A way to think about finetuning v. prompt eng (+ any technique w prompting like RAG) that I've been playing with:

Who takes on more of the burden of nudging the model to produce consistent good outputs for a domain - the model consumer at prompting time or the model trainer via examples at training time?

Tinkering. Co-Founder @ Livepeer. https://www.yondonfu.com/

how does zero shot/one shot/n-shot prompting fit into this? whats the tipping point for where the burden should ideally lie?

fan of the scientific method; enjoyer of Socratic Dialogue; never been wrong in my life; btw what is a 'sarcasm'?

It seems to me like the onus should be on the developers to minimize the dependency of the models on n-shot prompting. Do you know of any good measures for consistency req?

Seems that n-shot prompting puts burden on model consumer but then it becomes a question of how much burden the app takes on with prompt templates vs. the end user.

Good q re: tipping point - prob depends on domain and maybe discovered by pushing n-shot prompting until results fail to meet your consistency req.