Which of the following sounds more reasonable?
-
I shouldn’t have to pay for the content that I use to tune my LLM model and algorithm.
-
We shouldn’t have to pay for the content we use to train and teach an AI.
By calling it AI, the corporations are able to advocate for a position that’s blatantly pro corporate and anti writer/artist, and trick people into supporting it under the guise of a technological development.
I see it like this:
Our legal system has the concept of mechanical licensing. If your song exists, someone can demand the right to cover it and the law will favor them. The result of an LLM has less to do with your art that a cover of your song does.
There are plenty of cases of a cover eclipsing the original version of a song in popularity and yet I have never met a single person argue that we should get rid of the right to cover a song.
Sure, you have the legal right to cover someone else’s song without asking permission first, but you still have to pay them royalties afterwards, at fair market rates.
If an LLM was trained on a single page of GPL code or a single piece of CC-BY art, the entire set of model weights and any outputs from the model must be licensed the same way. Otherwise this whole thing is just blatant license laundering.
This depends on how transformative the act of encoding the data in an LLM is. If you have overfitting out the ass and the model can recite its training material verbatim then it’s an illegal copy of the training material. If the model can only output content that would be considered transformative if a human with knowledge of the training data created it, then so is the model.
IMO content created by either AI or LLMs should have a special license and be considered AI public domain (unless they can prove that they own all content the AI was trained on). Commercial content made based on content marked with this license would be subject to a flat % tax that should be applied to the product price which would be earmarked for a fund distributing to human creators (coders, writers, musicians etc.).
I think the cleaner (and most likely) outcome is AI generated work is considered public domain, and since public domain content can already be edited and combined and arranged to create copyrighted content this would largely clear up the path for creators to use AI more prominently in their workflows
Honestly, I’d personally prefer the latter, but there is the argument made by artists, coders and content creators. Their work is being scraped to train these AI’s, which in turn makes their future work less valuable. Hence, the thought of enforcing a tiny “royalty”/tax on commercial products based off of AI generated content and funneling that money back to human creators of intellectual works.
So I can make derivative works from commercial works, make something from that material, then release the result as public domain? I would think not.
What about LLM generated content that was then edited by a human? Surely authors shouldn’t lose copyright over an entire book just because they enlisted the help of LLMs for the first draft.
If you take open source code using GNU GPL and modify it, it retains the GNU GPL license. It’s like saying it’s fine to take a book and just change some words and it’s totally not plagerism.
Public domain is not infectious like GPL is. That being said, it seems like the parent comment has already mentioned this case, now that I’ve read them again:
public domain content can already be edited and combined and arranged to create copyrighted content
That’s fine by me. The important thing is that humans can still use AI as a legally recognized productivity tool, including using it as a way to use ideas and styles generated by other humans.
I think it’s the same reason the CEO’s of these corporations are clamoring about their own products being doomsday devices: it gives them massive power over crafting regulatory policy, thus letting them make sure it’s favorable to their business interests.
Even more frustrating when you realize, and feel free to correct me if I’m wrong, these new “AI” programs and LLMs aren’t really novel in terms of theoretical approach: the real revolution is the amount of computing power and data to throw at them.
The funniest thing I’ve seen on this is the ChatGPT CEO, Altman, talking about how he’s a bit afraid of what they’ve created and how it needs limitations – and then when the EU begins to look at regulations, he immediately rejects the concept, to the point of threatening to leave the European market. It’s incredibly transparent what they’re doing.
Unfortunately I don’t know enough about the technology to say if the algorithms and concepts themselves are novel, but without a doubt they couldn’t exist without modern computing power capabilities.
I can tell for a fact that there’s nothing new going on. Only the MASSIVE investment from Microsoft to allow them to train on an insane amount of data. I am no “expert” per se, but I’ve been studying and working with AI for over a decade - so feel free to judge my reply as you please
nothing new going on
Uhhhh the available models are improving by leaps and bounds by the month, and there’s quite a bit of tangible advancement happening every week. Even more critically the models that can be run on a single computer are very quickly catching up to those that just a year or two ago required some percentage of a hyperscaler’s datacenter to operate
Unless you mean to say that the current insane pace of advancement is all built off of decades of research and a lot of the specific advancements recently happen to be fairly small innovations into previous research infused with a crapload of cash and hype (far more than most researchers could only dream of)
all built off of decades of research and a lot of the specific advancements recently happen to be fairly small innovations into previous research infused with a crapload of cash and hype>
That’s exactly what I mean! The research projects I’ve been 5-7 years ago had already created LLMs like this that were as impressive as GPT. I don’t mean that the things that are going on aren’t impressive, I just mean that there’s nothing actually new. That’s all. IT’s similar to the previous hype wave that happened in AI with machine learning models when google was pushing deep learning. I really just want to point that out.
EDIT: Typo
nothing new going on
I can’t think of anything less accurate to say about LLMs other than that they’re a world-ending threat.
This is a bit like saying “The internet is a cute thing for tech nerds but will never go mainstream” in like 1995.
Cool, thanks for your take
Yw!
The concepts themselves are some 30 years old, but storage capacity and processing speed have only recently reached a point where generative AI outperforms competing solutions.
But regarding the regulation thing, I don’t know what was said or proposed, and this is just me playing devil’s advocate: but could it be that the CEO simply doesn’t agree with the specifics of the proposed regulations while still believing that some other, different kind of regulation should exist?
Certainly could be, but probably an optimistic take. Most likely they’re just trying to do what corporations have been doing for ages, which is to weaponize government policy to prevent competition. They don’t want restrictions that will materially impact their product, they want restrictions that will materially impact startups to make it more difficult for them to intrude on the established space.
I think if you fed your response into ChatGPT and asked it to summarize in two words it would return,
“Regulatory Capture”
And what are they doing? To remind, OpenAI is non-profit.
I thought they moved to for profit back in 2019?
Wikipedia lists them as non-profit https://en.m.wikipedia.org/wiki/OpenAI
They’re a non-profit managed by a for-profit, who’s received most of their funding from another for-proft.
LLMs are pretty novel. They are made possible by invention of the Transformer model, that operates significantly different compared to, say, RNN.
Even more frustrating when you realize, and feel free to correct me if I’m wrong, these new “AI” programs and LLMs aren’t really novel in terms of theoretical approach: the real revolution is the amount of computing power and data to throw at them.
This is 100% true. LLMs, neural networks, markov chains, gradient descent, etc. etc. on down the line is nothing particularly new. They’ve collectively been studied academically for 30+ years. It’s only recently that we’ve been able to throw huge amounts of data, computing capacity, and time to tweak said models to achieve results unthinkable 10-ish years ago.
There have been efficiencies, breakthroughs, tweaks, and changes over this time too, but that’s just to be expected. But largely its just sheer raw size/scale that’s just been achievable recently.
LLMs aren’t really novel in terms of theoretical approach: the real revolution is the amount of computing power and data to throw at them.
This is 100% true. LLMs, neural networks, markov chains, gradient descent, etc. etc. on down the line is nothing particularly new. They’ve collectively been studied academically for 30+ years.
Well LLMs and particularly GPT and its competitors rely on Transformers, which is a relatively recent theoretical development in the machine learning field. Of course it’s based in prior research, and maybe there even is prior art buried in some obscure paper or 404 link, but if that’s your measure then there is no “novel theoretical approach” for anything, ever.
I mean I’ll grant that the available input data and compute for machine learning has increased exponentially, and that’s certainly an obvious factor in the improved output quality. But that’s not all there is to the current “AI” summer, general scientific progress played a non-minor part as well.
In summary, I disagree on data/compute scale being the deciding factor here, it’s deep learning architecture IMHO. The former didn’t change that much over the last half decade, the latter did.
Now as I stated in my first comment in these threads, I don’t know terribly much about the technical details behind current LLM’s and I’m basing my comments on my layman’s reading.
Could you elaborate on what you mean about the development of of deep learning architecture in recent years? I’m curious; I’m not trying to be argumentative.
Could you elaborate on what you mean about the development of deep learning architecture in recent years?
Transformers. Fun fact, the T in GPT and BERT stands for “transformer”. They are a neural network architecture that was first proposed in 2017 (or 2014 depending on how you want to measure). Their key novelty is the method of implementing an attention mechanism and a context window without recursion, which was the method most earlier NNs used for that.
The wiki page I linked above is admittedly a bit technical, this articles explanation might be a bit more friendly to the layperson.
Thanks for the reading material: I’m really not familiar with Transformers other than the most basic info. I’ll give it a read when I get a break from work.
by that logic there was nothing novel about solid state transistors since they just did the same thing as vacuum tubes; no innovation there I guess. No new ideas came from finally having a way to pack cooler, less power hungry, smaller components together.
Okay, I’m glad I’m not too far off the mark then (I’m not an AI expert/it’s not my field of study).
I think this also points to/is a great example of another worrying trend: the consolidation of computing power in the hands of a few large companies. Without even factoring in the development of true AI/whether that can or will happen anytime soon, the LLMs really show off the massive scale of both computational power consolidation AMD data harvesting by only a very few entities. I’m guessing I’m not alone here in finding that increasingly concerning, particularly since a lot of development is driving towards surveillance applications.
We all remember SmarterChild…right?
I remember Tay
No, I have clearly forgotten: What was that?
AOL chatbot that did basic stuff.
If you ask Siri, “Do you sleep?” Siri will respond, “I don’t need much sleep, but it’s nice of you to ask.” Meanwhile, if you asked SmarterChild the same question, he would respond, “No, but I dream. I dream of a better world. A world where man and machine can coexist in peace and happiness.”
It also plays into the hype cycle they’re trying to create. Saying you’ve made an AI is more likely to capture the attention of the masses then saying you have a LLM. Ditto that point for the existential doomerism that they ceo’s have. Saying your tech is so powerful that it might lead to humanity’s extinction does wonders in building hype.
Agreed. And all you really need to do is browse any of the headlines from even respectable news outlets to see how well it’s working. It’s just article after article uncritically parroting whatever claims these CEO’s make at face value at least 50% of the time. It’s mind-numbing.
The fear mongering is pretty ridiculous.
“AI could DESTROY HUMANITY. It’s like the ATOMIC BOMB! Look at it’s RAW POWER!”
AI generates an image of cats playing canasta.
“By God…”
We could say that the human brain isn’t novel in terms of biological composition: the real evolution is the size increase compared to the body.
The fact that insects exist doesn’t make us less intelligent.
But I agree with the sentiment of the argument.
both sound the same to me IMO. Private companies scraping ostensibly public data to sell it. No matter how you word it they are trying to monetize stuff that is out in the open.
I don’t see why a single human should be able to profit off learning from others but a group of humans doing it for a company cannot. This is just how humanity advances at whatever scale.
I had a comment about the morality of it at first but I pulled it out. This is not an easy question to answer. Corporations gate keeping knowledge seems weird and dystopian but the knowledge is out there and they are just making connections between it. It also touches on copyright and fair use.
I agree it’s much more complicated an issue than most people give it credit.
I’ll note that there are plenty of models out there that aren’t LLMs and that are also being trained on large datasets gathered from public sources.
Image generation models, music generation models, etc.
Heck, it doesn’t even need to be about generation. Music recognition and image recognition models can also be trained on the same sort of datasets, and arguably come with similar IP right questions.It’s definitely a broader topic than just LLMs, and attempting to enumerate exhaustively the flavors of AIs/models/whatever that should be part of this discussion is fairly futile given the fast evolving nature of the field.
Still, all those models are, even conceptually, far removed frow AI. They would most properly be called Machine Learning Models (MLMs).
The term AI was coined many decades ago to encompass a broad set of difficult problems, many of which have become less difficult over time.
There’s a natural temptation to remove solved problems from the set of AI problems, so playing chess is no longer AI, diagnosing diseases through a set of expert system rules is no longer AI, processing natural language is no longer AI, and maybe training and using large models is no longer AI nowadays.
Maybe we do this because we view intelligence as a fundamentally magical property, and anything that has been fully described has necessarily lost all its magic in the process.
But that means that “AI” can never be used to label anything that actually exists, only to gesture broadly at the horizon of what might come.They would but that doesn’tv sound as sexy to investors.
That’s what it all comes down to when businesses use words like AI, big data, blockchain etc. Its not about whether it’s an accurate descriptor, its about tricking dumb millionaires into throwing money at them.
Fair enough.
I’m not sure what you’re trying to say here; LLMs are absolutely under the umbrella of AI, they are 100% a form of AI. They are not AGI/STRONG AI, but they are absolutely a form of AI. There’s no “reframing” necessary.
No matter how you frame it, though, there’s always going to be a battle between the entities that want to use a large amount of data for profit (corporations) and the people who produce said content.
True, and this is the annoying thing about people unqualified to talk about AI giving their opinions online. People not involved in the industry hear “AI” and expect HAL-9000 or Ava from Ex Machina rather than the software that the weather service uses to predict if it will rain tomorrow, or the models your doctor uses to help determine your risk of Heart Disease.
This is compounded further when someone makes a video simplifying what an LLM is and mentioning that the latest models use it, which leads to the chimes of “bUt iT’S jUsT aN Llm BrO iTs nOt AI” and “ItS jUsT a LOaD oF DaTa aND aLGorItHMs, tHaTs NoT AI”. A little bit of knowledge is a dangerous thing.
or that people are only exposed to trivial/childish publicly available examples.
This is actually exactly what I mean. Most people hear AI and envision something much, much more complex. It’s easier to argue that HAL-9000 is like a human and should therefore be allowed to freely view book content like a human, versus argue that a sophisticated LLM is like a human and should be allowed to freely view books like a human. That’s moreso where I’m coming from. And politicians are stupid enough to pass laws envisioning these as HAL-9000.
Right, where I’m coming from is that I don’t think the personhood arguments you see for why content should be free for it really hold any water. Whatever the case on its intelligence, it isn’t comparable to humans for copyright law
On the flip side, the same battle is also fought between giant corporations that amass intellectual property and the people who want to actually use that intellectual property instead of letting it sit in some patent troll’s hoard until a lawsuit op presents itself. Seeing as there are quite a few reasonably decent open-source LLMs out there like Koala and Alpaca also training on data freely available on the Internet, I’m actually rooting for the AI companies in this case, in the hopes of establishing a disruptive precedent.
That’s absolutely not correct. AI is a field of computer science/scientific computing built on the idea that some capabilities of biological intelligences could be simulated or even reproduced “in silicon”, i.e. by using computers.
Nowadays is an extremely broad term that covers a lot of computational methodologies. LLM in particular are a evolution of methods born to simulate and act as human neural network. Nowadays they work very differently, but they still provide great insights on how an “artificial” intellicenge can be built. It is only one small corner of what will be a real general artificial intelligence, and a small step in that direction.
AI as a name is absolutely unrelated with how programs based on the methodologies are built.
Human intelligences are in charge of all copyright part. AI and copyright are orthogonal, people are those who cannot tell the 2 and keep talking about AI.
There is AI, and there is copyright, it is time for all of us to properly frame the discussion on “copyright discussion related to <company>'s product”
What I’m getting at moreso is that comparisons to humans for purposes of copyright law (e.g. likening it to students learning in school or reading library books) don’t hold water just because it’s called an AI. I don’t see that as an actual defence for these companies, and it seems to be somewhat prevalent.
You can absolutely compare AI with students. The problem is that, legally, in many western countries students still have to pay copyright holders of the books they use to learn.
It is purely a copyright discussion. How far copyright applies? Shall the law distinguish between human learning and machine learning? Can we retroactively change copyright of material available online?
For instance, copilot is more at risk than a LLM that learned from 4chan, because licenses are clearer there. Problem is that we have no idea on which data big llms were trained, to know if some copyright law already applies.
At the end it is just a legal dispute on companies making money out of AI trained on data publicly available (but not necessarily copyright free).
My argument is that an LLM here is reading the content for different reasons than a student would. The LLM uses it to generate text and answer user queries, for cash. The student uses it to learn their field of study, and then apply it to make money. The difference is that the student internalizes the concepts, while the LLM internalizes the text. If you used a different book that covered the same content, the LLM would generate different output, but the student would learn the same thing.
I know it’s splitting hairs, but I think it’s an important point to consider.
