Key Stable Diffusion Researchers Leave Stability AI as Company Flounders
forbes.comSD3.0 looks really good and will probably displace a lot of closed image generation models once it's out.
I don't understand why Stability gets so little support from the community. They released the first usable open-source models and their models are the foundation of the most interesting AI-bashing workflows out there - VC funded or otherwise.
Stability AI had massive support when Stable Diffusion 1.x and SDXL each hit the scene. But the Gen AI industry has evolved so rapidly that forks and iterations on the models outpaced them, especially commercially.
That's a feature of open-source development, not a bug. But it's a reason (along with the general financial issues which are the company's fault alone) why Stability is switching to a "need a membership to use commercially" business model, and IMO it won't work.
its a (minor, not permanent) blow to open source ai that stability, and now mistral, felt the need to switch to closed for their leading models. openwashing is a crowdpleaser but everyone goes closed when things get serious. basically only meta is the leading hope. i hope more players emerge - but i also dont have a ton of ideas on how to fund them. these efforts take serious resources.
that said i think its impt to acknowledge how much stability has shared in its research, just the other day they were on HN for Stable Video 3D, not to mention hourglass diffusion and other Stable* models. may not be the overwhelming SOTA but its real open source AI work that pushes the frontiers. you have to give them credit for that.
> these efforts take serious resources
Meta just published their new optimization results [1]. According to them
In this context a GPU is an NVIDIA A100, which you can buy, if you can buy, for $10000.> training a 7B model on 512 GPUs to 2T tokens using this method would take just under two weeks.And this is after an explosion of ideas that lead to unthinkable optimizations just two years ago.
If someone did train such a model 2 years ago, it would have cost hundreds of millions. Now it's 5 million. Maybe in 2 years it's going to be only $50k. Should you start a startup now and invest $5 million, an risk someone stealing the show for pennies in 2 years? If you do, I really can't see if you can afford to open source the results of your training.
[1] training a 7B model on 512 GPUs to 2T tokens using this method would take just under two weeks.
There is nothing you can run on your computer that is even remotely as good as stability products.
Which means, there is nothing with even remotely the same fine tuning ecosystem.
And for that - stability is way ahead of the competition.
Yeah. It is weird. On one hand, v3 looks really good, the arch is sound, and the work is solid. On the other hand, selling Clipdrop is not a good sign (a growing startup divest is uncommon). Emad likes to talk about how they were able to retain researchers and now this, so it is hard to know what's going on.
The CEO being a fraud could be a possible reason. Just saying.
You shouldn't judge people by what Forbes says, but what people do. So far, Emad delivers on what he promises (with some delays here and there), and Stability.ai is the only one that publish image / video generation models at production quality.
Not just Forbes. Read about him on Wikipedia.
>However, according to him, he did not attend his graduation ceremony to receive his degrees, and therefore, he does not technically possess a BA or an MA.
Oh wow, he's probably lying about his education.
Good. I trust people who lie about their education and exceed the expected ability of a BA, vs. the huge amounts of barely articulate and completely disinterested CS graduates entering the workforce as of late
I received them by mail. He didn't?
actually he recently received his BA and MA now.
I mean, you get a nose for people like that after a while. They all have common characteristics like self aggrandizement, making bold claims and then changing tune when caught, a history of burnt out and salty ex co workers. It’s not that hard to spot in this guy’s case.
Did you read the article?
I saw everyone repeating this over and over but then I actually read the article and couldn't even understand what the big deal was... hard to find a founder who hasn't done all the things he was accused off none of which were really a big deal.
Felt completely overblown and honestly like a weird hit piece but where the journo didn't actually find any real dirt to smear.
Certainly a shady player. Doesn't seem to be in good hands.
FWIW, they lost my interest entirely when they launched SD 2.0 and the model was worse, presumably (this was what people at the time were coming up with as the core problem) as they went overboard in their attempts at preventing it from generating naked photos (by removing so much from the training set that it no longer seemed to understand as well what a human looked like). The wider community then just got stuck on SD 1.5 and so frankly it isn't clear to me if the work Stability was doing was even relevant anymore... even just a few weeks ago I saw some new thing using SD, said to myself "god I hope it is using 1.5", and it was in fact (thankfully) still using 1.5.
SDXL is amazing.
The community is entrechend in 1.5 because that's what everyone is now familiar with, IMO
>The community is entrechend in 1.5 because that's what everyone is now familiar with, IMO
That probably has some weight to the community's decision to still use 1.5. Other reasons (and more important IMO) why we're still stuck on 1.5 is due to nerfing 2.0, and the plethora of user trained models based on 1.5.
I'm continued to be amazed by the quality possible with 1.5. While there are pros and cons of each of the different offerings provided by other image generators, I haven't seen anything available to the public that can compete with the quality gens a competent SD prompter can produce yet.
SDXL seems to have taken off better than 2.0, but nothing so amazing to justify leaving all the 1.5 models behind.
Well, personally, SDXL just blows 1.5 out of the water for me. I haven't had a reason to even touch 1.5 in months.
But note that SDXL is really awful in automatic1111 or vanilla HF diffusers for me. You have to use something with proper augmentations (like ComfyUI or Fooocus(which runs on ComfyUI)).
>You have to use something with proper augmentations (like ComfyUI or Fooocus(which runs on ComfyUI))
Yeah, comfy was given a reference design of the sdxl model beforehand so it would be supported when sdxl was released. I should probably switch to comfy, but I don't touch the tech very frequently as I don't have a practical use case besides the coolness factor.
