I'm kind of surprised by this. I understand there's basically two views: Content creators say "Hey, this is my stuff, you don't have permission to use it, you need to pay me" and there's Silicon Valley: "I'm allowed to look at your images, feeding your images into my machine is the same thing". Legally, the first view seems probably correct, but from a "history of silicon valley" view, breaking the rules in order to gain a competitive advantage in the market has always been the better strategy.
It'd be like if Uber launched by applying for NY taxi medallions. So this seems like a crazy risk:reward ratio here. Adobe is going to end up with a massive bill, a weak model and the hope that the guys who steal everything will face consequences. We've never seen Silicon Valley face consequences like that in the past so i don't see why you would bet it's going to happen this time.
The whole differentiator between firefly, the adobe image generator, is that it's trained on their licensed stock image library and commercially safe. Besides that, even openai is licensing some proprietary high quality content from proprietary sources. Even if the open Internet is deemed fair use, there's still a lot of content locked away to license
Adobe is an established player with real customers who don't want to deal with copyright issues. Telling them "this is safe to use because we paid for it" solves that issue.
With Uber, no one's suing individual riders or drivers over an illegal taxi ride.
This is such a tectonic shift for what the concept of media even means that any assumptions are on uncertain ground.
The courts could find that it isn't possible to sign away these rights. There are many other rights that the law does not permit to be signed away for much more tame reasons than "enables further creations of 'your' work (or in the case of actors, literally 'you') not only without your actual involvement, but even long after your death". I would expect that it at least comes with a time limit. "Artists can sign a license to use these rights but only until their death + 10 years, after that the licensee can't use anything trained on their work. Artists also have an unrevokable right to cancel the license at any time with 90 day notice." might be one possible outcome.
That seems like a crazy stretch to me. Not a lawyer but the rights courts typically don’t let you sign away are things like your freedom. Signing away your right to profit or control a video you created is super common already.
Content licensing is really well established and already allows for licensing for specific use and different royalty structures based on usage. An example of this is streaming vs theater vs syndication.
So, now there is a new venue to take into account and have the lawyers add a few pages to their contracts moving forward.
>courts typically don’t let you sign away are things like your freedom.
This is exactly my point. An AI that can replace you, personally, is closer to signing away your identity or freedom than the rights to display a specific already completed work.
The court telling an artist they're not allowed to sell their videos to the highest bidder seems like a greater infringement on their freedom. What if no one wants to pay for this person's videos except for an AI company? You're basically telling them they're not allowed to profit off of their work.
i was at adobe summit a couple weeks back, the push was ai heavy— the strategy is 100% around no copyright issues.
observationally they’re in an interesting position, balancing their artistic clients and their executive clients. a clean model is the easiest way to hedge their portfolio and reputation.
Artist don't want their work taken for free. And executive clients can know that court systems can be extremely fickle... It can go one way or an other depending jurisdiction and even one big enough going wrong can be expensive.
The Artists it was trained on already signed rights away to Adobe when they put their work for sale on Adobe Stock so they’ll get what they’re given really.
I am struggling with the Silicon Valley's latest business model that seems to be based on stealing all content in order to train AI to replace the very creators who created that content. If we then replace white-collar workers with AI and blue-collar workers with robots... and most of the population are jobless who's going to be able to pay for the content, the services, and the goods produced by AI and robots? Is it why the VC are in favour of universal basic income? But if we all go on UBI then what's the point to selling to us if that money could go to the VCs... but then... what do they do with the money if it ceases to circulate and incentivise people to work and trade?
> If we then replace white-collar workers with AI and blue-collar workers with robots...
...we would be living in post-scarcity and everything would be free. But that doesn't happen in the absence of AGI, what actually happens is that technology replaces some jobs and then people do the remaining jobs, which are now in higher demand because the things done by technology become cheaper and the money that had gone to pay for labor there now gets spent on something else, increasing demand for the other thing.
Technology has been replacing jobs for hundreds of years and we still have low unemployment.
The bullshit jobs are allowed by automation, not required. If you had that many bullshit jobs in the old days, everybody would starve because there wouldn't be enough labor going to producing and distributing food.
Many of the bullshit jobs are, in fact, destroying the cost/efficiency benefit of automation by imposing bureaucratic waste on things that would otherwise have become less expensive. If someone is doing a job which is useless, we could just give them a portion of someone else's productive work to do, have them each work less and reduce prices by the same amount and then average purchasing power would be the same but there would be less busy work and more recreation/creativity/entrepreneurship.
Big TVs used to be thousands of dollars, now they're $300. People used to pay money for long distance voice calls. 30 years ago 500GB of flash memory would have made you a millionaire, now it's less than $50. Prices come down when there is competition.
Bullshit jobs are the inverse of this. Some large organization has insufficient competition so it can waste resources without losing business. The law is excessively complicated so everyone in the jurisdiction has to hire compliance bureaucrats. Government contractors and employees doing unnecessary work want to keep their jobs and legislators don't want to lose their votes/money. Then everybody has to pay higher prices because they don't individually have the choice to avoid it.
> "Hey, this is my stuff, you don't have permission to use it, you need to pay me"
I see a lot of the first and second parts, but nowhere near as often the third part: The rights holders aren't seeking financial growth, just wants control in perpetuity. I suspect that's the part proving difficult to solve.
