Spotify for foundation models
Thoughts on what a product like this might look like
Welcome to Infinite Curiosity, a weekly newsletter that explores the intersection of Artificial Intelligence and Startups. Tech enthusiasts across 200 countries have been reading what I write. Subscribe to this newsletter for free to receive it in your inbox every week:
Hello friends,
Foundation models are being built at a rapid pace. The cost to train a foundation model is going down whereas the performance is going up. It's a good trend! As more foundation models are getting built to cater to the needs of different developers, it made me think of a structure that has worked well for music lovers -- the Spotify model.
I've been noodling on the idea of Spotify for foundation models. Imagine a platform where users can use a foundation model as much as they want. Or as little as they want. The builder of that foundation model gets paid as a proportion of overall usage. The platform takes care of the infrastructure for both sides.
I realize that there might be companies/founders out there who are already building something like this, but a clear winner hasn't emerged yet. So I wanted to put this thought out there and create a space for discussion.
What's the point of having a product like this?
There are so many products being built using foundation models. The applications include copilot products for salespeople, marketers, developers, and more. Builders need infrastructure to build these products. They also need access to foundation models that are oriented towards a specific goal.
OpenAI has been a pioneer in building foundation models. But developers are using their APIs to build a variety of applications. As the surface area continues to increase, the world might benefit from verticalized foundation models.
What's the problem?
On one end of the spectrum, we have OpenAI. They have a developer platform, but you can only access OpenAI's foundation models. What about all the other amazing models out there? Too bad. You need to build your own infrastructure to access those models.
On the other end, we have Hugging Face or Replicate. They have a developer platform where you can access open source models. But the people who develop those open source models are not getting paid. I know that the point of open source is to be open, but indulge me for a second here.
What about developers who want to develop verticalized foundation models and get paid for it? What about companies who want to use those verticalized foundation models in their applications and pay for it? You need a platform that connects the needs of these two groups. That's where a Spotify-like platform comes into play.
What are verticalized foundation models?
If you look at the foundation models being built, they usually go for a specific modality e.g. text, images, video, audio. Let's say there's a developer who wants to build an AI-infused application for the healthcare vertical. Why can't they just use the big foundation model that’s trained on all the text data? Because it's too broad! And expensive too.
They will get better performance from a model that's trained specifically on healthcare data. These verticalized foundation models will act as a starting point to build vertical-specific applications. There's the added benefit of cost. It will cost way less to run a smaller foundation model because you don't need this model to do well on any other task.
What should this product do?
This product should provide infrastructure to generate content using AI. The users can use any foundation model they want. It should provide infrastructure for inference. When I use a foundation model for a given task, it performs inference to provide an answer. This costs money. For example, let's say you type something into ChatGPT. They need run inference to provide an answer. Each API call to the inference engine consumes compute power. This platform needs to account for that when they create their pricing and revenue sharing plans.
Will it work?
When Spotify first started, people said that no artist will ever allow their work to be published on Spotify. Why would they? The music industry hadn't changed in a hundred years. Artists have all the leverage in terms of creation. Record labels have all the leverage in terms of rights/distribution.
Despite these constraints, Spotify became a juggernaut! They have over 500 million users. They have every major and minor artist on their platform. And they convinced the music industry that their business model is the way the world is going to work.
So how does Spotify work? In the older days of physical sales, record label companies paid artists a fixed price per song/album sold. But Spotify changed this model. They went out and gathered a large number of music lovers. These music lovers are willing to pay a monthly fee to use Spotify.
Spotify keeps 30% of the total revenue generated and gives out the remaining 70% of its total revenue to artists (or record labels). They pay royalties based on the number of streams as a proportion of total number of streams. For example, let's say the total number of streams is 2,000. If you're an artist and your song is streamed 20 times, then you'll 1% of the total royalties paid out by Spotify.
Now replace songs with foundation models and artists with foundation model builders. This is the platform we’re talking about. There’s a lot to be fleshed out here. I’m excited to see how this shapes up.
If you are getting value from this newsletter, consider subscribing for free and sharing it with 1 friend who’s curious about AI: