a lot more agentic editing so you can just delegate to the AI and say, you know, this five slide presentation should really be 20. And can you expand more on section three and maybe redo the visuals in this other part just by talking to it? And then a big one that I'm really excited about is that we're launching an API. So we've had a really surprising amount of interest in this.

我认为有太多公司在工作流程中需要基于某些环节生成可视化输出。举个简单例子：销售团队获得新潜在客户时，需要准备推介方案。如果能直接从CRM系统导入所有客户沟通记录，立即生成符合公司模板的预制演示文稿，省去销售人员一小时的人工整理时间（更不用说可能存在的随意拼凑情况），这该多好？让我惊讶的是，很多投资者告诉我，他们的投资组合里约有10家初创公司都需要这类工具——因为太多初创公司做着极具价值的工作，只是缺乏展示手段。

I think there are so many companies that, as part of their work, just need to create some kind of visual output based on some sort of workflow. So, a simple example is a sales team where a new prospect comes in, and you wanna tee up a pitch to that prospect. And so what if you could just take your CRM, pipe in all the notes you've had from them, and get a pre canned pitch deck in your template that you can just go and take to that prospect without the hour of salesperson time and, let's just say, mingling that might have gone on, when they did it? And I'm actually surprised by how many investors I hear from that say, actually, I have like 10 startups in my portfolio that need something like this. Because there's so many startups that do really valuable work for people and just need a way to show it off.

所以我们对其即将上线也感到非常兴奋。

So we're really excited for that to launch, too.

太令人振奋了。记得Karpathy曾发推问：幻灯片界的Cursor是什么？

Super exciting. Remember Karpathy tweeted, What is the cursor for slide?

没错。

That's right.

所有人都在评论区高呼Gamma。能听到你们如何通过远见卓识、灵活应变，以及在模型编排和实验层所做的种种努力取得今日成就，同时了解你们对代理式编辑未来发展的展望，实在精彩。非常感谢今天分享Gamma的故事，我迫不及待想见证你们继续改变视觉传播的未来。

Everybody showed up in the comments saying gamma. It's really cool to hear the story of both how you guys have gotten so far, the combination of vision, of scrappiness, of everything you're doing in the model orchestration and experimentation layer, and also kind of how you see the future of agentic editing and what's to come. So thank you so much for sharing the story of Gamma today, and I can't wait to see you guys continue to change the future of visual communication.

非常感谢邀请。能来参加真是太好了。

Thanks so much for having me. Was so great to be here.