We now note that that we do not waive our attribution requirements for Spritely Goblins, which are present in Apache v2 (and for that matter, CC BY, CC BY-SA, and most FOSS licenses) https://codeberg.org/spritely/goblins/pulls/998/files
-
I expect the world looks very different from where you’re sitting than from where I am. Every time I look at my (FOSS-heavy) fedi feed, I’m struck by the contrast between the AI pessimism & skepticism here versus everywhere else I look.
LLMs have become part of everyday life for vast swaths of ordinary people. Skilled software developers routinely use them to produce code that’s far from slop. And probably most importantly, the AI industry is widely viewed as the most significant driver of U.S. economic growth and competitive advantage going right now.
All of those things surely sound negative to you, but they’re descriptively true. And against that backdrop, I see very little possibility that US courts will cut OpenAI et al off at the knees, even though we’ll surely see adverse rulings in cases of obvious misappropriation (Anthropic’s trove of pirated books, Sora reproducing entire songs from its training data).
@copiesofcopies @bkuhn @cwebber evidence would indicate that pessimism on AI is the norm at this point. Studies show massive unpopularity of AI. People are trying to shoot folks that build AI data centers.
It's not the same pessimism but there's sure a lot of it.
-
@copiesofcopies @bkuhn @cwebber evidence would indicate that pessimism on AI is the norm at this point. Studies show massive unpopularity of AI. People are trying to shoot folks that build AI data centers.
It's not the same pessimism but there's sure a lot of it.
@wwahammy makes a really good point, though. I think *outside of tech*, the response to LLM-gen-AI is very negative as companies inappropriately apply it to non-tech fields. At least that's what I see anecdotally with folks who don't work in tech.
The one exception I've seen is both human doctors & veterinarians — all of whom seem to be thrilled to let #LLMGenAI chart for them.
I give our vet permission to record the session but not my human doctor, FWIW.
-
@wwahammy makes a really good point, though. I think *outside of tech*, the response to LLM-gen-AI is very negative as companies inappropriately apply it to non-tech fields. At least that's what I see anecdotally with folks who don't work in tech.
The one exception I've seen is both human doctors & veterinarians — all of whom seem to be thrilled to let #LLMGenAI chart for them.
I give our vet permission to record the session but not my human doctor, FWIW.
@bkuhn @copiesofcopies @cwebber I think there are a smattering of use-cases that average folks like:
* Similar to your examples, for first pass notes of a low stakes meeting, it's pretty good.
* Getting a quick answer to easy questions. Sure,it's often wrong but no one is living or dying based on the result.
* Editor for writing. I wish people didn't use it as much but if you're looking for some stylistic help, it's not bad.
* It's not near as common as the other three but vibe coding little projects is also up there.There are downsides to all of those but they have something in common: they're low risk and they're situations where the individual is in control.
If that was the end of it, AI would be relatively popular.
-
@cwebber @bkuhn @trashheap IMO the real challenge to FOSS enforcement against LLM producers is the same one everyone else has: if the training is fair use because the resulting model is transformative, the producer doesn’t need a license (and doesn’t have to respect the terms of any existing license). It’s still early early days in terms of court decisions, but this looks like the direction we’re headed in the US (though the outcome will likely vary depending on how much memorization a model can be shown to do).
@copiesofcopies @cwebber @bkuhn @trashheap Compiling code is transformative too. Does it matter if I store your plain text in ASCII format or in lossy floating point format on the disk? Obviously the transformation to floating point is fair use since the floating point blob is learning from the input. What if I chain a bunch of find, grep and sed commands and pipe the output into a different file? Is the output of this chain even copyrightable?
-
@copiesofcopies @cwebber @bkuhn @trashheap Compiling code is transformative too. Does it matter if I store your plain text in ASCII format or in lossy floating point format on the disk? Obviously the transformation to floating point is fair use since the floating point blob is learning from the input. What if I chain a bunch of find, grep and sed commands and pipe the output into a different file? Is the output of this chain even copyrightable?
While IANAL and TINLA, my understanding that indeed the compiled version of source code requires copyright permission to create/distribute/etc. because it derives from the source code.
