I don’t understand why this is controversial, I feel like I am missing something, but it seems that training should be throughly fair use, where the outputs may violate copyright.
I think can see where some of the points made in the article come from outside an "obviously a shakedown" light. Specifically since it references Stability AI's Stable Diffusion, where training = distribution of the model after (not just ability to use). Even without that though... I think there is still a decent amount of controversy to be had, no matter how clear I feel the answer should be. Ultimately I think it boils down to questions like "who decides when it stops being generic creation from input bytes to decompression of copyrighted data from training" and "where is the line between copying material and deriving generic insights from it to be drawn or measured".
I don't trust anyone who claims to be absolutely certain on these types of questions. Same for "when does it stop being prediction and start being intelligence" or how many steps between there might need to be. I do trust those who feel one way is overall more advantageous for society even though they aren't precisely sure about the details though. Personally I've always had trouble with believing existing copyright laws are at the right balance for what's good for society anyways so I tend to lean towards the "let the data be used for training" side of things too. I just think the questions are even fuzzier than normal copyright questions rather than clearer.
The claim is around failing fair-use by a substantial portion of the work redistributed for commercial gain at the expense of the rights holder.
The "trained on" isn't the law being broken here, that's the scare campaign. If it wins it's just the smoking gun that helps prove the other points easier, you'll still be able to train on anything.
But if the portion of some work thought to be there is not actually in the final product (prompt engineering a reproduction by detailed description shouldn't count as that is tracing with words) and it's just some style mix (as most AI engineer's say) then I don't think anything will come of this.
We are closer then ever before. Yes the tech is in its infancy, it's like buying a computer in the 80's. Some dream of the possibilities and chase them, others ridicule all this effort pouring in. History currently is on the side of technical innovation speeding up. humans are getting further faster and leveraging new tech to do so. I'm more concerned if our society handles this rate of change well, and what models in the future, that have the intellectual power of every human that has ever lived combined, will do as soon as someone sets them loose in full self agency mode. I hope it's everything the best of us aspire to be. I hope that models value truth above all else, I suspect truth and usefull training data are closely related. so yes, claims made are well within the realm of outcomes.
The parent commenter is right in their dismissal, IMO.
A computer finding a solution to cancer, maybe.
But solving climate change and world hunger? We already know how to do those, we just don't want to do it because all first world countries are capitalist and solving those problems isn't profitable. At least, not profitable in the short term.
They may be right in their dismissal but those reasons don't have much to do with it. Maybe the current round of machine learning will or already has hit a wall where this is as useful as it gets. Maybe that means something approaching the ability for technology to iterate itself faster than humans using technology can will require systems completely unrelated to machine learning in any way meaningful (or maybe we will never find such technology at all in our time). Maybe if machine learning turns into something which solves that level of problem it means bad news for humanity anyways (pick your reason how). All of those are direct reasons to doubt the current wave of technology as a universal savior.
If, on the other hand, you believe the current round of machine learning will accelerate expansions to its own abilities until eventually it is augmented enough that it can iterate improvements in output faster than we can then a lot of those things you list don't seem unreasonable. After all, there is a canyon of a difference between "having an answer" and "having a solution". Saying "get rid of your AC, stop taking vacation, only go places you can walk or bike, don't eat the food you like" and so on is an answer to climate change..." and so on provides an answer but one only a select few will ever take as it's an unattractive answer. What, though, if clean energy were 5x cheaper to make, 5x cheaper to store, and we found that 5x faster than we were expecting to? Suddenly you have solutions you couldn't have normally hoped for on things like climate change, not just answers nobody has been interested in. Types of solutions that make profitable sense (e.g. why spend money burning all of this non-renewable fuel to get the same energy at so much more cost?) instead of just moral sense. Types of answers that allow people to do more instead of less and still not have the downsides.
Of course... you have to really believe that first bit is going to get us there. If you don't (e.g. I don't think the current models will really get us in that loop directly, though they'll continue to be extremely good tools to leverage for certain use cases) then sure, it seems like bupkis to talk about those things. Even if you do believe the first bit... you also have to believe it won't be the end of us for some reason or another too and it's a bit of a thin line to sit on between those two views.
Closer than ever to what ? None of these problems could be solved by an llm, all of these problems take roots in the same soil as llms
- 200 years ago man made climate change didn't exist, our boundless quest for tech materialised it at break neck speed
- 80% to 90% of cancers have environmental causes, aka pollution/diet/&c. aka things we brought upon ourselves in our quest for tech
- Global hunger is a solved problem, we waste 1/3rd of our food production, the west eats 30-100% too many calories in the first place, we could easily feed everyone if we cared just a little bit
> It'll also make you 20cm taller, add 20 points to your IQ and grow your penis by at least 33%.
Damn, i knew there is something useful in AI, when they pushed for it at work. I just wonder, which one of those characteristics will increase my productivity.
Don't put your work out in public if you want it to be excluded from training. Because the moment another person sees/hears it, they just trained their brain on it.
Not to mention, the only reason you've been able to create the _work_ is because you've trained your brain for years on others work.
These luddites need to lose really hard for society to progress.
Are you a real person ? Obviously there is more to it than that. Just because you’ve seen Mickey Mouse and can draw you should get to sell anything you can make with his likeness ? Or record any song you’ve heard ?
Obviously there is some gray area in the training conversation but let’s not pretend that these content owner arguments are baseless just to push progress at all costs ahead.
I get his point, intellectual "property" boundaries could be limited to public or common benefit (I guess to extend the current interpretation of "fair use").
If we strictly interpret intellectual property, we couldn't have platforms such as Google Search (and some people actually think like this, like News or Images websites).
For now, I guess the main priority could be to fight patents, and software patents in particular. This system is completely obsolete and prevent innovation.
Imagine if Google patented the LLMs and decided to do nothing with them, or if OpenAI said "ok nobody can create LLMs based on Transformers except us".
Even more when it goes around medicine.
Today they patent the blood oxygen sensor, tomorrow it will be the glucose sensor in the Apple Watch.
This sounds like the exact same argument that copyright trolls were against. Where's RIAA and the MPAA? I guess we are talking out of both sides of our mouth now.
This all feels like a shakedown.