[License-discuss] DIscussion on FMLL v1.0: A License for Free Artificial Intelligence

[Mihail Poteha]

Dear David Woolley,

Thank you for your feedback.

I believe that a license should protect the rights holders of datasets, and
withholding datasets is unacceptable (for example, they might contain
private data). Also, users have the right to know what data a model was
"trained" on. Since datasets and weights are not code, existing licenses
cannot affect this in any way.

As for reproducibility, I did not intend to make "AI Body" fully
reproducible in practice, which is why I used the term "Theoretical
Reproducibility" – this term is primarily meant to encourage dataset
openness. Perhaps I should refine the definition of "Theoretical
Reproducibility."

Sincerely,
Mikhail Potekha

пт, 17 июл. 2026 г., 06:29 David Woolley <forums at david-woolley.me.uk>:

> In general, I don't think it is close to being well drafted, but I'll
> pick some points specific to the application.
>
> On Thu, 2026-07-16 at 16:28 +1000, Mihail Poteha wrote:
> > I am an independent developer from Russia, and I am creating my own
> > AI license. To date, I have not found a license for Open AI. All
> > existing licenses, including OSAID, allow datasets to be withheld,
> > either partially or in full, thereby creating a “black box.” Such
>
> For most applications of AI training on datasets that you can freely
> redistribute will give them inadequate diversity to be useful in the
> real world.
>
> > licenses also fail to protect the rights holders of the datasets. I
> > disagree with this approach.
>
> Nothing in the main text stood out as achieving that, and it seemed to
> me that the licence only really works if none of authors of the
> training material intend to exercise any of their rights.
> >
> > Theoretical Reproducibility. A property of the model’s source file
> > that, in theory, makes it possible to recreate a complete copy of the
> > language model’s source file on the basis of the instructions and
> > tools specified in the AI Body. The Author must provide complete
>
> AI models are trained using random numbers, and training tends to be
> massively parallel.  Even with explicit seeds, programming languages do
> not specify the algorithm used to create the random sequences, so they
> could differ over time and between machines.  The parallelism also
> introduces the possibility of different scheduling orders.
>
> At inference time, simpler user interfaces won't include the ability to
> set seed values, and many users will not want exact repeatability.  For
> training that converges and inference that converges, you can only
> define statistical properties of the result, not exact values.
>
> I'm not sure if the OSD forces you to publish private modifications,
> but it is generally not part of the open source culture, and it is
> unenforceable, because it is happening in private.
>
> _______________________________________________
> The opinions expressed in this email are those of the sender and not
> necessarily those of the Open Source Initiative. Official statements by the
> Open Source Initiative will be sent from an opensource.org email address.
>
> License-discuss mailing list
> License-discuss at lists.opensource.org
>
> http://lists.opensource.org/mailman/listinfo/license-discuss_lists.opensource.org
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.opensource.org/pipermail/license-discuss_lists.opensource.org/attachments/20260717/9f0a6c40/attachment.htm>