Artists aren't going to like this technology because it competes with them, but it competes with them regardless of whether it was their work or someone else's in the training data.
This leads to a visceral response where they want to call this "stealing" and hope that the creators of the technology can be sued into non-existence so their competition can be eliminated. But as Adobe is demonstrating, that isn't going to happen anyway. So the question isn't whether the technology will exist, it's if it will be locked up behind the walls of major corporations. The latter doesn't do artists any good but harms the public -- including artists who want to leverage the technology in their art. So why should the law protect Adobe's moat from the public?
The reason I think the first view seems more correct is that it's like downloading a song from spotify. Yes you would think streaming a song from spotify is technologically identitical as downloading it, but legally there is a distinction. If you found a way of ripping a copy of a song from spotify there would be a record company ready to sue you and a law they could use to do it.
It's theoretically true that the models could be trained with someone else's training data, but there's a flaw in that argument. If you can train with other data without these legal issues, why don't they? And the answer is actually because they do assign some value to that training data, and there's not an infinite supply and it's actually quite difficult to get large sets of quality data.
I think it's a pretty open question how this will resolve, it could be like stremaing music where companies like spotify are little more than puppets for the major record labels. It could end up like youtube where the model started with "We're going to steal stuff" and ended up "We're going to strongly enforce copyright now we're the encumbent" or some other third way. But I don't expect the "We're going to take everything with no regard to the existing legal framework" will sustain long term.
> The reason I think the first view seems more correct is that it's like downloading a song from spotify. Yes you would think streaming a song from spotify is technologically identitical as downloading it, but legally there is a distinction. If you found a way of ripping a copy of a song from spotify there would be a record company ready to sue you and a law they could use to do it.
But the law would probably be DMCA 1201 for circumventing the DRM rather than normal copyright for making the copy, which is much more ambiguous. Also, record companies like to sue people, that doesn't mean they're right and provides no indication of what the law should be. You could just as easily pick some other example, like whether you can rip a music CD or vinyl you bought to put it on your iPod, which the record companies might not like to be allowed, but that doesn't mean that it isn't.
> It's theoretically true that the models could be trained with someone else's training data, but there's a flaw in that argument. If you can train with other data without these legal issues, why don't they?
To which the answer is that they do. A lot of models are trained on arbitrary content from the internet.
As to why Adobe in particular is doing this, think about it. It's because that interpretation benefits them rather than the artists, by creating a moat where companies who already have licenses to bulk stock images etc. are the only ones who can create a model, rather than having lots of competitors because anyone can create one and many of them are free and can be run locally.
> I don't expect the "We're going to take everything with no regard to the existing legal framework" will sustain long term.
Publishing companies don't like public libraries. So anybody can go there and borrow a copy of any book for free? That doesn't mean that libraries are bad or are or should be illegal.
Also, laws are created through the political process, which is not always great. The outcome "individual artists somehow benefit from this" isn't even in the room there. The two most plausible outcomes are that "tech companies" win and anybody can train a model on anything they can get their hands on, and that "content conglomerates" win and then this technology gets locked up as a service from only megacorps and the artists still don't get anything meaningful, but now the world has another abusive cartel imposing arbitrary censorship and using control over this to cement control over adjacent markets etc. Of these, the second is clearly worse.
Weird spot. $3 per minute seems like a lot more than most AI companies are willing to pay, and a lot less than most creators who are not making slop would be willing to take.
If I’d posted 10 minute videos weekly for 5 years and someone offered me $7500-ish to use them as training data, I’d think that was a pretty good deal. YMMV.
Depending on the content that’s a terrible deal. Low ball is it takes an hour to produce a single minute of video. You’re being paid 3 bucks an hour. If it’s animation it could take a day to animate a single minute, in which case you’re being screwed even harder.
Suckers who don’t properly value their time is why BigCo generally wins.
Did they say they’re buying only high quality videos? What if it’s just a video of me talking through my coding sessions? It would be rather low effort and sounds like “free money” to me
would it though? nothing would be stopping them. the idea something would relies on an assumption that we'll soon be able to generate long, coherent, and useful instructional videos with fully resolved text on demand, with such high quality and low cost that no one will be able to compete. but we already have people out there who can do this instruction/review live and off the cuff, and who would certainly be able to make use of this stuff in their own work
Because that style isn't that unique, so the difference is between getting paid nothing and now there is a bot that can do the same thing, or getting paid something and now there is a bot that can do the same thing.
It hasn't even been established that they're required to pay you at all.
My style is unique. If I'm teaching folks something I'm skilled in I wouldn't want to be learned from $3 especially for a $$$ company that abuses trust.
Style is how teachers make learning happen. If teachers follow the general generic book mundaneness, you learn far less than of you apply your own style.
Besides, if they were using the content without my permission I should be allowed to seek costs for such.
It doesn't matter if the thing is using your exact style or one which is enough of a substitute for it that the difference isn't going to make up the difference between your $50 fee and the $0.01 in electricity it takes to have the AI do it.
> Besides, if they were using the content without my permission I should be allowed to seek costs for such.
On what basis? How is it different than someone teaching their students your style, so the students can make their own original works in the same style? It's directly analogous to classroom use, which is an explicit example of fair use from the copyright statute.
This is moot. For sake of sanity, and that I said what I wanted to say and I'll agree to disagree.
This just shows that anyone is willing to cloned for less than their actual worth which is calculated on their own basis.
If you'd rather be ripped off, having a class taught for $5 from some AI bot from your own teaching style earning a single $3 than yourself teaching and earning $5 from each class, be my guest.
Edit: I'm now post capped, so can't comment/reply on HN for another four hours anyway. Old news.
The AI is the thing being taught, not the thing teaching a class. Once you have a model, $0.01 is the correct order of magnitude for the cost of generating an image from a prompt. If anything it's an overestimate.
it's barley short of what a 1920x1080 image costs from openai, but we're in a thread about instructional video, which is neither economical nor available yet
Video would cost more than still images for the obvious reasons, but still likely much less than the cost of however many frames per second times that number of images, because nearly all of the frames will be minor variations on the previous frame. Meanwhile it's going to be a couple years before that technology exists because you'd have to develop something that can sync video with audio etc., by which point the hardware would be more power efficient.
So now we're speculating on the cost of something that doesn't exist yet, but it's highly likely that hardware is going to get more power efficient over time, so the question then isn't whether "AI can do this for a lower price than humans" will happen, it's just a question of how long before it does.
it restricts one from making more videos? how? i understand your comment as reiterating the assumption i was responding to then reframing that assumption with a generalization. of course fair pay is good, and my point is we don't have a solid foundation to assume ai will impact that more any other creative technology we've seen on computers in the last few decades
i reckon nearly zero of these creators are making videos for a single viewer. and as a film student, or even as a professor, you wouldn't spend $300 per movie just to reference in study. demanding even more than that can only be hedging against the idea that one will be out of a job
"Low ball is it takes an hour to produce a single minute of video."
As some of us like to say, "Fuck it, we'll do it live." As in, that's the only take you get and it goes right to production zero editing and processing. Everything gets set once, fired once, and it gets uploaded immediately.
If I learn through the press that Adobe is willing to pay amount A and this information is very peblic, I can at the very least assume they are making 10x more so I'm willing to sell it 20x more and see where this goes
These AI video systems will need huge amounts of random videos showing things like leafs rustling in the wind, car driving, etc. so you don't even need to put in 10 hours of effort to create a 10 minute video, you can just grab your camera and go for a walk and publish that as it is, and it will be valuable training material.
$3 per minute would be a pretty high price to pay for "man walks outside" video.
As you mentioned models need _random_ videos. Just 'walk outside' will produce more or less the same. My guess Adobe is more interested in 'family' sort of videos, with humans. This is what most users will be asking for. Getting them in other ways will be either illegal or really expensive.
I'm not suggesting that anyone here will make $3 per minute shooting random videos, whether those videos are "man walking outside", or "family gathers at the fireplace". The distinction between those types of videos was not pertinent. I was pointing out that making videos for the purpose of training AI doesn't require you to put 10 hours of effort into making a 10 minute video, like for example, Youtube creators might do.
Yes. And there are many forgotten accounts on youtube, their owners now have another way to monetarize. Not sure why Adobe didn't talk to Youtube directly.
If I was in that situation, the last thing I'd want is to provide an AI video generator the means of creating videos like mine and that would compete with me.
It doesn't affect creators life as they guarantee (likely) it will not be made public. So it's just side income. Likely creators make a lot of fragments which don't get into the final version. The problem for Adobe is most of these videos will be similar which, if not filtered, will make AI model biased. In other words they will pay and throw away significant part.
So are we just going to let OpenAI, Google, Facebook, Microsoft and others to just steal the whole internet, including videos on YouTube, and train their models for free without compensating creators? Or will they get away with it because there were no laws that prohibited their behaviour?
Adobe is pretty much banking on the "scrape everything" approach getting too bogged down in legal problems to be commercially viable in the long term. They're going out of their way to only train on licensed material so they (and Getty Images) will be the last ones standing if the worst case scenario happens for the companies who thought they could just take everything, and in the meantime they get the business of risk-adverse customers who don't want to chance it until it's settled one way or the other.
It's a decent lobbying tactic too - companies like OpenAI claim their technology can only exist with unrestricted scraping, and requiring them to license training data would kill the whole industry in the crib, but Adobe and Getty can point to their own products as existence proofs that it can be done but OpenAI just doesn't want to.
Still, it's possible to end up in a situation where people are getting paid and in theory it's an improvement over a previous lawless IP situation, but in practice it still leaves creators in the cold.
That's basically what happened with Spotify. Artists are getting paid something for streaming, but it's fractions of pennies on the dollars they used to make from music sales.
Didn't they quickly change terms before people knew what was going on with the Stock AI stugg, and only do any press for it after it was available but too late to opt out?
And by being the first mover they get to dictate prices as well. Content creators will sell for whatever because the alternative is $0. Once more companies get involved and start bidding the value of these catalogs will undoubtedly go up.
Adobe’s stuff isn’t only trained on licensed material. They are as fucked as everyone else.
