HN Explorer

GitHub Copilot has been trained on private repos?

4 points by LunicLynx ↗ HN
So copilot just gave me this link:

https://stackoverflow.com/a/10445101/828022

While i was writing this comment:

// based on th [cursor here]

The interesting thing is, this url can't be found via github search. Not parts nor the whole url. Which leads me to belief that either the relevant repository was deleted or it is private.

5 comments

[ 2.6 ms ] story [ 82.1 ms ] thread
florescudev
You should leave more context for this. Is the SO topic related to your current context where did you used copilot?
orbz
Two things here: the numbers at the end are random. You can replace with anything else and get the same answer.

The other is that it might have been trained from a corpus of web scrape data in addition to public GitHub repositories.

cratermoon
> the numbers at the end are random.

In which case I don't believe Copilot is offering a relevant link based on any information in the context, but just generating a link that looks like a SO link, based on pattern matching https://stackoverflow.com/a/[0-9]+/[0-9]+

al2o3cr 

    just generating a link that looks like a SO link, based on pattern matching
That's exactly what it's doing: it's an LLM. That's what they do.
cratermoon
You and I agree there, but people out there are declaring LLMs have "sparks of AGI" and self-awareness.