0: magnet:?xt=urn:btih:3a4a612d75ed088ea542acac52f9f45987488d1c&dn=sd-v1-4.ckpt&tr=udp%3a%2f%2ftracker.openbittorrent.com%3a6969%2fannounce&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337
So a paywall. I agree it's a annoyingly misleading term for the concept and would be happy to hear better alternatives[0], but I haven't found one yet.
0: like eg "passphrase" instead of "password" or "assuming the conclusion" instead of "begging the question" for their respective concepts
Seconded, actually. I do have a bad habit of assuming people already know this.
> Official sources are recommended for this reason.
Very not seconded; see for example comments elsewhere in this thread about untrustworthy sources for popcorn time, and recall that the GP was specifically discussing the risk of Stability AI deciding to kill this.
You mention Popcorn time. I wonder if torrents in general could be a great example of how something like this plays out? Torrenting took the world by storm and had an amazing "product-market fit" for the early internet days. Of course, downloading copyrighted material was always illegal but that didn't stop many.
Over time, legal but paid alternatives rose up: Spotify, iTunes, Netflix. These players found their place in the market by balancing the interest of copyright holders and the needs of users looking for cheap and easy access to entertainment.
Just as Netflix acquired large content libraries, same here. With enough money, large training datasets could be acquired in a legally solid manner.
It's interesting to think where this analogy might fail as well, and how the paths of these technologies could differ. For one, torrenting was mostly for entertainment, and thus impacted B2C first. On the other hand, language models are more so for media _creation_ and the B2B sphere.
Like torrents, you first have to resort to random websites who get randomly taken down as they acquire reputation. If a person takes the face and responsibility for something, he gets litigated into oblivion.
So you get to the point where trustworthy and untrustworthy sources are indistinguishable
. Now what they do is create untrustworthy sources. Like time for popcorn. Sow discord.
Fork several times, create intentionally malwared versions of both the program and the website. Keep kicking off the trustworthy sources of search engines, while magically skipping takedown requests for the less trustworthy websites.
Find ways to break old versions if possible, just to force them to keep moving. (they can make gradio randomly change APIs just to break the old trustworthy versions)
All of this can happen.
Before this gets flagged to oblivion, this is obvious. You just have to recognise that the "regulators" and industry insiders Emad is trying to "shield" you from are enemies and ask yourself, how do I hurt them?
And reasonably technical people have zero issues, as it should be.
Media piracy has been abundant my entire life. It's never slowed down or become inaccessable.
You didn't mention one of the largest (perhaps even the largest) distributor of copyright content (which happens to also be free, for now): YouTube.
You can watch/listen to endless amounts of copyrighted content (and other types of content) on there completely for free, and to say it's tremendously popular would be an understatement.
Google has made it work through ads. Perhaps something like that will happen with image-generating AI.
Tit-for-tat. Regulators and artists don't want this? Okay, include in all open source software licenses that regulators and artists are now barred from using them without payment.
https://www.gnu.org/licenses/license-list.html
> [...] is a nonfree license because it extends the four freedoms only to some kinds of organizations, not to all. Such a restriction in a software license, in the name of any cause whatsoever, imposes too much power over users. Please don't use this license, and we urge you to avoid any software that has been released under it.
https://opensource.org/osd-annotated
> 1. Free Redistribution
> The license shall not restrict any party from selling or giving away the software as a component of an aggregate software distribution containing programs from several different sources. The license shall not require a royalty or other fee for such sale.
> Rationale: By constraining the license to require free redistribution, we eliminate the temptation for licensors to throw away many long-term gains to make short-term gains. If we didn't do this, there would be lots of pressure for cooperators to defect.
I've favoured the MIT licence for what little OSS I've published thus far. But, I'm becoming increasingly concerned that ruthless profit-above-all-else driven companies can include my (benign) work in systems that causes real harm.
