Viet Le’s post earlier this week on what it takes to build a defensible machine-learning based company is a great read. There’s a lot in the post, but I’ll pick out three key points that will be most important to anyone looking to build something useful and valuable with generative AI:
- The core of what is being delivered has to be a value driving product. In other words, it’s the value delivered by the service that’s important, not whether or not it uses AI.
- The products that will succeed in the current cycle are those which have an interface with the user (and, by inference, a relationship with the user). This is in part, because many companies will have access to the same foundational models and in part because this is the best place to understand user queries and their satisfaction with responses.
- It’s critical to be able to capture data and models (or, said another way: you need a “data engine“). This means that as your service does its work, it is critical to capture what works, what does not and continue training models.
Almost all good product development and growth to scale is the result of a well-thought-out instance of a Jim Collins flywheel:
The key is that the value necessarily drives usage, which necessarily drives improvement, and with each loop around the flywheel, momentum and usage grows. In a bakery, this would be innovating on the formula for a particular cake. The better it gets, the more people talk about it and come to buy. The more people buy, the more can be invested in experimenting with the recipe and improving the ingredients.
In a standard tech product, the cycle revolves around solving a particular use case well, attracting the first customers, learning from their needs, and using the revenue from their usage to improve and expand their usage. If the product has a network effect, such as a messaging network, the product effectiveness improves just with an increase of users, with no product change even needed.
In a generativeAI context, what does it take to deliver value over the long term? The flywheel looks a little like this:
The key point here is that there must be some genuine utility in the service that an audience is willing to ultimately pay for. Then flywheel improvement may well come from product feature improvements. However, more critically, it comes from capturing user input data and training the models which provide the service. The faster this flywheel spins, the better data and models will become. This benefits the users and will likely make those companies that are fastest at tapping into user demand (assuming they are able to build the right data engine) runaway leaders in their categories.
It’s important to note the difference between this user-driven model building and the model training which is being done by the likes of OpenAI and others at the foundational level. Foundational models are hugely expensive to train (though Stable Diffusion initial training costs were potentially only around $600,000, which is much less than what one would expect). Even at Stable Diffusion’s potentially low price point, given the skills needs, it is very unlikely it makes sense for most companies to build their own foundational model.
What many of the foundational models and services (OpenAI, Midjourney, Dale-E, Stable diffusion, etc.) are now enabling is the ability for many other companies to make that first step in value on the flywheel and begin serving a need. The next steps are up to you though:
- Deliver specific value.
- Reach people in order to use it.
- Learn from that usage to improve your own models.
- Build a data / learning engine which drives the flywheel as fast as possible.
In order to develop something really valuable, the long term objective has to be to reach more and more users and grow/learn the best models.
As a side note: Viet Le posits that Google Search will lose much in the new era of ChatGPT since it is also going to be strong in such AI solutions. I think the threat is bigger that it may it appear and it’s more to do with the risk that interfaces (and hence relationships with users) may now fragment more.
Images generated by Midjourney AI:
- /image a tiny robot doing a benchpress with a city backdrop –aspect 3:2