Jan-Erik Asplund
Background
Chris is co-founder of Copy.ai, an AI-assisted copywriting app. We talked to Chris to learn more about the unit economics of building on top of GPT-3 and where generative AI is going in the future.
Questions
- What is Copy.ai, what is your core customer profile, and how do they use the product?
- Copy.ai went from $0 to $2.4M ARR in one year. Could you help us understand where you found early traction, what customer segments and use cases saw rapid adoption, and what your core growth loop looked like?
- What does the AI tech stack driving Copy.ai’s growth look like, and how has it changed since you first started?
- People have this generic understanding that there’s a colossal GPT-3 model, and individual startups are adding a layer of fine-tuned models on top of it. Can you help us understand this better? How does Copy.ai dip into GPT-3, and what fine tuning entails in terms of training data, training the models, model ownership, etc?
- The AI-assisted content generation market is rapidly getting crowded with companies like Jasper, Rytr, Writesonic, Writer, and Peppertype.ai. How is Copy.ai positioned vis-à-vis these other players, and where do you see differentiation coming from?
- With OpenAI and others making it cheap and friction-free to build AI applications, it’s a common refrain that all incumbents like Microsoft, Google, and Canva need to do is add AI to their existing apps to win this space. What do you think about the competitive threat posed by these kinds of incumbents? How do you see the AI pie being split between incumbents and startups?
- Do you see application-layer products like Copy.ai being competitive with companies at the model layer like OpenAI?
- Can you give us your sense of where generative AI goes from here? Do you see content generation as a wedge to go deeper into markets like word processors (Google Docs) and image editing (Adobe Photoshop), or do you see it expanding horizontally to cover markets like search and social media?
- OpenAI costs $16 for generating 100,000 words, and Copy.ai costs about ~$99 for the same number of words. So with storage and compute costs included, is it accurate to estimate that your gross margins are 60-65%? Can you generally talk about the margin profile of building an AI-assisted copywriting tool?
- What does your cost structure look like? Is OpenAI the biggest cost for you? What’s the next biggest cost? Do you spend more on training models on OpenAI or calling APIs for content generation in production? What all does OpenAI provide apart from GPT-3 (like compute, storage, etc.)?
- How do you see the competition at the foundational layer between proprietary model like OpenAI, open source models like Stable Diffusion, and people making smaller, narrow use case models? Do you think the models are becoming larger and smaller at the ends at the same time?
- There have been concerns raised about the cost and speed of deploying large-scale models in data centers. Do you see that or any other challenges that can slow down generative AI’s growth?
- There's a narrative out there that a lot of Copy.ai’s initial users were people who were doing some form of labor arbitrage—essentially reselling content produced via Copy.ai on Upwork at a much higher rate. Was that actually an important part of the growth of Copy.ai?
Interview
What is Copy.ai, what is your core customer profile, and how do they use the product?
Today, our core customers would be small business owners and enterprises looking to supercharge their creativity.
Many times when you have to write a lot, it's really hard to get started. What we do is we help you get there. We help you get to a first draft.
In the future, the goal is to really go after a much larger market—what we imagine is that everybody in the world would be a Copy.ai user, and it would basically be an AI virtual assistant that will allow anyone to become a successful entrepreneur. That's the goal.
Copy.ai went from $0 to $2.4M ARR in one year. Could you help us understand where you found early traction, what customer segments and use cases saw rapid adoption, and what your core growth loop looked like?
What people don't know is that when GPT-3 launched, my co-founder Paul (Yacoubian) and I built about five MVPs in the course of a month: all kinds of different things, just trying to explore and experiment.
What really got traction was the marketing use case, specifically the copywriting use case. At the time, OpenAI had a lot of limitations on what you could do with it, but they were open to these shorter form uses, and so we were able to really nail down the copywriting marketing use case.
We launched on Twitter. We had a viral tweet that helped us get a lot of our initial customers, and it's just been growing from there.
We got a good domain name, and then Paul was really, really public about Copy.ai for the first six months to a year. We didn’t have much of a network, or prestige, and we weren't based in the Bay Area at the time because it was during COVID.
We took the one thing we were good at: being extremely transparent, pushing the boundaries of what's possible, and just really going at Twitter as hard as possible.
Even today, a lot of people still find us through Twitter. But since then, word of mouth and all these other things have taken off. And then we've also been able to own a lot of search results too. And it's been really great since.
What does the AI tech stack driving Copy.ai’s growth look like, and how has it changed since you first started?
Initially, the stack started off as just raw GPT-3. People were like, "Oh, aren't you just a wrapper around GPT-3?"
At the beginning, I would've been like, "Yeah, we are. We launched it in two weeks. What do you expect?"
