Hacker News new | ask | show | jobs
by brap 1009 days ago
With everyone talking about LLMs being glorified autocomplete, I actually would like to see how well they perform as autocomplete. Because most built-in ones are pretty bad.
3 comments

I’m assuming this is how whisper works.

For some evidence, say a sentence, then say a sentence with the words scrambled. The performance nose dives, relative to audio levels.

That's essentially what GitHub Copilot is.
I mean classical autocomplete, the ones we have on our phones that give us one word as a time (on iOs you get 3 options to choose from)
iOS 17 is basically doing that: https://www.macrumors.com/guide/ios-17-keyboard/
stupid chatgpt...could be the new "stupid autocorrect"

I used to love those posts.