Transcript: ‘His GPT Wrapper Has Half a Million Users—And Keeps Growing’

The transcript of AI & I with Vicente Silveira is below.

Timestamps

1. Introduction: 00:00:35
2. AI PDF’s story begins with an email to OpenAI’s Greg Brockman: 00:02:58
3. Why users choose AI PDF over ChatGPT: 00:05:41
4. How to compete—and thrive—as a GPT wrapper: 00:06:58
5. Why building with early adopters is key: 00:20:49
6. Being small and specialized is your biggest advantage: 00:27:53
7. When should AI startups raise capital: 00:31:47
8. The emerging role of humans who will manage AI agents: 00:34:53
9. Why AI is different from other tech revolutions: 00:45:25
10. A live demo of an agent integrated into AI PDF: 00:54:01

Transcript

Dan Shipper (00:00:35)

Vicente, welcome to the show.

Vicente Silveira (00:00:36)

Hey, Dan. Thank you for having me here.

Dan Shipper (00:00:39)

Thanks for coming on. So, for people who don't know, you are the CEO of AI PDF. It’s one of the biggest AI PDF readers in the world. You have about 500,000 registered users and it's only been live since the end of last year. You've done over 2 million conversations in the GPT Store and you just started monetizing and you have about almost 3,000 paying subscribers. So, it's a really, really cool business, and what I think is most interesting is there is this narrative, especially in the sort of earlier days of AI, but I think it's still happening now, which is like, oh yeah. All of these AI PDF companies are fucked and they're not going to do well and whatever. And I think you're actually building a really interesting business and you're also part of this broader wave of people who are making things in this new AI economy with really small lean teams, who run the company break-even without raising a ton of money. I know you raised a friends-and-family round, but you haven't raised a ton of VC. I think there's a lot of overlap between the kinds of things that you're building and the kinds of things we're building internally at Every. And I just want to learn about how you're thinking about it. We can swap stories.We've done a lot of thinking too about fundraising and all that kind of stuff and how to build this. Because I think people assume, oh you can't build a good business like this, but I think you can actually build a sneaky, really, really great business this way with small teams, especially with AI. So yeah, tell us about that. Tell us about your business and how you're thinking about running it.

Vicente Silveira (00:02:22)

Yeah, it's funny you mentioned the whole wrapper thing. We almost have a wrapper death countdown clock, which is this many days since we died the last time. From time to time, we're pronounced dead. In reality, the business keeps growing. It's kind of interesting. At some point, when OpenAI first allowed to upload PDFs to ChatGPT, it was, oh, all this stuff is dead. No. Some of our competitors actually just gave up at that time, but we kept plugging it. And I think at some level this is an interesting industry because everyone is pointing at each other saying, you are a wrapper. It's like, I don't know if NVIDIA is a wrapper around math. Everything goes from there. 

But yeah, we started on this because we tried different things actually. When we first started, when ChatGPT was just coming out, and I was basically trying to do property injection against ChatGPT and Sydney back then. And I found some stuff. I sent an email to Greg Brockman at OpenAI and he replied, oh, that's kind of interesting. You should talk to this guy here. And at the time, I just got out of Meta and I was angling for a job at OpenAI. I talked to this guy, he's like, well you have an interesting background. Right now we're not hiring for this, maybe in a couple of months, so we’ll keep you in mind. And I'm like, well, maybe you can give me an API key or let me into this new developer program, and that's how we started. And we tried different ideas, but the PDF one really took off immediately. And in reality it is because this is one of the first things that people are trying to figure out with AI. It's a lot of pain dealing with lots of documents, and PDF is the main kind of document across platforms. So people just gravitated to that.

Dan Shipper (00:03:21)

That's really interesting. And so people gravitated to that. And you've said people have declared your death multiple times. What do you think—? I mean, if a PDF reader is in Chat, why are people using you?

Vicente Silveira (00:04:39)

Great question. So the thing is when we looked at this and we started building it we— Actually at first, we didn't even have a place for people to upload PDFs. We were just like, okay you can just give us a link and our server would go there and fetch the content and because we're growing so fast at the time, people were giving us Google Drive links and Dropbox links. And Google and Dropbox started rate-limiting our IPs because they saw us as an aggressive bot. And this was a bad experience for our users because it would get an error. And we were like, well what are we going to do? And we're like, well, I guess we could try to play the cat-and-mouse game with those guys but we actually know how that works and that would be kind of pretty bad. So we're like, well, what if we just let them upload their files? And we thought no one would do it. And so the first version of our website, My cofounder, Karthik, he was like, you look scary. And to our surprise, in a week, that domain for our website became the number one domain for the links passing the Google domain and the Dropbox domain.