Perhaps something like the OpenAI approach to their GPT-3 deal with Microsoft is better. That is, if the work Microsoft do with GPT-3 goes in a direction OpenAI doesn't like, OpenAI reserves the right to veto the work [1].
[1]: https://www.ted.com/talks/the_ted_interview_the_race_to_buil...
It's not an easy question. However, as the creator of the software I guess I feel that my opinion should count in how it's used. As a simplistic example, if in some dystopian timeline my OSS were used to facilitate a holocaust I'd like to be able to do something to halt that. It doesn't matter that the perpetrators feel that what they're doing is right.
The Slaughterbots campaign argued, rightly, I think, that advanced autonomous lethal weapons should be suppressed because they enable unethical uses and unscrupulous actors far more than legitimate defense.
It can't really be seen in isolation from the environment (social, economical, etc) it's going to come into I suppose, but in the real, concrete world we have creating them is not a neutral act, and some of the consequences are reasonably predictable.
>The Software shall be used for Good, not Evil.
AIUI, it was put it mostly as a joke.
https://www.cnet.com/culture/dont-be-evil-google-spurns-no-e...
A terrible idea for a number of reasons (in terms of legal enforceability, unintended side effects, and more). The following two articles do a good job of explaining why such a license really isn't practical:
https://perens.com/2019/09/23/sorry-ms-ehmke-the-hippocratic...
https://www.gnu.org/philosophy/programs-must-not-limit-freed...
Your first assumption is that your inventions are important enough to be of use to “bad people”.
The other is your assumption that you have the objective ability to determine good from bad uses of a benign invention.
I’m increasingly looking for the psychological reasons why these ML models and their outputs cause such an emotional reaction in certain individuals.
For example, the language of opponents of Copilot speaks in absolutes. And when presented with the history of copyright when applied to software the opponents seem to not register that copyright (logically) does not extend to the non-expressive parts of a work.
“In computer programs, concerns for efficiency may limit the possible ways to achieve a particular function, making a particular expression necessary to achieving the idea. In this case, the expression is not protected by copyright."
https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...
This allows for verbatim copies if they are utilitarian in nature!
As for why we should allow verbatim copies of utilitarian features... First, let's preface this with the substantial similarity of the structure, sequence and organization as established in Whelan v. Jaslow which amongst other things says that you cannot merely change the variable names if the expressive structure of the code remains the same. Now let's imagine 10,000 software developers who all implement Dijkstra's algorithm in C and then run it through clang-format. Aside from variable names, isn't it safe to assume that many of the implementations are going to be exactly the same?
Now, this doesn’t mean that GitHub is not in violation of other copyright claims, such as clearly expressive parts like comments and more!
I'm not trying to be contrarian here, I was curious why not and why this isn't a thing already. I'm just more of a programmer guy and less of a lawyer guy
I don't argue that such licenses are bad (though the FSF might), just that they are neither open source nor free.
Also, discriminating like you suggest would make those licenses closed source by definition.
Long live clopen source!
Definitions are not what matters in the end. Why doesn't the viral and restrictive element of GNU's GPL license make the license "non-open"?
"Open source" has a definition (https://opensource.org/osd) and the GPL meets it, because it doesn't prevent derived worries from being distributed under the same license.
We can barely get governments to use open source even today, without restrictions. Hell, we can barely make them manage source code for commercial products they commission and pay for. I've walked into govt shops that were 100% binary dependent to the original software author, which never delivered source code and charged them trough the nose for the basic servicing.
Like it or not, the government and regulators represent us, we need individual accountability but harming the govt. directly harms ourselves firstly. The bureaucrats and the corrupt hardly care.
We'll go thru the same with AI but ultimately it won't be stopped. As long as there's no world wide coordination limiting its impact, AI will continue its course.
The iPod would not have had the impact it had without piracy.
Is it legal? I don't know. I guess they have the fig-leaf of taking down copyrighted content when asked.
Fig-leaf not withstanding, if Google (YouTube's owner) didn't have such deep pockets I'd be amazed they didn't get sued in to oblivion like Napster.