Today, we have something like 20 or 30 distinct fine-tuned models running and they all help out with different parts of the process. We’ve basically had to build out an entire monitoring and A/B testing stack to allow us to experiment a lot faster, deploy new models faster, and then get feedback on it really quickly.
We have the advantage of a massive user base, which means we can get feedback really quickly. You run a test for a day and you'll get to statistical significance.
People have this generic understanding that there’s a colossal GPT-3 model, and individual startups are adding a layer of fine-tuned models on top of it. Can you help us understand this better? How does Copy.ai dip into GPT-3, and what fine tuning entails in terms of training data, training the models, model ownership, etc?
When somebody's using Copy.ai, we track things like whether they copied the text, saved it, or did anything else with the results, because that shows signal that what we produced might actually have been useful.
We take all of that data—whether they tried to rewrite it, save it, or whatever—and we use a lot of that data to basically retrain a new model that has that same input pairs, but with data that we know humans like.
That then results in a new model that is more human friendly.
The AI-assisted content generation market is rapidly getting crowded with companies like Jasper, Rytr, Writesonic, Writer, and Peppertype.ai. How is Copy.ai positioned vis-à-vis these other players, and where do you see differentiation coming from?
Paul and I never started Copy.ai to be a content writing app, to be fully transparent and honest. What we did care about is enabling people—really getting people to that 100x or 1000X in value. Where we're going with this looks completely different.
Today, we're more like Figma—you have to actually come in here and do the work and put in effort in order to get value.
Our guiding principle is how you 100X every single person, and there's a chance here with generative AI.
With OpenAI and others making it cheap and friction-free to build AI applications, it’s a common refrain that all incumbents like Microsoft, Google, and Canva need to do is add AI to their existing apps to win this space. What do you think about the competitive threat posed by these kinds of incumbents? How do you see the AI pie being split between incumbents and startups?
I'm going to start from a slightly different angle, which is this: AGI's coming. Everyone hears it. You are starting to see early signs of it. But if you go to most AI researchers and you say, "Hey, when you have AGI, what do you do with it?" Their first response is, "I'm going to have it make me money." It’s like, "Well, how?" And the first answer is, "I'm going to connect it to the stock market and it should be able to day trade better than everyone else. And then I'll become extremely rich."
The second answer is, "I'm going to let it run its own business and I'm just going to own it." It’ll just bring the go-to market of a company to zero.
The real question is this: what makes an AI company different? Why does it matter that a company is AI-native? In my eyes, what’s important is that it’s deeply rooted in the actual blood of the company. You have to live, breathe it, really use it yourself, and be constantly improving it.
Our belief is that larger companies won't be able to do that—they just won't be able to be completely AI-native. Only when the AI has full access to all the information and all the systems and processes, and all the context of the business, will it be able to just constantly generate more value itself.
A great example of this is chess playing AI. It'll make what looks like a really dumb decision now, but it pays off six moves later. AI models can have a similar type of mindset when it comes to building businesses. It's a very similar play.
For Paul and I, our entire vision is this: what can you do where AI becomes this valuable, where you just plug it in and it starts creating value on its own. That's stuff that is going to be really hard for incumbents with systems built out to really just add in as an additional value add.
Do you see application-layer products like Copy.ai being competitive with companies at the model layer like OpenAI?
If OpenAI wanted to launch their own copywriting app, they could, and then they could try to cut everyone else out and be like, "Oh, we're going to own the AI copywriting space." That is a possibility. I don't think that's what they're going to do. They have bigger ambitions. They're going for AGI.
It's our job at the application layer to make AGI useful. They want to build it, but once you build it, still having it create a business for you means you have to plug it into all the different tools and things that you need to do in order to run a business. There are other players out there like Adept that are kind of approaching it in a similar way, but they still are assuming that there's going to be an interface.
They’re imagining a world where you can describe the business, and then it'll just go to Chrome and open up tabs to go to Stripe, apply to Stripe Atlas, open a Ramp account, create a website, and so on.
We’re trying to tackle business AGI, because that's what we think is going to be the most valuable to the world. That's what speaks to Paul and I the most. That's where we are the most creative, and what we are passionate about, so if we can stick to this mission and do it for eight to 10 years, I think some massive things are going to come out of it.
Can you give us your sense of where generative AI goes from here? Do you see content generation as a wedge to go deeper into markets like word processors (Google Docs) and image editing (Adobe Photoshop), or do you see it expanding horizontally to cover markets like search and social media?
Why do you write documents in the first place? It's to communicate context, right? It's to communicate information. But if that happens automatically, if that happens between businesses automatically, you won't really need an editor. The way we think about our business is we're selling two things.
