[License-discuss] OpenMDW license

[Brian Behlendorf]

Hi all,

Asking about this here, rather than submitting to license-review, as I am 
not the license steward, nor am I (any longer) involved in the Linux 
Foundation.

About a month ago, the Linux Foundation released a license applicable to 
LLM models called the OpenMDW License:

https://openmdw.ai/license/

The website describes this as an "open source" license, yet I see no 
attempts (by reviewing the archives of license-discuss and license-review) 
by the authors to bring this to OSI for formal approval. Does anyone see 
anything in the license that would hinder such approval?

I'm personally unclear on the problem this solves. The purpose as stated 
on the OpenMDW's FAQ reads a bit like "but it goes to 11". It seems like 
one could have written a guide to distributing a model with existing 
permissive licensing, as typically the purpose of the software is not 
really relevant to its licensing.

It seems to be emphatic about things that don't need stating, like outputs 
of the model aren't covered by the license - but nor are "outputs" of gcc 
or LibreOffice.

It also doesn't require, as OSAID did, that the underlying data used to 
train the model weights be published. This is "fine" from a 
permissive-license POV - I imagine with some digging we can find 
permissive-licensed-works that contain binary blobs, and we've long 
accepted closed-source binary firmware updates as a part of the Linux 
kernel project. So it's still unclear to me that weights couldn't just be 
distributed under current permissive licenses.

The real stand-out portion for me, however, is the second-to-last 
paragraph, disclaiming responsibility for any IP rights that may some day 
be associated with the weights or other model materials, even if derived 
from data not included in the distribution. While the courts have not 
ruled decisively on this, you would not see AI companies signing deals 
with content companies to scrape their data if there wasn't at least some 
OIP risk involved in not doing so. It also seems to ignore that the global 
policy train seem heading in the direction of limiting the ability to 
disclaim liability in a software license, and that doesn't seem to have 
changed under the current US administration. The disclaimer seems 
extraneous, compared to current disclaimers in most permissive licenses. 
Furthermore to the degree that people rely on that disclaimer, it seems 
like it can create novel risks that OSS licenses are supposed to be 
mitigating rather than adding to. This is because if presented with a 
claim of infringement, there may be no way for the end-user or distributor 
to quickly modify the model by modifying the training set and rebuilding.

But none of these concerns are really about potential violations of the 
OSD.

Therefore, for the sake of clarity, if there isn't anything in this 
license that clashes with the OSD, I humbly suggest it should be proposed 
by the license steward and considered for approval. That way the use of 
the term "open source" on the OpenMDW pages is legitimized, and the AI 
community can be reassured that their "unique" needs are being met - even 
if OpenMDW is duplicative of existing permissive licenses. Right now I 
sense a schism emerging between generations that threatens to sideline OSI 
in the minds of developers, and this could be an bridge between the two.

Thoughts?

Brian