Countries that don't do that will be just as successful in the world marketplace as are countries that don't respect copyright.
Doesn't this ultimately result in local maxima? All the biases get reinforced and all the novelty (things the system hasn't seen/produced yet) goes away.
A tiny example: Dall-E (and SD) both struggled with eye positioning, for example. Wouldn't training a model on their output then reinforce that particular bias of poorly positioning eyes? Now multiply this by every existing quirk in the models.
Irony is the "NSFW" moral concerns, when the media companies put out such negative and filthy content as it is.
What I would value much more is the writing, directing, editing, and acting... and you can't yet get very good quality of any of that through AI yet.
Maybe someday, but not today.
There is a legal machinery that works behind the scenes which we aren't always aware of.
Sure, it's not on the most visited homepages of the world - but it hardly went away. Even on the most visited homepage it's just few clicks away.
LLMs have the unfortunate limitation of being both powerful and lending themselves to centralized control choke-points due to how resource intensive they are to train. Under this paradigm, I fear commercial entities will be able to easily navigate the legal landmines and continually improve while open efforts perpetually lag far behind.
There are many vested interests who want this control for various reasons they justify as: protection from x-risk, keeping it out of the hands of abusers and bullies, economic advantage. Their reasons for want of control are either well intended but wrong-headed or profit-motivated and disingenuous.
Rather than challenging the likes of GPT-3 and Copilot enabling freedom, I fear folks will be forced to send all their videos, pictures, text and code to the servers of Microsoft, Amazon and Google or lose access to advantages as LLMs continue to improve at a rapid clip.
It was hard to accomplish, but you can finetune SD on your computer. They are working on instruction-tuning LLMs as well. In general ML models are not closed boxes inaccessible to us - they can be finetuned, reprompted, you can even average two versions to get a mix of two models. In the last 2 years lots of papers were written on finetuning and prompting, all of them geared towards low resource AI adaptation to new tasks.
- lexical filtering by applying a blacklist of artist names on the original prompt
- perceptual filtering - drop all generated images that look too close to copyrighted images in your training set
- re-captioning based filtering - use a model to generate captions for an image and apply filters on the captions; you can also filter by visual style
- CLIP based filtering where you use embeddings to find nearest neighbours, and if they are copyrighted then you can drop the image
- or train a copyright violation detection model that takes generated images and compares them to images from the original authors
Copyright enforcement struggles are going to be interesting to watch in this decade. But I think it will slowly become irrelevant, because anything can be generate again slightly different until they finally pass the filters.
This is part of the motivation for “foundation models”.
There’s another paradigm called student/teacher models where a randomly initialized model updates it’s weights according to another pretrained model. This could (maybe?) be used to achieve the desired effect of a model that learned in a “clean room”.
Because it's already good enough to have made it's way into many of my workflows.
I do feel that many companies will, ironically, use "ethical" as a pretext to not be open.
I mean this isn't even speculative anymore after what happened with - hilariously named - OpenAI
> that many companies will, ironically, use "ethical" as a pretext to not be open.
Yes, weaponized ethics as sleight of hand for control is a common historical pattern.
This is exactly why I pay $30 per month for MidJourney. The output is just phenomenally better than most of the images coming out of SD, and the UI is much better as well. It's just not worth my time fiddling with SD if the results are so bad in comparison.
If/when SD catches up, I'd jump ship to using it in a heartbeat.
While it'd be difficult to improve upon the model, it might be easy enough to finetune it if needed, and it's certainly worth it to USE it as is.
There is a limited number of models costing 6 digits in dollars in train time and are freely available. There is certainly value in preserving them, in a world of artificial scarcity.
Is it possible to crowdsource AI training with something that looks similar to folding@home?
AI thrives and depends on large amounts of clean, well labeled data.
Large corporations understand this and have hoarded data for a long time now. Some of them have also managed to label this data by millions of people through things like Recaptcha, or just by hiring lots of people to do it.