There doesn’t exist a LLM trained only on expressly licensed data that works well enough for the image diffusion conditioning task. So unless they made some kind of huge scientific discovery, which they didn’t, they are using a text model like T5 or CLIP to achieve the “text” part of “text to image”, which of course is trained on not expressly licensed data.
I'm guessing they will then lobby to ban open source AI models, especially their usage in commercial applications. One would need a huge bank account to create models in the future. I think this would have a chilling effect on the AI ecosystem. In essence it's a form of regulatory capture.
You already need a huge bank account to create models, the (quality) open source ones are all hand-me-downs from companies burning through huge stacks of investor cash and giving away the spoils. It should go without saying that that won't last.
I mean it's in the crown sourcing range for diffusion models and going to get cheaper. It currently costs 50k to retrain stable diffusion 2 (https://www.databricks.com/blog/stable-diffusion-2). So yeah, mandatory licensing of data on the open web would absolutely kill university projects and small startups. The road to hell is paved with good intentions, as they say
That's 50k if you already know the exact architecture you're going to use, and the exact training data you're going to use, and nail it all on the first try. Multiply that by however many failed attempts it takes while experimenting with new ideas, especially when the corporate players stop being so eager to publish the finer details of their research, and that 50k could very quickly turn into millions.
Yeah, it's not going to be SOTA but existing models are already good enough for a reasonable portion of use cases. Even if they lag behind it's important to have free models.
Data Laundering is an old term... I was using in more than a decade ago in talks about how unethically gathered private data was being sold to Israeli companies, who then licensed it back to US corporations. It was (probably still is) a way to side step privacy laws.
This is why open models trained with disregard for IP are important.
It’s going to happen anyway, the tech can either be free to everyone so we all benefit or behind a paywall. Those are your two options there isn’t a 3rd if we’re talking reality not fantasy.
Most of a creator’s content is consumed for free as is by their audience. Their compensation arrives in the form of payments from a few large subscribers or from an ad distribution platform or other such sponsorship deals.
I’m not sure why you and others think that using someone’s work as inspiration for creating some new work requires payment. Are we required to pay someone when sharing a meme, or using a sound for a reel or a TikTok? A lot of these creators are just creating shit they saw someone else create.
Satire of the work is allowed, but I don't think using the work, itself, in satire of something else, is allowed.
I think that the example a lot of folks use is the "Pissing Calvin" decal.
Watterson never allowed any commercialization of his characters. Every "Pissing Calvin" decal is actually a copyright violation, but I think that they never enforced it, so it may be considered "OK" to use, as the copyright has been allowed to wither.
I don’t think there’s any legal issue with using data for training. I think the problem is how do you subsequently prevent your model from creating copyright violations. If you watch a Spider-Man movie and take some inspiration from it, you know that you go and make a movie with similar themes or whatever, but you can’t just go out and create your own Spider-Man movie. An AI model doesn’t know this, and I don’t know how you’d teach it this concept. Especially when properly educated humans frequently have disputes about what is/isn’t allowed.
Most likely these models are going be hidden behind the network. Aside from the issues of size, which will probably run into the terabytes, I can't see companies willing to risk including them in downloadable code them given how expensive they are to generate, for fear of being copied.
A single network transfer alone is enough to qualify "sharing".
Heck some of the content creators are even arguing that the models are essentially a form of lossy compression, because no one quite knows exactly how they work.
Also the users probably won't be the ones with deep pockets, unless it was a studio. And the lawyers are generally going to after the ones with deep pockets.
Spiderman was released 1962. Copyright runs out in 2057.
What you could do is what some of of the original comic book creators did when they ripped off each others work (Doom Patrol/Xmen, Quicksilver/Flash, etc.) - create an alternative version that was close the original but was different enough that it wasn't a straight up copy. So it wouldn't be Spiderman, but Venom or something, and the suit wouldn't be red/blue but like maybe purple/green or something.
I think the strategy is to succeed at all costs now then pay out a fine or settlement as a cost of doing business later when it's a fraction of their profits. Seems to be how things go.
Google owns youtube. I'd be surprised if their user agreement doesn't allow them to train on the uploads. Same with Apple and Facebook.
I'm not sure a dozen companies (before consultation and licensing) having control of 90-something percent of their world's content is a good idea either
Personally, Photoshop’s AI tools are the weakest among all that I’ve tried. Even free tools like Remove.bg and Canva are running circles around Adobe’s background removal
Lenoardo’s AI tools are even better and the native background removal is just 10x better than Adobe
yea those guys completely missed the bus. I worked on some AI image editing tool a few years back with a partner and we were constantly worried Photoshop was going to suddenly make us obsolete, which is now kind of funny looking at where they're at
They're not buying slop. they're buying high quality content like animations. It can take hours to create one minute so it's really the other direction , division from 3/min down to something like 0.50c/hour of work, not multiplication from 3/min to 180/hr .......
It's also not true so the whole idea is nonsense. Youtube alone in 2022 was taking in 30,000 hours of video per hour. Every day youtube receives more video than anyone could watch in their entire lifetime.
I don’t think this is a great line of argument. It hinges on choice, which rapidly degrades into an argument about free will and the nature of whether anyone truly makes a choice or if our apparent “choices” are just a rationalization of deterministic response to external stimuli.
