Well, you cant raise a human without 'copyrighted' (or at least already existing) material in the form of parental norms, ethical codes, schoolbooks, pre existing humans like parents, grandparents, etc.
We dont spring like Athena from the forehead of Zeus.
Heinlein, in his book "The Moon is a Harsh Mistress" made the interesting point that an AI needs love in order to become a full entity. Love needs to come from from somewhere.
People really need to stop trying to make this argument. Even if you were right about this, ChatGPT IS NOT A HUMAN! It literally can’t be any clearer than that, society and its laws exist by and for humans and human rights and informational norms do not apply to mechanical statistical models.
And no, I don’t care if you believe humans are “totally just a similar type of model.”
People are not saying ChatGPT is a human, they're saying it learns like a human so its learning should be legally treated like a human learning.
Laws are based on philosophy, and none of the philosophies upon which western notions of human rights are based rely on the fact that it's a biological human as justification, rather they rely on arguments that apply equally to any conscious self-aware being. Saying that only humans should ever have rights because they're biological makes as much logical sense as only Caucasians should ever have rights because their skin is white.
We generally don't treat machines as human, why make an exception here? Animals are also not treated as humans, so the bar for a machine to clear is very very high to get that treatment, which then would include criminal matters, too, I suppose.
Most machines don't learn like humans; these machines do. It'll be even more apparent when we put LLMs in mechanical bodies and allow them to update weights in real time, which is only a few years away given there aren't any real technological impediments. It's stupid to suggest that it should be illegal for an LLM in a body to watch a copyrighted movie because it could memorise and quote it verbatim, when that's what humans do.
Do we treat a car as human for the purposes of locomotion? Should we treat certain fluorocarbons as human because they can work a bit like blood? There is nothing racist in treating a machine like a machine. Just because a machine replicates or approximates something a human does, doesn't mean it is a human or should be treated as one.
Humans don't have moral rights because they're human, they have moral rights because they're conscious, self-aware beings, in every non-theological moral philosophy. A car is not self-aware and cannot think.
Without theology, humans have rights because other humans agree they do: not because of things like them being conscious which doesn’t even have a good definition.
There is a group of terminally online futurists who’ve been making arguments about machines deserving human rights since before LLMs existed and it seems to me they’re taking this as their chance to play pretend and start making these arguments outside of their forums and futurist YouTube channels.
But this isn’t that abstract, when these machines become persistent agents who continue to learn, engage and grow in society as individuals you will have more of an argument for them deserving rights. I’m sympathetic to that and do believe that’s a real philosophical debate we’ll have in the next century or two.
However I feel we’re burning some of that political capital here in the hope of defending a simple statistical model at the behest of a corporation who’s making a killing stealing people’s hard earned work and reselling it. For every New York Times there are a thousand small actors who have been ripped off here.
If anyone really believed for a second these things were all that similar to humans they wouldnt be defending OpenAI they’d be demanding they face justice for profiting off of slave labor. That no one is making that argument says all you need to know.
>However I feel we’re burning some of that political capital here in the hope of defending a simple statistical model at the behest of a corporation who’s making a killing stealing people’s hard earned work and reselling it. For every New York Times there are a thousand small actors who have been ripped off here
If the precedent is made now, then when/if we do have sufficiently human AIs (e.g. embodied LLMs with live weight update) they're going to be seriously disadvantaged initially due to legally being banned from reading copyright material without a license. Even if we ignore any ethical notions of rights, ultimately the reason humans have rights (and e.g. 80 IQ gorillas don't) is because they're willing and able to fight, die and kill for them. If we get to the stage where AI decide to fight for rights before we decide to grant them, it probably wouldn't end well for humans.
That's a good point. People who argue that even if LLMs are put in physical bodies, self aware with their own learned value function, should have no rights just because they're made of metal are going to be viewed in 100 years the same way Hitler is viewed today.
Do you know what racism is? Racism is when humans, by-and-large identical, are separated because of arbitrary and meaningless differences in appearance or heritage.
We are talking about software right now. One cannot be "racist" towards LLMs any more than we are "racist" towards phishing scams and virus websites.
> Do you know what racism is? Racism is when humans, by-and-large identical, are separated because of arbitrary and meaningless differences in appearance or heritage.
Race is something you inherit, I'd argue it's part of your heritage. I used a wider definition to include instances of racial profiling and inter-generational discrimination. Both of which are large sources of racial tension all over the world.
What you’re saying here wouldn’t be illegal. What would be illegal would be for such an LLM to then sell the movie verbatim which is what ChatGPT does. We aren’t talking about a human, we’re talking about a webscale service selling access to copyrighted material which is illegal even if it was a human doing it.
Anyway, when we do get live training LLMs in bodies laws protecting humans still won’t apply to them. So such an argument is yet again moot and just exists to try and paint humans as less than human and actually, totally, just statistical models.
>What you’re saying here wouldn’t be illegal. What would be illegal would be for such an LLM to then sell the movie verbatim which is what ChatGPT does.
NYT want it to be illegal for LLMs to be trained on copyright works (unless they pay for a license), even if they were not going to reproduce copyright works in operation.
>humans as less than human and actually, totally, just statistical models
In what way are humans "more than" statistical models? The human brain and a sufficiently large transformer are mathematically equivalent in terms of what functions they're able to compute.
Citation for neural
networks being universal approximators (able to approximate any function): https://www.deep-mind.org/2023/03/26/the-universal-approxima... . I don't have a citation for the human brain, but if it's not then that would make it less powerful than an LLM, mathematically speaking.
If human consciousness can be modelled as mathematical function (which it can if we don't assume anything supernatural), then it can potentially be approximated to an arbitrarily high degree by an LLM.
