5 ms·

It only seems to if you give it no or very little "source" input, like an empty file with a comment that says "// X algorithm". There's been a lot of bikeshedd

by meibo 5y ago

It only seems to if you give it no or very little "source" input, like an empty file with a comment that says "// X algorithm". There's been a lot of bikeshedding on this, but GitHub decidedly hasn't given enough information on how it works and what the training dataset is, and the fair use question definitely needs to be answered, maybe even in court - it's just a matter of time.

Hamuko 5y ago
>what the training dataset is All non-private repositories on GitHub.
- brutal_chaos_ 5y ago
  Is it though? I'd assume that too, but we don't really know, do we? I mean to ask, what have Microsoft stated that leads you to believe this? (Maybe there's a press release I missed?)
  - teraflop 5y ago
    https://twitter.com/NoraDotCodes/status/1412741339771461635 https://twitter.com/NoraDotCodes/status/1412741339771461635
    - ben0x539 5y ago
      (this doesn't say they didn't use other sources too, though)
      - tyingq 5y ago
        "source code from publicly available sources, including code in public repositories on GitHub." (from the copilot home page) Pretty much says they did also use other sources.
MadVikingGod 5y ago
What I think is interesting and not talked about in the copilot cases is there is actually two different copyright actions that come from using it. The first is if GitHub had a license to distribute the code they have learned on. This is assuming that the code it produces is a derivative, which I don't think would be much of a stretch. For this I think most of the licenses that are used, GPL MIT etc, allow for someone to distribute in such a way. The second is the user of copilot. They would be getting and distributing code where they wouldn't know the original license, but that wouldn't be a viable defense of infringement. To actually comply with most licenses they would have to follow the requirements. In both cases I don't know if a fair use would really apply. Maybe GitHub could stretch the research aspect, but if you just use the code in your product there is no fair use.
- wlesieutre 5y ago
  > This is assuming that the code it produces is a derivative, which I don't think would be much of a stretch I don't think that's easily settled. Is the code I write a derivative work of all prior code that I've learned from? The difference (IMO) is that Copilot has a much more precise memory than I do; I couldn't recall the exact text of a code example that I've encountered repeatedly even if I wanted to.
  - ric2b 5y ago
    > I couldn't recall the exact text of a code example that I've encountered repeatedly even if I wanted to. If you could it would be copyright infringement anyway.
- orwin 5y ago
  easy solution: license all the code you write with copilot under AGPL or GPLv3, and you're good.
  - cassonmars 5y ago
    Not quite, there are far more aggressive licenses on repositories on GitHub, such as SSPL, which in theory by providing the SSPL licensed code, Co-Pilot should be open sourced under the same license to comply.

by meibo 5y ago

Hamuko 5y ago

>what the training dataset is All non-private repositories on GitHub.

brutal_chaos_ 5y ago
Is it though? I'd assume that too, but we don't really know, do we? I mean to ask, what have Microsoft stated that leads you to believe this? (Maybe there's a press release I missed?)
- teraflop 5y ago
  https://twitter.com/NoraDotCodes/status/1412741339771461635 https://twitter.com/NoraDotCodes/status/1412741339771461635
  - ben0x539 5y ago
    (this doesn't say they didn't use other sources too, though)
    - tyingq 5y ago
      "source code from publicly available sources, including code in public repositories on GitHub." (from the copilot home page) Pretty much says they did also use other sources.

MadVikingGod 5y ago

What I think is interesting and not talked about in the copilot cases is there is actually two different copyright actions that come from using it. The first is if GitHub had a license to distribute the code they have learned on. This is assuming that the code it produces is a derivative, which I don't think would be much of a stretch. For this I think most of the licenses that are used, GPL MIT etc, allow for someone to distribute in such a way. The second is the user of copilot. They would be getting and distributing code where they wouldn't know the original license, but that wouldn't be a viable defense of infringement. To actually comply with most licenses they would have to follow the requirements. In both cases I don't know if a fair use would really apply. Maybe GitHub could stretch the research aspect, but if you just use the code in your product there is no fair use.

wlesieutre 5y ago
> This is assuming that the code it produces is a derivative, which I don't think would be much of a stretch I don't think that's easily settled. Is the code I write a derivative work of all prior code that I've learned from? The difference (IMO) is that Copilot has a much more precise memory than I do; I couldn't recall the exact text of a code example that I've encountered repeatedly even if I wanted to.
- ric2b 5y ago
  > I couldn't recall the exact text of a code example that I've encountered repeatedly even if I wanted to. If you could it would be copyright infringement anyway.
orwin 5y ago
easy solution: license all the code you write with copilot under AGPL or GPLv3, and you're good.
- cassonmars 5y ago
  Not quite, there are far more aggressive licenses on repositories on GitHub, such as SSPL, which in theory by providing the SSPL licensed code, Co-Pilot should be open sourced under the same license to comply.