- it can be automated. CGI still required a lot of human work.
- AI generated content requires way less skill. You had to train a lot to become good at CGI. We already see people getting Stable-Diffusion-proficient in a few hours.
- in the 90, CGI was introduced gently, through selected medias. AI content has access to the full internet right out of the box.
- producing CGI required a lot of hardware, but with SaaS, AI will be available to anybody.
- CGI was first exposed to professionals, who produced for the public. AI is already been used by hobbyists, and soon the public, directly.
- CGI was mostly about visual techs, but AI generated content can be applied to text, sound, molecules, basically any structured data.
In summary, the very nature and context in which AI generated content is taking form means the impact will be larger, stronger, and way more intense.
People are hardly capable at dealing with CGI intelligently, even after 20 years. They trust photos and videos a lot, and react strongly to them. Even when they know they are fake, like in a movie, some emotions will forge they judgement like if it were real.
AI generated content? Forget about it. People are not ready. The emotional impact it's going to have will make you feel like the cancel culture was a whisper in the wind.
I was underwhelmed by the last decade of AI dev, but 2022 is for AI what 2000 was for the dotcom revolution.
Question is, what the equivalent of domain name for AI, something cheap you can buy to sale later?
I think it's still comparable if you view it as "Generative AI is to CGI what CGI was to practical effects".
- CGI can be automated, practical effects required a lot of human work
- CGI content requires less skill. You had to train a lot to become good at sculpting / painting / makeup / carpentry / stunts / whatever to become good at practical effects. We already see people getting $MODELLING_PROGRAM proficient in a few weeks.
- producing practical effects required a lot of hardware, material and tools but with CGI you just need a powerful computer. CGI will soon be available to anybody as computers become cheaper
Your last two points are probably only true for AI generated content but there are still some barriers to people setting up their own AI models (technical know how, hardware) and there are certainly limitations on what type of content can be produced (you need a big and reliable dataset of examples).
Overall I think this will have a similar impact but probably more widely distributed in both terms of access and impact.
Having worked in CG for several decades, this is a pretty misleading summary. CG takes a lot of manual work, and there are other reasons and benefits to using CG than how much manual work it takes. It’s often less work than practical but not always. There are certain kinds of things that can be automated (with a lot of manual setup work!) but most of what you see in movies also involves lots and lots of manual work for every single shot. It’s not unlike how machines get used to automate certain parts of practical effects work, using for example, compressors, electrical rigging, vehicles, pneumatics & hydraulics, gas & explosives, etc..
> We already see people getting $MODELLING_PROGRAM proficient in a few weeks.
This is similar to saying we already see people getting C++ proficient in a few weeks. Some people can write very useful & interesting programs within a few weeks, but nobody understands all of it by then, and nobody’s getting a high paid career on just a few weeks of noodling. It takes a long time to understand all the ins and outs, and independently it takes a long time to understand complementary computer science fundamentals. Same is true for Maya, Blender, Houdini, etc.. Someone might have something cool to show after a few weeks, but nobody is ready to make a CG movie by then, and nobody is landing the good jobs in CG. (By “nobody” I’m talking about statistical relevance, not that it has never happened.)
Another reason a few weeks is insufficient for landing good jobs is that the CG industry is highly competitive; you have to be better than the guy who’s gone to art school and played in Blender for a few years if you want that job.
True, I just think that because of all the leverages in place for AI content generation, the effects will be orders or magnitude bigger. But it's just speculation, let's get the pop corn and watch the show.
> AI generated content? Forget about it. People are not ready. The emotional impact it's going to have will make you feel like the cancel culture was a whisper in the wind.
Forget about it? People are not ready? When has that ever been a blocker? This is like a new era of 80s computing where some people will get a home computer and their dork kids will dork on it and produce some unbelievable stuff in the decades to come. I'm scared, slightly optimistic, but what is available right now as commodity, there ain't no stopping it.
Sub-textual thought experiment: how big can Hugging Face get?
Can they achieve a Monopoly on Art Production into the Future?
To build their moat: they build the largest supercomputer for image training, acquire the largest image training set (~10B), and then institute an IDF Unit 8200 style Garin Lotam recruitment effort of the best artistic talent on earth to feed new forms into the machine, and everyone else simply forgets they ever knew how to make art in the first place as the drawing prompt becomes as ubiquitous as the google search box ;)
I am afraid that rise of generative AI will result in a lot of junk being released and advertised as quality.
It shouldn't be like so, I think generative AI can yield great results if there is enough human supervision, I think it can be great if used to augmenting creative work. But I think many will chose the shortest path towards money and they will put in the least amount of work possible.
Imagine tens of thousands of movies spitted by AI based on sub mediocre screenplays, imagine millions of songs generated by AI based on sub mediocre lyrics.
Somewhere, in that ocean of junk, there is something valuable. Good luck finding it.
If we use CGI analogy I can say that all pre CGI Tom and Jerry cartoons are works of art but many of the cartoons released after aren't so great, while some are. And CGI still requires some effort, so not everyone is be able to release movies. Generative AI will require far less effort and knowledge.