I don’t think there’s a subjectively true answer to that question, so we just end up bickering over free will instead of AI.
If you just want to declare humans to be special, I would just go straight to that. Humans are already legally special in all kinds of ways.
I think making humans legally special is a good argument though.
The whole point of copyright, is to allow artists to be able to generate a little income with exclusive rights to the works they produce. It's not just there because it's some inherent property of artwork: us humans decided that courts should punish people that threaten the livelihood of artists (in a very narrow scope of ways) via the judicial system.
Of course the "spirit" of the law is meaningless, but I at least don't think understanding copyright law as a purely legal tool, divorced from the human element, is participating in the same conversation artists are having.
Should people that just make art, eat? Early human history the answer was no. Then we advanced to a point where people can specialize in skills, and art made life worth living, so the answer was yes. Now where are we?
Now that generative AI is here, there will be artists who will try to differentiate themselves from AI, and artists that will embrace it to explore new possibilities. It's not a matter of protecting artists, but of choosing what artists to protect, the first or second group? Should we ban photography as well, how about synthesizers, should home owners band new construction in their neighborhood to protect their investments?
In my opinion any attempt at restricting AI training is also restricting people who would use it to create new things. And it will backfire spectacularly, by scanning for derivative works in AI we automatically need to do the same for every work claimed to be created by a human, because we don't know who's using AI secretly.
And it will create a chilling effect, incidental similarity would be too high a risk to take. Even old works will be revealed, their secret influences laid bare for everyone to see. Excessive scrutiny could deter artists from experimenting with new forms of expression, possibly leading to a homogenized art world where innovation is curtailed.
I can get that. There's never a path forward that doesn't leave something on the table. I of course want artistic freedom + technology to advance alongside each other.
But we're also talking about MAKING generative AI models. Not USING generative AI models. Making a model is so far outside of the scope of a single artist it's almost comical. We're talking about a task that, given no bounds, would include the subtask "ingest all human creative output". That's hard to do so therefore it's restricted to the realm of well funded companies.
But even as well off and talented as these companies are, they did very little of the actual work. The total scale of effort that goes into these big lists of coefficients, is beyond human comprehension at this point. We've literally NEVER had to contend with tasks of this magnitude before. Many thousands of centuries of human life hours distilled. Copyright as it's legally inscribed precisely, is ill-prepared to achieve it's end goals in this situation.
But considering:
1) that's kind of cool we can do this
2) it makes interesting things that could better people's lives
3) we still think that humans should be able to make art (ai or not) and not die of starvation ideally
... we need to figure out what to do. Maybe it's not copyright! That was created to solve a specific problem. Now we have new problems. I just think people latch too hard onto the legalese as if it's some magic property of the universe, and NOT a set of systems humans made, for very human goals. I don't want a system where we just say "well look at copyright, it says this is fine, so we will just use everything for free". It's missing the point.
I agree with you that it's asking the hard question of "which artists do we protect". That, I'm not feeling like I have a good answer for right now. I definitely heavily lean towards the talented artists who've honed their craft. That's just more interesting. But I wish more people we're thinking about it, in those terms.
> We've literally NEVER had to contend with tasks of this magnitude before. Many thousands of centuries of human life hours distilled.
And one poem or a page of text is but a drop in this ocean, an infinitesimal contribution. The model is much smaller than the training set. For example Stable Diffusion packs 5B images in 5GB, that's a compression factor of 100,000:1, about 1 byte per training image, not even a pixel. Taking so little from everyone, I am amazed how strong a reaction it creates. The final model doesn't steal almost anything, it doesn't rely on any training example too much, just a compressed abstraction of the training data. You have to overfit it and also prime it to make it regurgitate some training examples, and most attempts fail, resulting in novel outputs.
> In my opinion any attempt at restricting AI training is also restricting people who would use it to create new things.
Won't need to rely on any bunk statistical methods that don't work to detect and suppress ai. Supply and demand and cultural rejection will work just fine. I don't have enough toes and fingers to count the number of creators I respect. I can't name a single AI artist.
Just like a camera. Which is why the copyright for photographs gets assigned to the photographer, it doesn't just disappear into the ether. In this case the copyright for a computer-generated image would tend to belong to whoever typed in the prompt.
Amusingly this policy might end up meaning that the AI model they produce by training on this data will never be able to produce video that will ever be worth more than 3$ per minute. They are probably unintentionally filtering for content created by people willing to sell for that price or below, and that bias will be present in any downstream model. You get what you pay for I guess?
I wonder if this is an exclusive deal for Adobe? If not, it's just an additional $3 per minute. Though I agree that they need to be flexible on the pricing.
How much will AI generated video cost per minute? Not from Adobe, just generally, where it's obtainable.
There's a good scam if you can automate video generation and pass it to Adobe. Them setting a price of AI training video, may have just set the floor price for buying AI generated video.
What if users started poisoning the data set by automating and submitting AI-created videos? lol
What's the price for production? Set up a pipeline right through to Adobe. AI-laundering?
"The software company is offering its network of photographers and artists $120 to submit videos of people engaged in everyday actions such as walking or expressing emotions including joy and anger, according to documents seen by Bloomberg. The goal is to source assets for artificial intelligence training, the company wrote."
