This is a great point. It appears the success of OpenAPI has validated their approach, specifically around (1) using the web as a training set and (2) using transformers.

I imagine a lot conversations with in-house AI folks is around deploying these methods.