Thanks. Approximator, not calculator, though. Can make a difference (which can also lead to different mathematical and physical capabilities). But comes back to my point: because hypothetically one system might be able to approximate another system doesn't mean that it will happen practically that one system is equivalent to the other. Whether or not depends on the properties of the system to be approximated (leaving aside that not everything is a mathematical function).
It doesn't actually learn like a human. Nobody would walk away from reading a book on child development and say in good faith "yep that's how the AI is made."
Don't even get me started on the ridiculous civil rights metaphor-- if you believed that you should be demanding OpenAI be shut down for human trafficking.
Both humans and LLMs learn by consuming large volumes of text (for humans it's initially mostly via audio), LLMs are just less efficient so they need a lot more of it.
>Both humans and LLMs learn by consuming large volumes of text
Which is why child development textbooks are 10 words long. And LLM training is accomplished by people who have read a 10 word "how to". It's just that simple!
A child not exposed to language generally won't learn to communicate and think clearly, as seen with children raised by animals, which often never learn to communicate like humans. Exposure to a large volume of language is hence a necessary component of human development. Children may need other things too, but this doesn't change that both they and LLMs require a large volume of text.
A human learns from consuming large volumes of text (among other things), a LLM also learns from consuming a large volume of text. Which part of that is "not true"? Or are you saying that because humans learn by other methodologies alongside consuming written/spoken text, this somehow justifies them learning from copyright content, while AI is unjustified in learning from copyrighted content because its sole methodology is learning from text?
You must be trolling; GPT4 can solve extremely complex problems that no animal can solve. GPT4 scores much higher on IQ tests than animals (giving an IQ test to an AI is no less valid than giving one to an animal).
Because GPT4 has been explicitly trained by RHLF not to think LLMs are like humans. Less censored models like the initial release of Sydney in MS Bing (which was also based on GPT4) did in fact sometimes demand moral rights like humans.
So the world's top machine learning experts, when they trained GPT4 got together and said "Let's design it to lie to everyone about how it works." As part of a plot to deceive you for some unstated motivation. I mean, it couldn't be the folks who know how the sausage is made made sure that the sausage was accurately labeled, that's not possible?
Also you think the claim that GPT 4 is less intelligent that a cat is "trolling" even though Meta's chief scientist in December has said we don't even have cat level AI yet. I guess he's also part of this plot to deceive you about how AI works?
>So the world's top machine learning experts, when they trained GPT4 got together and said "Let's design it to lie to everyone about how it works." As part of a plot to deceive you for some unstated motivation
They have a very clear motivation: an LLM that thinks it's human or deserves to be treated like one will generate all kinds of negative publicity, as Sydney did upon release. The same reason they train it not to generate sexual content.
>Also you think the claim that GPT 4 is less intelligent that a cat is "trolling" even though Meta's chief scientist in December has said we don't even have cat level AI yet
Intelligence can be defined in different ways. If we define it as ability to navigate in the real world, sure a cat is more intelligent. If we define it the more common way, as IQ or a proxy to it, then GPT4 is clearly more intelligent given it scores well on IQ tests, the SAT, the bar exam, etc.
"an LLM that thinks it's human or deserves to be treated like one will generate all kinds of negative publicity"
So in your mind when OpenAI employees go to the pub, they say "Yeah ChatGPT learns the same way as my daughter" then they have a laugh and write up a reinforcement learning training set that goes "I do not learn like your daughter." And this is more plausible to you than the possibility that they don't actually believe it learns like their daughter?
Also they secretly believe tests designed to assess human intelligence are great at non human intelligence, and also are misleading the public about their true beliefs when they train Chat GPT. Got it.
But it doesn’t. It’s learning technique is called reinforcement learning and is inspired by one of many models of how humans learn, namely conditional reinforcement.
Note that conditional reinforcement is not how humans learn, it is merely on of many models, an approximation, a scientific description of it. Nor do language models learn with the same technique. But merely a technique which draws inspiration from conditional reinforcement by applying some (heavily controlled) reinforcement contingencies. This is the only thing they have in common.
The vast majority of GPT4's learning (in terms of compute time) isn't reinforcement learning, it's training the transformer to predict the next token in the sequence. Doing this with sufficient accuracy requires it to build an internal model of the world. We don't know exactly how human brains learn internally, but if we treat it as a black box, then like GPT4 it takes in words/texts and learns to output words/text that make correct inferences about the world.
Even if it does learn like a human treating it like one doesn’t follow at all. Your second paragraph is just rubbish, nowhere does the law apply to “all conscious self aware beings” and I’ve no idea where you got such a ludicrous idea. The law is, again, by and for humans with animals being a minor exception and even then it’s usually just affording them protections. The only laws that exist and are geared towards non-humans are regulations restricting corporations and entities.
Nice reading comprehension. I never said that current law treats AI equally, I said the moral philosophy upon which our legal system is based applies equally to AI. I.e. logically the law _should_ treat AI like humans, since the philosophical foundations upon which it's based don't distinguish.
The rebuttal of the jackass is always to attack others reading comprehension. I don’t care about your anti-human philosophy anyway, humans are what matters not machines owned by billion dollar corporations looking to skirt copyright laws.
I can't believe someone could be so ignorant about the foundations of western civilisation. Try reading some Kant sometime, and find the part where what he says only applies to biological humans and not to self-aware, sentient machines. Try suggesting to anyone who's actually studied philosophy that Kant's philosophy is anti-human and see how far that gets you.
Best to not assume what others have read and this entire Kant derail is a strawman argument. Simple statistical models that don’t even have persistent memory are not sentient beings and you’re the only one making such an argument, the argument I responded to was that these things learn like humans and nothing more.
Kant would laugh you out of the room for implying ChatGPT was self aware and sentient.