Ok, I'll try SDXL? But, I continue to believe that it was the botched release and attempts to push people with SD 2.x that led to whatever is being talked about in this thread for why Stability gets "so little support from the community": I lost interest in what they were working on well before they released SDXL, as I was no longer convinced that their newer stuff would be better than their older stuff due to 2.x.
FWIW, "everyone had gotten so used to 1.5 that they just didn't want to bother with 2.x" might provide a similar mechanism, if a very different place for the blame: if people aren't paying attention to the new stuff you are building, it is going to hurt your "support".
I was looking for something about how the company actually made money, but only found this:
> It’s a dramatic exodus that comes less than 18 months after Stability’s 2022 fundraise that valued the company at $1 billion. Now, the company is facing a cash crunch, with spending on wages and compute power far outstripping revenue, according to documents seen by Forbes. Bloomberg earlier reported that the company was spending $8 million a month. In November 2023, CEO Emad Mostaque tweeted that the company had generated $1.2 million in revenue in August, and would make $3 million in November. The tweet was later deleted.
This sounds a lot like the two-year period leading up to the dot com crash. Insane valuations and no revenue model. Meanwhile those insanely-valued companies bought Sun Microsystems servers like there was no tomorrow. When the games ended, a lot of those insanely-valued companies went to zero and left a massive overhang of Sun hardware in their wake. Sun began its long nosedive not long after that.
This many people leaving and key investors resigning from the board is definitely interesting. I wonder if there's some undisclosed scandal that will break because this is pretty wild.
Every Forbes article should have the warning that they are not real publication anymore. Everything they publish is written with an intention that you believe an authors opinion as fact.
The model for Stability ends in a likely way, they will be swallowed up in a payday. We have recent examples of this.
For now, they’re the most open area in a garden of closing gates. And I wish them all the best.
Yup, that’s why the HN code shows the username after the domain now, same as Twitter users.
I just noticed. Thanks for this friend. Much appreciated!
I was contacted by a Forbes writer for a hit piece, and his pitch on why he wanted to talk to me was clearly a lie.
They feel that they are entitled to be dishonest while writing articles about how CEOs are dishonest.
*Every legacy media article
No argument from me, but Forbes is special in this regard.
It takes $500 to be a “featured contributor” and they will post whatever nonsense you like. The brand is diluted.
See Wikipedias guidelines for sources.
(ctrl-f for forbes)
https://en.wikipedia.org/wiki/Wikipedia:Reliable_sources/Per...
Indeed.
Articles can be bought outright.
Archived non paywall copy of the article: https://archive.is/iRIlp
Cheers to you, good citizen. I wish more people started with such links rather than waiting for someone else to make it easy to have a discussion.
It makes searching for repeated or similar articles difficult.
For example - https://news.ycombinator.com/from?site=forbes.com/sites/iain...
Linking to archive (which incidentally gets into loops for me with the captcha) means that I can't see the original article either (or use other methods to try to find reproductions).
Meanwhile, even though I hit the "please get a subscription" for Forbes, I can click the reader mode and read the page in its entirety.
That's a violation of HN guidelines.
Is it really? TIL, thanks.
edit: Arguable. I assume your referring to:
Please submit the original source. If a post reports on something found on another site, submit the latter.
...and I might argue that an archive of the original is just as good, judging by the verbiage referring to "If a post reports on...". Archive links don't change, remove, or add any context, they just make it accessible.
I suppose if we wanted to know this in specifics, it'd be a question for dang.
Posting paywalled links should be, as well. It's borderline marketing. If I pay for the source, I don't need to find it on HN.
not true. hn surfaces quality and adds discussion. and the people who wrote the content need acknowledgement/attribution, and some conversion to paying, or they'll eventually stop. do you think good reporting/writing just falls from the sky?
It comes from imputing the facts into GPT 3.5 and asking it to write an article.
Sorry, what is? It isn’t to post or ask for archival links. It’s only a contravention if someone complains about paywalls.
When I try posting archived links, they are auto-dead.
They should consider selling flounder
How can we help save Stability AI?
The open source community should pitch in.. goFundMe
If Stability was a nonprofit, sure. Otherwise, lol.
the models came out of an academic CS lab...can always fall back to those....
For my use cases OpenAI's DallE is often good enough. No more Discord nonsense.
You're thinking of Midjourney. Stable Diffusion is awesome, you can run it locally on consumer GPUs.
Parent comment is classic HN: strongly held opinions with 100% confidence, 75% of the information.
Appropriately for the topic, also classic AI.
Of course the verb they're looking for is founders. A flounder is a fish. Founder (as a verb) means "to fail."
No, Stability hasn't foundered yet. They're floundering: https://www.merriam-webster.com/dictionary/flounder
Per wiktionary (https://en.m.wiktionary.org/wiki/founder) Usage Notes: Frequently confused with flounder. Both may be applied to the same situation, with the difference being the severity of the action: floundering (struggling to maintain position) comes first, followed by foundering (losing it by falling, sinking, or failing).
I think “flounder” is fine in this context.
Confidently incorrect is certainly the worst kind.
You have floundered in analyzing the headline
Stability always seemed like a dark horse thanks to the reputation of their founder and their open source business model. I do get some schadenfreude from seeing them like this considering the damage many feel they've done to the art community.