Open datasets tend to be much smaller and dirtier than small, open projects have access to.
I suppose it would be possible to, over time, collect lots of data and crowd-source some project to clean it up and label it well enough to be useful, then crowd-source the AI model training itself, but it would probably take a long time and by then corporate-owned AI models will already dominate (as they do now with MidJourney, for example, being way better in my experience than Stable Diffusion, but with time the difference will only get starker).
I'd also be concerned with such ostensibly open projects eventually going closed and commercial as IMDB did after getting lots of work by volunteers freely giving their time to writing reviews.
More recently the open LAION data sets have become widely used by both tech giants and independent researchers.
The problem is DL is really sensitive to dirty data, disproportionately so.
At $DAYJOB once we cleaned the dataset, removed a few mislabeled identity/face pairs (very few, about 1 in 1e4) and the metrics goes up a lot.
In fact, generally DL is quite tolerant to label noise, especially using modern training methods such as SSL pretraining.
A fair compromise is any marketplace for clusters with good interconnect but a lot cheaper than the cloud. Tuning distributed training and network transport layer for settings not as homogeneous as the cloud will also help on top of generally good interconnect. Security is a concern.
Building on points raised by pmoriarty, being able to scrape data makes up for lacking labeled data in the era of self-supervised training. IP-hawks are now putting a damper on that option, which is why I worry this might backfire from a freedom perspective.
I even think there will be multiple initiatives like this, and there will be at least 1 big repository that accepts inputs and retrains periodically for anyone who wants the model.
The quality of the output you can get with the models right now have perpetual utility IMO. If you use it to create patterns, backgrounds, or even just for inspiration creations right now, it might be a shame if it didn't progress (depending on your position) but it's fine as-is if you put in the work to compose and refine the raw output.
I wonder if that will continue.
My understanding is that's partially because it currently relies on GPUs, which until relatively recently there was a limited demand for, and the market is basically controlled by a single company.
Will we see cheaper special purpose AI accelerators? Like happened with crypto mining ASICs.
i was actively following torrentfreak at the time and there was genuine excitement with something incredible but that only lasted a week :-(
why do you say they lost the battle? the original team threw in the towel within the week but there are people who have taken the fight
https://github.com/popcorn-official/popcorn-desktop/releases... here, the latest release was on 04 Sep 2022 so it is very much in active development with a lot of people contributing https://github.com/popcorn-official/popcorn-desktop/graphs/c...
so while the original team might not be working on it, like a true free software, the code lives.
Encapsulates it all well I like this statement, total pottery
The blog post says they are worried about the ability to use the model to "use it for illegal purposes and hurting people". I think that they are referring to the ability to create all kinds of compromising pictures (porn) with celebrities, kids, etc. Am I misreading that? They don't mention copyright anywhere.
The conspiracy theorist would say that if you were doing something you shouldn’t, you wouldn’t mention it. Instead, you’d give a more palatable excuse to buy yourself some time while you figure out how to get away (legally) with the thing you shouldn’t be doing.
That's the only possible interpretation, really. AI models algorithmically remix input intellectual property en masse, without any significant amount of human creativity, the only thing copyright law protects. As such, the models themselves are wholly derived works, essentially a compressed and compact representation of the artistic features of the original works.
Legally, a AI model is equivalent to a huge tar.gz of copyrighted thumbnails: very limited fair use applies, only in some countries, and only in certain use contexts that generally don't harm the original author or out-compete them in the market place - the polar opposite of what AI models are.
It reminds me to follow the datahoarder maxim that if you don’t admin then servers, you don’t have the data. So now I clone stuff to a local drive.
https://texaslawreview.org/fair-learning/
Here is a short quote from an IP lawyer:
“In terms of the ingestion of publicly accessible code, Ochoa said, there may be software license violations but that's probably protected by fair use. While there hasn't been a lot of litigation about that, a number of scholars have taken that position and he said he's inclined to agree.”
https://www.theregister.com/2022/10/19/github_copilot_copyri...
