The business of wheeling and dealing ML models#
While OpenAI is a unique case study of a company, what most people forget (or don’t know) is that their business model is not particularly novel. Companies have been selling access to ML models for a while (probably 5+ years?), ranging from your typical big company stuff to smaller, more specialized startups.
Imagine you’re a software engineer at an e-commerce company, and you want to build a basic recommendation engine to recommend relevant products after someone buys something. What are your options?
- You can do it manually for each product. This can work early on, but doesn’t really scale.
- You could try and build it yourself, but you don’t have a Data Science background. You could _probably _pick it up in a few weeks? But then you wouldn’t be doing other work your team needs.
- You could hire a Data Scientist, but how much would that cost? How long would it take? You’re probably looking at at least 6 months before you’d see anything ready to test.
Many teams do opt for one of these options. But your most straightforward bet is to buy a model from someone instead of building it yourself. Even for something as specific as a recommendation system, you’ve got a few options in the market:
So what are you actually _buying _here?
When you buy an ML model, you’re not really getting the model itself – the provider needs to keep that proprietary so they can (a) make money on it, and (b) continually improve it. What you’re really getting is access to that model, usually via an API interface. Here’s what it looks like to use Algolia’s Recommend API:
- You need to train the model on your data. Algolia gives you guidelines for what they need; it’s basically all of your order events and which products people bought together in the past.
- You pick which type of algorithm you want to use: they provide a few different options, each optimized for a different use case.
- You preview the results and make sure they look half decent.
- You integrate the model into your app.
If you’re buying this model, why do you need to train it? Doesn’t that defeat the purpose? Good question. Algolia has trained the model to be good at recommending things, but it needs to know your sales history to recommend your items. In other words, the core ability to recommend things based on other things is what you’re paying for, but the specifics of which things need to be inputted into the model.
The end result of this process is a series of API endpoints that you use in your app to display recommended products. After someone adds something to their cart – or maybe just when they’re viewing an item – you might make a request to the Recommend API[^1] that looks like this:
GET [https://api.algolia.com/recommend/v2](https://api.algolia.com/recommend/v2)
{
# the product the user is currently viewing
product_id: “23o8hri3u2tr8721yg”
}
And you might get back something like this:
{
recommended_product_ids: [2,45,68,211]
}
And then you’d take those product IDs, get the necessary information about them (name, pictures, price, etc.), and display them in the “you should check these out” section. In case you were curious, pricing is something like $0.60 per 1,000 requests.
This is the basic gist of what buying an ML model is like. Beyond recommendation systems, you can buy models for computer vision, natural language processing, and a lot more.
OpenAI as an ML selling business#
OpenAI sells ML models, just like Google, or Algolia, or AssemblyAI, or any other number of startups in this space. Very simple! The main difference (outside of the research thing – see below) is that OpenAI’s models are really, really good. There are 3 “big” models they advertise but more are available when you dive into the documentation.
OpenAI GPT#
You’ve most likely heard of this one. GPT is a text generation model – you give it a prompt, and it gives you back a sentence, paragraph, essay, whatever you ask for. The most recent version is GPT-4, released a few weeks back.
You can use GPT-4 through OpenAI’s ChatGPT interface – where you write a prompt almost as if you were sending the model a text message – or via API.
OpenAI DALL-E#
Instead of generating text, DALL-E generates images. Like GPT, you prompt it and it responds in kind.
Also like GPT, DALL-E is available both as a simple UI from OpenAI, or as an API for developers. I asked DALL-E to generate a new logo for Technically…not great.
OpenAI Whisper#
Whisper is OpenAI’s third and final model available to the public – it’s built for transcribing audio to text. It works for a bunch of different languages, and can even transcribe from a language other than English to English. Whisper seems to be a lot less popular than GPT or DALL-E – I even noticed a sort of “advertisement” for GPT within the Whisper UI:
Whisper is the only major OpenAI model that’s open source. You can check out the GitHub here.
Other OpenAI models#
Beyond the big 3 advertised on their homepage, there are a few more OpenAI models you can use via API:
- Embeddings – a sort of “middleware” model to convert text into numbers you can use in other models.
- Moderation – a set of models to detect whether text is threatening, sexual, stuff like that.
- Codex – a now deprecated (read: old and unsupported) model for generating code for software engineers. Now part of GPT.
OpenAI has also open sourced a bunch of models for anyone to use, including Whisper above.
OpenAI as a research company#
A confession: I’ve slightly misrepresented OpenAI thus far. Because while they are at least technically a company that makes money by selling ML models, there’s a lot more to the puzzle. Consider how OpenAI describes themselves:
Or the Wikipedia description of OpenAI:
A bit of history: back in 2015, a combination of researchers, entrepreneurs, and companies like AWS formed a non-profit called OpenAI, with the goal of doing AI research. The group committed $1B in donations; here’s the original blog post. From then on, they basically focused on recruiting the best possible talent in AI research, competing with other labs like Google Brain, The Allen Institute, and Deepmind (now also part of Google). It’s a popular adage that AI researchers are more expensive than NFL quarterbacks (I cannot verify this statistically).
Every few months, OpenAI would release some sort of free playground to work with the research they were doing, like OpenAI Gym for reinforcement learning back in 2016. GIFs from Gym would go viral because they were this hilarious stick figure who often looked like he was hurtling across space and time:
Their primary cost started to become compute power: these state of the art models were just incredibly expensive to train. They spent $8M just on training in 2017! Anyone who was in the AI scene around this time can tell you that the “OpenAI vibes” were very unique: tech hadn’t really seen a research lab this well funded that had attracted such powerful talent. Maybe since Bell Labs?
Things started to change in 2019 when a few things happened:
- Microsoft announced they were investing $1B into OpenAI (!)
- OpenAI transitioned from a non-profit to something called a “capped-profit company”
- They announced that they were going to start commercializing their models
In other words, they began transitioning – or at least expanding – from a research company to a company company. The capped profit thing basically stipulates that investments into OpenAI can only start making you once they’ve generated 100x returns for OpenAI. In other words, if you invest $1M, once that $1M makes OpenAI $100M, then you can start earning returns on that $1M. It’s an unusual structure, but it allowed OpenAI to start raising a lot more money.
The Microsoft relationship was key, because it allowed OpenAI to manage their compute spend. We don’t know the figures, but the assumption is they had really steep discounts on Azure, Microsoft’s cloud platform. In January of this year, there was a really big announcement: Microsoft was investing a lot more money (maybe $10B, nobody is sure) into OpenAI, a lot in the form of Azure credits so OpenAI could keep spending tons on training.
Today, most of OpenAI’s new releases aren’t open source, and there’s a clear focus on commercialization. Pricing is generally on a per use basis, and they’ve even got a page focused on customer case studies. Oh and by the way, if all of this wasn’t interesting enough: Elon Musk was on the board.
So what does OpenAI do? Well, it’s sort of a research company, sort of an ML model provider, and some other things too.
