4 ms·

I think that’s mostly my point. GP release source code, presumably under a license that permitted this use. It’s okay to not be happy that it was used this way…

I think that’s mostly my point. GP release source code, presumably under a license that permitted this use. It’s okay to not be happy that it was used this way… but that’s life

lvass 3y ago
No license except CC0/unlicense permits laundering authorship like Copilot does.
- rectang 3y ago
  Yes, this is exactly true among the popular FOSS licenses. The attribution licenses such as ALv2, BSD, MIT, etc. which are generally considered to be "permissive" still require attribution, and this condition is not upheld by popular LLMs. Of course, the copied material has to be judged substantial enough for the license to apply in a court of law, which is a human-arbitrated threshold. It will most likely be a copyleft author who eventually brings a court case, but the attribution licenses do not allow for LLM usage either unless the LLM provides attribution and preserves notices as spelled out in the licenses.
bee_rider 3y ago
It is a new angle of attack that the authors of the code didn’t know was coming. I think it is pretty obvious that lots of GPL authors wouldn’t want their code to be productized in ways that don’t contribute back to the community, because they had the ability to use a BSD or MIT license and didn’t take it.
- moonchrome 3y ago
  If you don't want people to learn from your code then don't open source it. Learning via LLM vs person learning is a distinction that sounds unenforceable in practice.
  - bee_rider 3y ago
    The stealing open source code is generally extremely difficult to prevent in practice; nobody knows by default if you copy-paste some open source code into a closed code base. But this shouldn’t prevent us from calling out bad behavior when we do see it. Legally, of course, LLMs are not people. They don’t have the same rights as people, and it isn’t obvious whether or not they can legally generate new IP. They operate by a complicated but essentially mechanical process, and it is pretty novel to say that such a process could be used to remove copyrights.
  - rectang 3y ago
    There will eventually be a court case where an LLM reproduces creative code verbatim which clears the de minimis threshold. For example: https://twitter.com/mitsuhiko/status/1410886329924194309 https://twitter.com/mitsuhiko/status/1410886329924194309 But the defendant won't be Microsoft — they just provided a tool, and that's legal. No, the defendant will be the downstream consumer who incorporated the code spat out by the LLM. It doesn't matter whether the LLM "learned" — intent is irrelevant and the defendant will have committed copyright violation regardless. The LLM can't copyright-wash the code — nobody can.

I think that’s mostly my point. GP release source code, presumably under a license that permitted this use. It’s okay to not be happy that it was used this way… but that’s life

lvass 3y ago

No license except CC0/unlicense permits laundering authorship like Copilot does.

rectang 3y ago
Yes, this is exactly true among the popular FOSS licenses. The attribution licenses such as ALv2, BSD, MIT, etc. which are generally considered to be "permissive" still require attribution, and this condition is not upheld by popular LLMs. Of course, the copied material has to be judged substantial enough for the license to apply in a court of law, which is a human-arbitrated threshold. It will most likely be a copyleft author who eventually brings a court case, but the attribution licenses do not allow for LLM usage either unless the LLM provides attribution and preserves notices as spelled out in the licenses.

bee_rider 3y ago

It is a new angle of attack that the authors of the code didn’t know was coming. I think it is pretty obvious that lots of GPL authors wouldn’t want their code to be productized in ways that don’t contribute back to the community, because they had the ability to use a BSD or MIT license and didn’t take it.

moonchrome 3y ago
If you don't want people to learn from your code then don't open source it. Learning via LLM vs person learning is a distinction that sounds unenforceable in practice.
- bee_rider 3y ago
  The stealing open source code is generally extremely difficult to prevent in practice; nobody knows by default if you copy-paste some open source code into a closed code base. But this shouldn’t prevent us from calling out bad behavior when we do see it. Legally, of course, LLMs are not people. They don’t have the same rights as people, and it isn’t obvious whether or not they can legally generate new IP. They operate by a complicated but essentially mechanical process, and it is pretty novel to say that such a process could be used to remove copyrights.
- rectang 3y ago
  There will eventually be a court case where an LLM reproduces creative code verbatim which clears the de minimis threshold. For example: https://twitter.com/mitsuhiko/status/1410886329924194309 https://twitter.com/mitsuhiko/status/1410886329924194309 But the defendant won't be Microsoft — they just provided a tool, and that's legal. No, the defendant will be the downstream consumer who incorporated the code spat out by the LLM. It doesn't matter whether the LLM "learned" — intent is irrelevant and the defendant will have committed copyright violation regardless. The LLM can't copyright-wash the code — nobody can.