So that was a lesson for us because it told us that the users that were gravitating to ChatGPT, and they're going to all the trouble to enable plugins, that those people were actually risk takers and early adopters. So we built that for them. And then, going back to your question, why do they keep using us even after this was created on ChatGPT proper is because they don't want to just upload one file. When it works, they basically want to upload their whole collection of files. And even to this date, I think our ChatGPT is limited to 20 files. And for us, we have people with more than 150,000 files in one account. We have people with multi-level folders. No one else supports that kind of stuff. But it's important because people need that low-friction and even respecting the intelligence they put in creating their folder structure to be part of this onboarding. So this is part of the product experience.

Dan Shipper (00:06:42)

That's interesting, but I want to push you a little more. So it sounds like one of the reasons people are still using it is because you're sort of staying one step ahead of what ChatGPT will do. Do you have a theory about why they won't eventually do a multi-folder upload? An example might be a NotebookLM where it does let you open a lot of files. Are you worried about that at all? Or is there some strategic reason why you think you're going to go deeper than a ChatGPT or like a Gemini or whatever.

Vicente Silveira (00:07:20)

I mean, this is interesting because I feel like these guys, especially if you look at ChatGPT. Their focus is to kind of race towards AGI and create a product that's good enough to have enough usability there so that lots of people use it, they can collect the training data, and then feed their machine. I don't think they're going in one particular specific direction. They're mostly kind of touching on what is the minimum for core use cases. So I think, is there a possibility that something like ChatGPT or Claude can actually compete with us and basically there's no need for this kind of platform? Yes. But that is always the question for startups when things start.

So when you think about a startup like Loom, Why does Loom exist and will sell, I think, for almost $1 billion to Atlassian. Loom is just recording video. But they did that use case so well that even though YouTube had all the technology to do it, they didn't do it. Vimeo had technology to do it, they didn't do it. All of the major providers had the technology. They didn't do it, but Loom actually nailed the use case. So, for us, we're nailing the use case of getting a collection of documents that you have and being able to do end-to-end workflows with those documents.

Dan Shipper (00:08:46)

That's really interesting. I think it dovetails with some of the things that I think about. One of the things that I have in my mind when we build things internally at Every is that ChatGPT and Claude, I think, like you said, rightly, they're building for the most broad use cases possible. They just want anyone to go on and be able to do whatever they want, basically. And in that way, it's sort of a little bit like Excel, where anyone can go into Excel and you have a blank page, a new sheet full of lots of cells and you can just start typing numbers and anyone can do that. And then people are going to discover as they're using ChatGPT and Claude that they have more specific— They're going to discover use cases for themselves that they didn't know existed.

So for example, we have a product called Spiral that lets you automate a lot of creative work. It helps you do headlines and come up with tweets and all this kind of stuff and that's the thing you can do with Claude. But Claude is not purpose-built for it. So our thesis is people will discover use cases for AI and discover problems to solve with AI by using these more general-purpose tools. And then that will create demand for other players to peel off some of those use cases for particular kinds of people for particular kinds of workflows. For us, it's marketers and creators who have a very specific need for that kind of workflow. And having a product that's purpose-built is going to serve those people better. 

And I'm sort of curious, for you, because a key part of my thesis is you need to have a particular persona in order to be powerful enough for that kind of workflow. But it sounds like you're kind of going a level up, which is more general. But do you have a particular persona in mind or how do you think about it?

Vicente Silveira (00:10:46)

We get this question quite a bit and it's interesting because our persona is an early adopter of technology, a risk taker that has an actual job to be done involving lots of documents. So all of those things are important. One is they're not just an early adopter because an early adopter— And a lot of ChatGPT and Claude use people that sign up to just see what's possible, they don't have an actual job they're going to do there, but they just want to get that familiar with technology, be able to talk to someone about it, those kinds of things. So a lot of that kind of use case that's one component of our user base. The other component of our user base is they have an actual job that they need to do today. And they have a lot of documents. So this is the combination of things that creates our—so like, who are these people?

We have a law firm. The partner is an 80-year-old lawyer And he found us. He's like, I'm using this every day and we're going to have it within our firm and I’m a decision maker— We're going to have it adopted here. And we have people that are researchers, we have accountants, we have writers. So there's a range of these kinds of different types of profiles. But what you'll find in common is that they bring a lot of documents to the platform and they are basically trying to get some job done today. And they want to do this in a new way, which is the AI-first way. So that's one thing about the persona. The other thing that I want to kind of highlight is we also differentiate from a platform like Claude or ChatGPT in a way that I think they're not really going there, which is we're building this from the perspective of, it is kind of like we're giving a cloud drive to an AI agent. 

That's very different than you allow an AI agent to access some files. So what do I mean by that? And we can show you a little bit about that as well. But the agent that we have in AI Drive is capable of doing things like creating new files, updating metadata in files, and going through the file structure. So it's really kind of driving that effectively that cloud drive to be able to accomplish a job for the user. Because a lot of the job involves manipulating a lot of documents. So that's another very important point, which I think it's not really the focus for these other platforms.

