The actually issue according to another comment [0] is this[1]:

> Around iOS 17 (Sept. 2023) Apple updated their autocorrect to use a transformer model which should've been awesome and brought it closer to Gboard (Gboard is a privacy terror but honestly, worth it).

> What it actually did/failed to improve is make your phone keyboard:

> Suck at suggesting accurate corrections to misspelled words

> "Correct" misspelled words with an even worse misspelling

> "Correct" your correctly spelled word with an incorrectly spelled word

Which makes me wonder: is Transformer model good with manipulating short texts and texts with errors at all ? It's kind of known that open weight LLMs don't perform well for CJK conversion tasks[2], and I've also been disappointed by their general lack of typo tolerances myself as well. They're BAD for translating ultrashort sentences and singled out words as well[3]. They're great for vibecoding, though.

Which makes me think, are they usable for anything under <100 bytes at all? Does it seem like they have a minimum usable input entropy or something?

0: https://news.ycombinator.com/item?id=47006171

1: https://thismightnotmatter.com/a-little-website-i-made-for-a...

2: The process of yielding "㍑" from "rittoru"

3: No human can translate, e.g. "translate left" in isolation correctly as "move left arm", but LLMs seem to be more all over the place than humans