No, generative AI will raise the quality bar of user-produced content. Pros will have even more to work with. I suspect that we may be somewhat overwhelmed by all the amazing images and videos that will be available. I get that sense already while browsing through user generated images.
This is already happening - SD unleashed an absolute torrent of content that is technically amazing but artistically boring.
Hopefully it all eventually settles down and becomes just another tool for skilled artists, rather than thousands of nerds churning out the visual equivalent of elevator muzak.
To be fair, visual muzak from nerds has at least one redeeming quality: it's fun popcorn-munching while watching artists go on tirades about how it violates their consent.
The whole point of programming is automation. Having a tool that would make programming much more efficient is not a failure condition for programmers.
It's called no-code(this time around at least). Quickly turn out a decent-looking app, then hitting a brick wall when the least amount of unexpected requirement turns up. Scares you the first time you see it, but both devs and artists will be fine long term.
The consequences of anyone being able to create any software they want cheaply are so enormous that I think coder employment would be among the least of our concerns.
It already has, and as with art AI, I've seen worse from beginners and it's got certain failure modes and there are copyright/moral rights[0] issues and it is still obviously useful.
[0] I saw this in a few of my UK job contracts in the context of signing away that right, and it seems to be "the right to be identified as the author of a work", but IANAL.
The great irony is that the most popular and powerful curators and docents will be AIs. Private models used to guide users for commercial purposes. In all honesty, if you haven't already begun working on such curation models, it's probably too late, because I'm positive the big guys have been working on them for a while. While Facebook, Amazon, and Netflix may come out with some decent docent models, I'd be willing to bet that Google and Apple will have the best ones by a big margin. And unlike stable, they won't share.
If this is the case then we'll just find a way to bundle the curator AI with the art-generator AI to improve the quality of the art generated in the first place.
They’d first need to create a meta curator that will help you find a curator that is at least somewhat aligned with your taste.
The only things that are working for me are Apple music and YouTube recommendations, both algorithmic (but also both with very low downside of recommending bad content due to short length in general)
sure people will flood YouTube with their creations, that no-one will ever watch, but isn't that already happening kinda? Early with Stable Diffusion every creation was just "wow", now it's hard to even get someone to look at your image.
The quality bar will be raised, but if you want an audience, you will still need to produce something that stands above the rest. And the way to do that is the same as today, put in more effort than the others. Make something that's better, that's worth peoples time.
Humans won't be able to tell the difference. If another AI is able to tell the difference then we'll just use this as in adversarial model to make sure the bar is raised even more.
I don't think its comparable. Generative AI is a far far greater leap than CGI is.
CGI is a mere evolutionary/sustaining innovation, movies/shows pre and post-CGI are largely made by the same studios. CGI just lowered costs compared to physical sets, and slightly expanded what was possible to film.
Generative AI, on the other hand, lowers costs so radically, that it will spur entire business models. Within a few years, we can expect feature length animations/TV shows to be created by teams of less than 5 people. Live action shows, requiring photorealism, will be a bit harder, so maybe 5-10 years.
This will in turn change the business model radically, with indie creations competing against big budget productions, it'll turn Netflix in the direction of youtube, etc etc. It will be very similar to how livestreaming and youtube affected TV-talk shows, as large production staff are no longer needed.
This is no exaggeration, I've seen animation created with diffusion models, that would have taken a traditional animator a year to animate by hand, now doable by a single person, within a week. It's been less than a 1 month since the model's release (novelai), imagine what would happen in a year.
In theory sure, but it seems to me that if the cost and effort to create become lower so will the content. There will always be better and worse, and just like CGI, the content would be good or bad depending on what humans are behind it and the amount of effort they put on.
Im thinking generative AI will indeed make animations ubiquitos to the point we’d ingest so much content to the point of becoming ‘alergic’ to it or seek something other kind of entertainment.
There a huge demand for animation, that is currently being unsatisfied.
Youtube used to have a large genre of short-animations. They all died out, because their production costs are way higher than what weekly-ad-revenues can sustain, compared to the incredibly cheap costs of talking heads, which can update daily.
Not everyone will like animation, but there'll probably at least 50x animation on youtube in 2 years than today.
There's also the quality issue. Animation scales very linearly with budget. With low budget, the best you can do is flash animation, which limits the formats of story you can tell very heavily (99+% comedy)
With AI, individuals can match the quality of professional animation teams, which drastically expands the range of ideas they can explore. So that eliminates a lot of the boredom.
Note this wouldn't displace existing animation studios, it'll just allow them to push their production values way way higher. For hand-drawn animation, even top-budget films, still have moments where they cheap out. In the future, the standard is going to get pushed insanely high, where every frame will be screenshot worthy, because they are literally AI illustrations.
Aka, instead of one Mona-Lisa, expect 1000 Mona Lisa paintings, formed into an animation of Mona-Lisa going through her day in renaissance Italy.
The actual threatened jobs, aren't in animation, but in live action. Because its far harder to make perfect photo-realistic images than stylized images. This gives animation an advantage that live action doesn't have.
For example, the recent Rings of Power cost 1 billion, but an AI animation would both look better, and cost far far less. Sci-Fi and fantasy shows are especially vulnerable to competition from animation.