I think on the other items you raised, basically "no one knows". New frontiers of technology often don't fit existing policy and regulatory frameworks precisely and the regulation lags decades (and sometimes a generation!).
-
@bkuhn @copiesofcopies @cwebber I think there are a smattering of use-cases that average folks like:
* Similar to your examples, for first pass notes of a low stakes meeting, it's pretty good.
* Getting a quick answer to easy questions. Sure,it's often wrong but no one is living or dying based on the result.
* Editor for writing. I wish people didn't use it as much but if you're looking for some stylistic help, it's not bad.
* It's not near as common as the other three but vibe coding little projects is also up there.There are downsides to all of those but they have something in common: they're low risk and they're situations where the individual is in control.
If that was the end of it, AI would be relatively popular.
@wwahammy wrote:
> “If that was the end of it, AI would be relatively popular.”*That* part is not new. Every technology invented in our lifetimes has been overhyped, overdeployed, & overused to its illogical conclusion. While that's happening ∃ (fortunately) some folks along the way who will say the Emperor is only half-dressed.
It's years later that people just quietly pretend they weren't taken in by the foolishness of the new and shiny.
Think “blockchain”!
-
While IANAL and TINLA, my understanding that indeed the compiled version of source code requires copyright permission to create/distribute/etc. because it derives from the source code.
I think on the other items you raised, basically "no one knows". New frontiers of technology often don't fit existing policy and regulatory frameworks precisely and the regulation lags decades (and sometimes a generation!).
@bkuhn @yo8rzz @cwebber @trashheap
“Transformative” in the context of fair use doesn’t necessarily follow the word’s colloquial meaning. A work is transformative with respect to another if it adds “new expression, meaning, or message” to the original work. Compilation is more like translation, which is not considered transformative—a translation is a derivative work that must be licensed.
Ultimately though, this situation was arrived at not through careful reasoning about definitions but through court cases presenting specific facts before specific judges. So object code is basically deemed the same work as source code as a matter of law, and how we arrived there has nothing to do with the definition of “transformative.” It’s just the way the cookie crumbled.
-
We are more in agreement than you realize:
I guess now is a good as time as any to “come out”: I'm using #Anthropic's #Claude on a regular basis to help me figure out what types of policies we should set for reasonable & ethical use of #LLMGenAI in #FOSS, & also to actually improve FOSS codebases.
I believe it is reasonable to use proprietary technologies as long as our eyes remain on the prize of defeating those technologies.
@cwebber, I suspect, doesn't agree.
@bkuhn @copiesofcopies While I appreciate being forthcoming, you are correct, I don't agree with feeding the torment nexus to defeat the torment nexus. I think using Claude is just helping them build more torment nexus, while normalize a situation that puts us in a legally dubious situation where we were previously relatively legally strong, and eroding community trust and shared understanding in the process.
-
@bkuhn @copiesofcopies While I appreciate being forthcoming, you are correct, I don't agree with feeding the torment nexus to defeat the torment nexus. I think using Claude is just helping them build more torment nexus, while normalize a situation that puts us in a legally dubious situation where we were previously relatively legally strong, and eroding community trust and shared understanding in the process.
@bkuhn @copiesofcopies I think the situation is more acceptable when using fully "local models", it doesn't erase all the concerns, but it at least improves user agency and autonomy. But overall, "local models" are mostly for "localwashing"; people like to bring them up as an alternative that nobody uses. Likewise, using LLMs to *find problems*, I don't have a problem with; search-space usage is one thing, but generative is another, in terms of social impact.
I also think that there are ways to improve the reliability of these systems, make them cheaper, and understandability of what they are doing, with neurosymbolic approaches, but that work hasn't advanced yet, and the current situation is pretty dismal. I really don't believe that we can vibe our way out of this mess, and I'm afraid we're going to see important projects, including Linux, head down the vibe bobsled https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/
-
@bkuhn @copiesofcopies I think the situation is more acceptable when using fully "local models", it doesn't erase all the concerns, but it at least improves user agency and autonomy. But overall, "local models" are mostly for "localwashing"; people like to bring them up as an alternative that nobody uses. Likewise, using LLMs to *find problems*, I don't have a problem with; search-space usage is one thing, but generative is another, in terms of social impact.
