As competitors to OpenAI emerge, I think an obvious advantage would be response-codes for any kind of canned (non-completion) response.
"I'm sorry, but as an AI language model I can not do X, however" --> R982352.
You could also do that for requests. Because the requests will be parsed anyway, you could maintain a library of pre-parsed/tokenized requests for your user account and then not waste tokens on sending them to the API.
"You are a polite customer service agent who seeks to calm the customer and make the customer happy while prioritizing the needs of the company." --> S09438523.
This shit needs to die. I just want completions from a language model, log probs of the tokens, and more control over the generation.