> I've seen animation created with diffusion models, that would have taken a traditional animator a year to animate by hand, now doable by a single person, within a week.
Do you have a link? That's pretty hard to believe.
It uses diffusion to post-process a 3d video frame-by-frame.
You can see the massive improvements in facial expressions, from bug-eye alien face to something quite pleasant.
3d animation is great at allow for complex movements and dynamic camera angles, but horrible at facial expressions. That's why Pixar had to go with the exaggerated proportions 'cartoons' look, to ensure the faces are large enough to convey emotions, even in 3d.
2d animation is superb at facial expressions, but can't do complex movements. Drawing a static face is easy, drawing a rotating character is very hard, drawing a rotating character, while the camera itself is rotating, is brain meltingly hard, so almost never done.
Diffusion based workflows will have none of the downsides, and all of the upsides, compared to either approach.
Not exactly, tracing over the 3d is called 'rotoscoping', and has been used in animation for a few years now, basically hand-drawing animation over a 3d video.
There are two problems with rotoscoping.
1. The cost, its not much cheaper than just drawing it by hand.
2. The effects don't compare to diffusion. If you pay close attention, you'll find that diffusion not only added facial expressions, it even added SHADING and crevasses to the dresses. That stuff is extremely difficult even for a human to do repeatedly.
Indeed, diffusion is superior to 2d animation, even in raw quality. Because diffusion will soon be able to produce illustration level pictures for each individual frame.
Whereas 2d animation had to rely on detail-compressed simple artstyles AND keyframe-inserts to cut down on costs. Soon every frame will be a keyframe, and it can execute artstyles previously impossible to animate.
In theory it could empower amateur storytellers to punch above their weight by creating semi-professional animated films. Though to be fair, there was a similar realization when it was clear the Quake engine could be used for filming custom narratives rather than just recording gameplay. Not too much came out of that era. Professional artists still better understand what media humans want to consume.
Amateur storytellers will be replaced too… as this thing goes exponential, it will create a glut of content where previously we had scarcity.
In about 10 years from now, nearly all the activities - playing music, composing music, animation, live action footage etc. and yes storytelling — will all be mostly automated. And curation of them, too. It will be like you knitting your own sweaters… most people wouldn’t see the point. They could just buy an endless variety of shirts on Amazon.
In Kim Stanley Robinson's Mars trilogy, written during the rise of photoshop (such things were already being predicted back in the 90s), people eventually reject tv/movie style media and go back to plays since the generated media is just too easy to pump out an chintzy.
This one resonates with me a bit.
We already see this in goods to an extent "artisinal" vs "cheap dropshipped crap".
People find value in seeing other people do something well (see the popularity of maker videos, or the number of people commenting on someone's performance in a movie rather than the plot, etc). It also captures the idea that art isn't just a product, but a method of communication and is inherently human.
I think that will be true for "mass produced content" things like music for specific activities or uses e.g. workout music, background music for TV/movies/games/restaurants. I think there will still be a lot of people involved in higher value productions. They will definitely be using AI generated content as a multiplier but will likely add their own composition and tweaks to it in order to make it fit their vision. The first step in any design process will probably be to generate a bunch of related content kind of like 3D modelling artists dig up a bunch of reference material before starting their work except the reference material will probably be able to be directly incorporated into the finished piece. There will also be a bunch of purists who do everything themselves, this is obvious since there are still a lot of people who knit their own sweaters.
Not only that but it may turn out story tellers would still want to use words and the listeners to prefer using their imagination sometimes too. Not that I have something against CGI or animations but when somehing becomes too much humans get bored.
Light novels (From Japan) already balance these two demands.
Illustrations depict the core characters, and images are far better than words in depicting appearances.
Words do the rest.
AI will just mean most fiction start following this same model. Where characters and key environments get detailed designs and images, but the rest remain in words.
Are you sure it isn't the complete opposite? I'm pretty sure that the amateur kids who played around with Machima in the age of Quake went on to become the professionals in the video game industry.
Yes, but it didn't break down the divide between the professionals and the amateurs. CGI is still not common in the amateur world. People aren't producing successful movies in game engines in their homes. That is still up to Hollywood, even so Hollywood is using the same game engines that people have access to at home.
Compare that to what Youtube did to TV, where Youtube did in fact replace a lot of peoples TV watching with Youtube content. A similar shift hasn't happened with movies. Good low budget indie movies are still extremely rare, and generally aren't done in CGI, but filmed classically with cameras.
Cameras, editing and compositing getting cheaper and completely changed the game for Youtube-style video productions. Machinima just didn't have the same impact. Creating CGI/game assets is still a costly and timely endeavor out of reach for most amateurs at a movie-scale. You see a few 5min shorts every now and then, but no two hour movies.
I think the thing is once you get good enough at CGI you make enough to become a small studio
Think about RocketJump who went on to make Video Game High School after making cool looking Youtube videos for a while.
Where are you going to get the training data, once copyright law is amended to protect artists from this new kind of intellectual property theft?
Without enormous training data from human victims, your "Artificial Intelligence" doesn't exist.
