how can one avoid building a chatgpt wrapper?

For LLM, grab a model from 🤗 and fine tune the head. But I still think you should make a wrapper that is interoperable with chatgpt, just in case they drop a v4.5 that blows everything away.

I just mean to abstract away from the model in code, that way you can plug-and-play with different models—including one that wraps the chatgpt api.

i see. yes, agree. also the notion of the user 'picking the model' and having control over which to use seems to be a nice track.

although i wonder if in 10+ years it's gonna be one mega base model, or every human with its own local model. like a tinybox.

That’s the question! I’m hoping for Hotz’s bet to pay out. 

Even if true that error rates decrease only by increasing model parameters, we just need a point were the error rates are sufficiently low to not justify increased spending (usually better-than-human performance).

Then we wait for Huang's law.

i see. yes, agree. also the notion of the user 'picking the model' and having control over which to use seems to be a nice track.

although i wonder if in 10+ years it's gonna be one mega base model, or every human with its own local model. like a tinybox.
https://tinygrad.org/