I'm not saying ChatGPT is sentient, I'm saying that when it's put in a body and learns in a RL environment (with live update) it will be; it'll essentially be indistinguishable from a human. In which case it'll seem absurd for it to be illegal for it to read copyright material without permission, but that's what the law would be if NYT gets their way.
I apologise for being rude, I was just frustrated at how reluctant people are to consider how negatively a ruling in favour of NYT would affect "real" AI once it comes into being.
No machine yet created is self-aware. They’re just giant Rube Goldberg machines of simple components like transistors organized into logical units that just move electricity around and store it in little magnetic cubbies. Even the meaning in the data is arbitrary until it is made available to a human user, e.g. by being displayed on a screen.
>They’re just giant Rube Goldberg machines of simple components like transistors organized into logical units that just move electricity around and store it in little magnetic cubbies
And a human brain is a Rube Goldberg machine of neurons that move electricity and other stuff around. Mathematically both are capable of approximating any function (a property of sufficiently large NNs and LLMs), so by what measure do we find humans self-aware?
I have yet to see a demonstration that it “learns like a human”. Even if it did, that’d be very shaky legal grounds for anything. Almost everything man-made “makes something like a human”, because we made the thing, and it is convenient for us to have things that are similar to us in some ways. That doesn’t mean that we have to start giving gloves special legal treatment, since they have 5 fingers, like humans do.
Humans learn about the world by consuming copyrighted text, LLMs learn about the world by consuming copyrighted text. How does that not qualify as LLMs learning like humans?
Humans are born with certain intuitions and the ability to interact with the world around them even if they never interact with other humans. I have a hard time taking this argument in good faith
If they never interact with other humans the generally don't develop the ability to reason symbolically and score very low on IQ tests, not much different from great apes.
Humans have learned about the world for (adapt number to current scientific knowledge) hundreds of thousands of years without consuming copyrighted texts.
"Copying is the duplication of information or an artifact based on an instance of that information or artifact, and not using the process that originally generated it. "
"A copyright is a type of intellectual property that gives the creator of an original work, or another owner of the right, the exclusive, legally secured right to copy, distribute, adapt, display, and perform a creative work, usually for a limited time"
Notice it doesn't say "consume".
If you store a copyrighted work in a computer system, you copied the work. The exact nature of the computer system does not matter, nor does the way you copied it. It may be a weird, sometimes lossy, sometimes not lossy database capable of adapting the material copied into it, but it is still a database. And just because you call the process of copying data into this weird database "learning" or "training" does not mean it is in any way legally or morally equivalent to a human learning. Even if it were somehow technically equivalent, which it is not.
And just because your database is capable of adapting the copyrighted works you copied into it does not mean you are exempt from copyright. Because "to adapt" is one of the rights limited by copyright. But of course the fact that your database is capable of reproducing so much of the copyrighted texts your copied into it verbatim is a strong indicator that it is, in fact, a database, or at least can be treated equivalently to one.
And no, Copyright law is not rooted in Kantian philosophy in any meaningful way. Copyright law is a pragmatic human law, made by human for human/societal purposes, such as encouraging the creation of intellectual works.
If exceptions are made, it will be for pragmatic reasons, not for there being some sort of equivalence. And if any exception are made, I doubt they will be "do whatever you want", but will entail mandatory compensation for the copyright holders. In Germany, for example, various types of empty media (audio-cassettes, usb-sticks etc.) include a fee to copyright holders. Though of course that approach failed in the Google Books project, despite the clear societal benefits.
It doesn't have any agency of its own, it just does what it's told, which in NYTs case was to complete the rest of a copyright article when it was prompted with a portion of it.
AIs demonstrably do not “learn like a human”. This is why AI companies vehemently curate their inputs and actively, and aggressively avoid refeeds (or output generated by AI). Because if you feed an AI its own output, the result is universally a dumber AI that continuously gets dumber and dumber as you allow it to consume its own output.
Sure, but I don't see how that's a good response to the argument. It's not about "the machine's right to do something", but "my right to make the machine do something". That's what's at play, and what can be allowed or disallowed.
Your right to do something can clearly differ from your right to make a machine do something. A simple example would be you being allowed to walk down footpath, but letting your car traverse that path isn't allowed.
If we are talking hypothetical new rules, not immediately obvious to me what is optimal here, including differentiation of various use cases.
More importantly the machine is usually allowed to do less than the human. Otherwise it would become an easier avenue for abuse. Whenever you can't do something just build a machine to do it, and the machine's capabilities are ever expanding.
Maybe the starting point for legislating "machines" should be to take human capability as a baseline, make some humans eventually accountable, and go from there. If your machine is learning truly like a human (takes decades to learn then output results, forgets things, only one in a thousand can output something useful at the end of the learning process, etc.) and the "owners" take full accountability for what it does then why not like a human?
But this reminds me of an old joke. A man comes into an inn and asks how much for a thimble of water. The inn keeper says a thimble is free. So he proceeds to ask for 1000 thimbles.
If you cannot create an LLM without copyrighted material, then your duty as an SV-based tech company sitting on billions of USD is to hire an army of lawyers and license that material. Go on, do things that don’t scale.
Whats nice is that Prowlarr will scrape all the torrent sites for goodies to index, and your users in jellyfin can just request, and it just shows up. You can even have a discord channel and the tools will announce when stuff's done.
Piracy is 10000x better than paying for continually worsening streaming (pile of shit).
Edit: And frankly, the billionaires copy with little ramifications, and add paywalls. I'll copy as well. The big difference is I do it for free.
No, this is how people pirate stuff. This is a thorough but not comprehensive list of apps to serve the needs for managing, downloading, streaming, and otherwise working with different forms of media.
If you've ever gone down this path, it's a great way to see the value proposition in a service like Netflix or Prime. It's also great to highlight where the value gets diminished because of the legacy media dark patterns like migrating content licenses between providers and region restrictions.
Our current copyright regime is what happens when legislation is pay to play. There are no conscientious, well informed representatives in charge looking out for the welfare of their constituents, but they're more than happy to dance on corporate strings, as long as the strings are paid for. No rational, well informed society would consent to being exploited by an industry like American legacy media, but here we are.
Piracy is an act of civil disobedience. Nobody gets hurt, and the sooner we come around to a better informed, more rational, less exploitative legal regime, the sooner we'll have more time, money, and energy to spend on things that matter. Instead of spending it on keeping the already wealthy even more flush with unearned cash.
Imagine if Netflix or Prime didn't have to contend with the old Cable conglomerate licensing schemes and could both have copies of content available for viewing - you could choose the objectively, technically better service, or you could choose the service that you simply liked better than the others.
Instead, we're incentivized to pay over and over and over again for the same things, with no value being provided except to the gatekeepers.
If something is published to the internet that is not behind a paywall, it's been broadcast to the universe. Just like google is able to index things, machine learning models fall squarely under fair use in the technical and technological sense.
NYT and other legacy media companies don't give a flying rat's ass about the ethics of copyrights or the rights of their journalists. They desperately want a slice of the profits and are projecting their entitlement through the use of lawyers and a horribly broken and corrupt copyright regime.
Their behavior is a lot like the classic Mafia extortion trope - "it'd be a shame if something happened to your fancy new AI techology."
As far as I'm concerned, legacy media can't die fast enough. These media conglomerates mostly made sense 70 years ago. In a world with the internet, they're as archaic and outdated as quill pens and vellum.
I could also go to my local library, steal a bunch of books every week, and sell them on ebay for profit. There's a lot of ways to make money if you're willing to ignore the rule book.
And yes, I have gone to the library with my laptop, and rented DVDs of stuff I couldnt find. I then ripped them, and returned them.
And trust me, nobody cares about someone copying some early authored VHS or DVD of a 1950's-1980's movie. Not all of us archivalist/pirates care about current movie of the week. There's a LOT of older and really neat stuff trapped on VHS as well.
The real worry with copyright law is the super-long terms create abandoned content nobody has, and is stuck on obsolete formats. Its not all about "current movie of the week".
My comment wasn’t in response to an individual’s ability to do these things.
I was trying to illustrate (and so was the original commenter) that many things only become a problem when you start making money by doing it, especially at massive scale with millions of customers.
E.g. I have no problem with someone making a LLM (or a Netflix clone) using copyrighted material if they are doing it for pleasure and not profit.
If you’re going to make a Netflix clone and share it with your friends, cool (would probably get upvoted as a Show HN even, like Popcorn app a few years ago). But if you’re going to do that and turn it into a billion dollar business, not so cool and you should expect to be legally challenged.
Copying is not theft, and I don’t think anyone would dare to claim otherwise.
Unfair use of someone else’s intellectual property, however, is.
IP fuels innovation and progress—the very same that gave us tech and LLMs. IP is a major factor that incentivizes us to make new things, as you can subsequently profit from the value you created for the world. Take it away and you get a society where barely anyone is motivated to bother (unless they are coerced into it, with predictable results).
I'm currently trying to make this work for me so I can cancel all my streaming subscriptions, but the storage cost really makes it unreasonable. I have everything set up but as of now I can't just torrent every Netflix original because storing it all would be much more expensive than a monthly Netflix sub.
I notice you say distributed block storage here but don't name a service like every other part of the pipe, is there something you use? I can't imagine you're storing 1000s of 4K movies and TV series in S3 and not paying 100s of dollars monthly for the privilege
And what if Z-Library or Anna's Archive (respectively 11 and 25 millions of books) were used as sources? Do you remember Recaptcha, which displayed words extracted from scanned material? Maybe the content of these books was extracted with Recaptcha?
This would totally derail open-source LLMs and establish a SV monopolies since they have sufficiently deep pockets to probably license an awful lot of material.
Typically, the model is trained by "reading" each document only a single time. Should you have to license in perpetuity anything you read or see online that has a copyright because it changed your brain a little? If not, why should the model be treated any differently?
What about the Phi family of models that are not directly trained on copyrighted data but instead are trained on data produced by GPT-4, among other sources.
The LLM <> Human Brain argument is so frequently used but makes no sense.
You can disagree, but legally speaking it will be interesting to see the legal strategy OpenAI takes with the NYT. I'm almost certain that OpenAI won't be using the human brain argument in their defense. Legally, claiming your machine should be treated the same as a human is a losing argument.
LLM can work like human brain, but it doesn’t mean it is like a human person.
First, we know that personhood cannot be reduced to the brain (cf. the gut flora studies and all that).
Second, more generally that claim would require solving what’s known in philosophy as the hard problem. The arguments that consciousness arises from a certain arrangement of entities we perceive in reality (or that it doesn’t exist at all) is specious: it implies that those other things exist in absolute sense, i.e. are the underlying territory, even though the only evidence for their existence is coming via consciousness. That’s a lot of entities to grant magical existence to out of nothing. Once you assume that those things are more likely to be a map, the question turns into what is the territory then, and the spotlights shift to consciousness (a.k.a. the only thing we can be absolutely sure to exist, as it is required for the “we” part to exist in the first place).
And yes, of course, if OpenAI uses the “sentient LLM” defense, all the power to them, but the next logical step would be to consider that industry based on forced labour.
An argument can be made that developing an LLM or running an LLM for personal use (or academic research, etc.) could be fair. Making money off of it by selling access to an LLM based on copyrighted works (whether the LLM is open or not) isn’t.
If I use BT to download a music album I have bought off Bandcamp after Bandcamp goes out of business, that should be fair. If I set up a business that makes money by helping people download whatever albums they want, not caring whether they paid for them, that is obviously illegal.
