Not sure if this is the ideal section… This was just a random thought and may be technically infeasible. Would it be possible to assist in the creation of new skills by using GPT-3/GPT-J/PROSE/Openai Codex to attempt to use consume api docs and render new skills?
Hi @kratosok,
Welcome to the community!
This sounds certainly like an interesting research direction. Generally speaking, the biggest hurdle of developing a skill is acquiring training data for the NLU part. With the Genie toolchain, that boils down to preparing API schemas annotated with a few natural language phrases for each input/output parameter, because the tool can synthesize training data from it.
In our AutoQA paper we showed how you can generate the natural language phrases associated with a schema using a pretrained BART paraphraser. I suggest you read this paper, as it seems very close to what you’re suggesting.
As a follow up, it would be quite interesting to explore whether a different model, perhaps a GPT3, would fare better or worse at the task of building a skill given the schema.