[orost]

I really want a high powered instruction-following LLM that's capable of assisting with writing fiction. GPT-3.5-turbo is fantastically capable at this if you can get it to drop its stilted AI-assistant voice, but doing so is too hard and too unreliable to be practical. I presume GPT-4 is even better but even harder to convince to try.

It's frustrating to see through those glimpses that this capability exists, but has been trampled over and made mostly inaccessible in pursuit of making sure the tool doesn't ever offend by producing something that's not baby food. And all the open replication efforts are aping this approach. Sigh.

[james-revisoai]

I agree.

It's like instruct = Gain cosistentcy, lose creativity. Base = No consistency (really), but creative.

Could a model be trained or positioned to be in the middle of these two?

[orost]

I suspect you could train a model to just shut up and follow instructions. I.e. instead of "Do X -> Sure! As a large language model, I'd love to help you with X!", just "Do X -> X".

This would avoid giving the model a dull default voice. But it wouldn't be sufficiently hedged and "safe".