In the end, your "AI" is just huge interpolator, using matrix multiplications and max() or gelu() to interpolate the work of real intelligence: the human artists.
Once it's illegal, only criminals will use it. No first-world company will break the law to save money on artists. You think anyone who matters will secretly use a neural net trained in violation of copyright? Think again.
I say this as a person who writes GPU BLAS kernels professionally at a company you've heard of. I directly profit from this theft. But I also know it's wrong.
It will be a hellscape if copyright law is extended to include artistic style. Copyright is already overreaching and broad. To extend it further would be disastrous for the public and for the small to mid tier artist.
I could imagine that if copyright law were to be extended in this way, artists would either be small enough to fall under the radar or select for popular enough artists that they could sub-license the 'copyrighted style' in question. Regardless, only large corporations and a select few 'superstar' artists would benefit at the expense of the vast majority of artists and the public.
If this were to pass (and I very much doubt there's a good chance of it happening), all is not lost because there is a wealth of public domain imagery that we can still train on. This could then snowball into providing richer datasets to train on.
I don't like the framing of a lot of these arguments. I'm paying taxes to provide a state apparatus that provides copyright legal protection for artists. I'm happy to do this, up to a point, as it's essentially a contract that provides incentives to innovation by rewarding artists, with the understanding that, eventually, the artists works get to be used by the public (that is, given to the public domain/commons).
Copyright law, at least in the US, specifically forbids ideas and "styles" to focus on specific realizations. I understand we're in a shifting landscape where technology is changing this narrative and that copyright law was created with different technology in mind but copyright law exists, presumably, as a compromise to incentivize artistic creation while serving the public good. Copyright law is not meant to be corporate welfare nor is it meant to protect the artist at the expense of the public, in perpetuity.
It would be nice to have a more reasoned argument instead of this knee-jerk reaction of theft.
We don't have to even get to the point of extending copyright to artistic style. If using copyrighted art to train a model is found to not be fair use, the technology will no longer be above board. Companies in countries with these rules won't be able to ignore them.
What could happen if this interpretation of fair use comes to pass? Model trainers may have to license images from companies like Getty. In a sense, it should be cheaper than using an image individually. In some countries, music organizations deal with licensing large sets of songs, no need to negotiate with each label or artist individually. That sort of arrangement could come to pass. Perhaps there'll be a simple option on image upload sites to select the license for the image - allow or disallow model training. Or perhaps everyone will simply use below-the-board models surreptitiously.
Should we request that all human artists verify that the weighting between the neuron connections in their brain wasn’t influenced by any copyrighted media?
Or similarly, if a company trains a model and then bins the training data. How would you verify any artists work was in there?
Humans cannot learn and spit out arts or animations at the same scale, so there is no problem.
Your first question is based on a technicality (human can learn too, should we then …?), but the idea of right is not based on physical properties or similarities, but on realistic desired outcomes of its application. (Of course we see flaws of laws all the time, but that is tangential.)
company trains a model and then bins the training data. How would you verify any artists work was in there?
Like any shady company it will drag a set of risks with it forever. It may fly under the radar for some time, but will be snitched on after few internal disagreements. This well-known dynamic disincentivizes big companies from choosing going this route, and small companies are not a big deal.
How do you know. If I stick up an image on the internet. Its legal if I made it myself in photoshop, but illegal if it was mostly generated? But what if I just say I made it myself?
This is different to most copyright cases where all that matters is the end result, not the process.
- Looking at how others have done art, imitating them to learn, and then combining what we learned with everything else we've seen in our life is how humans produce art. And it's the same way that stablediffusion produces art: train on data, reproduce based on what it's learned, combine everything its seen with a prompt to create something new.
I personally believe that both training and generating images through AI is fair use. It's transformative. You're creating completely new ideas. Consider that plus the fact that styles aren't copyrightable.
Also keep in mind that, at least in the US, you can't legislate your way out of fair use since it's a constitutional right. (It's the result of the tension that exists between the copyright clause and the first amendment)
The output is transformative. But in order to use it as an input to a training set, you're effectively making a copy that may or may not be authorized, not to mention at some level the intermediate representation might well present a copyright issue. "Style" isn't the issue.
I think on an ethical level we're just going to need the expectation of separate agreements / legal conceptions for having ones work go into a training set. It's novel territory, providing not only work-for-hire but meta-work-for-hire. Might even be that there's no true compensation without equity.
I can only potentially see an issue with that if you're selling the model which is essentially a collection of copyrighted works (organized in a new way).
However, the output of that model is the epitome of derivative work.
The trained model has to be viewed as an artistic tool used by the artist to create art, imo.
Derivative work isn't permitted under current copyright though, you know that?
Like I can't release a sequel to a Star Wars movie without getting a license to do so by Disney, because they own the original work's copyright, I'm not allowed to produce a derivative of it.
Similarly this is why there needs to be an explicit open source license that allows you to produce derivative work from it.
That's the crux of it, the trained model is a derivative work, literally it's mathematically derived almost deterministically (ignoring SDG being stochastic).
The question is, even though it's derivative, is it still fair use to use it without permission?
