How are LLMs so good but autocorrect is so retarded?
Mostly size and speed. The LLMs people are impressed by run on racks of GPUs in a datacenter and can take a second or more to answer. Autocorrect has to run on your phone, offline, on every keystroke, in milliseconds, without draining the battery, so it gets a tiny model by comparison.
Apple actually says this in its own iOS 17 announcement: autocorrect got "a transformer language model", same family of tech as LLMs, but "on-device" and built for word prediction. https://www.apple.com/newsroom/2023/06/ios-17-makes-iphone-more-personal-and-intuitive/
The other catch: autocorrect only sees the few words before your cursor and has to guess before you finish the sentence. An LLM reads your whole message first. If it keeps "fixing" the same word, adding it to your keyboard dictionary or text replacements usually stops it.
Apple actually says this in its own iOS 17 announcement: autocorrect got "a transformer language model", same family of tech as LLMs, but "on-device" and built for word prediction. https://www.apple.com/newsroom/2023/06/ios-17-makes-iphone-more-personal-and-intuitive/
The other catch: autocorrect only sees the few words before your cursor and has to guess before you finish the sentence. An LLM reads your whole message first. If it keeps "fixing" the same word, adding it to your keyboard dictionary or text replacements usually stops it.