E: -4 in 10 minutes. Stay classy hackernews. I hope this company and OpenAI choke on the algorithmic disgorgement when the law catches up.
The "art community" did most of the damage to itself.
To expand upon this, the majority of the complaints I see are coming from the online community where the largest part of the material seems to be based of IP/copyright violation by the artists. I've been part of multiple communities who have decided to ban AI art because of a moral clause against using other's peoples works without their permission, but yet which don't like having it pointed out that the IP holders have withheld their permission of their characters being used for unlicensed artwork.
While some of this are people practicing their artwork and I don't see any reason we should care what artwork someone practices on, this is also the general trend for artwork being sold. Go to any convention where artists sell work and look at how much artwork is sold of characters the artists do not have license to. While I think one can take a philosophical stance against the current IP laws that outlaw this, such a stance would make it quite hard to oppose the use of content in training an AI.
In short, if those making the AI stole IP to train the AI, it was stolen from a community that was fine with IP theft that benefitted them. And if the claim is that it wasn't IP theft because the law was generally tolerating it (as long as no one became so much a target they received a C&D), then unless there are some lawsuits won against the AI it would be equally allowed.
(And of course individuals will have their own philosophical stances which might be much more consistent, I'm speaking of the generalized view I have developed from overall interactions with parts of the community and as such it is not meant to be strongly prescriptive to any specific member of the community).
Personally I think framing this as some kind of double-standard regarding copyright or even a copyright-specific issue misses the forest for the trees.
The problem I see is that the generative AI economy hinges on an injustice: the presumption that all art on the internet - no matter the medium, or means or relative notoriety of the artist - shall be candidate training data, and no burden of attribution whatsoever shall be laid upon those who leverage it.
Most graphic artists that I know bemoan copyright. But, it's a tool that the law has given them.
Also: most graphic artists that I know exist under low economic circumstances - some near poverty - relative to most of the people I know who are building the next great wave of technological innovations with generative AI.
I don't see a struggle over copyright. Artists, who exist towards the bottom of the economic ladder as it is, are doing what they can to survive.
>no matter the medium, or means or relative notoriety of the artist - shall be candidate training data
Why is this a special case? When people were viewing it, reposting it, using it to learn... those were all accepted uses. Passing it off as your own wasn't allowed, but outright plagiarism has pretty consistently been unaccepted.
The problem seems to stim from using it in a way that directly competes with the artist, and given your other point about their financial position, is a direct financial threat to them. The morality of the situation seems to be that it is wrong because of the financial harm, but recognizing that such an argument is rarely accepted, it must instead be justified by some other argument, any other argument, that condemns the outcome.
I don't think this is anything particularly unique. How often do we find things wrong because of a logical argument as to why it is wrong, and how often do we find a logical argument to justify our felling that something is wrong?
There is also a element of helplessness. No matter what the government does, pandora's box has been opened and it can't be closed. While it might slow down the development of better AI, it isn't going to stop it and banning existing software isn't going to be possible. The damage has been done, and even if the artists have an overwhelming victory, they are only going to recover a fraction of lost ground only to eventually lose it again.
The tricky part is that the ongoing training of ever-more-interesting generative AI *depends* on the ongoing labor of artists.
Proponents of the current economic model like to frame the artist rejection of AI as an obvious case of Luddism. Of course the artists reject this, it threatens their economic station! And: it's not even wrong.
But, it is a high modernist foible: at some point the raw resource is fully exploited and the wave of companies that rode high on its vast-but-unrenewable quantity will reckon with reality. Their businesses are unsustainable (who could have foreseen it!).
In the mean time, artists won't disappear. Most likely what will happen is that they will continue to subsist - they are essential in this economic loop, whether fairly compensated for their labor or not - but with an even lower economic posture than before.
I don't think there is a moral crisis here, but an economic one. Incidentally, an injustice is perpetrated upon an entire class of laborers. I'll leave it to others to decide the morality of that, considering all the trade-offs.
>The tricky part is that the ongoing training of ever-more-interesting generative AI depends on the ongoing labor of artists.
I think there is an ongoing issue. Much like how the privatization of the public domain has led to an ongoing issue of a large percent of our culture being privately owned. I'm not sure the fix to this.
I am by no means happy with the current situation, but I do find the moral reasoning behind the outrage at AI questionable at best as it doesn't seem to be consistent and instead based on what is economically beneficial to those showing outrage. By that same standard, AI is great because it lets me create things at a much cheaper cost.
Artist creating art of popular characters and AI using publicly posted art both seem pretty acceptable to me. Then again I'm the weirdo who goes to conventions to buy originals, the ones actually painted on canvas and not just easily reproducible prints, even though that does mean paying far more than the prints cost.
We are booking our tickets to Comic-Con. We tap through the form, maneuvering past the cumbersome AI-generated takeover promoting a new prestige TV show. The AIs are getting pretty good, and seem to have moved beyond the glassy-eyed doll faces characteristic of the Stable Diffusion era. Nevertheless, the ads are obnoxious and we can't wait to submit the form and get off of this website.
A memory rises unbidden: we once made comics and posted them to the web, free for all to read. We would even browse the web just to find and read them. Wild.
We find ourselves circuiting the convention hall. It is a brightly lit maze, festooned with endless AI-generated promotions for Marvel supers and yesteryear reboots. The cast of Friends is back, youthful as ever, and apparently we're getting at least three more seasons. We round a corner and..