Dan Shipper (00:13:29)

Why do you think it's not the focus?

Vicente Silveira (00:13:32)

Because I think it's not necessary to accomplish what they're trying to do. And it also introduces other types of considerations and risks that they may not be interested in dealing with. So this opens up our platform for a tinker type user to be able to do things like— So think about this: You have chat history in something like a ChatGPT. But that's something that you can go to, you can look at the chat history. In our product, chat history is actually made of files that— I can access those files on your behalf and use the same tools, like a search tool to be able to go into those kinds of files. So these are all kinds of the same primitives that we're building on.

Dan Shipper (00:13:42)

So one of your core users, you said early adopters, but it's I mean, I'm thinking of myself, information nerds like I was like, ooh, your chats are files. That's amazing. That's so cool. It's like feeding back into itself, but that's a certain kind of nerd that cares about that.

Vicente Silveira (00:14:35)

I was just going to illustrate what you're saying. So you see that we have this little system folder here.

Dan Shipper (00:14:40)

So basically you're sharing AI PDF and the product is AI Drive. And basically you're showing me on the left, it looks like you have a list of files. So that's your drive. That's like a Google Drive. And then you have a chat window which I assume allows you to sort of chat with those files, sort of like a NotebookLM. And then on the right, it looks like there's a reader view where, for any particular file that you're talking to, you can also see the PDF—that's basically what I'm looking at?

Vicente Silveira (00:15:09)

Yes, that's right. And you can see on the left side here, we have shortcuts here. So if I click on this little clock, you can see a history. So this is very similar to what you have in ChatGPT. There's a history of the chats that you had today and that's all you get in a typical platform. That's not really open for people that are kind of tech enthusiasts like you and many others. But what that means for us is that we're building this. So if I go back to files, you can go and find these as files under the chat history folder. We're building these as basically kind of an open platform where everything from your chats to and we're going to be adding memory—all of that will be basically just a file in the system.

Dan Shipper (00:15:54)

That's interesting. I want to go back to what we were talking about earlier. So one persona is information nerds like me, and maybe that cuts across lots of different industries. There’s probably information nerds who are lawyers or accountants. It doesn't seem like any of your marketing is specifically about that. Your marketing is a lot more general AI PDF stuff. How do you feel about that?

Vicente Silveira (00:16:16)

Yeah, and we're flipping that because our trajectory is going through this transformation. So if you go back to when we built this, and I give this advice to people when they're thinking about what they're going to do. It’s like what we did with the plugin. That was the least effort thing that we could have done at the time. It's just an API. At the time I had a server running on Replit. Replit’s amazing. And that was everything. And with that, we discovered the market. But then at that point, we're just a plugin. So we couldn't operate independently.

Dan Shipper (00:16:57)

What do you mean by it was just a plugin?

Vicente Silveira (00:16:58)

What I mean by there was no web, no web app, no place for you to create an account. We had no direct relationship with the user. So you would go to ChatGPT, you would enable plugins, you would find us.

Dan Shipper (00:17:05)

I see. It was a ChatGPT plugin. I just forgot about that whole era of ChatGPT.

Vicente Silveira (00:17:11)

It wasn't that long ago, but it feels like that. Yeah, so at that point, we were a product that only made sense as an add-on to ChatGPT. But what we realized is that people needed an actual environment where they could do the work end-to-end with the AI and the files. You are able to verify the work of the files, which is another thing that we as you can see on the right side of our AI Drive screen, you have the actual files that are the source material. But beyond that, the other thing that we realized, and I think it will continue to be true for the foreseeable future, is because of this arms race between the main providers, the ones that can play the game right now, probably Google and Anthropic and OpenAI, I guess, is coming up as well as a potential one. So you have maybe four or five providers there that can actually provide unique capabilities when it comes to the models. And our users, because they are like you, and I heard you say in your podcast, you tried this on Claude or this on o1. They want to be able to have the latest and greatest. Now, if you're going to upload your files to ChatGPT, and then tomorrow, Claude has a better reasoning model. Then now you get to upload your files over there as well. So we bring that one place where you can use all the models combined. So that's the other aspect of that.

Dan Shipper (00:18:42)

That makes a lot of sense. I mean, it's funny. We do a lot of work with big companies where we help them figure out what to use and sometimes train their employees and that kind of thing. And I think mostly they're not in this category. They were just excited to see that ChatGPT had a new feature, but none of them had heard of Claude. They're like, what? We're like, it's the best. And so, yeah, I think there is that sort of a different market. The early adopter market is a bit different for people who really want to use the latest and greatest-type thing and, like I said earlier, that sort of cuts across industries and there's room for that kind of nerd and what's a nerd as a customer.