Seems a bit steep? But maybe the idea is that as the models improve in quality and sophistication, price will go down. Still, creators might jump at the chance, but Adobe shareholders will probably have plenty of questions.
I was just at a party where a guy told me he was hired by a SF startup to covertly download many hours of YouTube videos without paying either Google or the original creators. When he mentioned the word "copyright", they fired him, but only after assuring him that they had worked at Google before the startup and it was completely legal.
There will be some creators who will take this deal…tends to be the high volume producers who will see this as a pure numbers game. As a creator myself who licenses content for AI, this would be a lot lower than our going rate, so I’d pass.
154 comments
[ 5.5 ms ] story [ 175 ms ] threadIt'd be like if Uber launched by applying for NY taxi medallions. So this seems like a crazy risk:reward ratio here. Adobe is going to end up with a massive bill, a weak model and the hope that the guys who steal everything will face consequences. We've never seen Silicon Valley face consequences like that in the past so i don't see why you would bet it's going to happen this time.
With Uber, no one's suing individual riders or drivers over an illegal taxi ride.
The courts could find that it isn't possible to sign away these rights. There are many other rights that the law does not permit to be signed away for much more tame reasons than "enables further creations of 'your' work (or in the case of actors, literally 'you') not only without your actual involvement, but even long after your death". I would expect that it at least comes with a time limit. "Artists can sign a license to use these rights but only until their death + 10 years, after that the licensee can't use anything trained on their work. Artists also have an unrevokable right to cancel the license at any time with 90 day notice." might be one possible outcome.
Content licensing is really well established and already allows for licensing for specific use and different royalty structures based on usage. An example of this is streaming vs theater vs syndication.
So, now there is a new venue to take into account and have the lawyers add a few pages to their contracts moving forward.
This is exactly my point. An AI that can replace you, personally, is closer to signing away your identity or freedom than the rights to display a specific already completed work.
observationally they’re in an interesting position, balancing their artistic clients and their executive clients. a clean model is the easiest way to hedge their portfolio and reputation.
*you paid for it.
...we would be living in post-scarcity and everything would be free. But that doesn't happen in the absence of AGI, what actually happens is that technology replaces some jobs and then people do the remaining jobs, which are now in higher demand because the things done by technology become cheaper and the money that had gone to pay for labor there now gets spent on something else, increasing demand for the other thing.
Technology has been replacing jobs for hundreds of years and we still have low unemployment.
We have more bullshit jobs than ever before in order to maintain that unemployment rate though
Many of the bullshit jobs are, in fact, destroying the cost/efficiency benefit of automation by imposing bureaucratic waste on things that would otherwise have become less expensive. If someone is doing a job which is useless, we could just give them a portion of someone else's productive work to do, have them each work less and reduce prices by the same amount and then average purchasing power would be the same but there would be less busy work and more recreation/creativity/entrepreneurship.
Bullshit jobs are the inverse of this. Some large organization has insufficient competition so it can waste resources without losing business. The law is excessively complicated so everyone in the jurisdiction has to hire compliance bureaucrats. Government contractors and employees doing unnecessary work want to keep their jobs and legislators don't want to lose their votes/money. Then everybody has to pay higher prices because they don't individually have the choice to avoid it.
I see a lot of the first and second parts, but nowhere near as often the third part: The rights holders aren't seeking financial growth, just wants control in perpetuity. I suspect that's the part proving difficult to solve.
It's not obvious why that would be.
Artists aren't going to like this technology because it competes with them, but it competes with them regardless of whether it was their work or someone else's in the training data.
This leads to a visceral response where they want to call this "stealing" and hope that the creators of the technology can be sued into non-existence so their competition can be eliminated. But as Adobe is demonstrating, that isn't going to happen anyway. So the question isn't whether the technology will exist, it's if it will be locked up behind the walls of major corporations. The latter doesn't do artists any good but harms the public -- including artists who want to leverage the technology in their art. So why should the law protect Adobe's moat from the public?
It's theoretically true that the models could be trained with someone else's training data, but there's a flaw in that argument. If you can train with other data without these legal issues, why don't they? And the answer is actually because they do assign some value to that training data, and there's not an infinite supply and it's actually quite difficult to get large sets of quality data.
I think it's a pretty open question how this will resolve, it could be like stremaing music where companies like spotify are little more than puppets for the major record labels. It could end up like youtube where the model started with "We're going to steal stuff" and ended up "We're going to strongly enforce copyright now we're the encumbent" or some other third way. But I don't expect the "We're going to take everything with no regard to the existing legal framework" will sustain long term.
But the law would probably be DMCA 1201 for circumventing the DRM rather than normal copyright for making the copy, which is much more ambiguous. Also, record companies like to sue people, that doesn't mean they're right and provides no indication of what the law should be. You could just as easily pick some other example, like whether you can rip a music CD or vinyl you bought to put it on your iPod, which the record companies might not like to be allowed, but that doesn't mean that it isn't.
> It's theoretically true that the models could be trained with someone else's training data, but there's a flaw in that argument. If you can train with other data without these legal issues, why don't they?
To which the answer is that they do. A lot of models are trained on arbitrary content from the internet.