Here. Yes, here. We remember it now. This whole row was once filled with tables showcasing prints and original works of art. Behind the tables: a spouse, a friend, or the artist in the flesh. Artists, who were remarkable in their day for their contributions to the great pop culture that drew us to the convention. Artists who, despite their labor and their infamy among certain fandoms, never appeared in a legible place on the credit roll. Artists who worked a day job for years, stocking shelves, packing boxes, approving Disney licensee merchandise, so that in the evening they might bend their weary backs, put pen to tablet and spill their imaginations across the screen. Artists who did all that so that we could come to Comic-Con today and appropriate for ourselves an original work of their art.
Where once there were artists, now there is Hello Kitty. Sanrio has taken over the whole row. You can walk up to Hello Kitty and ask it for any combination of officially licensed characters, with optional accessories if you have a few more dollars to spend. A 3D printer somewhere behind the booth's facade fabricates the bespoke toy on-demand in food-safe ABS. An original work of art.
>ongoing training of even-more-interesting generative AI depends on the ongoing labor of artists. What does that mean? That you need artist to produce "even-more-interesting" art to train new iterations of generative AI? Could you point out some of these superfresh new art artist are doing that are completely different than things already done to exhaustion in the 1970s (other than anime)?
I don't think you need artists for that anymore. Certainly you don't need them for commercial purposes. If they are going to survive as an artist professionally, it will be because of the people that refuse to use AI art for whatever reason, but I don't see how that won't be short lived in the market.
Artists will survive, not professionally, but because they are doing it for the arts sake, even if that doesn't offer them any financial reward.
Yeah one thing that's often missed in these discussions is that the AI companies still need the artists! When you overexploit the commons, there will be consequences. We'll see if the synthetic data is good enough or not for future models I guess.
> the largest part of the material seems to be based of IP/copyright violation by the artists
What do you mean by this? Fan art is pretty well known to be fair use, particularly because it is transformative and it has no commercial intent. AI training is transformative alright, but the commercial intent part is a huge factor in the analysis. The fair use analysis is very much not clear at this point.
> it has no commercial intent.
Under what possible analysis does art sold at a convention, by full time professional artists, not count as being done for a commercial purpose?
Does this actually happen at large scale without the permission of the owner of the IP?
I can see someone making a few hundred bucks with their fan art, but doing tens of thousands of dollars of business is a different story.
The main commercial factor (per the courts - see the recent Warhol lawsuit) is whether the derivative work competes in the market with the original work. I sincerely doubt that even if there is large-scale selling of fan art at conventions, that fan art is meaningfully competing with (ie reducing) the market for the original IP.
> Does this actually happen at large scale
Yes?
Go to any convention center. Go commission a piece of art on the internet.
It is almost all infringing "fan art".
> I can see someone making a few hundred bucks with their fan art
Its not some rando person doing this stuff for a hobby. Instead, I am talking about the entire industry.
All you'd have to do is go to any gaming/media/comic convention and this is immediately obvious.
> that fan art is meaningfully competing with (ie reducing) the market for the original IP.
I mean, ok? Then if thats your metric, then you can't complain about the entire open source industry of people making AI art on their home PCs.
If you are giving that gigantic, large hole to slip through, then you have now allowed almost the entire open source AI art industry to exist.
> All you'd have to do is go to any gaming/media/comic convention and this is immediately obvious.
Most of the gaming/media/comic conventions I have been to have involved game/media/comic artists selling their own work, not people selling fan art. In fact, the presence of the original artists selling copies of their original works is generally a big draw for the convention. Maybe we go to different conventions or something (I have never been to an anime convention - so maybe that's what you're referring to). The little third-party art I have seen at these conventions is sold with the explicit permission of the original artist/IP holder. So no, it is not "obvious" to me, as someone who has actually gone to a few gaming/comic conventions before, that fan art is a huge industry or that it undercuts demand for the original art.
The fan art I have seen is generally drawn by (professional/high-end amateur) artists for free on deviantart because they like the characters or want to practice their skills.
Also, nobody is currently suing (or particularly upset) over people making art on their home PCs. People are suing over companies selling AI art generators for $billions that directly compete with the artists and stock photo libraries that were used to train these art generators.
> Also, nobody is currently suing (or particularly upset) over people making art on their home PCs
Nobody is suing because few people have the resources to sue anyone, yes, but people are absolutely upset about all the AI art that is now on the internet.
Much of which is entirely non commercial, in the same way that any other piece of online fan art is.
But hey sure, if your position is that almost all of the online AI art stuff is totally fine (it's mostly all non commercial), then great. You support almost all AI art.
So I guess that means that both fan art and AI art are basically the same anyway, using that same definition of non commercial, which was my entire point.
Also, it doesn't really matter if stable diffusion, or 1 or 2 other big companies go out of business at this point.
Their models are already available for anyone to use, and other people are training them even now.
A couple companies being sued doesn't stop any of this technology even a little bit, because lots of amazing models are available.
To be very clear: I support the prompters generating AI art under the same terms as fan art - you can do it for non-commercial purposes, but if you want to make money you have to get a license. I do not support the companies selling AI art models for commercial purposes without compensating the people whose work they used to train.
The NYT's lawyers and Getty's agree with me, by the way - they aren't suing users (and there are big users out there who could be worth suing).
> I do not support the companies selling AI art models
Then you should be happy to know that the large majority of AI art models that matter are open source, or fine tuned open source models.
That's most of the space.
You support basically all AI art, which is much different than what most artists are complaining about.
