Re: Open Source Initiative: Open Source AI Definition rc1 published
A couple more thoughts, having followed more of the recent discussion
about this:
I'd anticipate and/or have encountered two forms of
fear-uncertainty-doubt misinformation in response to arguments (that I
agree with) that say that an open label also requires open training
data:
* A misinformation claim that "it's not possible to produce the same
results during large model training" -- I think this could be
countered by demonstrating that it's possible to retrain a nontrivial
model for something like the Opus codec deterministically (I haven't
yet confirmed doing so, though).
* A misinformation claim that "because the OSI allows/encourages
reproducibility requirements to be added alongside OSAID rc1/rc2 when
relevant, the open label isn't lowering the bar" -- I think this is
untrue, because I believe that any training data that inadvertently or
intentionally included (non-pseudo)randomness would be
difficult-if-not-impossible to rebuild, certainly to the extent that
it would exclude good-faith contributors with limited resources.
(I don't think I'm stating anything new or previously-unsaid with
these thoughts, but wanted to express them)
Regards,
James
On Sun, 20 Oct 2024 at 16:35, James Addison <jay@jp-hosting.net> wrote:
>
> Hi folks,
>
> I'm not subscribed to this list, but felt that this might be relevant to you:
>
> The Open Source Initiative has recently announced[1] version 1.0rc1 of
> their Open Source AI Definition, and are collecting community
> endorsements[2] in addition to community feedback on hackmd[3] and
> their forum[4].
>
> The text of the rc1 definition can be found at:
> https://opensource.org/deepdive/drafts/the-open-source-ai-definition-1-0-rc1
>
> Regards,
> James
>
> [1] - https://opensource.org/blog/the-open-source-ai-definition-v-1-0-rc1-is-available-for-comments
>
> [2] - https://opensource.org/osaid-endorse
>
> [3] - https://hackmd.io/@opensourceinitiative/osaid-1-0-RC1
>
> [4] - http://discuss.opensource.org/
Reply to: