how can one avoid building a chatgpt wrapper?

Everything in LLMs is built on top of a base model.

There are just different strategies to achieve what you need.

I guess what people call a "wrapper" is when prompt engineering is used.

Maybe tuning will get you away from the "wrapper" moniker?

The open web is dead, long live the open web. ⌐◨-◨

Building /warpcast and /farcaster

— gabrielayuso.com

I've been wondering this question:

How much "data" is needed to tune the base model so you have something proprietary?

built projects in architecture, video games, web 1 & 2 | currently building mettle.coach | playing in web3

It depends on the use case, both the nature of the data and how the tuned model will be used.
You'll have to try different methods and check what produces the best outcomes for your needs.

In many cases just a few-shot prompt over the base model will do. Might be slow and expensive though.