And even if a few of those big companies go away, those models are still out there and available for users to generate from.
There is also the fact they have driven themselves mad trying to "detect" or distinguish AI art, and now consistently attempt to tear each other down over claims of AI use.
Yes, blaming the victim, classic argument.
The entire AI industry is powered by piracy at a massive scale. Very little training data is properly licensed or compensated. It's just more obvious with open models because we can investigate them. Closed models are sausages and we don't know what went in.
Download a movie and you can get sued or your Internet connection terminated, but pirate the entire collective output of humanity and sell it back to us from behind a paywall and that's fine.
I have more sympathy for Stability here because at least they opened the models. IMHO models trained on not-properly-licensed (pirated) data should at the very least not be copyrightable and should be public domain. (These piracy enterprises are aware of this as a possible legal outcome in some jurisdictions, so the whole AI safety bullshit performance is an attempt to scare people about open models to head off the potential of questionably-trained models being declared uncopyrightable and forced to be released.)
> IMHO models trained on not-properly-licensed (pirated) data should at the very least not be copyrightable and should be public domain.
My understanding is that ML model weights cannot be copyrighted as an original creative work. They are trade-secrets and protected through contracts but once leaked to third parties it’s not a copyright violation to use/distribute.
Whether the model is actually a derivative work of the training data is another interesting question.
Or is my theory off here?
The main argument I have seen (which is also OpenAI's in their legal briefs) is that it is fair use. The idea of "fair use" is that you are conceding that you are infringing by creating a derivative work, but it's still okay. Implied in the fair use argument is that it is a derivative work.
> Implied in the fair use argument is that it is a derivative work.
You can get all LLMs to spit out almost exact copies of known IP visuals from movies and games. For instance, with Dalle-E and Midjourney, it's relatively easy to get similar pictures from film and game studios. Those are copies with minor changes. It would be hard to argue otherwise in court. The same happens with ChatGPT spitting out verbatim passages from New York Times articles.
>and sell it back to us from behind a paywall and that's fine.
That's the sticking point. If it's an open tool for humanity's benefit being created given back to us, that's one thing... but to sell it back to us...
With that said, piracy is close to what's happening... but I think we should be careful classifying where/what exactly is the matter. I reason I think that matter's lies may be down the end of a slippery slope, or it may be straight ahead of us... the future is hard to know. If we classify it poorly we may unintentionally cause human(post/trans-human) right's issues {if I upload my consciousness to a digital mind, I don't want archaic laws to dominate what I can see/compute based on the material of which I'm made}.
These companies need to make money because they took VC money and training these things take 10s if not 100s of millions of dollars.
Also nice to see the complete nonsense of digital mind uploading on hackernews vis a vis this discussion. If that happens we'd need to change a lot of laws anyway.
The mind uploading is just one of many potential outcomes.
To me the more interesting concern: we can't seem to agree on the bare minimum requirements for sentience/experience. Maybe the 'bare minimum' is 'electricity runs through it'. It may be that these LL/SD/ML models are having an 'experience' without the proper memory/state/internal-control to achieve sentience/consciousness.
Law's need to change, that's for sure (look at copyright).
And taking VC money to do anything Open seems like a trap. There are government grants... but yeah... there exists a whole host of (related or similar) problems in that.
Thought crime in 2084: the name given for a crime for which the only evidence is a scan of your brain.
e.g. "You imagined someone naked! That's a non-consensual deepfake of intimate personal imagery!"
>The entire AI industry is powered by piracy at a massive scale.
ARRRRR..
This is a grey area still for me. It's a neural network. It works similar to our brains work, but more consistent. It's doesn't seem like piracy to me. If an artist was really into Salvidor Dali, and happened to imitate his surrealist style, it would not be considered piracy. In fact, this is how art has evolved over the centuries. Each relevant artist in the past has incrementally contributed to what we call art today.
I feel like the people unwilling to accept that AI may impact their career are more worried about putting food on the table than anything else, which is very understandable, but it's just the cost of progress.
The bigger problem we need to deal with is how to retrain and provide job placement who are affected by disruptive technologies. We've really failed the public on this in the past and I don't think it's worth nerfing emerging tech just to keep people employed. This is not the first or last time this has happened, and it's going to be more frequent as technology advances.
> It's a neural network. It works similar to our brains work, but more consistent.
Irrelevant and incorrect.
> It's doesn't seem like piracy to me.
It's pretty indisputably piracy, whether or not it's legal/fair use/whatever. Many of the training sets included material like the books3 corpus which was downloaded to a server somewhere. That is simply piracy, doesn't matter why they downloaded it.
I believe many artists rightly refuse to accept this threat to their livelihoods because it was built on their labor. It's so fucking rich to see people patronizingly suggest that this is just an economic problem and those artists better just figure out a new profession.
You built a commercial product on unlicensed data. Do you actually think the law is going to agree that that's fair use?
> It's pretty indisputably piracy, whether or not it's legal/fair use/whatever.
Ah, this is obviously some strange usage of the word 'indisputably' that I wasn't previously aware of.
> I believe many artists rightly refuse to accept this threat to their livelihoods because it was built on their labor.
This model is trained from scratch using only public domain/CC0 and copyright images with specific permission for use: https://huggingface.co/Mitsua/mitsua-diffusion-one
Does it change anything?
If all the other models were deleted, and this was the only one left, and all future models also had to be similarly licensed, would it change even one single point?