Incorrect. In the absence of new legislation case law is how these things get worked out, but new statutes could be passed and could void prior case law.
https://www.copyright.gov/title17/title17.pdf
Let me expand on this a bit... if you read through the text (feel free to search for the below terms) of the current laws around copyright you will notice that there is no discussion of:
https://en.wikipedia.org/wiki/Structure,_sequence_and_organi...
https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...
https://en.wikipedia.org/wiki/Idea–expression_distinction
These legal doctrines are the result of the details are hashed out in case law.
This is distinct from countries that use civil law jurisprudence. Common law jurisprudence relies heavily on case law.
Funny way to admit that your original statement was flat-out wrong.
There is no separate "civil law jurisprudence" and "common law jurisprudence". Common law by itself is by definition not jurisprudence. Civil law combines aspects of statutory and common law (as one form of precedent) into a single system. Some aspects of common law even make their way into criminal law. Instead of just Googling for buzzwords, learn what they mean before you try to bluster your way through an argument with them.
The law moves slow. Even were that to happen eventually, the laws will very likely be challenged in courts, and those will take a while to be resolved.
Finally, even if the US outlaws this, there'll be plenty of other countries where it'll be legal. There's plenty of infringement of US copyright in China, for instance. The same is likely to happen in regards to AI that's illegal in the US but legal elsewhere.
By the time the law catches up, model creation may become so easy to create by individuals instead of just by deep-pocketed corporations, it may be practically impossible to stop.
But sure, let's talk about outputs as well. From the second source we can see this from Tyler Ochoa:
"If there's only one good way to do it, OK, then that's probably not eligible for copyright. But chances are that there's just a lot of code in [the training data] that has used the same open source solution, and that the output is going to look very similar to that. And that's just copying."
I have seen some probable copyright violations from the output of Copilot, such as comments and some certain structural similarities that might be protected, although it is hard to say. But focus on the first part of what Mr. Ochoa is saying here, which is also laid out in this quote:
“In computer programs, concerns for efficiency may limit the possible ways to achieve a particular function, making a particular expression necessary to achieving the idea. In this case, the expression is not protected by copyright."
https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...
This allows for verbatim copies if they are utilitarian in nature!
As for why we should allow verbatim copies of utilitarian features... First, let's preface this with the substantial similarity of the structure, sequence and organization as established in Whelan v. Jaslow which amongst other things says that you cannot merely change the variable names if the expressive structure of the code remains the same. Now let's imagine 10,000 software developers who all implement Dijkstra's algorithm in C and then run it through clang-format. Aside from variable names, isn't it safe to assume that many of the implementations are going to be exactly the same?
As for why it was carefully selected... more often than not when I bring these things up people who feel upset about Copilot go off to cherry-pick some random quote out of context in order to support their upset feelings. Therefore I'm highlighting the important parts as to help people look beyond their upset feelings.
This is a complicated and nuanced matter. Attempting to channel everything through the lens of "this makes me personally feel bad and must be completely wrong" does not help the discourse. It may make you popular to a certain crowd but it might be unpopular to the public at large and it might also be incoherent from a legal standpoint, akin to bashing your head against a wall at a weekly meetup of the local heads-bashing-against-walls club.
There is plenty of room for discussion on what constitutes not only the legal interpretation of fair use and the idea/expression dichotomy but also the bigger picture. The knife always cuts both ways. Would it be acceptable to the open-source community if Microsoft could stop anyone from publishing Dijkstra's algorithm in C# because they wrote it first?
That's a very petulant way to defend cherry picking. I wasn't asking you to support one particular view; in fact that's the problem I was identifying. Your sources presented a balanced view, which you misrepresented by citing only the part that supported your own conclusion.