These models wouldn't produce a sequel to Star Wars unless you only fed it Star Wars. The output works would be original works, derived from many sources. It'd be more like it outputted Star Wars because you trained it on Hidden Fortress and Buck Rogers.
You can't say it's derived from many sources and then say it's original work haha.
The output work would appear original in most cases (if you filtered out for left over watermarks and too much memorization, and avoided it copied any style or visibly recognizable characters, etc.). That said, even though it appeared original, it would be known to be derived unless you'd hide the fact you used a ML model for it.
I'm not sure you can argue it's not derived when it's literally being mathematically derived lol.
I think "fair use" is really the crux of it. Like do we feel this is a valuable derivativion to allow, because it benefits society in other ways that are worth it.
C'mon; we all know the 50 Shades of Grey was Twilight fan fiction, but it is also original work... Likewise all of fantasy, which is footnotes on Tolkien. You're suggesting a much higher bar than we apply for humans, which leads to bad places... Do we raise copyright to the point that no one can produce new work, or accept a world where generative art lowers the bar for making art?
> While Meyer has not granted permission for James to create a derivative work from Twlight, she has not yet initiated any public litigation directed at the 50 Shades of Grey series. Still, because Meyer's copyright lasts for her lifetime plus 70 years, James and her heirs may face litigation down the road.
This is in fact an unknown, it could appear that Meyer would have at least a case against it if she wanted too.
We know for sure people are capable of original thoughts. If there were no prior movies to learn from, a human could figure out how to make one, just as the pioneers in film making did.
On the other hand, models cannot, they need large quantities of existing source material, so it's unclear if they're capable of original work, and likely it's derivative.
> Do we raise copyright to the point that no one can produce new work
You don't actually have to raise copyright at all, it only takes for a court to say that ML training isn't fair use, that's all. It would only exclude ML training from fair use, meaning the people training the model would need to get a license approval from the copyright owners.
> or accept a world where generative art lowers the bar for making art?
Yes, that's the question of Fair Use. We might consider the infringement to be a net positive and want to include it as Fair Use.
Normally that's what Fair Use looks at, it's the balance between what's right by the authors, and what's best by society.
I think it's totally fair to say it be best here to allow this use because the benefit to society of that technology are worth it.
I also think it be fair to say that the benefits are not worth it compared to the loss inflicted on artists, or to think that most companies investing in AI would actually have the mean to compensate the artists anyways so it wouldn't even stifle ML innovation while also being fair to the artists.
Personally I'm still debating it with myself and not sure where I stand.
But I think it's very misleading and disingenuous to pretend that the ML model isn't even using copyrighted material or deriving work from it, and to pretend like it's similar to an artist simply looking at other people's work or being inspired by other people's work. It's very different in many ways, the process, the mechanisms, the nature, the function, the ethics, the financial applications, the capabilities, there's just so many differences here.
>Derivative work isn't permitted under current copyright though, you know that?
Yes it is? There's nothing that blanket bans derivative works in the west. Parody is often derivative, and explicitly protected by the fair use doctrine in the US, for example.
> Once it's illegal, only the criminals will use it. No first-world company will break the law to save money on artists. You think anyone who matters will secretly use a neural net trained in violation of copyright? Think again.
Step 1. Hire some outsourced artist in the rest of the world, where people don't care about AI art regulation.
Step 2. Watch as those outsourced artists outproduce the domestic artists, since they have the twin advantage of lower costs AND better quality due to being able to use insanely powerful AIs.
Step 3. Watch as the domestic art industry gets wiped out, just like manufacturing did.
Politicians can understand this 1.2.3, therefore they'll never ban art AI.
> Where are you going to get the training data, once copyright law is amended to protect artists from this intellectual property theft?
The same places as before?
What's stopping anyone from training these models?
The cat is out of the bag and artists need to learn they will have to cede on these copyright issues.
Similar to musicians having to cede on digital reproduction of music.
This argument every time I hear it is so ridiculous.
It's like saying that the CNC machine should also be given mandatory time off each year just as all the other factory workers.
One is a machine, the other a human, what rights and laws apply to one don't mean they make sense and should be applied to the other.
The better framing is, should art students also be allowed to make derived work from copyrighted images without permission?
The output image from a model is basically a derivation from it's entire dataset, I mean this is mathematically demonstrable. If it wasn't for a few stochastic parts of the model and explicitly added randomization, it be deterministic even.
It takes copyrighted work + prompt and outputs derived image.
I don't think the question of derivation is being debated that much though, the question is if it's fair use. As in, deriving from copyrighted work a ML model, maybe that's fair use. And that's what has to be figured out.
The comparison between a ML model and a human is often used but it is not very useful.
The problem is in most of these cases is that a powerful entity is profiting off of other people's work without their consent and gives nothing to the exploited members in return. Sure, an individual human learns from copyrighted works and reincorporates to make something slightly new all the time. And then they may also profit from it and not give anything in return to those that came before.
The problem here is the scale and the power that enables that scale. This is industrial level mining of non-consenting humans, exploiting their life's work in many cases.
The very nature of AI is that anyone can make their own model with enough training. There are already open versions of Copilot available, such as Fauxpilot, or GPT Neo for GPT-3, so I never understand the "only rich and powerful will have AI." I mean, sure, some will, but just like programming, the effort will eventually be distributed to open source versions.