One is a second brain. Literally imagine a brain that knows everything on the internet and knows everything about your business. That's pretty valuable. Right?
The second thing that we're selling is extra hands. Hands that never get tired. Hands that can do things. And when you put these two together, you're basically creating a digital twin of your business as an AI assistant that can literally just run your business.
OpenAI costs $16 for generating 100,000 words, and Copy.ai costs about ~$99 for the same number of words. So with storage and compute costs included, is it accurate to estimate that your gross margins are 60-65%? Can you generally talk about the margin profile of building an AI-assisted copywriting tool?
OpenAI charges for input and output tokens, and if you're really smart about it, you can actually save a lot.
Today we sell words, but not all words are created equal. The perfect headline for a company that's three words long could be a lot more valuable than a 1,000 word blog, and so per-word is not a really great way of characterizing value unless you're in blog writing.
The margin profile is pretty solid.
What does your cost structure look like? Is OpenAI the biggest cost for you? What’s the next biggest cost? Do you spend more on training models on OpenAI or calling APIs for content generation in production? What all does OpenAI provide apart from GPT-3 (like compute, storage, etc.)?
Well, people are the biggest cost for our business, and then OpenAI is second because it's a core component of our business. Next biggest cost is probably the infrastructure of just running everything.
Training models is actually really cheap now. Even with the amount of data that we have, it’s actually a drop in the bucket compared to the APIs we run in production.
OpenAI has the best models. That's not necessarily always true because other providers are now allowing fine tuning and we just have a really, really large amount of really high quality data that we can draw from, so we can actually fine tune a lot smaller models to do really well.
How do you see the competition at the foundational layer between proprietary model like OpenAI, open source models like Stable Diffusion, and people making smaller, narrow use case models? Do you think the models are becoming larger and smaller at the ends at the same time?
They're getting smarter with less size, and that's a really good thing because that means OpenAI can reduce costs on the hosting of these models while providing a better product. I'd be really happy about that if I were them.
OpenAI dropped their costs by 66% the other day. How many times has Amazon dropped their costs? 107 times in the history of AWS. We expect similar dynamics in this space.
As far as competition at the layer below, we don't really care. We just are going to go with the best models. It benefits us that there's more competition there: it increases the pace of research, it increases the pace of investment on that layer and we're going to see more and more of that. Then it just comes down to talent for them.
In the AI space right now, the extremely talented people are able to produce stuff that's just absolutely mind-blowing. It pushes us way further along and both sides have really good talent.
There have been concerns raised about the cost and speed of deploying large-scale models in data centers. Do you see that or any other challenges that can slow down generative AI’s growth?
Regulation could end up being a challenge. I don't know if regulators are going to fully appreciate the bigger vision. That's an issue. Other than that, there's really nothing stopping us.
There's an insane opportunity here. The way we see it is like the invention of the internet all over again. These last few weeks have shown that the stalwarts of Web 2.0 are no longer investing as heavily into R&D. They're doing hiring freezes, they're not going to be able to innovate as quickly, and that opens up a massive opportunity for an AI-native startup to come in here and reinvent the entire system.
Web 2.0 was built on a very specific set of assumptions, and these generative AI models break a lot of those assumptions. That’s where the opportunity is going to lie.
If you look at our app today, you would say, "Hey, this seems like any other Web 2.0 app" And you'd be right. And all the apps in our space are basically the same. But when you start unlocking what is possible with generative AI, we have barely gotten started.
There's a narrative out there that a lot of Copy.ai’s initial users were people who were doing some form of labor arbitrage—essentially reselling content produced via Copy.ai on Upwork at a much higher rate. Was that actually an important part of the growth of Copy.ai?
It’s possible, but I wouldn’t say it was an important part of the growth. Basically, people came to us with a need: they needed to write something, and they had deadlines.
We would help them create something really good. Market demand was stronger in other countries or with people whose writing capabilities are a little worse, who don't speak English as their native language, and people who otherwise really needed that extra help.
That being said, a lot of people in the United States use Copy.ai as well. We have a case study with Airtable who use it for brainstorming. For a lot of these companies, the amount of copy they need to write is insane. It's literally just never-ending, so being able to have something that can really just produce new ideas in one-click helps so much.
Disclaimers
This transcript is for information purposes only and does not constitute advice of any type or trade recommendation and should not form the basis of any investment decision. Sacra accepts no liability for the transcript or for any errors, omissions or inaccuracies in respect of it. The views of the experts expressed in the transcript are those of the experts and they are not endorsed by, nor do they represent the opinion of Sacra. Sacra reserves all copyright, intellectual property rights in the transcript. Any modification, copying, displaying, distributing, transmitting, publishing, licensing, creating derivative works from, or selling any transcript is strictly prohibited.
