- Joined
- Sep 11, 2024
Performing exorcism on source codes. Up to 30% of KDE krashes are caused by malevolent spirits.
Open Source Software Community - it's about ethics in Code of Conducts
- Thread starter CrunkLord420
- Start date
-
🔎 If you see an alert for analytics.usips.org, this is a self-hosted system I am running as a test. It is GDPR compliant and collects no invasive information.
-
Want to keep track of this thread?
Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
Create account
UNIXLover
kiwifarms.net
- Joined
- Aug 31, 2023
I just saw that the AI Moratorium for Krita ends this month, October. It's worth keeping an eye out for potential discussion to see if they still think people will get "furious."
Since the Krita developer community cannot currently find consensus on whether the use of AI tools to assist development is acceptable or not, we have decided to put a moratorium on the use of these tools until October 2026.
The reason for pushing this off to October are:
The reason for pushing this off to October are:
- if we'd allow the use of AI in development right now, the backlash from our users and supporters will be furious.
- KDE as a whole does not have a policy in place, and if our policy would conflict with KDE's we'd potentially have to revert contributions made with AI
- The development of AI itself might change for the better or the worse in the next period. Microsoft is already backtracking from their CoPilot brand. Acceptance of AI can change for better or worse. The environmental impact of AI could lessen.
- Joined
- Jan 27, 2022
yeah this is why copyright law needs to be abolished, it gives unfair advantage through government monopoly to proprietary application developers that open source doesn't have. I think people forget that the principle goal of open source licensing is to use copyright against itself and hopefully, in the end, legally destroy it.
- Joined
- Dec 23, 2023
And now we know why so many programmers are taking their saved up tech boom salaries, buying land in parts of the country that county surveyors forgot existed, and disappearing to raise banana peppers or goats or something.
To comment on the copyright vs. GPL vs. LLM sperging a few pages back, I am surprised nobody has tried to come up with a new license with a copyleft virality like the GPL specifically to combat AI company scraping. Can you imagine: a license that requires any LLM trained upon the data licensed to be released with open weights, along with its entire training corpus and and any software used to train it under the same copyleft license? Can you imagine the potential for deliciously hilarious seething butthurt in all directions over such a thing?
- Joined
- Dec 12, 2022
That's just the GPL that already exists.
Since these "models" are very obviously derivative works, this is already the case. The proponents of loose licenses—loose like a whore—such as the MIT license didn't even get the one thing they request, credit.
They argue that none of the licenses apply to them, so it wouldn't matter.
Slurred
kiwifarms.net
- Joined
- Jan 29, 2022
Either GPL already does this or it can't be done. FSF posted an article (archive) arguing for this years ago.
I have the (admittedly self-serving) position that I'd love for OpenAI, Anthropic et al. to be forced to put their weights on HuggingFace. But the idea that output of the models is de facto copyright infringement/automatically a derivative work of all input doesn't sit well with me, it just seems like a massive expansion of the concept of copyright. When you steal from one person, that's ripping them off, but if you steal from thousands of people in tiny ways and mix it up into your own thing, well, that's just the creative process.
I find it so funny how anti-AI this project and most of its community spaces are considering that it is also the (unwilling) host to the Krita AI Diffusion plugin, arguably the best currently existing tool for AI-assisted image creation.
- Joined
- Oct 2, 2023
No, GPL doesn't solve for this. GPL only says you have to give away your source code if you combine a GPL'd work with your work and distribute it. Nothing in the GPL stops you from training a model on GPL code as long as you don't reproduce and distribute the actual GPL-licensed code.
It's not clear you could even solve this via a license tho, as the whole concept of fair use is that you can copy and infringe someone else's content if the use is transformative. You're only talking about fair use when the training isn't licensed. A funny symmetry is that the models themselves aren't protected by IP except for trade secret law. Congress needs to weigh in but they'd only fuck this all up to keep out chinese competitors
- Joined
- Mar 23, 2020
AI companies are state-sponsored laundering machines that cannot be held accountable to any standard whatsoever because muh Chyna o algo will beat us in le AI wurr if they are impeded in ANY way!!! Surely the next round of bone chilling billion dollar subsidies will help them truly own them chinks or something to that effect, totally man, don't you know AGI is just around the corner or something????
- Joined
- Jan 31, 2021
No one has tried because the United States Copyright Office has already weighed in on the issue of LLM training, as have the courts, and the consensus is clear: Insofar as you have lawful access to the material in question, nothing can stop an LLM from being trained on that material, insofar as no protected expression is incorporated or otherwise reproduced. What you're advocating for would essentially entail allowing for a license that restricts your ability to use facts and ideas derived from a piece of software.
