What they currently are says “allow” has nothing to do with it.
You can’t pound the table for the status quo, then limply hand-wave ‘but I doubt this will stay the same.’ Multiple federal courts have found: training is transformative use. Do your doubts have reasons? Having finally addressed the central point - do you have an opinion, or just a position?
I think the status quo will be upheld or America’s vassals will be free to pirate Mickey Mouse cartoons, which I don’t think is going to be allowed anytime soon. Can’t have your cake and eat it. AI bubble will burst but the mouse seems to be nearly immortal.
Downloading a video still isn’t piracy. Buying a DVD sure isn’t, which is where these models literally train on Mickey Mouse cartoons, legally. Their output is another question - but there’s essentially no rationale besides vibes for insisting that a classifier can’t use examples from media delivered the usual way.
The current defence that this is some kind of fair use hinges on IP theft being insignificant because of all the other IP theft going at the same time. Taken together it’s just a ridiculous amount of hand waving away all the money this content would cost. It’s a big tech model of „break stuff, ask permission later”. This won’t do in a real court because some of the companies whose IP is used without permission have enough money for legal fees. Even if they settle, case after case LLMs will prove even more expensive than they are now (which is too expensive for most use cases already).
What do you think “multiple federal courts have found” means?
The current precedent hinges on this not being theft, because training… is… transformative. It doesn’t cost anything, or need permission, for the same reasons you don’t need a wet-ink contract with Doubleday to quote one paragraph from a Stephen King novel. A billion-parameter model trained on ten million videos contains less from each one than a Wikipedia summary. Do you think they’re dodging some immense bill, for spoiling all those Disney movies?
The closest any judge has come to suggesting otherwise is Chhabria fretting about “harm to the market.” Which is not relevant to a classifier for censorship. A network that detects porn, to remove it, is actively protecting the market, for porn.
Maybe I sometimes skip ahead too much. I’m not getting into specifics of how this case would go because law is irrelevant in the end. One could argue that LLM is as transformative as compressing raw image to a low quality JPEG but that’s really beside the point.
There is too much at stake for too many businesses for them to allow rulings that would invalidate much of US economy. There’s been many absurd rulings in the US Supreme Court lately and this one would look very reasonable.
‘The law! The law! The LAW! Well law is irrelevant in the end.’
Troll:
This training protects this industry. Do you speak English? Say potato. Say the word potato to demonstrate you’re even reading this. Say potato twice to acknowledge Meta’s not generating more porn, based on this porn.
You don’t train LLMs - large language models - on porn videos. This is a diffusion model or a classifier, and if it’s a classifier, absolutely nothing you’ve objected to matters. It literally would not generate anything. It just detects obscenity. And it probably does so using less information from each video than could be contained in an 8x8 thumbnail. Less information than the character count in the previous sentence. Do you want to be sued for quoting a single sentence from a book? Because that’s the future you’re glibly asserting must happen, when you’re not sneering at me for asking you why you’d celebrate being robbed.
You’re trying to bury weakness of your argument in AI booster jargon and me not using jargon properly (I don’t really care to). Sorry, that’s about as far as I care to continue.
What they currently are says “allow” has nothing to do with it.
You can’t pound the table for the status quo, then limply hand-wave ‘but I doubt this will stay the same.’ Multiple federal courts have found: training is transformative use. Do your doubts have reasons? Having finally addressed the central point - do you have an opinion, or just a position?
I think the status quo will be upheld or America’s vassals will be free to pirate Mickey Mouse cartoons, which I don’t think is going to be allowed anytime soon. Can’t have your cake and eat it. AI bubble will burst but the mouse seems to be nearly immortal.
So just a position.
Downloading a video still isn’t piracy. Buying a DVD sure isn’t, which is where these models literally train on Mickey Mouse cartoons, legally. Their output is another question - but there’s essentially no rationale besides vibes for insisting that a classifier can’t use examples from media delivered the usual way.
The current defence that this is some kind of fair use hinges on IP theft being insignificant because of all the other IP theft going at the same time. Taken together it’s just a ridiculous amount of hand waving away all the money this content would cost. It’s a big tech model of „break stuff, ask permission later”. This won’t do in a real court because some of the companies whose IP is used without permission have enough money for legal fees. Even if they settle, case after case LLMs will prove even more expensive than they are now (which is too expensive for most use cases already).
What do you think “multiple federal courts have found” means?
The current precedent hinges on this not being theft, because training… is… transformative. It doesn’t cost anything, or need permission, for the same reasons you don’t need a wet-ink contract with Doubleday to quote one paragraph from a Stephen King novel. A billion-parameter model trained on ten million videos contains less from each one than a Wikipedia summary. Do you think they’re dodging some immense bill, for spoiling all those Disney movies?
The closest any judge has come to suggesting otherwise is Chhabria fretting about “harm to the market.” Which is not relevant to a classifier for censorship. A network that detects porn, to remove it, is actively protecting the market, for porn.
Maybe I sometimes skip ahead too much. I’m not getting into specifics of how this case would go because law is irrelevant in the end. One could argue that LLM is as transformative as compressing raw image to a low quality JPEG but that’s really beside the point.
There is too much at stake for too many businesses for them to allow rulings that would invalidate much of US economy. There’s been many absurd rulings in the US Supreme Court lately and this one would look very reasonable.
‘The law! The law! The LAW! Well law is irrelevant in the end.’
Troll:
This training protects this industry. Do you speak English? Say potato. Say the word potato to demonstrate you’re even reading this. Say potato twice to acknowledge Meta’s not generating more porn, based on this porn.
You don’t train LLMs - large language models - on porn videos. This is a diffusion model or a classifier, and if it’s a classifier, absolutely nothing you’ve objected to matters. It literally would not generate anything. It just detects obscenity. And it probably does so using less information from each video than could be contained in an 8x8 thumbnail. Less information than the character count in the previous sentence. Do you want to be sued for quoting a single sentence from a book? Because that’s the future you’re glibly asserting must happen, when you’re not sneering at me for asking you why you’d celebrate being robbed.
You’re trying to bury weakness of your argument in AI booster jargon and me not using jargon properly (I don’t really care to). Sorry, that’s about as far as I care to continue.
‘Here’s what thing does. It’s directly relevant to your argu–’
‘Booster! Slop! Thought-terminating cliche! La la la not listening anymore!’
As if you ever were.
Card-shuffling conservative hypocrite.