Even if it was the only remaining model and this kind of licensing a requirement for all future work, artists would still be automated out of their highly skilled yet poorly paid profession. It still sucks. There's still no nice way to convey that.
> You built a commercial product on unlicensed data. Do you actually think the law is going to agree that that's fair use?
What do you think the Google search engine is, if not a commercial product built on unlicensed data?
The courts go both ways on this specific question with Google depending on the exact details, because nothing in law is as easy or simple as the clear-cut, goodies-vs.-baddies, black-and-white morality play you want this to be.
The fact that Stability AI have not yet been sued out of existence in a simple open-and-shut court case about copyright infringement ought to have demonstrated both this point, and also that the question "is this piracy?" is, in fact, disputable.
https://huggingface.co/datasets/P1ayer-1/books-3/discussions...
It seems incredible to me to suggest that piracy wasn't involved in the collection of training data, regardless of your view on the morality or legality of it. Datasets like books 3 indisputably contained copyrighted content that was being distributed without permission from the rightsholder. That's just the definition of piracy. If we can't agree on that then I'm not sure what we're doing here.
More materially to this discussion, yes, it would absolutely make a difference if the AI was only trained on licensed content. I wouldn't use it but I wouldn't have a problem with it. The issue is specifically that much of the work being used without permission is being used to replace the people who made that work, and is being used without permission. If the model is based on ethically acquired data, it would be less able to reproduce the style of specific artists. Imo, there would be more room for both kinds of art in this case.
I'm also aware that it's not a clear cut case legally but I think AI advocates and tech enthusiasts think it's a lot more likely that AI will win in court than the actual chances. Napster took years to litigate and was eventually shutdown. There's a really good discussion about this on the decoder podcast between actual lawyers.
> https://huggingface.co/datasets/P1ayer-1/books-3/discussions...
https://transparencyreport.google.com/copyright/overview?hl=...
> It seems incredible to me to suggest that piracy wasn't involved in the collection of training data, regardless of your view on the morality or legality of it. Datasets like books 3 indisputably contained copyrighted content that was being distributed without permission from the rightsholder.
Is the Google search engine piracy?
https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....
https://en.wikipedia.org/wiki/Perfect_10,_Inc._v._Amazon.com....
https://en.wikipedia.org/wiki/Field_v._Google,_Inc.
https://9to5google.com/2016/04/27/getty-images-google-piracy...
https://www.reuters.com/article/idUSN07281154/
> That's just the definition of piracy. If we can't agree on that then I'm not sure what we're doing here.
It literally isn't the definition of piracy.
Piracy exists only with regard to the legal definition: "Copyright infringement (at times referred to as piracy) is the use of works protected by copyright without permission for a usage where such permission is required, thereby infringing certain exclusive rights granted to the copyright holder, such as the right to reproduce, distribute, display or perform the protected work, or to make derivative works."
Even this definition annoys a lot of people, but I will ignore the whole "it's not theft because you're not depriving the original owner of anything" as a case of taking an analogy too literally.
> More materially to this discussion, yes, it would absolutely make a difference if the AI was only trained on licensed content. I wouldn't use it but I wouldn't have a problem with it. The issue is specifically that much of the work being used without permission is being used to replace the people who made that work, and is being used without permission. If the model is based on ethically acquired data, it would be less able to reproduce the style of specific artists. Imo, there would be more room for both kinds of art in this case.
Congratulations on being consistent, almost all the artists and authors are still permanently out of work.
Even ignoring that style isn't covered by copyright (because you could reasonably argue instead that it's a trademark and/or design right issue), most artists are already extremely poor due to oversupply by other humans.
> I'm also aware that it's not a clear cut case legally but I think AI advocates and tech enthusiasts think it's a lot more likely that AI will win in court than the actual chances. Napster took years to litigate and was eventually shutdown. There's a really good discussion about this on the decoder podcast between actual lawyers.
FWIW, I know better than to trust my own beliefs[0] about law, as (free) ChatGPT is simultaneously bad, and yet vastly better at it than me.
Likewise, I think (but hold the view weakly) the mere existence of AI at even the level it was before ChatGPT's first release, is going to force a radical change in the nature of IP laws — even then these models were too good-and-cheap for countries to not allow them, while also breaking a lot of the current assumptions about everything: https://benwheatley.github.io/blog/2022/10/09-19.33.04.html
[0] I really ought to get a T-shirt printed with "Wittgenstein was wrong!"; there are so many different ways I don't accept one of his famous quotes: https://philosophy.stackexchange.com/questions/72280/first-p...
> The entire AI industry is powered by piracy at a massive scale.
Forget about AI. Instead it is almost the entire art industry, wholesale!
The semi-professional online art commissioning market is almost entirely copyright infringing fan art works, being sold without permission of IP owner.
Yes, fan art is infringing. Especially when it is sold. And if you go to a convention center, to the artists section, you will see that over half of the booths are straight up selling other people's IP without permission.
This is the case for conventions, online art commissions, etsy/handmade items, all of it.
Its all illegal, all infringing, and the only reason why anyone cares now is because someone else can do the same thing that others have been doing for decades, but quicker and cheaper.
I'm glad someone brought this up. Artists, especially fan artists, will only hurt themselves if they advocate for classifying transformative works as infringing. Fan art has been so normalised that people have forgotten that it used to be considered legally dubious. Better to advocate for reskilling and social safety nets; automation affects everyone, not just them.