As an OSS contributor you can nicely ask people building some OSS equivalent of Copilot to not use your work as training data, and they would comply because this is culture. Good luck doing the same with a megacorp
Don't large companies already have access to "legal" training data? They can use the images you uploaded. Most ToS have language to allow them to do that.
We’ve argued this for centuries. Fair use is a fine line, but it exists. And it doesn’t matter whether the generative agent is human or machine. If a sufficiently unique work of art is the result, it’s not stealing.
The only outcome of this would be that all commercial image/video AI creation tools are Chinese, but illegal to directly use in the West, so everyone wanting to do this legally has to pay contractors in China instead that semi-unofficially use these tools, but pretend not to.
There's no putting the genie back into the bottle.
> Without enormous training data from human victims, your "Artificial Intelligence" doesn't exist.
Copyright only concerns with expression, not ideas. You can still learn the concepts from copyrighted content without violating copyright. There is a limit to what can be protected from AI with copyrights.
When they trained Stable Diffusion they used 5B image-text pairs, but the final model fits in 5GB. So they didn't actually take too much, about 1 byte per input example, which is in line with the idea of learning general concepts rather than copying.
An interesting parallel - the human DNA is about the same size, 3.2B nucleotides. The SD model is like a cultural DNA, and like DNA is compact, easy to replicate and powerful. At this level copyright is an anti-pattern, genes and memes like to travel and evolve.
An exception should be recognized: open models - free of charge, with no strings attached.
I think this is the best approach because:
(1) Models are public, so everyone can benefit from them.
(2) Such models can be trained on public data for improvements.
(No complexity involved over paying for data, no copyright risks regarding data, etc)
(3) Companies can sell services.
(model-as-a-service, customization, apps/tools, education, etc)
(4) Companies may lose competitive edges, but they can enjoy the bigger market.
(I'm looking at Matlab and Mathematica. They are really good, but Python took the market eventually.)
It feels like many of FOSS pros/cons can be plugged in here as-is.
> Where are you going to get the training data, once copyright law is amended to protect artists from this new kind of intellectual property theft?
For me, countries that don't enforce that or old models from before the ban. Then again, I'm a nerd who wouldn't release my creations commercially—they'd be AI-generated shitposts or images done for the joy of exploring what my new tools can do.
You know piracy (actual theft, not just pure unsubstantiated hypotheticals) exists and yet the game, music, and entertainment industries seem to be doing just fine. I think artists will be OK.
I would love to change my position if there's any evidence. Perhaps it could start with a single example of any non-trivial job position being removed due to these image generation models.
There's a few games, albeit often indie, where they used generative music makers for their soundtrack, saving them having to hire a music producer.
I mean, if you've used or tried any of those models, you know that very soon it will allow an artist to be way more productive, what needed 50 artists might only need 30 now.
I'd really be surprised if this didn't cost jobs.
The only possibility it doesn't is that the projects get even bigger in scope, so increase productivity would go into more am ambitious projects instead of job cuts.
shutterstock, like what openai/dalle-2 is doing? The only impact this kind of legislation will have is to ensure that only big companies get to train large models.
Wait till you find out how brains work. Will you be mandating copyright enforcement chips in the frontal lobe? Your position is untenable and will be destroyed thoroughly, if not in places where Disney and co. are entrenched then everywhere else first and eventually in Disneylands too.
Powerful interests of media giants got copyright law where it is now, they mostly decide where it goes. If training models on illegally scraped indie artist imagery allows the likes of Disney to save money then there is no one to lobby for copyright law to be strictly applied in this case.
Yes, but it's pretty hard to definitively say that a particular work is not copyrighted, given the life+X years term, and that works don't have to be marked as copyrighted. I believe that's why books have had fewer illustrations these last 20 years: authors can't afford to indemnify publishers against copyright suits over illustrations, and it's really hard to make sure either you have the correct permission or it's out of copyright.
I think you underestimate the exponential scale of what's happening and how the training data can make and dictate digital worlds to then use to train itself again to then be directed and prompted by humans again and then rinse repeat except we also train an AI to be the human who directs and prompts but throw in a random seed but also still take in some human interaction and use that to build whole visual worlds. Flat or in Unreal Nanite 3D. Characters and places. Normal light and light we haven't seen... Train again. Use the previous human AI but now the human prompts create a set of human AIs to feed back into that and then that is put out into a space to make a set of paintings as controls that are painted by robots and those are scanned in intense detail by cameras that also pick up near IR and UV and those are used as a training set to seed the new environments which is then told to make me an episode of Duckman from a half remembered drunk memory voice memo and the output is better than the original show probably ever was.
Then turn that AI loose and let the internet use it to make Duckman episodes but people figure out how to tweak the prompts to make Toy Story at scale but also use it to train their own AIs.
We have reliable secure GPU multi tenancy now. Next year we'll start seeing AI designed silicon photonics creeping in. All this GPU training is now pretty accessible to be passed all over the internet and back securely and streaming.
