Ancestors

Toot

Written by Alexander Hanff on 2024-12-23 at 12:08

My latest article on unlawful training of AI models and what we should do about it just dropped on TheRegister today:

https://www.theregister.com/2024/12/22/ai_poisoned_tree/

[#]ethics #ai #genai #llm #trainingdata #privacy #dataprotection #copyright #commons #law

=> More informations about this toot | More toots from thatprivacyguy@eupolicy.social

Descendants

Written by katzenberger on 2024-12-23 at 12:48

@thatprivacyguy

Putting LLMs in the #PublicDomain looks like a good idea at first, but consider @RainerMuehlhoff's and @HannahRuschemeier's thoughts on purpose limitation, as well.

We'd be handing over a loaded weapon to everybody who wants to use it, for whatever purpose.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4711621

=> More informations about this toot | More toots from katzenberger@mastodon.de

Written by Alexander Hanff on 2024-12-23 at 13:22

@katzenberger @RainerMuehlhoff @HannahRuschemeier as opposed to a loaded weapon that can only be used by elites... I think I would prefer everyone to be on a level playing field otherwise we just remain underfoot.

=> More informations about this toot | More toots from thatprivacyguy@eupolicy.social

Written by Alexander Hanff on 2024-12-23 at 13:26

@katzenberger @RainerMuehlhoff @HannahRuschemeier

Plus I would argue that purpose limitation already applies to LLM's trained on public personal data and is in breach of GDPR as a result - but we are not seeing any significant penalties to deter such behaviour (which was the point of my article).

They already broke the law - the issue is enforcement of the law and consequences for breaking it have so far been utterly ineffective.

=> More informations about this toot | More toots from thatprivacyguy@eupolicy.social

Written by katzenberger on 2024-12-23 at 14:07

@thatprivacyguy

A lack of enforcement does not seem to be a good reason for simply opening the gates, though.

@RainerMuehlhoff @HannahRuschemeier

=> More informations about this toot | More toots from katzenberger@mastodon.de

Written by Winfried 🏳️‍🌈 on 2024-12-23 at 13:38

@thatprivacyguy

The models, open source or not, are useless when you don't own a massive data center. So I would say delete them to avoid further further environmental harm.

=> More informations about this toot | More toots from winfriedtilanus@mastodon.nl

Written by Alexander Hanff on 2024-12-23 at 13:50

@winfriedtilanus that isn't true - running the models is not as resource intensive as training them.

I run many LLMs on my own systems at home (not even specialised AI systems and tech which is 4+ generations old) I even run stable diffusion models on my phone.

And open source communities have historically filled the gap by providing hosting solutions - i expect we would see the same happen here (or similar solution to Folding@Home using p2p or distributed system models).

=> More informations about this toot | More toots from thatprivacyguy@eupolicy.social

Written by Winfried 🏳️‍🌈 on 2024-12-23 at 14:00

@thatprivacyguy

Fair, the smaller models can certainly be run on a simple laptop. But I see wild estimates of the costs of a single prompt of the larger models, ranging from cents per prompt (still wild) to thousands of euros for a 'brute force mode'. Probably needs more investigation...

=> More informations about this toot | More toots from winfriedtilanus@mastodon.nl

Written by Daniel Darabos on 2024-12-23 at 13:52

@thatprivacyguy I like the idea! It's hard to argue this would stop progress. Open source models are already making fantastic progress. If you want to keep your model private, you better source your training data legally.

After the first few paragraphs I thought this dude must be on Mastodon. And here we are! 😀

=> More informations about this toot | More toots from darabos@mastodon.online