[johnsimer]

Or at the very least would you have to finetune some 3B or 7B model to do this?

I want to create some locally hosted model that can do this in real time with a low memory footprint

Let people mash the keyboard at 2x-3x their typing speed and the unscramble it for them in real time to enable typing at 200-300wpm

[arnaudsm]

Contextual autocorrect has been on phones for a decade. It's still far from perfect. Many ambiguities are not trivial, even for an LLM

[MacsHeadroom]

Mistral-7B does this reliably with no special finetuning, while being 1000x smaller than GPT-4.