Original: 2022-09-17
Updated: 2022-09-17
There has been a lot of conversation lately about ML models being the goal. Let's talk about that.
Recent works such as ACT-1 by Adept are proclaimed as huge advances - and they are! But these are research advances. The version of ACT-1 that will revolutionze knowledge workers is not one which requires user input. For the same reason, CoPilot works so well precisely because it does not require explicit user input - the suggestions are instead generated based on the context implied from their work environment, which in this case happens to consist of source code and comments.
Consider cloud computing in the 2010s. The common thread is that the market focuses on sensationalizing the provider (AWS, GCP, Azure, etc.), but tends to ignore that they serve as a strong foundation for the products built upon them. Take a moment to consider how many companies wouldn't have been possible without the incredible iteration speed that cloud providers enabled. Products like Netflix ($107Bn), Stripe ($95Bn), AirBnB ($75Bn), or Figma ($20Bn) were all built on AWS alone. One would be wise to consider whether the economics for these companies would have allowed for the requisite pace of iteration in the early days, had AWS not existed. So many products exist today thanks to rapid feedback loops between customers and development teams, and cloud infrastructure enabled that rapid iteration.
Similarly, machine learning models are going to enable an entire generation of products, most of which we probably can't even imagine today. The innovation will lie at a layer above the models. Much like how building on AWS is not a competitive advantage between different SaaS companies today, building on GPT-3 or other large models will not be a competitive advantage between products tomorrow. What will are the layers built on top, that solve problems real people face.
The model is important, because it enables products, but the model itself is not the product we should be focusing on.
This is a call to build. It is rare to experience such a dramatic opportunity to radically re-think how we create value and improve lives through software. Never before have we yielded the immense power of language, nor the power of machine creativity, in the way we do today. We are entering an era where computers fundamentally understand people, instead of people exclusively understanding computers.
You read this far? Cool! You must be excited to get building. Why not build with me so that we can enable a billion people to build useful software?