I also think that there are ways to improve the reliability of these systems, make them cheaper, and understandability of what they are doing, with neurosymbolic approaches, but that work hasn't advanced yet, and the current situation is pretty dismal. I really don't believe that we can vibe our way out of this mess, and I'm afraid we're going to see important projects, including Linux, head down the vibe bobsled https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/
@bkuhn @copiesofcopies
I think we should be pushing back on these things. I'm worried and dissappointed that FOSS leadership isn't. I'm really worried we might never recover from the damage that's already been done. -
@bkuhn @yo8rzz @cwebber @trashheap
“Transformative” in the context of fair use doesn’t necessarily follow the word’s colloquial meaning. A work is transformative with respect to another if it adds “new expression, meaning, or message” to the original work. Compilation is more like translation, which is not considered transformative—a translation is a derivative work that must be licensed.
Ultimately though, this situation was arrived at not through careful reasoning about definitions but through court cases presenting specific facts before specific judges. So object code is basically deemed the same work as source code as a matter of law, and how we arrived there has nothing to do with the definition of “transformative.” It’s just the way the cookie crumbled.
@copiesofcopies @bkuhn @cwebber @trashheap The derivative aspect of LLM output seems pretty clear cut to me, after all, without being derivative how else can it reproduce facsimili of the original? There is no other mechanism to explain it sufficiently other than derivation and statistical amalgamation with numerous other sources by means of intermediate storage in an appropriate format for such computation, namely floating point.
-
@copiesofcopies @bkuhn @cwebber @trashheap The derivative aspect of LLM output seems pretty clear cut to me, after all, without being derivative how else can it reproduce facsimili of the original? There is no other mechanism to explain it sufficiently other than derivation and statistical amalgamation with numerous other sources by means of intermediate storage in an appropriate format for such computation, namely floating point.
@yo8rzz @bkuhn @cwebber @trashheap
Again, it’s a mistake to apply the colloquial meaning of “derivative” to its use as a term of art in copyright law. Transformative works are necessarily “derivative” in the colloquial sense, and often even in the legal sense, but may nonetheless be deemed a fair use (depending on how the court applies the various factors of the fair use analysis).
And where models reproduce substantial portions of their reasoning data verbatim, courts are less likely to find fair use. But plaintiffs have had a harm time showing that in many cases, including the CoPilot case, because when you train on enough data, the likelihood of reproducing a single reasoning source verbatim is low.
-
@bkuhn @copiesofcopies I think the situation is more acceptable when using fully "local models", it doesn't erase all the concerns, but it at least improves user agency and autonomy. But overall, "local models" are mostly for "localwashing"; people like to bring them up as an alternative that nobody uses. Likewise, using LLMs to *find problems*, I don't have a problem with; search-space usage is one thing, but generative is another, in terms of social impact.
I also think that there are ways to improve the reliability of these systems, make them cheaper, and understandability of what they are doing, with neurosymbolic approaches, but that work hasn't advanced yet, and the current situation is pretty dismal. I really don't believe that we can vibe our way out of this mess, and I'm afraid we're going to see important projects, including Linux, head down the vibe bobsled https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/
Indeed, we should all be looking at more at local models. @ossguy suggested to me just yesterday that perhaps big systems that rely on these soon-to-go-away gratis “token credits” should be run in parallel on local models and output compared to verify that the local models can serve as a drop-in replacement for the proprietary network LLM-gen-AI services. It's an excellent idea that should be pursued with rigor.
Cc: @copiesofcopies
-
@bkuhn @copiesofcopies While I appreciate being forthcoming, you are correct, I don't agree with feeding the torment nexus to defeat the torment nexus. I think using Claude is just helping them build more torment nexus, while normalize a situation that puts us in a legally dubious situation where we were previously relatively legally strong, and eroding community trust and shared understanding in the process.