Fair use has 4 factors. Transformative is one of them. Recently, courts have gotten much more interested in a different factor, the "commercial intent" factor. While fan art is less transformative than AI training, it's not commercial and it's not competing with the original work (if anything, it enhances the market for the original work). Generative AI models are both commercial products and very successful competitors with the original works they used for training.
> it's not commercial and it's not competing with the original work
Yes it is and yes it does.
"Fan art" is "fan" in name only.
If you read back on my original post, you will see that I am talking about almost the entire online professional art commissions market.
From online, to convention centers, and more.
All of this is commerical and all of this competed with the IP owners.
People just sell other people's IP in all of these places.
I'm perfectly fine with getting rid of AI... if all fan artists paid the statutory and actual damages for their infringing activities.
What actual damages are there for fan art? Can you prove that work done by a fan artist would otherwise have been done by the original creator and that fan artists are costing sales of the original works they create fan art for?
> Very little training data is properly licensed or compensated.
Could it ever be the case, I wonder, if we could trust/enforce/believe that a model had so abstracted what it learned from the training inputs such that the model was not a derived work from them?
I've seen the examples where the model is able to reproduce recognizable characters from popular media. Those look like they might be "just" overfitting? While I can see that as desirable from the point of view of being able to create a picture of "Robocop shopping for diapers". But maybe we could compromise and converge to a point where AI art isn't quite so demonized and instead is seen as a useful tool.
I think it's obviously problematic that these companies are deriving value from millions of people without compensating them, while creating a product that competes with those masses.
You are describing the original meaning of "cultural appropriation", like when jazz and rock & roll were copied from Black American culture and sold.
I am describing "copyright infringement"
If you are selling something, and no one is buying it, the value you have generated is zero. If you put something online and you did not bother to understand this material can potentially be used by a third-party on account of its loose licensing, then who's to blame?
But the licensing isn't loose in many (most?) cases we're aware of. Merely making an image publicly available online doesn't give the viewers rights to do whatever they want with it under our copyright laws.
Well, I suppose the keyword here is "most?" because the burden of proof lies with the prosecution, the legal gymnastics of coming up a reasonable argument to this will be interesting.
They could at least make an effort to purchase licenses from all non-open content, comply with open licenses, and exclude content otherwise. They aren't making anything more than the most lame token effort because they don't care.
But that's just it - if we believe that what the model learns from the training material is abstract enough, they shouldn't license the content at all. Humans learn from and are inspired by art all the time. They create new works that are not considered derived works, despite there being obvious influence. Could we conceive of the same circumstance being possible with machine learning?
If we go down this road right now, we are allowing superintelligent AI powered corporations to front-run the entire human race and sell everything we think back to us.
It's not about theory of mind stuff. It's about just compensation of living human beings.
Well, with the status quo, there's no license required to train on the greatest works of art from centuries past.
I recognize some of the concerns about AI but I don't think pinning hopes on copyright law will deliver anything remotely resembling a remedy to the problems you bring up.
Are you talking about training a human or training an artist?
Downloading copyrighted data at huge scales to use in your commercial software product is pretty substantially different than an art student studying a reference.
> Are you talking about training a human or training an artist?
Neither: I am talking about training a machine learning model. Unless that's what you meant by "artist"?
> Downloading copyrighted data at huge scales to use in your commercial software product is pretty substantially different than an art student studying a reference.
You may have misunderstood my comment. My comment was stating that there's only a portion of human art - the most recent decades of works - which are protected by a copyright. Models like Stable Diffusion could be re-trained instead on centuries of artworks and not infringe at all. So the problem described as "AI powered corporations to front-run the entire human race and sell everything we think back to us" - this problem is here regardless of whether licenses were purchased.
You're right on both counts. Things are a bit fraught in this thread so I apologize for misunderstanding you.
1848 had a publication that might interest you.
I think the manifesto is missing some important aspects about game theory and human nature, and for some of that theory of mind is indeed very important, and that's why this particular political experiment didn't work out in the end despite the good intentions and that several aspects have become globally accepted.
I’ve read it but I think this case is much more unambiguous. Workers are paid; Marx would argue they are systematically underpaid and disempowered.
In this case the workers are not paid at all. Their work is not even acknowledged. It’s closer to cultural appropriation but quite a bit more unambiguous than that as well since this isn’t people learning from people. This is mass uncompensated value harvesting.
The number of hands benefiting here are incredibly tiny. In theory you could have one human owning the entire human mind and renting it back. This is the danger of present generation AI, not Skynet scenarios, and it anything the sci-fi stuff distracts us from this.
It’s like an information theory equivalent of today’s shoplifting epidemic except there are tiny gangs of only a few shoplifters able to run at Mach 10 and shoplift from every store in the country in days.
> The entire AI industry is powered by piracy
Just like all art. When you draw something you don't cite every single thing you've seen and experienced in life that inspired your drawing and style. Nor did you own or pay royalties to all that inspiration either.
>When you draw something you don't cite every single thing you've seen and experienced in life that inspired your drawing and style.
Oh please. There is an astounding degree of nuance and context missed in your example here.
I'm kind of down to let people follow this dead end reasoning since it's legally irrelevant. Makes it much easier to disregard.
> questionably-trained models being declared uncopyrightable and forced to be released
I think uncopyrightable is a likely outcome, but where are you coming up with forced to be released?
If there is a model which is, for the sake of the argument, absolutely definitely and unambiguously powered by piracy at a massive scale, then the act of forcing it to be released is going to necessarily entail all of the possible harms from that specific act of IP piracy.
