Conversation

Notices

1. Embed this notice
   Greg Wilson (gvwilson@mastodon.social)'s status on Wednesday, 27-Nov-2024 02:51:59 JST Greg Wilson

   GitHub Copilot

   • Haelwenn /элвэн/ :triskell: and snacks like this.
   • Skadi and Kevin Beaumont repeated this.
   • Embed this notice
     Dr. Anna Latour (anna@mathstodon.xyz)'s status on Wednesday, 27-Nov-2024 02:51:59 JST Dr. Anna Latour

     in reply to

     @gvwilson

     Okay so I tried this, and initially it didn't look so bad, until I hit enter again and copilot came up with a new suggestion...

     I then deleted the test file, created a new empty file, and opened it with a different editor to see what would happen if I prompted it with "calculateManSalary(baseSalary". Got an opposite result, of sorts.

   • Embed this notice
     Anže :python: (anze3db@fosstodon.org)'s status on Wednesday, 27-Nov-2024 03:40:12 JST Anže :python:

     in reply to

     @gvwilson I've tried it in Python and it's even worse (0.77) 🙈

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Wednesday, 27-Nov-2024 10:17:44 JST Rich Felker

     in reply to

     @gvwilson This is indicative not only of discriminatory garbage in LLM training sets, but of cross domain contamination that's inherent in the models being overly large. Even if it were correct or valid, this kind of "social convention knowledge" does not belong anywhere near "programming knowledge". Copylot mixing that up is a huge disservice.