@cwebber @bkuhn @copiesofcopies the masters tools will never dismantle the master’s house
-
@bkuhn @copiesofcopies I think the situation is more acceptable when using fully "local models", it doesn't erase all the concerns, but it at least improves user agency and autonomy. But overall, "local models" are mostly for "localwashing"; people like to bring them up as an alternative that nobody uses. Likewise, using LLMs to *find problems*, I don't have a problem with; search-space usage is one thing, but generative is another, in terms of social impact.
I also think that there are ways to improve the reliability of these systems, make them cheaper, and understandability of what they are doing, with neurosymbolic approaches, but that work hasn't advanced yet, and the current situation is pretty dismal. I really don't believe that we can vibe our way out of this mess, and I'm afraid we're going to see important projects, including Linux, head down the vibe bobsled https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/
-
@cwebber @bkuhn @copiesofcopies the masters tools will never dismantle the master’s house
-
@yo8rzz @bkuhn @cwebber @trashheap
Again, it’s a mistake to apply the colloquial meaning of “derivative” to its use as a term of art in copyright law. Transformative works are necessarily “derivative” in the colloquial sense, and often even in the legal sense, but may nonetheless be deemed a fair use (depending on how the court applies the various factors of the fair use analysis).
And where models reproduce substantial portions of their reasoning data verbatim, courts are less likely to find fair use. But plaintiffs have had a harm time showing that in many cases, including the CoPilot case, because when you train on enough data, the likelihood of reproducing a single reasoning source verbatim is low.
The CoPilot case¹ was ill-constructed from get. 'Twas brought by folks whose enthusiasm (for a payday from FOSS in a hastily & ill-conceived class action) exceeded their desire to build the correct & ethically sound case for adjudicating the problems with LLM-gen-AI with respect to FOSS.
As a matter of fact, one *must* make a copy (and, typically, receive distribution) to train LLM-gen-AI, and one needs a license for that.¹ https://sfconservancy.org/news/2022/nov/04/class-action-lawsuit-filing-copilot/
-
@cwebber @bkuhn @copiesofcopies the masters tools will never dismantle the master’s house
But FOSS' success up until this point has shown the exact opposite.
FOSS was bootstrapped by using proprietary software every single day for more than a decade, at least.
Most FOSS developers regularly used proprietary technologies daily for their work for the entirety of FOSS' history.
Plus, what @copiesofcopies said:
https://fedi.copyleft.org/@copiesofcopies@social.coop/117457049015278425Cc: @cwebber
#LLMGenAI #LLM #AI #FOSS #SoftwareFreedom #OpenSource #FreeSoftware
-
The CoPilot case¹ was ill-constructed from get. 'Twas brought by folks whose enthusiasm (for a payday from FOSS in a hastily & ill-conceived class action) exceeded their desire to build the correct & ethically sound case for adjudicating the problems with LLM-gen-AI with respect to FOSS.
As a matter of fact, one *must* make a copy (and, typically, receive distribution) to train LLM-gen-AI, and one needs a license for that.¹ https://sfconservancy.org/news/2022/nov/04/class-action-lawsuit-filing-copilot/
@bkuhn @yo8rzz @cwebber @trashheap
I agree about the CoPilot case. One of several purely conceived case hurriedly filed in a gold rush.
Copying is an essential step to training, but there’s strong support for that copying being fair use if the ultimate use is. That’s what the judge found re: book ripping in Anthropic v. Bartz, on the basis of decisions in multiple cases against Google in the early 2010, where plaintiffs alleged that intermediate copying to enable caching, image search, etc. was infringing.
-
We are more in agreement than you realize:
I guess now is a good as time as any to “come out”: I'm using #Anthropic's #Claude on a regular basis to help me figure out what types of policies we should set for reasonable & ethical use of #LLMGenAI in #FOSS, & also to actually improve FOSS codebases.
I believe it is reasonable to use proprietary technologies as long as our eyes remain on the prize of defeating those technologies.
@cwebber, I suspect, doesn't agree.
@bkuhn @copiesofcopies @cwebber
imo, it is not beneath anthropic (or other AI companies) to steer their models - either generally or on a targeted basis - toward outputs that represent their own best interests rather than that of FOSS or of the wider public, even if only in subtle ways.do you think working out policy with humans exclusively, instead, has particular downsides that can't be overcome?
