My cynical mind says they broke autocomplete to encourage voice to text which is a much more rich data mining avenue than text only. Think of what all information about yourself can be determined just from your voice, which passes through Google servers when you use their voice to text and they most definitely keep.
-Initially it was terrible, but it was not imposed on you, so you just ignored it and kept typing as if it weren’t even there, unless you legit could not figure out how to spell a word.
-Then it got better and gave decent suggestions occasionally, so people started integrating into their typing workflow as an assistant to quickly complete common words, or to figure out longer/more complex words.
-Somewhere along the way it began to impose itself on your typing to the point where it was expected that you would accept its suggestions the majority of the time. So even if it’s right 98% of the time, having to manually fight with it every other sentence is a MAJOR hassle
As a simple illustration of the last point, I remember the default behavior was that if you started to backspace an autocorrected word, the system assumed it got it wrong and let you fix it. Now the default behavior is to ignore you and keep autocorrecting until you tap a special key to insist you know what you are doing (at least on my Samsung phone)
Just a shot in the dark here, but maybe they aren’t accounting for code switching between different apps on your phone and that is skewing your prediction algorithm. You’re phone is guess at what word you’re typing based on everything you’ve ever typed and not just what you text in the messenger app.
I remember a time when autocorrect was terrible then it became good and then, terrible again. What happened?
They switched to using gpt2 lol
I’m not fluent in neural network systems, but I thought it was NLP in the early 10s and not GPT.
Yea, then they switched.
When? I don’t see that any where.
I may be mixing up correction and prediction, although I think the two are related. https://jackcook.com/2023/09/08/predictive-text.html
That makes sense. I’m not sure how they are related, but it’s not my field. I wonder what prediction used before gpt.
This was an interesting read.
My cynical mind says they broke autocomplete to encourage voice to text which is a much more rich data mining avenue than text only. Think of what all information about yourself can be determined just from your voice, which passes through Google servers when you use their voice to text and they most definitely keep.
I don’t ever remember it being good, and I predate cell phones.
I think what happened is:
-Initially it was terrible, but it was not imposed on you, so you just ignored it and kept typing as if it weren’t even there, unless you legit could not figure out how to spell a word.
-Then it got better and gave decent suggestions occasionally, so people started integrating into their typing workflow as an assistant to quickly complete common words, or to figure out longer/more complex words.
-Somewhere along the way it began to impose itself on your typing to the point where it was expected that you would accept its suggestions the majority of the time. So even if it’s right 98% of the time, having to manually fight with it every other sentence is a MAJOR hassle
As a simple illustration of the last point, I remember the default behavior was that if you started to backspace an autocorrected word, the system assumed it got it wrong and let you fix it. Now the default behavior is to ignore you and keep autocorrecting until you tap a special key to insist you know what you are doing (at least on my Samsung phone)
Just a shot in the dark here, but maybe they aren’t accounting for code switching between different apps on your phone and that is skewing your prediction algorithm. You’re phone is guess at what word you’re typing based on everything you’ve ever typed and not just what you text in the messenger app.