They're only modestly worse than text-davinci-003 in my experience, and you can finetune them cheaply to do simple tasks e.g. triage human input and decide wher

They're only modestly worse than text-davinci-003 in my experience, and you can finetune them cheaply to do simple tasks e.g. triage human input and decide where to send it. But yeah, if you can afford to run a larger model or pay OpenAI for gpt4 calls, that's gonna work a lot better. If you're relying on prompting for the 7B models IMO you're gonna have a bad time — they're mostly toys at that: interesting output but not consistently useful. But finetuning gets better results, and it's cheap to finetune.