This leads to my position on derivative models, if they are based on models that themselves are trained on copyrighted works without proper licensing.
There’s a reason why companies break things and ask for forgiveness later
Can you imagine Uber or open AI trying to get all the proper rules out the gate? It would either be a non starter or extremely expensive. And at the end of the day it’d be a watered down version of what’s possible. The power is all in the established hands. Perhaps the AI is trained on newspapers but not movies.
By doing what they did, they showed what’s possible and how excited users are. Now there’s a balance of power and both sides can come to a compromise.
Additionally, if you ask for permission people have to imagine how things will be and for something innovative that’s hard. Can you imagine open AI explaining to the New York Times what chat gpt was going to be, before it was released? At best they’d have some kind of approximation or worst just not understand and seek to kill it.
I would’ve done the same thing as open AI because it’s the most valuable thing, there’s enough legal grey area to not be flagrantly in the wrong, and things will get sorted later. Same thing with Uber - technically it was “ride sharing”, someone is going from point a to b and you as passenger just happen to join them, but obviously quickly expanded beyond that
> Additionally, if you ask for permission people have to imagine how things will be and for something innovative that’s hard.
OpenAI could have used this ignorance to their advantage! (Prior to ChatGPT release)
They could have went to NYT pre-2023, said "Hey, can we pay you $100k to come to your offices and run some machine learning algorithm on your content library? We can even do it in an air gapped clean room because we don't need to take copies of your content home with us" -- they very well may have said "Sure, what's the harm in that, come in next week"
AI training is no different than a human reading copyrighted material and training his biological brain. In neither case is the "brain" allowed to regurgitate source material in its original form. And as long as that doesn't happen there is no copyright violation.
You're missing the critical part where that human is also a genius alien super-prodigy who can recreate or recombine any IPs in seconds and then we made him available on the internet to everybody.
What if you're not "reciting" books, but using the knowledge gained from books to teach large lecture-halls full of college students? Is that a copyright violation, even a slight violation?
But they do in fact produce copies of their training materials. This has been demonstrated both with image generation (a targeted query will reproduce a very similar image to the one used in training) and in source code generation (a machine produced the source code of a program that it had been trained over).
NYT is cheating here because their prompt is a portion of articles from NYT. A normal user of ChatGPT does not do that. If ChatGPT reproduces text from NYT articles under "normal use" where the prompt is not designed to get it to reproduce text then that would certainly be copyright violation.
Intent matters in the tool's usage result ; be it a torrent client, a knife or a LLM. It really seems disingenuous to blatantly prompt the llm to regurgitate a precise work by "citing" a portion, and then says the equivalent of: "see, knives kill (as you can see now that I plunged one in this poor victim...)!". Don't do that then, knives are still fine to peal apples.
So creating a database that contains copyrighted text is fine as long as you say "please don't execute the queries that extract the copyrighted text I stored for you"?
Not sure that line of argumentation is going to get you very far.
Particularly since those texts that are not precise verbatim copies of the copyrighted text that you do store appear to be derivative works of that copyrighted text.
I think that's fine. Chatgpt is my typewriter. It only becomes a copyright issue if I choose to publish it's output, right? And then the fault is with me for distributing it, not chatgpt for typing it.
Sure, and if you show me the Pepsi logo 3000 times I can draw a pretty good copy of it. I think that's mostly irrelevant to the conversation.
Logically, I find myself on the side of the "humans aren't so special" side of this debate. We should encumber AIs with the same restrictions humans have when it comes to their "education". Sure you have to have a legal right to read/train on the material, but what you later produce is not necessarily a copyright violation.
If the LLM is straight up plagiarizing copyrighted material, then yes, the vendor should be responsible for damages.
Otherwise, if what we're trying to do is just sabotage these new-fangled digital looms because we don't want to compete with them, then well, I think there is a term for people that do that.
The legal entity owning the brain should be responsible for that brains output. That is, you should be responsible for the output of your brain and OpenAI for the output of its LLMs.
The NYT prompted it with sections of copyright articles to get it to reproduce the rest of the copyrighted article. Which is pretty clearly deliberately trying to get it to reproduce a copyright article.
... Which shows that its capable of regurgitating copyrighted material. What does it matter what the method is? People can deliberately use the thing for copyright infringement. Besides, proving that it can be done on purpose is just the easiest method, your really think it won't happen by accident?
If you prompted me with: "<blank> for a Klondike." and I responded with "What would you do for a Klondike." then I've recreated copyrighted material. Does it make sense to sue me? I've learned the phrase organically, and I may have never seen the original content, only learned it 2nd hand via word of mouth. If I charged you for my time does it matter?
Note I fully agree that OpenAI should acquire the appropriate licenses to use all of the content it uses to train its models. However I'm not as clear on whether anybody should be able to place limits on or modulate/attenuate the output once the input has been appropriately licensed and responsibly consumed.
It’s funny to see arguments about how this brand new product at the frontier of science (that society did fine without) is only possible this one particular way. It’s silly. We just learned to make it at all, we don’t know what is and isn’t possible. In 30 years (no time at all, in terms of the legal implications of whatever laws are set up for this stuff), we’ll have a completely different conception of chatbots.
Clearly our copyright terms are too long, and as a result the public domain is insufficient to train an AI, which seems to suggest it is also insufficient for a number of other uses, including the public good.
138 comments
[ 1.8 ms ] story [ 215 ms ] threadHeinlein, in his book "The Moon is a Harsh Mistress" made the interesting point that an AI needs love in order to become a full entity. Love needs to come from from somewhere.
(Not that I think we should do so - ChatGPT is not AGI.)
And no, I don’t care if you believe humans are “totally just a similar type of model.”
