3 ms·

Curious to know if this extends to LLMs and if so how they would define open source. Specifically it would be nice to see repudiation of Meta's "Open" BS by a n

by andy99 1y ago

Curious to know if this extends to LLMs and if so how they would define open source. Specifically it would be nice to see repudiation of Meta's "Open" BS by a nation state.

b112 1y ago
I wouldn't call data "source", whether a book, a sound track, video, etc. In my view of the world, the code to train, the software to run, that's open source joy. Now... should the trained, and vectored data be free? Maybe so. But I bet this UN thing doesn't cover that.
- andy99 1y ago
  I didn't call the data the source and in the past have explicitly argued that training data is not necessary to exercise the freedoms normally associated with open source. Llama models have usage restrictions that go against any mainstream definitions of open source.
  - b112 1y ago
    The model is part of the data, agreed? Anyhow, I wasn't trying to put words in your mouth, simply stating my thoughts. And I focused on data because I see OSS code everywhere, so presume there is no issue there.
shakna 1y ago
They're just those eight guidelines. Not particularly precise, with intent mattering more than any definition. This isn't a policy, just a goal.
bzg 1y ago
https://www.comparia.beta.gouv.fr/modeles https://www.comparia.beta.gouv.fr/modeles compares models and Llama different licenses are not mislabeled as "open source". Also, https://opensource.org/ai/endorsements https://opensource.org/ai/endorsements shows code.gouv.fr in the list.
- andy99 1y ago
  Cool, thanks! Cette licence permet d'utiliser, reproduire, modifier et distribuer librement le code avec attribution, mais impose des restrictions pour les opérations dépassant 700 millions d'utilisateurs mensuels. Interesting they only mention the 700 million users thing and not the other restrictions on use. Personally I could regard the prohibition against basically Google and Microsoft using it to be a minor transgression, it's the larger list of unacceptable uses that's the big problem.
  - bzg 1y ago
    Agreed. If you feel the need to report inconsistencies, the source code is here: https://github.com/betagouv/ComparIA https://github.com/betagouv/ComparIA
  - drexlspivey 1y ago
    It's actually targeting Apple since both Google and Microsoft have their own models.
    - Zambyte 1y ago
      I'm a little confused by the context of this since I don't speak French, but it seems like you're unfamiliar with the OpenELM family of models. https://huggingface.co/apple/OpenELM https://huggingface.co/apple/OpenELM
      - drilbo 1y ago
        that looks like a dinosaur now
- pabs3 1y ago
  OSI's OSAID is complete and utter bullshit. Look to Debian if you want a real definition of open source AI. https://opensource.org/ai https://opensource.org/ai https://salsa.debian.org/deeplearning-team/ml-policy https://salsa.debian.org/deeplearning-team/ml-policy
BlueTemplar 1y ago
Related : https://elevenfreedoms.org/ https://elevenfreedoms.org/ > The traditional Four Freedoms of free software are no longer enough. Software and the world it exists in have changed in the decades since the free software movement began. Free software faces new threats, and free AI software is especially in danger.

by andy99 1y ago

Curious to know if this extends to LLMs and if so how they would define open source. Specifically it would be nice to see repudiation of Meta's "Open" BS by a nation state.

b112 1y ago

I wouldn't call data "source", whether a book, a sound track, video, etc. In my view of the world, the code to train, the software to run, that's open source joy. Now... should the trained, and vectored data be free? Maybe so. But I bet this UN thing doesn't cover that.

andy99 1y ago
I didn't call the data the source and in the past have explicitly argued that training data is not necessary to exercise the freedoms normally associated with open source. Llama models have usage restrictions that go against any mainstream definitions of open source.
- b112 1y ago
  The model is part of the data, agreed? Anyhow, I wasn't trying to put words in your mouth, simply stating my thoughts. And I focused on data because I see OSS code everywhere, so presume there is no issue there.

shakna 1y ago

They're just those eight guidelines. Not particularly precise, with intent mattering more than any definition. This isn't a policy, just a goal.

bzg 1y ago

https://www.comparia.beta.gouv.fr/modeles https://www.comparia.beta.gouv.fr/modeles compares models and Llama different licenses are not mislabeled as "open source". Also, https://opensource.org/ai/endorsements https://opensource.org/ai/endorsements shows code.gouv.fr in the list.

andy99 1y ago
Cool, thanks! Cette licence permet d'utiliser, reproduire, modifier et distribuer librement le code avec attribution, mais impose des restrictions pour les opérations dépassant 700 millions d'utilisateurs mensuels. Interesting they only mention the 700 million users thing and not the other restrictions on use. Personally I could regard the prohibition against basically Google and Microsoft using it to be a minor transgression, it's the larger list of unacceptable uses that's the big problem.
- bzg 1y ago
  Agreed. If you feel the need to report inconsistencies, the source code is here: https://github.com/betagouv/ComparIA https://github.com/betagouv/ComparIA
- drexlspivey 1y ago
  It's actually targeting Apple since both Google and Microsoft have their own models.
  - Zambyte 1y ago
    I'm a little confused by the context of this since I don't speak French, but it seems like you're unfamiliar with the OpenELM family of models. https://huggingface.co/apple/OpenELM https://huggingface.co/apple/OpenELM
    - drilbo 1y ago
      that looks like a dinosaur now
pabs3 1y ago
OSI's OSAID is complete and utter bullshit. Look to Debian if you want a real definition of open source AI. https://opensource.org/ai https://opensource.org/ai https://salsa.debian.org/deeplearning-team/ml-policy https://salsa.debian.org/deeplearning-team/ml-policy

BlueTemplar 1y ago

Related : https://elevenfreedoms.org/ https://elevenfreedoms.org/ > The traditional Four Freedoms of free software are no longer enough. Software and the world it exists in have changed in the decades since the free software movement began. Free software faces new threats, and free AI software is especially in danger.