Ancestors

Written by Jeff Geerling on 2025-01-28 at 03:28

The irony that 'Open'AI is so closed that a Chinese open model can come in and deflate the OpenAI bubble in less than a week...

=> More informations about this toot | More toots from geerlingguy@mastodon.social

Written by Warriormaster on 2025-01-28 at 06:50

@geerlingguy Is it really open-source though? I went to their repository and mostly saw lfs files that were called ”models-x”. So it could be the same case as with Meta where the model is open-source but everything else isn’t.

=> More informations about this toot | More toots from warriormaster@mastodon.social

Written by CyberFrog on 2025-01-28 at 06:51

@warriormaster@mastodon.social @geerlingguy@mastodon.social that is commonly what 'Open Source' means in the context of AI, even given the training data you'd be unable to reproduce the same model, so the checkpoints and weights are the next best thing to modify and build on top of.

=> More informations about this toot | More toots from froge@social.glitched.systems

Written by dieTasse on 2025-01-28 at 07:26

@froge not really, I think this was finally defined a few weeks ago and in case of AI the source upon which it was trained needs to be also fully known for it to be really open source.

=> More informations about this toot | More toots from dieTasse@floss.social

Written by CyberFrog on 2025-01-28 at 08:02

@dieTasse@floss.social Defined where? the training of AI is non-deterministic, it uses random data at every single step for optimization reasons, the models can't be created without a bit of uncertainty with the final results.

Even if you literally had exactly the same data, and exactly the same software stack, and exactly the same servers - even if Deepseek literally re-trained their own model the exact same way, the result would be a slightly different model that doesn't behave the same still.

That's just the nature of AI training in general, especially for LLMs, so I'm not sure why access to their data would be that important... beyond just having another high quality public dataset I guess (which is broadly good for many reasons unrelated to AI).

=> More informations about this toot | More toots from froge@social.glitched.systems

Written by dieTasse on 2025-01-28 at 08:05

@froge I will try to find it, I heard it somewhere, can be bogus of course. The thing is about showing sources - so a collection of links/articles that were used for the learning of the ai. So in your case thats the part "exactly the same data" in both cases the sources would be that data (+ code of the learning algorithm + things I am forgetting for sure)

=> More informations about this toot | More toots from dieTasse@floss.social

Written by CyberFrog on 2025-01-28 at 08:07

@dieTasse@floss.social yeah, I guess it would be nice, but even with all that data I could never re-create a copy of deepseek tbh

So I think the model weights being open are the most useful thing for most people in general

=> More informations about this toot | More toots from froge@social.glitched.systems

Toot

Written by Aurélien Pupier on 2025-01-28 at 22:29

@froge @dieTasse argh it was already mentioned (it was shown in another thread and missed it, sorry for the noise then)

=> More informations about this toot | More toots from apupier@fosstodon.org

Descendants

Written by dieTasse on 2025-01-29 at 08:02

@apupier

No prob. Better twice than not at all 😊

@froge

=> More informations about this toot | More toots from dieTasse@floss.social