As to why Adobe in particular is doing this, think about it. It's because that interpretation benefits them rather than the artists, by creating a moat where companies who already have licenses to bulk stock images etc. are the only ones who can create a model, rather than having lots of competitors because anyone can create one and many of them are free and can be run locally.
> I don't expect the "We're going to take everything with no regard to the existing legal framework" will sustain long term.
Publishing companies don't like public libraries. So anybody can go there and borrow a copy of any book for free? That doesn't mean that libraries are bad or are or should be illegal.
Also, laws are created through the political process, which is not always great. The outcome "individual artists somehow benefit from this" isn't even in the room there. The two most plausible outcomes are that "tech companies" win and anybody can train a model on anything they can get their hands on, and that "content conglomerates" win and then this technology gets locked up as a service from only megacorps and the artists still don't get anything meaningful, but now the world has another abusive cartel imposing arbitrary censorship and using control over this to cement control over adjacent markets etc. Of these, the second is clearly worse.
Suckers who don’t properly value their time is why BigCo generally wins.
That then takes away your next video. So you got paid a low ball amount only to be ripped off.
Because it restricts you from creating such style videos. Your being paid a pittance for a bot to learn your style.
Why wouldn't you want more money?
Because that style isn't that unique, so the difference is between getting paid nothing and now there is a bot that can do the same thing, or getting paid something and now there is a bot that can do the same thing.
It hasn't even been established that they're required to pay you at all.
Style is how teachers make learning happen. If teachers follow the general generic book mundaneness, you learn far less than of you apply your own style.
Besides, if they were using the content without my permission I should be allowed to seek costs for such.
"You're unique, just like everybody else."
It doesn't matter if the thing is using your exact style or one which is enough of a substitute for it that the difference isn't going to make up the difference between your $50 fee and the $0.01 in electricity it takes to have the AI do it.
> Besides, if they were using the content without my permission I should be allowed to seek costs for such.
On what basis? How is it different than someone teaching their students your style, so the students can make their own original works in the same style? It's directly analogous to classroom use, which is an explicit example of fair use from the copyright statute.
This just shows that anyone is willing to cloned for less than their actual worth which is calculated on their own basis.
If you'd rather be ripped off, having a class taught for $5 from some AI bot from your own teaching style earning a single $3 than yourself teaching and earning $5 from each class, be my guest.
Edit: I'm now post capped, so can't comment/reply on HN for another four hours anyway. Old news.
So now we're speculating on the cost of something that doesn't exist yet, but it's highly likely that hardware is going to get more power efficient over time, so the question then isn't whether "AI can do this for a lower price than humans" will happen, it's just a question of how long before it does.
Even if that were not true, releasing the rights for so little is dumb.
As some of us like to say, "Fuck it, we'll do it live." As in, that's the only take you get and it goes right to production zero editing and processing. Everything gets set once, fired once, and it gets uploaded immediately.
You shoot it "live" and let the AI deal with the editing and processing for you.
It takes an hour to produce 1 minute of scripted content and fancy graphics for a professional YouTube show.
$3 per minute would be a pretty high price to pay for "man walks outside" video.
They do not want theatrical productions, they want raw footage of the real world.
It's a decent lobbying tactic too - companies like OpenAI claim their technology can only exist with unrestricted scraping, and requiring them to license training data would kill the whole industry in the crib, but Adobe and Getty can point to their own products as existence proofs that it can be done but OpenAI just doesn't want to.
That's basically what happened with Spotify. Artists are getting paid something for streaming, but it's fractions of pennies on the dollars they used to make from music sales.
There doesn’t exist a LLM trained only on expressly licensed data that works well enough for the image diffusion conditioning task. So unless they made some kind of huge scientific discovery, which they didn’t, they are using a text model like T5 or CLIP to achieve the “text” part of “text to image”, which of course is trained on not expressly licensed data.
You create the synthetic data, move it through 10s of entities and then buy it back cheaply from an entity.
Break the law so you can undercut the real producers, and then buy the rights when they're broke and everything works out for you!
It’s going to happen anyway, the tech can either be free to everyone so we all benefit or behind a paywall. Those are your two options there isn’t a 3rd if we’re talking reality not fantasy.
I’m not sure why you and others think that using someone’s work as inspiration for creating some new work requires payment. Are we required to pay someone when sharing a meme, or using a sound for a reel or a TikTok? A lot of these creators are just creating shit they saw someone else create.
I suspect that, if the image in the meme is copyrighted, the answer is "yes."
I haven't come across too many memes with Mickey Mouse, but I do see the occasional Marvel one (I have forwarded these, myself).
It can get nasty.
Satire of the work is allowed, but I don't think using the work, itself, in satire of something else, is allowed.
I think that the example a lot of folks use is the "Pissing Calvin" decal.
Watterson never allowed any commercialization of his characters. Every "Pissing Calvin" decal is actually a copyright violation, but I think that they never enforced it, so it may be considered "OK" to use, as the copyright has been allowed to wither.
But IANAL. Things get sticky, here.
This is more akin to buying multiple books on a topic and then writing a new book on the topic using the aggregate learnings from the original books.
But even if you think that’s fine it makes using AI models for your own commercial applications more risky.
A single network transfer alone is enough to qualify "sharing".
