Ask HN: Is it just me, or GitHub Copilot is much worse than it was last year?

2 points by huqedato ↗ HN

2 comments

[ 3.5 ms ] story [ 12.4 ms ] thread
This is the kind of question that calls for a text body expanding on the premise. Much worse in which ways? At what tasks? In which languages?

Since you didn't provide any concrete examples, I'll just note that my experience has been the opposite. I have found Copilot to be better and better as time goes on. In particular, it's gotten better and better at learning from the context of the project as a whole.

As just one example: I was working on a Python script which needed some icons as external files. I dragged the icons into a folder and then, when it came time to wire them up into the python script, I prompted Copilot on a whim and was surprised to find that it used the correct directory name. I had never seen something like that before—not only does copilot now have context for code in other files in the project, it has information about the directory structure! Improvements like this are what keep making it more useful, far more so than improvements to the model.

It's my experience based on coding in Elixir, Julia and JS. Particularly for Elixir suggestion is getting worse and worse. I spent a lot of time to re-re-refine and re-ask a question. Buggy code. Hallucinations. (instead of Elixir code it's suggesting snippets in Erlang and Haskell). Probably it's RLHF... In summer gave me brilliant resolutions, quickly, without "insistence". Now is getting dumber. Senile?