It doesn't, and it can't. The FSF brings up the matter of Github Copilot regurgitating elements of its training data in its output, arguing that this means they're guilty of copyright violation unless this can be determined to be some specie of fair use. Well, the courts weighed in on that in Doe v. Github and determined that, with regards to claimed DMCA Section 1202 violations, because the code was reproduced in a probabilistic manner there was no copying of an existing work, and hence no removal of Copyright Management Information (the basis of the 1202 violations), and so those claims were dismissed (which was recently affirmed by the Appellate Court for the Ninth Circuit). What remains now for purposes of this discussion are the breach-of-contract claims for violating the terms of the licenses of the works that were ingested. Those claims, however, are almost certain to fail since as Judge Jon Tigar noted in his original decision:
Plaintiffs’ new allegations fare no better. Plaintiffs allege that, should a user elect to notuse the duplication-detection feature, a user could conceivably view an identical match of 150 characters, and use it without attribution. See ECF No. 201 ¶¶ 147–151. While Plaintiffs aver that the duplication-detection tool “by definition establishes Copilot’s ability to reproduce verbatim copies of code,” ECF No. 234 at 18 (emphasis omitted), they “do not explain how the tool makes it plausible that Copilot will in fact do so through its normal operation or how any such verbatim outputs are likely to be anything beyond short and common boilerplate functions.”[source]
That last sentence is important. "Common boilerplate functions" would constitute what's called scènes à faire, which are elements that are expressly outside the scope of copyrightable subject matter. This would almost certainly render it not contractual subject matter, and this is an improper attempt to create a pseudo-copyright over public domain material, failing to establish the extra element needed to avoid a ruling that their state law claims are pre-empted by federal copyright law.
TL;DR Unless they can show that Copilot is producing protectable expressions rather than unprotectable expressions, their entire case collapses. They've already lost their DMCA claims, and we're just waiting on the state law claims to fail too.
They're not derivative. They would be derivative if and only if they returned protected expression. As Doe v. Github shows.
It's almost like GPL partisans just parrot the first thing they heard and never bother to keep abreast of developments or something.
We now return you to your regularly scheduled faggotry already in progress.
- Joined
- Jan 28, 2018
On the one hand I hear from the usual suspects that AI is shit, not usable for anything, garbage, a next word predictor, slop machine etc. and soon every AI company will go out of business and the bubble will pop and nobody will use AI for anything anymore, on the other hand this endless worry that AI has training material. Why does it matter if it's shit and has plateaued and will never be good for anything? Which one is it?
Last edited:
- Joined
- Feb 8, 2021
They are different groups, at least for these extreme standpoints.
- Joined
- Dec 12, 2022
Are they? Treating the tech simulatneously as retarded and incapable of improving yet also threatening and will possibly steal your jobs isn't uncommon. Though it's mostly heads in the sand at this point and a product of people sitting all day on social and trying to make quips for easy "owns".
The two standpoints being opposites of each other doesn't matter when every smuck practices mental gymnastics.
- Joined
- Dec 12, 2022
I don't feel like arguing about this shit any longer, retard, so I'm just going to leave it at this insult. Have fun defending corporations.
- Joined
- Feb 8, 2021
The people who are worried about training data don't think that it is literally useless for anything. They likely think it's inferior, but with that there's no contradiction. Replace AI with Indians and you'll see.
- Joined
- Jan 28, 2018
I don't know, I read both of these coming from the very same people before.
- Joined
- Mar 23, 2020
It is a very small percentage of people that have real cognitive dissonance, but the vast majority I believe are either paid shills and or bots that shit up every place of public discourse with contradictory information as a form of cognitive warfare. That, and people who just cannot articulate what they mean well enough and so default to something that sounds dissonant.
- Joined
- Mar 4, 2019
I feel attacked, but I can't say why.
- Joined
- Dec 3, 2023
Really didn't expect you to get slapped down so hard but they made a great post and you wussed out. Such is life sometimes. Maybe you can report them for a COC violation or something.
- Joined
- Oct 2, 2023
It's useful to think about what would happen in that world, it would probably look like a compulsory licensing regime, like how ASCAP does mechanical rights for songs. I think the French will try to get there. But even if you blocked training as not fair use, if Anthropic or OpenAI had to, they could spend 5-10 billion paying people to create or buy code art and literature which the companies owned and could train on unimpeded. It would slow down AI training by 6-12 months but wouldn't change anything. What it would do is make it so startups without billions to pay for training data can't compete against the giants and that's bad for everyone, there would be no open weight models.
Last edited:
- Joined
- Jan 31, 2021
Honestly it really wouldn't change anything. If they tried to set up a compulsory licensing scheme for unprotectable expression then people will launder where they got their training data from. Moreover, it's a bit late to do anything of that sort when there's already models that people can run locally and have downloaded millions of time. The chance to lock this shit down has sailed. Changing the law won't stop anything at this point.