AFAICT, there is no authoritative information from OpenAI on how ChatGPT plugins work (please correct me if I'm wrong). However, is there any interesting external analysis on how it could be working under the hood?

I am aware of two approaches to building something like this: Either through a REPL-like prompt chaining approach (like langchain), or a toolformer-like approach where tool use becomes a trainable choice. Any sources/speculations on which it could be (or something else)? What do you all think? Thanks.