Re: The beast is live - Github Copilot

[Yuchen Pei]

On Sat 2022-06-25 13:47:24 -0700, Marc Sunet wrote:

>> Is copilot something only github can offer?  If the vast majority of
>> training data is from public repo that anyone can access, perhaps the
>> only thing stopping people from making something similar is the
>> computing power to convert the data to weights?
>
> Well, I guess that and your will to comply with law, right?

I'm just speculating whether it is feasible for free software orgs like
framasoft or software heritage to develop free software version of
copilot, that trains on data that is code under GPL-compatible licenses
and license the "free copilot" under (A)GPL.

>
> People's code on Github is copyrighted/copylefted, you can't just use
> the code however you please. And I am not too familiar with the Github
> TOS, but I doubt they can override a license. The TOS could also make
> scraping illegal, making it illegal for you to obtain the data en
> mass, but not for Microsoft. Facebook and Instagram do this.
>
> There is also the argument of whether GPL code output by a network
> should be considered a derivative work. After all, the network is
> basically just a different representation of its training inputs. I
> believe the FSF started working on this issue the moment Github
> Copilot was announced; I thought Copilot finally going live was of
> interest to the audience here.
>
> Legal issues aside, I'd argue it's also kind of shitty to take other
> people's work and re-sell it under a $10/$100 subscription, with the
> abuse of confidence and trust that that comes with. Microsoft bought
> the platform when it was already popular, then subverted it for its
> own profit-making. Similar to Google/Youtube.

I feel that one difference between the values of free software and open
source is that the free software does not have a problem with taking
people's work and re-selling it, as long as the work being resold is
free.  Whereas open source focus more on what authors get out of "open
source".

>
> Marc
>


Best,
Yuchen

-- 
PGP Key: 47F9 D050 1E11 8879 9040  4941 2126 7E93 EF86 DFD0
          <https://ypei.org/assets/ypei-pubkey.txt>