Hacker News Reader

Comment by andrewallbright | Hacker News Reader

This has probably been talked about but...

If most code is "bad" code (any definition works) and this AI was trained on all/most code on GitHub, does that mean that this AI mostly helps to produces bad code?

View on HN

It definitely can. Here it suggests a plausible looking but incorrect function for averaging integers: https://twitter.com/ridiculous_fish/status/14527512360594513...

I have been using copilot for some time... I'd say yes and no. It helps you a lot when you are writing repetitive code, so in a way it encourages you to write the repetitive BS instead of making a function for that or something. But it's also helpful for writing tests and nice error message. You just type

     if (x.length < 10) throw

And it figures out the rest. So while sometimes it encourages bad code, when you know how to use it well, it helps you write the good things I'd normally be too lazy to write

It depends. I have not collected data to prove my observations, but I find the rust suggestions better quality on average than the python suggestions. Some people do terrible things in Python.

one way to minimize this is to train on your own trusted code. You do need a reasonable amount of good code (ideally with good comments too) this is one of the options that we have here at Tabnine. Train on your GitLab, Bitbucket or GitHub repos.

Reply on news.ycombinator.com