On 8/10/26 7:32 PM, Theodore Tso wrote:
On Mon, Aug 10, 2026 at 05:35:33AM -0500, Pirate Praveen wrote:1. Public AI is AI as public infrastructure like highways, water, or electricity. An example of Public AI is https://publicai.co...Actually, publicai.cois not only an "example" but the *name* of the open-source project hosted at https://publicai.co. As such, your trying to use Public AI has a term of art, when its pre-existing use is as the name of an open-source project which you are trying to use as an example is likely to cause... confusion.
My idea was to base this definition on how Public AI actually works to a more easily understood general criteria. I have now made that clear in v3.0 of the draft,
"So to evaluate an AI web service, we will use the definition of Open Source AI by OSI, but some requirements may be relaxed."
...For AI to be accessible to people, it has to be available easily to people and not for profit collaborations between multiple organizations and public funding is essential to keep it sustainable.3. At this point, we don't want to insist on training data being freely available, similar to how we don't insist on Free Hardware Designs.So you *could* use the Open Source AI definition from the OSI institute, but it's not an overlap with what you are going for, since (a) the OSAID doesn't require that the LLM be publically funded, which you stated in your point (1) --- but since it's not clear whether is the key part of your definition of "Public AI", you might not consider this to be a problem, and (b) the OSAID has a much stricter set of requirements that would disallow Open Weight models that would otherwise be allowable.
I have clarified that point in v3.0. Public funding may be desirable, but not a requirement.
"For AI to be accessible to people, it has to be available easily to people, like Wikipedia or Lets Encrypt. Not for profit collaborations between multiple organizations and public funding is essential to keep it sustainable, though public funding is not a requirement. How we are going to fund such Public AI will evolve in the future."
So for example, some of the most commonly used Open Weight models such as Qwen 3 would run afoul of the OSAID requirements. And there are very few models indeed aside from the Apertus that would meet your publicly funded requirement. An example of a model that would meet the OSAID requirement might be the T5 model from the 2019 Google Paper, "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transforme", but which wasn't publically funded. It also isn't actively being developed anymore, since it was an LLM that was created just to explore some research ideas. But I'm not sure it really matteres if an academic LLM was funded by a company or by a grant from a government agency.2. Public AI must use Free Software models.If we strip out your publically funded model, then you could just simply use the word "Open Weights", although that's not exactly formally specified as a definition either. So for example, for Debian, do we care that the LLaMa has a restriction on commercial use? (After all, it's not at all applicable to how Debian would use the LLM.) What about the restriction in the LLaMa license which forbids using the model outputs to train competing LLM's? etc.
I have now used Open Weights as well.
In contrast, we might have the Gemma 4 model, which is released under the Apache license, which would allow its use to create derivitive LLM's, but (oh, horrors!) it was funded and released to the world by Google. (Again, I'm not sure why having an LLM funded by government agency makes it any difference from an LLM funded by a company. Ask someone who used to receive an NSF grant how capricious and subject to change government grant funding can be.)
I have clarified this as well.
This is why having a formal specification and definition of what is allowed is important, and why I don't think your proposal is quite ready for prime time. If it were passed, the amount of confusion and debate caused by the fact that the definition is not crisp and precise would be harmful to the Debian project, IMHO.
Does "Open Source AI definition with requirement for availability of training data relaxed" removes ambiguities?
I have also mentioned about preserving copyleft - though I'm not fully sure if we need that as how copyright law evolves is not really in our control and I could remove that requirement if people feel that should be removed as well.
Best regards,
- Ted
OpenPGP_0x8F53E0193B294B75.asc
Description: OpenPGP public key
OpenPGP_signature.asc
Description: OpenPGP digital signature