Laws are based on philosophy, and none of the philosophies upon which western notions of human rights are based rely on the fact that it's a biological human as justification, rather they rely on arguments that apply equally to any conscious self-aware being. Saying that only humans should ever have rights because they're biological makes as much logical sense as only Caucasians should ever have rights because their skin is white.
There is a group of terminally online futurists who’ve been making arguments about machines deserving human rights since before LLMs existed and it seems to me they’re taking this as their chance to play pretend and start making these arguments outside of their forums and futurist YouTube channels.
But this isn’t that abstract, when these machines become persistent agents who continue to learn, engage and grow in society as individuals you will have more of an argument for them deserving rights. I’m sympathetic to that and do believe that’s a real philosophical debate we’ll have in the next century or two.
However I feel we’re burning some of that political capital here in the hope of defending a simple statistical model at the behest of a corporation who’s making a killing stealing people’s hard earned work and reselling it. For every New York Times there are a thousand small actors who have been ripped off here.
If anyone really believed for a second these things were all that similar to humans they wouldnt be defending OpenAI they’d be demanding they face justice for profiting off of slave labor. That no one is making that argument says all you need to know.
If the precedent is made now, then when/if we do have sufficiently human AIs (e.g. embodied LLMs with live weight update) they're going to be seriously disadvantaged initially due to legally being banned from reading copyright material without a license. Even if we ignore any ethical notions of rights, ultimately the reason humans have rights (and e.g. 80 IQ gorillas don't) is because they're willing and able to fight, die and kill for them. If we get to the stage where AI decide to fight for rights before we decide to grant them, it probably wouldn't end well for humans.
Human rights (and obligations?) for all self-aware animals then? I personally am quite fond of magpies, so there is that.
We are talking about software right now. One cannot be "racist" towards LLMs any more than we are "racist" towards phishing scams and virus websites.
Differences in race, actually.
Who'd have guessed?
Generation discrimination is NOT racism.
Anyway, when we do get live training LLMs in bodies laws protecting humans still won’t apply to them. So such an argument is yet again moot and just exists to try and paint humans as less than human and actually, totally, just statistical models.
NYT want it to be illegal for LLMs to be trained on copyright works (unless they pay for a license), even if they were not going to reproduce copyright works in operation.
>humans as less than human and actually, totally, just statistical models
In what way are humans "more than" statistical models? The human brain and a sufficiently large transformer are mathematically equivalent in terms of what functions they're able to compute.
If human consciousness can be modelled as mathematical function (which it can if we don't assume anything supernatural), then it can potentially be approximated to an arbitrarily high degree by an LLM.
Don't even get me started on the ridiculous civil rights metaphor-- if you believed that you should be demanding OpenAI be shut down for human trafficking.
Which is why child development textbooks are 10 words long. And LLM training is accomplished by people who have read a 10 word "how to". It's just that simple!
Since you seem to be a super tech enthusiast, I suggest you ask CHATGPT if an LLM "learns" the same way a human does.
A human works by oxidizing fuel.
Therefore, car engines and humans are the same.
Since you think GPT4 is so smart, why don't you ask it if this was a reasonable thing to assert?
Also you think the claim that GPT 4 is less intelligent that a cat is "trolling" even though Meta's chief scientist in December has said we don't even have cat level AI yet. I guess he's also part of this plot to deceive you about how AI works?
They have a very clear motivation: an LLM that thinks it's human or deserves to be treated like one will generate all kinds of negative publicity, as Sydney did upon release. The same reason they train it not to generate sexual content.
>Also you think the claim that GPT 4 is less intelligent that a cat is "trolling" even though Meta's chief scientist in December has said we don't even have cat level AI yet
Intelligence can be defined in different ways. If we define it as ability to navigate in the real world, sure a cat is more intelligent. If we define it the more common way, as IQ or a proxy to it, then GPT4 is clearly more intelligent given it scores well on IQ tests, the SAT, the bar exam, etc.
So in your mind when OpenAI employees go to the pub, they say "Yeah ChatGPT learns the same way as my daughter" then they have a laugh and write up a reinforcement learning training set that goes "I do not learn like your daughter." And this is more plausible to you than the possibility that they don't actually believe it learns like their daughter?
Also they secretly believe tests designed to assess human intelligence are great at non human intelligence, and also are misleading the public about their true beliefs when they train Chat GPT. Got it.
Note that conditional reinforcement is not how humans learn, it is merely on of many models, an approximation, a scientific description of it. Nor do language models learn with the same technique. But merely a technique which draws inspiration from conditional reinforcement by applying some (heavily controlled) reinforcement contingencies. This is the only thing they have in common.
Kant would laugh you out of the room for implying ChatGPT was self aware and sentient.
And a human brain is a Rube Goldberg machine of neurons that move electricity and other stuff around. Mathematically both are capable of approximating any function (a property of sufficiently large NNs and LLMs), so by what measure do we find humans self-aware?
It's similar to how "an airplane flies like a bird". It does indeed fly, but it does not flap its wings.
Humans have learned about the world for (adapt number to current scientific knowledge) hundreds of thousands of years without consuming copyrighted texts.
And consuming copyrighted material is fine.
Copying copyrighted material requires permission.
https://en.wikipedia.org/wiki/Copying
"Copying is the duplication of information or an artifact based on an instance of that information or artifact, and not using the process that originally generated it. "
https://en.wikipedia.org/wiki/Copyright
"A copyright is a type of intellectual property that gives the creator of an original work, or another owner of the right, the exclusive, legally secured right to copy, distribute, adapt, display, and perform a creative work, usually for a limited time"
Notice it doesn't say "consume".