My take is that an LLM algorithm can’t freely consume any copyrighted work, even if it’s been reproduced online with the consent of the author. The company would need the permission of the author for the express purpose of training the AI. If there’s a copyright, it should apply.
You have me thinking though about the student comparison. College students pay to attend lectures on material that can be found online or in their textbooks. Wouldn’t paying for any copyright material be analogous to this?
Students and LLM do the same with data, simply in a different way. LLM can learn more data, student can understand more concepts, logic and context.
And students study to make money.
Both LLMs and students map the data in some internal representation, that is however pretty different, because a biological mind is different from an AI.
Regarding your last paragraph, this is exactly the point. What shall openai and Microsoft pay, as they are making a lot of money out of other people work? Currently it is unclear as openai hasn’t released what data they used, and because copyright laws do not cover generative AI. We need to wait for interpretation of existing laws and for new ones. But it will change soon in the future for sure
AI has been a blanket term for Machine Learning, LLMs, Decision Trees and every other form of “intelligence”.
Unfortunately I think that genie is out of the bottle and it’s never going back in.
And it has been the technical term used in academia since the 1950’s. If anyone is surprised by this usage then they have not studied it, only watched movies.
They are 100% AI. It’s an umbrella term. Simple pathing algorithms in games are also AI.
I think you are likely right, but it’s more general than just about training costs. The term “AI” carries a ton of baggage, both good and bad.
To some extent, I think we also keep pushing back the boundary of what we consider “intelligence” as we learn more and better understand what we’re creating. I wonder if every future tech generation will continue this cycle until/unless humanity actually does create a general artificial intelligence–every iteration getting slightly closer but still falling short of “true” AI, then being looked at as a disappointment and not worthy of the term anymore. Rinse and repeat.
If we’re unmasking tech, LLM’s right now are also all just Computer Vision models with a lot of more abstraction layers thrown at them. Nothing but fit assessment machine with a ludicrous amount of extra steps.
I am convinced this is all pedantry, and these models are going to become the de facto basis for true AI at some point. It was already weird enough that this type of tech got discovered from the goal of checking if an image has a cat or not.
Both of those statements are reasonable. You shouldn’t have to pay to utilize anything you scrape from the internet, so long as you don’t violate copyright by redistributing it
In fairness, AI is a buzzword that came out well before LLMs. It’s used to mean “tHe cOmpUtER cAn tHink!”. We play against “AI” in games all the time, but they arent AI as we know it today.
ML (machine learning) is a more accurate descriptor but blah doesn’t have the same pizzazz as AI does.
The larger issue is that innovation is sometimes done for innovation’s sake. Profits gets mixed up there and a board has to show profits to shareholders and then you get VCs trying to “productize” and monetize everything.
What’s more is there are only a handful of players in the AI space, but because they are giving API access to other companies, those companies are building more and more sketchy uses of that tech.
It wouldn’t be a huge deal if LLMs trained on copywritten material and then gave the service away for free. As it stands, some LLMs are churning out work that could be protected under copywrite law by humans (AI work can’t be copywritten under US law), and turning a profit.
I don’t think “it was AI” will hold up in court though. May need to do some more innovation.
Also there are some LLMs being trained on public domain info, to avoid copywrite problems. But works go into the public domain after 70 years past the copywrite holder’s death (disney being the biggest extender of that rule), so your AI will be a tad out dated in it’s “knowledge”.
We shouldn’t have to pay for the content we use to train and teach an AI
Wait people think that sounds reasonable?
I do. If it’s publicly available, individuals should be able to learn from it. Artists don’t pay their influences that helped develop their style, we don’t pay the programmers that answer questions on stack overflow
Hell, I’m not sure generative AI should have to pay for training data at all. It points to a weakness in the system, and it doesn’t fix it - the field is getting away from needing existing datasets. GPT4 swallowed everything worth swallowing, and it’s already training GPT4.5. This would only make it harder for new players to compete in the generative AI space
It can’t profit only the few, it’s too big a force multiplier. Paying up front doesn’t fix it, recurring payments don’t fix it… That’s nothing but a payoff to a few people as this starts to eat the best parts of the job market
We need to think much bigger - we need to look at how we handle ownership as a society
