It astounds me that we have AI shoulder into every aspect of our lives… except auto correct.
Edit case in point. I typed “shouved” and it corrected it to shoulder instead of shoved
Apple’s autocorrect is based on GPT, of chatGPT fame, but it’s the GPT2 model which came out around 2015. Back then it ran on supercomputers, now it fits in your pocket. ChatGPT uses the GPT4 model which requires really big supercomputers for now.
That’s to train it. Even the most modern version of AIs like deepseek or get can run on a raspberry pi.
Only if you distill it to the point of uselessness (well, even more useless). Otherwise you wouldn’t have headlines like this.
No? I litterally run a local deepseek model on a raspberry pi to use as an alexa-like personal assistant.
Yea I guess it works fine for anything that doesn’t require a real context window.
My phone will not let me type ‘was’ without trying to autocorrect it to SAS. I’ve told it not to suggest SAS multiple times, but no luck. Swiping always works, though, so it’s only when I press each key.
my phone occasionally decided to “correct” “that” to “thar”, it is deeply annoying
For me it corrects ‘that’ to ‘thst’ all the time, I wish the was some way to whitelist and blacklist words
you can technically unteach your autocorrect a word but you do have to catch it trying to “correct” it and longtap the culprit, which can be tricky when it only happens sometimes and you have a habit of typing fast
It just works!
This is the one use of LLMs I actually really want, smarter text prediction!
Whatever algorithm Google search uses is 1000% better than whatever GBoard does.
Sometimes I think heliboard uses the same algorithm