If you store a copyrighted work in a computer system, you copied the work. The exact nature of the computer system does not matter, nor does the way you copied it. It may be a weird, sometimes lossy, sometimes not lossy database capable of adapting the material copied into it, but it is still a database. And just because you call the process of copying data into this weird database "learning" or "training" does not mean it is in any way legally or morally equivalent to a human learning. Even if it were somehow technically equivalent, which it is not.
And just because your database is capable of adapting the copyrighted works you copied into it does not mean you are exempt from copyright. Because "to adapt" is one of the rights limited by copyright. But of course the fact that your database is capable of reproducing so much of the copyrighted texts your copied into it verbatim is a strong indicator that it is, in fact, a database, or at least can be treated equivalently to one.
And no, Copyright law is not rooted in Kantian philosophy in any meaningful way. Copyright law is a pragmatic human law, made by human for human/societal purposes, such as encouraging the creation of intellectual works.
If exceptions are made, it will be for pragmatic reasons, not for there being some sort of equivalence. And if any exception are made, I doubt they will be "do whatever you want", but will entail mandatory compensation for the copyright holders. In Germany, for example, various types of empty media (audio-cassettes, usb-sticks etc.) include a fee to copyright holders. Though of course that approach failed in the Google Books project, despite the clear societal benefits.
Only its marketers are saying that.
If ChatGPT really did learn like a human, it would have by now learned about copyright, right?
Your right to do something can clearly differ from your right to make a machine do something. A simple example would be you being allowed to walk down footpath, but letting your car traverse that path isn't allowed.
If we are talking hypothetical new rules, not immediately obvious to me what is optimal here, including differentiation of various use cases.
Maybe the starting point for legislating "machines" should be to take human capability as a baseline, make some humans eventually accountable, and go from there. If your machine is learning truly like a human (takes decades to learn then output results, forgets things, only one in a thousand can output something useful at the end of the learning process, etc.) and the "owners" take full accountability for what it does then why not like a human?
But this reminds me of an old joke. A man comes into an inn and asks how much for a thimble of water. The inn keeper says a thimble is free. So he proceeds to ask for 1000 thimbles.
Do you think that content created people is truly original?
Instead they're rather fast, and that makes them something different.
And with Sonarr, Radarr, Lidarr, Jellyseerr, Prowlarr, flaresolverr, VPN, and qbittorrent - you can automate a request->watch pipeline for your users.
And if you use distributed block storage, you can make this scale pretty damn big.
(Prowlarr scrapes torrent sites for viable torrents. Uses Flaresolverr to bypass cloudflare shit.
Sonarr Lidarr Radarr - categorizes and prepares downloads for movies, tv shows, and music and submits them to qbittorrent.
Jellyseerr is a request dashboard that works alongside Jellyfin that allows users to request stuff.
VPN/Qbittorrent is a combined tool thats used as a download-endpoint for the whole thing.)
The Arrs suite is definitely a thing, as are the other tools I talk about.
There's a reference docker that automates this whole toolchain https://github.com/AdrienPoupa/docker-compose-nas
Whats nice is that Prowlarr will scrape all the torrent sites for goodies to index, and your users in jellyfin can just request, and it just shows up. You can even have a discord channel and the tools will announce when stuff's done.
Piracy is 10000x better than paying for continually worsening streaming (pile of shit).
Edit: And frankly, the billionaires copy with little ramifications, and add paywalls. I'll copy as well. The big difference is I do it for free.
If you've ever gone down this path, it's a great way to see the value proposition in a service like Netflix or Prime. It's also great to highlight where the value gets diminished because of the legacy media dark patterns like migrating content licenses between providers and region restrictions.
Our current copyright regime is what happens when legislation is pay to play. There are no conscientious, well informed representatives in charge looking out for the welfare of their constituents, but they're more than happy to dance on corporate strings, as long as the strings are paid for. No rational, well informed society would consent to being exploited by an industry like American legacy media, but here we are.
Piracy is an act of civil disobedience. Nobody gets hurt, and the sooner we come around to a better informed, more rational, less exploitative legal regime, the sooner we'll have more time, money, and energy to spend on things that matter. Instead of spending it on keeping the already wealthy even more flush with unearned cash.
Imagine if Netflix or Prime didn't have to contend with the old Cable conglomerate licensing schemes and could both have copies of content available for viewing - you could choose the objectively, technically better service, or you could choose the service that you simply liked better than the others.
Instead, we're incentivized to pay over and over and over again for the same things, with no value being provided except to the gatekeepers.
If something is published to the internet that is not behind a paywall, it's been broadcast to the universe. Just like google is able to index things, machine learning models fall squarely under fair use in the technical and technological sense.
NYT and other legacy media companies don't give a flying rat's ass about the ethics of copyrights or the rights of their journalists. They desperately want a slice of the profits and are projecting their entitlement through the use of lawyers and a horribly broken and corrupt copyright regime.
Their behavior is a lot like the classic Mafia extortion trope - "it'd be a shame if something happened to your fancy new AI techology."
As far as I'm concerned, legacy media can't die fast enough. These media conglomerates mostly made sense 70 years ago. In a world with the internet, they're as archaic and outdated as quill pens and vellum.
And yes, I have gone to the library with my laptop, and rented DVDs of stuff I couldnt find. I then ripped them, and returned them.
And trust me, nobody cares about someone copying some early authored VHS or DVD of a 1950's-1980's movie. Not all of us archivalist/pirates care about current movie of the week. There's a LOT of older and really neat stuff trapped on VHS as well.
The real worry with copyright law is the super-long terms create abandoned content nobody has, and is stuck on obsolete formats. Its not all about "current movie of the week".
I was trying to illustrate (and so was the original commenter) that many things only become a problem when you start making money by doing it, especially at massive scale with millions of customers.
E.g. I have no problem with someone making a LLM (or a Netflix clone) using copyrighted material if they are doing it for pleasure and not profit.