No one has any clue what's coming in a very short few months. Well, we have a few ideas... but it's going to be surprising. These systems build systems to build more systems there and back again. The only constraint being energy availability and attention.
The powers that be captured all of human experience for a couple decades to seed this project. So off we go I guess! This moves so much faster than copyright law. This is our new world. The compression of information contained inside of models is something that existing bureaucratic systems cannot hope to regulate. I'm pretty sure no one that can is even slightly interested in regulating it anyway.
Our new way of life is gonna be pretty weird. I think some fun new media will come out of it though. Which is pretty cool. Also you know... Utopia or Oblivion.
> No first-world company will break the law to save money on artists
If the future of copyright law in the west moves to outlaw use in training data then the west will lose out to companies from China, Vietnam and anywhere else that doesn't care at all about western copyright. I think it will give them a competitive edge.
Not at all. I'm no great artist but I did about 440 episodes of a webcomic some years ago. That's hundreds of times what I'd need to train an AI on that style.
Now consider if I'd made that, trained the AI, and then DID NOT SHARE IT. Suddenly, I've got an AI thing that other people don't have (not directly, they can eventually have a second-order knockoff of it) made to express an intention that knockoffs won't have. So I'd have both the story to tell (hence the original training set) and the means to automate the creation of it.
You've already got artists/technicians making things like Clip Studio brushes, things that transform a stroke into a very different style or dump scatterings of clip art for texture. This is literally the same thing only way more sophisticated.
Just because you have a 'foliage brush' doesn't mean you can draw a woodland scene, it just makes the texture rendering trivial. Just because you can type 'woodland scene' doesn't make it a SCENE, it just makes the image rendering trivial. Does it serve a story purpose? What does the imagery do, beyond just existing?
The mere existence of 'imagery in X style' is rapidly going to become really boring. Once 'an image can look like anything at all' becomes commonplace, to be impressed by that will be about as sophisticated as being impressed that plastic is shiny. It becomes all about 'what did you do with it?' and that is where artists come in.
That's the trajectory if we stay on this path. Artists would essentially become AI power users. The question upstream of that is, do you think this is a good path?
Personally, I think AI has done about as much as crypto, maybe even less. I can't think of a single use of AI that I couldn't do without.
AI image recognition is widely in use for real products today. I use it pretty regularly to identify plants. Have even used it a few times to identify electrical or mechanical parts.
You underestimate people’s willingness to sacrifice other people’s work for the benefit of a few corporations and their cult like leadership. Already some argue passionately that ai stealing data is just like humans learning from other humans. Let alone that once a new trend starts ai will have to catchup - meaning that for a while it will lag so sufficient data is available for it to steal again. Also if we completely push artists out due to ever lower costs there will be little progress in original content.
Yes there will be a market for human artists as much as there is a market for horses to pull things. There are still jobs for horses now, but their number is way lower than just 150 years ago. Not great news if you are a horse (or a human with a job).
Yes but smaller. There is a lot of content outside acclaimed paintings that we will not benefit from, and most important, people wont get to practice. And art is all about practice. For every 10k mediocre artists we replace with ai we will miss that one great artist that was taken out of the pool because they couldnt practice because the job no longer exists. I see the benefits of the technology, but the price we will pay will be too large. Same could be said about photoshop but photoshop is a tool. Ai is a mind.
There is one star trek episode (think it was TNG but not sure) where the local civilization is totally dependent on some benevolent AI that their anchestors built. It broke but due to their reliance on it, they lost all expertise to repair it.
Great, now i have to watch the whole series again just to find that episode.
On a serious note i am now confused. Because at the same time i see ai generated art as a tool for creating new art, a facilitator. Still i dont agree people’s ip should be used without their consent. Oh boy progress is confusing sometimes.
The biggest shift that generative AI enables is that content becomes personal. Everyone will create and consume their own content. It's not just that the content will be AI-generated, but it is generated for each person individually, which is the 'big' thing.
Already with Stable Diffusion it's way more rewarding to play with it yourself than look at generations made by others. It's more like a game or a dream that is a personal experience, than 'art' to be shared with others.
Generative AI is CGI. But I think it's going to be a much more profound shift than was CGI. I don't think anyone really yet knows what it'll mean. I'll also claim that it won't really mean anything until it's combined with 3D ray tracing. So what's going to be "generated"? A 3D model, right? All the current AI CGI stuff that I've seen is 2D.
I agree, this far more profound than CGI. Ultimately I think this tech will be seen as allowing the exploration of all possible worlds. When the AI can generate convincing representations of anything we can think of and things we could never dream of - then what we are really looking at is a kind of inter-dimensional craft that can take us between realities.
From what I see, It seems the general theme of generative AI on HN and on the web is this:
Using an AI writer generator (GPT-3): "Looks great, I can make a great SEO blog and a storybook SaaS out of this now, I don't need a writer anymore. Writers adapt or die."
Using an AI artist generator (DALL-E, Stable Diffusion, Midjourney): "Looks brilliant, this IS the future. it's inevitable, I don't need an artist anymore. Artists must adapt or die, sorry."
Using an AI code generator (programmers) (Copilot, Replit Ghostwriter): "No thank you, Bad. legislation now, why are you violating my copyright / open source licence?"