Heck some of the content creators are even arguing that the models are essentially a form of lossy compression, because no one quite knows exactly how they work.
Also the users probably won't be the ones with deep pockets, unless it was a studio. And the lawyers are generally going to after the ones with deep pockets.
What you could do is what some of of the original comic book creators did when they ripped off each others work (Doom Patrol/Xmen, Quicksilver/Flash, etc.) - create an alternative version that was close the original but was different enough that it wasn't a straight up copy. So it wouldn't be Spiderman, but Venom or something, and the suit wouldn't be red/blue but like maybe purple/green or something.
Isn't $3/minute exactly the compensation for creators you're looking for?
I'm not sure a dozen companies (before consultation and licensing) having control of 90-something percent of their world's content is a good idea either
Disney, WB,etc etc
That'll show them
Lenoardo’s AI tools are even better and the native background removal is just 10x better than Adobe
Do they care what the content is?!
[1] https://en.wikipedia.org/wiki/Paint_Drying?useskin=vector
So I guess Adobe needs to raise their bid before filmmakers can afford the arbitrage!
I don’t think there’s a subjectively true answer to that question, so we just end up bickering over free will instead of AI.
If you just want to declare humans to be special, I would just go straight to that. Humans are already legally special in all kinds of ways.
The whole point of copyright, is to allow artists to be able to generate a little income with exclusive rights to the works they produce. It's not just there because it's some inherent property of artwork: us humans decided that courts should punish people that threaten the livelihood of artists (in a very narrow scope of ways) via the judicial system.
Of course the "spirit" of the law is meaningless, but I at least don't think understanding copyright law as a purely legal tool, divorced from the human element, is participating in the same conversation artists are having.
Should people that just make art, eat? Early human history the answer was no. Then we advanced to a point where people can specialize in skills, and art made life worth living, so the answer was yes. Now where are we?
In my opinion any attempt at restricting AI training is also restricting people who would use it to create new things. And it will backfire spectacularly, by scanning for derivative works in AI we automatically need to do the same for every work claimed to be created by a human, because we don't know who's using AI secretly.
And it will create a chilling effect, incidental similarity would be too high a risk to take. Even old works will be revealed, their secret influences laid bare for everyone to see. Excessive scrutiny could deter artists from experimenting with new forms of expression, possibly leading to a homogenized art world where innovation is curtailed.
But we're also talking about MAKING generative AI models. Not USING generative AI models. Making a model is so far outside of the scope of a single artist it's almost comical. We're talking about a task that, given no bounds, would include the subtask "ingest all human creative output". That's hard to do so therefore it's restricted to the realm of well funded companies.
But even as well off and talented as these companies are, they did very little of the actual work. The total scale of effort that goes into these big lists of coefficients, is beyond human comprehension at this point. We've literally NEVER had to contend with tasks of this magnitude before. Many thousands of centuries of human life hours distilled. Copyright as it's legally inscribed precisely, is ill-prepared to achieve it's end goals in this situation.
But considering:
1) that's kind of cool we can do this
2) it makes interesting things that could better people's lives
3) we still think that humans should be able to make art (ai or not) and not die of starvation ideally
... we need to figure out what to do. Maybe it's not copyright! That was created to solve a specific problem. Now we have new problems. I just think people latch too hard onto the legalese as if it's some magic property of the universe, and NOT a set of systems humans made, for very human goals. I don't want a system where we just say "well look at copyright, it says this is fine, so we will just use everything for free". It's missing the point.
I agree with you that it's asking the hard question of "which artists do we protect". That, I'm not feeling like I have a good answer for right now. I definitely heavily lean towards the talented artists who've honed their craft. That's just more interesting. But I wish more people we're thinking about it, in those terms.
And one poem or a page of text is but a drop in this ocean, an infinitesimal contribution. The model is much smaller than the training set. For example Stable Diffusion packs 5B images in 5GB, that's a compression factor of 100,000:1, about 1 byte per training image, not even a pixel. Taking so little from everyone, I am amazed how strong a reaction it creates. The final model doesn't steal almost anything, it doesn't rely on any training example too much, just a compressed abstraction of the training data. You have to overfit it and also prime it to make it regurgitate some training examples, and most attempts fail, resulting in novel outputs.
Won't need to rely on any bunk statistical methods that don't work to detect and suppress ai. Supply and demand and cultural rejection will work just fine. I don't have enough toes and fingers to count the number of creators I respect. I can't name a single AI artist.
heheheh, AI art certainly has more than enough duplicated fingers!
If you did that to Disney then you'd be in jail, but since it's the little guys getting fucked I guess it doesn't matter.
Also, as other people have pointed out, this just restricts training to well funded companies. Everybody forgets about the little guys on Huggingface.
There's a good scam if you can automate video generation and pass it to Adobe. Them setting a price of AI training video, may have just set the floor price for buying AI generated video.
As people with more money than me say: arbitrage!
"The software company is offering its network of photographers and artists $120 to submit videos of people engaged in everyday actions such as walking or expressing emotions including joy and anger, according to documents seen by Bloomberg. The goal is to source assets for artificial intelligence training, the company wrote."
0: https://www.msn.com/en-us/money/other/adobe-is-buying-videos...