IMO, if a model is deemed to be such, all copies of that model should be destroyed. Actual copyright law allows for the destruction of equipment used for copyright infringement, and those laws were written in the days where this meant "a printing press".
> the whole AI safety bullshit performance
The people who care about AI safety have been loudly warning about it for so much longer than these companies and models have existed, that they roll their eyes at newspapers using stock photos from Terminator to illustrate the discussion.
> The entire AI industry
Also includes self-driving cars, spam filters, medical diagnosis tools, …
reddit is down the hall and to the left.
Sure. It's clear this community isn't open to critical views on this topic.
1. It’s not just this community
2. It’s not just “critical views” it’s any view that deviates from the velocity or acceleration of the norm
3. Don’t take it personally, votes aren’t about what you said they’re about how the voter feels after having read it
Maybe this community is more open to objective analysis of the articles that are shared, and no so much of passional arguments that benefit no one.
Okay. Objectively, there are several ongoing lawsuits regarding this technology and its use of unlicensed data. Fair use is not as clear cut a case as many here would have you believe, and that's the opinion of many lawyers. Nilay Patel of the verge, who has a law degree, had a conversation about this with another lawyer on his podcast recently and it seems like the legal community is much less hopeful about it than people whose paycheck might rely on this technology.
Or maybe they are just not sure how making things open to everyone can make someone a "dark horse".
Defining dark horse: "a candidate or competitor about whom little is known but who unexpectedly wins or succeeds."
Compared to Altman, Emad Mostaque is a relative nobody and a somewhat controversial figure at the head of what is apparently one of the frontrunners of the AI industry. Additionally, releasing the model, something that was I assume very capital intensive to create, is definitely a bold business strategy, and not one we've seen succeed yet despite the popularity of Stability's models and its derivatives within the open source ai community.
> damage many feel they've done to the art community
If the "art community" can't understand what an insane gift SD1.5 and SDXL was to them then I don't know what to tell them.
Without those open models we could have easily ended up in a world where this tech existed but was only in the hands of people who could pay OpenAI or Adobe a month to use it, and I mean with the power of it what should that cost be? I mean to have such an advantage the monthly cost could have easily been in the hundreds a month like high end CAD/3D/VFX software is and only viable for huge studios leaving normal people in the dirt.
Emad's decisions mean for the rest of eternity a tool that could have ended up entirely locked behind an Adobe paywall can now be run on any machine you owned and tweaked entirely on your own hardware to work in a way specifically beneficial to your workflow.
I'm an artist and designer too, the fear of how fast these tools can replicate styles and take jobs becomes a lot less scary when I can take advantage of it myself or enhance my workflow with it myself without paying a subscription tax to do so. But if the "art community" can't understand or imagine how bad this situation could have been then I don't know what to tell them, some people just like being screwed over I guess...
>I'm an artist and designer too, the fear of how fast these tools can replicate styles and take jobs becomes a lot less scary when I can take advantage of it myself or enhance my workflow with it myself without paying a subscription tax to do so.
Have you tried to train SD on your artwork? Pretty curious about the results an artist can achieve when embracing this tech.
Yeah I've fine-tuned it on our company product aesthetics and now our product team uses it for rendering and concept work, something that wouldn't be possible using this tech via Midjourney, etc.
My understanding is the art community has two problems:
1) Brand destruction: when SD was new, lots of people put "Greg Rutkowski, trending on artstation" in their prompts in order to get better images. It's possible that Greg Rutkowski being the single most popular example of this means he personally lucked out on this (some reporting suggests so), and the exposure really did boost his career. Do you think everyone else this has happened to was so lucky?
If I image search for "Greg Rutkowski", I see some cool things yes, but I also see this: https://creator.nightcafe.studio/creation/gt4Z0uOIrrmop13OoU...
I suspect that many others have suffered from this association.
2) Substitution: the exact opposite problem.
Now that the image generators are pretty good, why should anyone hire an artist?
This image was generated in 267 milliseconds, for free: https://github.com/BenWheatley/AI-art/commit/d4e0322a30ab508...
That image is not perfect, but it's good enough for people like me, and that by itself is an economic risk to the future employability of that entire segment of the economy.
This really is important and does matter because all the talking heads were all busy confidently saying creative jobs like "artist" and "writer" were safe, and that it was truck drivers and factory workers who needed to re-skill, and thus we as a society have done basically nothing to prepare for or mitigate this economic disruption.
--
I don't know what's coming, not for me, not for anyone.
But I get why they feel scared, and I get why they feel this has taken something from them, even though the specific arguments about copyright and "parroting" that make it into public discussion (Gell-Mann amnesia warning) are often also deeply flawed and unconvincing.
To me, this is a trademark issue in the first case, not a copyright one; and in the second, the same disregard for workers that led to the creation of the actual literal Communist Manifesto.
> damage many feel they've done to the art community.
Boo hoo. This is not the first nor the last democratization of art. First people weren’t starving so many more could afford to become artists, printing presses could mass copy art, world wide shipping lanes moved styles, computer aids then photoshop, now AI. It’s always “been damaging to the art community”.
Now… to steelman the argument, it’s never been lower skill or easier to create your own modification or idea and get it in the style of some artist. In my opinion the low barrier to entry is obviously going to seem unfair - but - this is just going to make physical art more valuable.
If I were a sad-about-ai artist, I would jump in and see how new tools could improve my game.