Hacker News new | ask | show | jobs
by indoorskier 1447 days ago
This GitHub Copilot thing is pretty polarizing. According to one of the expert papers they sollicited, it's unlikely MSFT/GitHub is breaking any copyright laws.

What they're doing does feel icky, and they could have mitigated a few concerns by a) making the inventory of the full training set public and b) at least attempt to attribute if there is a direct copy (which by their own admission happens about 0.1% of the time). These seem very simple steps they could take, and takes away the "shady behavior" argument.
1 comments
Klathmon 1447 days ago
>at least attempt to attribute if there is a direct copy (which by their own admission happens about 0.1% of the time).

IIRC they just block that suggestion if they detect it's a copy now

link
dotnet00 1447 days ago
More specifically it's an option, the VSCode Copilot extension asks you to configure it and takes you to a GitHub page where you can choose what happens if a suggestion is a direct copy.
link
pas 1447 days ago
what constitutes a direct copy? do they upload the generated code and use it as input for a full text search?
link