They need to launch a model that includes knowledge of API documentation / libraries at the same time they launch the API. Two months back GPT could only provide irrelevant code.
Or use any model with large enough context limit to include relevant, or whole sale API documentation. No need to pre-train, or fine tune API stuff.
Large context limits really allow these models to do some crazy stuff. Ive been feeding the smaller gpt4 directions to coding repositories like ss64.com and getting very positive results in coherency, and reduced hallucination of code.