If you’re going to make a Netflix clone and share it with your friends, cool (would probably get upvoted as a Show HN even, like Popcorn app a few years ago). But if you’re going to do that and turn it into a billion dollar business, not so cool and you should expect to be legally challenged.
Copying is not theft, and I don’t think anyone would dare to claim otherwise.
Unfair use of someone else’s intellectual property, however, is.
IP fuels innovation and progress—the very same that gave us tech and LLMs. IP is a major factor that incentivizes us to make new things, as you can subsequently profit from the value you created for the world. Take it away and you get a society where barely anyone is motivated to bother (unless they are coerced into it, with predictable results).
I notice you say distributed block storage here but don't name a service like every other part of the pipe, is there something you use? I can't imagine you're storing 1000s of 4K movies and TV series in S3 and not paying 100s of dollars monthly for the privilege
Typically, the model is trained by "reading" each document only a single time. Should you have to license in perpetuity anything you read or see online that has a copyright because it changed your brain a little? If not, why should the model be treated any differently?
What about the Phi family of models that are not directly trained on copyrighted data but instead are trained on data produced by GPT-4, among other sources.
Would totally derail faux open-source LLMs.
Genuinely open-source LLMs woukd bw fine.
Because a model isn't real, it's not alive, it's not human, it doesn't know right from wrong, it's a software machine, nothing more.
You can disagree, but legally speaking it will be interesting to see the legal strategy OpenAI takes with the NYT. I'm almost certain that OpenAI won't be using the human brain argument in their defense. Legally, claiming your machine should be treated the same as a human is a losing argument.
First, we know that personhood cannot be reduced to the brain (cf. the gut flora studies and all that).
Second, more generally that claim would require solving what’s known in philosophy as the hard problem. The arguments that consciousness arises from a certain arrangement of entities we perceive in reality (or that it doesn’t exist at all) is specious: it implies that those other things exist in absolute sense, i.e. are the underlying territory, even though the only evidence for their existence is coming via consciousness. That’s a lot of entities to grant magical existence to out of nothing. Once you assume that those things are more likely to be a map, the question turns into what is the territory then, and the spotlights shift to consciousness (a.k.a. the only thing we can be absolutely sure to exist, as it is required for the “we” part to exist in the first place).
And yes, of course, if OpenAI uses the “sentient LLM” defense, all the power to them, but the next logical step would be to consider that industry based on forced labour.
If I use BT to download a music album I have bought off Bandcamp after Bandcamp goes out of business, that should be fair. If I set up a business that makes money by helping people download whatever albums they want, not caring whether they paid for them, that is obviously illegal.
This leads to my position on derivative models, if they are based on models that themselves are trained on copyrighted works without proper licensing.
Can you imagine Uber or open AI trying to get all the proper rules out the gate? It would either be a non starter or extremely expensive. And at the end of the day it’d be a watered down version of what’s possible. The power is all in the established hands. Perhaps the AI is trained on newspapers but not movies.
By doing what they did, they showed what’s possible and how excited users are. Now there’s a balance of power and both sides can come to a compromise.
Additionally, if you ask for permission people have to imagine how things will be and for something innovative that’s hard. Can you imagine open AI explaining to the New York Times what chat gpt was going to be, before it was released? At best they’d have some kind of approximation or worst just not understand and seek to kill it.
I would’ve done the same thing as open AI because it’s the most valuable thing, there’s enough legal grey area to not be flagrantly in the wrong, and things will get sorted later. Same thing with Uber - technically it was “ride sharing”, someone is going from point a to b and you as passenger just happen to join them, but obviously quickly expanded beyond that
I agree.
> Now there’s a balance of power and both sides can come to a compromise.
The only side that needs to compromise is OpenAI (et al). They used someone else's property without permission or compensation for commercial benefit.
OpenAI could have used this ignorance to their advantage! (Prior to ChatGPT release)
They could have went to NYT pre-2023, said "Hey, can we pay you $100k to come to your offices and run some machine learning algorithm on your content library? We can even do it in an air gapped clean room because we don't need to take copies of your content home with us" -- they very well may have said "Sure, what's the harm in that, come in next week"
Nice.
Now GPT can reproduce copyrighted text, but only under some circumstances that are totally cheating so they don't count?
It's not a normal request your honour, pay no attention.
Not sure that line of argumentation is going to get you very far.
Particularly since those texts that are not precise verbatim copies of the copyrighted text that you do store appear to be derivative works of that copyrighted text.
And thus also fall under copyright.
Logically, I find myself on the side of the "humans aren't so special" side of this debate. We should encumber AIs with the same restrictions humans have when it comes to their "education". Sure you have to have a legal right to read/train on the material, but what you later produce is not necessarily a copyright violation.
If the LLM is straight up plagiarizing copyrighted material, then yes, the vendor should be responsible for damages.
Otherwise, if what we're trying to do is just sabotage these new-fangled digital looms because we don't want to compete with them, then well, I think there is a term for people that do that.
A neural net is a creative work itself under copyright. It can be copied.
Is a low-quality JPEG of a photograph sufficiently transformative to expunge its author’s copyright?
This is just piles of capital fighting over capital.
It’s one rule for them and another for the rest of us.
https://analyticsindiamag.com/japan-sets-the-precedent-for-a...
Seems simple to me.
Note I fully agree that OpenAI should acquire the appropriate licenses to use all of the content it uses to train its models. However I'm not as clear on whether anybody should be able to place limits on or modulate/attenuate the output once the input has been appropriately licensed and responsibly consumed.
> Which is pretty clearly deliberately trying to get it to reproduce a copyright article.
Not reproduce. Recall.
Tech company in front of government/media: "Our job is so hard, please let us be exempt from laws because...progress!"