Why does it seem OK to obsolete complete industries by outright violating copyright laws, stealing work (even generating with a shutterstock watermark!), often without attribution and hail this as a new dawn for AI?
Or is this mainly VCs finding out that this is now a new way to 'provide shareholder value' despite all the costs that come with it.
So just throw cash at any startup that can generate a hot dog image in any style or generate a hot dog detector with copyrighted code, writing hotdog story no matter how bad the story is.
HN is pretty pro-Copilot in most threads I've seen. Honestly I think I have seen this comment -- "you programmers sure like generative models when it impacts other industries but you get awfully defensive when it comes to your own!" -- much more often than I have seen anyone actually push back against Copilot.
Frankly I think it's projection from non-SWEs who don't understand that automating your work away is the core mission of software development. There are good reasons to worry about how Copilot might impact the quality of software, especially if it's being used by someone who doesn't understand the generated code, but "what about my job safety??" is nowhere on the list.
Most of the comments I see are taking issue with exact replication of code - this is obviously copyright infringement. The same thing can occur with Stable Diffusion - if I output an image of Iron Man, this is also copyright infringement. Both are bad. Realizing this does not mean that you must inherently be against Co-Pilot or Stable Diffusion. It would be just as bad if I copied the code with a ctrl+c ctrl+v, or drew Iron Man with a pen and pencil.
There are of course some people in there against the very idea of Co-Pilot, but from my quick skimming they do appear to be the minority.
as with anything, reasoning is motivated by class interests. reasoning flows out of class interests. thinking exists in the first place to satisfy needs.
by the way, you forgot hypothetical fleets of self-driving trucks putting truckers out of work, and so on.
Comparing image generation to earlier technological breakthroughs has been very helpful for me in two ways. Photography is my go-to comparison.
First, it helps quell the extreme ends of anti-AI paranoid fear I sometimes feel, to think that yes, it could change the world significantly and make some things disappear or reduce to irrelevance (e.g. realist art) and make new things appear and rise to relevance (e.g. mass media), yet throughout all this the more fundamental aspects of humanity and human society and culture have endured and adapted just fine.
Second, it gives me more empathy for those people who were/are afraid of or simply opposed to other technological developments which I take for granted (again photography, TV, the internet, trains, nuclear power..), which various fears/oppositions are often an object of derision and ridicule.
I feel this (AI, starting roughly at ~alphago) is the first significant breakthrough that I'm consciously living through myself, and this does a lot to one's worldview I think.
Im looking forward to AI helping generate game art and assets so that developing games will be more practical for smaller teams and larger organizations will be able to make grand games.
The situation is so bad that even the latest Elder Scrolls games reuse assets from games from way back to speed up the process.
I'm wondering if we are going to see the same pattern than with AI driving:
a lot of enthousiasm at first because what is done if very impressive but after the initial phase the interest slow down a lot because at the end AI is still "stupid"
I predict all art on the web will soon include a tag that says "Permission is not granted to use this image to train an AI algorithm". Perhaps something similar also as a general legal notice for all content on all web sites. Stenography will be used to enforce this. If you find your stenographic tag in an AI generated image you sue.
Generative AI is much more than "generative". It doesn't just make it orders of magnitude cheaper to create things, but enables vast new areas of applications.
Calling Stable Diffusion etc AI art generators is kinda like saying our visual cortex is an AI art generator because it can render pictures for our consciousness.
It's perhaps not much more than this at the moment, but it will be, soon.
I read this whole interview and dozens of comments here before realizing that the intended reference was most likely Computer-Generated Imagery and not Common Gateway Interface. (I was imagining that the interviewee mentioned the latter as an example of a technology that somewhat unexpectedly ignited a whole industry -- in that case, web development.)
Just like CGI, the early use of generative AI experience isn't always there, which can lead to frustration. I'm working on a platform to help with this, can't give details just yet: https://inventai.xyz.
It will be much bigger, because:
- it can be automated. CGI still required a lot of human work.
- AI generated content requires way less skill. You had to train a lot to become good at CGI. We already see people getting Stable-Diffusion-proficient in a few hours.
- in the 90, CGI was introduced gently, through selected medias. AI content has access to the full internet right out of the box.
- producing CGI required a lot of hardware, but with SaaS, AI will be available to anybody.
- CGI was first exposed to professionals, who produced for the public. AI is already been used by hobbyists, and soon the public, directly.
- CGI was mostly about visual techs, but AI generated content can be applied to text, sound, molecules, basically any structured data.
In summary, the very nature and context in which AI generated content is taking form means the impact will be larger, stronger, and way more intense.
People are hardly capable at dealing with CGI intelligently, even after 20 years. They trust photos and videos a lot, and react strongly to them. Even when they know they are fake, like in a movie, some emotions will forge they judgement like if it were real.
AI generated content? Forget about it. People are not ready. The emotional impact it's going to have will make you feel like the cancel culture was a whisper in the wind.
I was underwhelmed by the last decade of AI dev, but 2022 is for AI what 2000 was for the dotcom revolution.
Question is, what the equivalent of domain name for AI, something cheap you can buy to sale later?