> focus on the first part
No, because the second part matters too. Here's Lemley and Casey again (emphasis mine):
<<<some purposes—say, ... a translation program that produces a translation of an entire copyrighted work—seem more substitutive than transformative, so that if they run afoul of the ever-broadening definition of similarity in music, fair use is unlikely to save them.>>>
Or the Register:
<<<"I actually think there's a decent chance there is a good copyright claim," said Tyler Ochoa ... the functional nature of the code means that reproducing it in a suggestion may not be seen as particularly transformative, which is one of the criteria for determining fair use>>>
Those are your own sources undermining - if not outright contradicting - your one-sided interpretation.
The limitation to market harms in the four-factor test for fair use should not be considered permanent. Law is, after all, a social construct. There's ample precedent for considering harms to the commons, to communities, and so on in other areas of law. Also, there might indeed be market harms. If a company open-sources some of their code but also hopes to profit by selling it in pre-packaged form or as a service, then AI-washed copying could constitute harm in even the most market-myopic terms. The "transformative" test is also pretty suspect in the context of AI-assisted copying, but this is getting long enough so I'll not go down that rabbit hole just yet.
> <verbosity about "utilitarian" copies which are not the issue here>
Enjoy your red herrings. I don't share your taste for them.
When Tyler Ochoa is saying that there is a decent chance of a copyright claim he is specifically talking about the output of the model.
Here is the full quote:
"[Machine learning] systems should generally be able to use databases for training, whether or not the contents of that database are copyrighted," they wrote, adding that copyright law isn't the right tool to regulate abuses.
But when it comes to the output of these models – the code suggestions automatically made by the likes of Copilot – the potential for the copyright claim proposed by Butterick looks stronger.
"I actually think there's a decent chance there is a good copyright claim," said Tyler Ochoa, a professor in the law department at Santa Clara University in California, in a phone interview with The Register.
The use of the word "but" marks the transition from a discussion around the model itself to the outputs of the model.
Oh, you mean the very first thing I had to explain to you at the beginning of this exchange because you seemed to be ignoring it? Very little of these discussions has been about the models. Most of the discussion is about the outputs, and there the fair-use cases is - as Lemley/Casey and Ochoa both concede - much weaker.
But by all means keep going on about feelings. We can all tell it's not others' feelings that are being hurt by mere disagreement.
IANAL but I would take any opinions on this right now with a huge grain of salt and treat them more as advocacy than actual predictions of any legal outcomes.
Whether there is a good case for it being considered fair use doesn't matter at all until its actually litigated and historically the result with fair use in relation to new technologies has always been a crapshoot.
The result could easily be affected by the actual cases that get litigated, and one well chosen lawsuit where machine learning software is shown to produce output that's too close to the material it was trained on could result in a completely different outcome.
Once patent trolls figure out there's gold in these hills, it's downhill at that point.
Get 200 interested people backing up 1 TB each and you have your 200 TB backup.
With redundancy and error correction data added to the mix, you should be able to lose a certain percentage of participants and still have access to the full, error-free backup.
This should be next on my list since my current project depends on SD model and having data backed up gives me confidence that I can get rid of all their stuff if needed.
You don't even need to store it all at once on your computer: stream it and generate checksums on the fly. Then distribute the torrent, and seed sections at a time. It can also be distributed on IPFS.
I've seen a lot of torrents being used for distributing neural network (mostly stable diffusion forks).
Or multiple IPFS CIDs. I think you can have a "directory" (CID) that contains multiple CIDs, and only need the content hashes to build it.
You can also publish multiple CIDs and ask people to seed random ones; that's how Libgen does it (and is similar to the multiple torrents concept).
The same file can be used to seed both torrents and IPFS.
Just feels absurd to me because how is this different from any Human artist who you could equally say was "trained" on copyrighted material.
>Get it now when there are trustworthy sources. Once these kinds of things go underground, it gets much harder to get a trustworthy version.
People have already reverse engineered most text2image models and given enough hardware can train their own. There is no need for this hysterical take. As long as the internet exists you will be able to train these models.
Personally, I think commercial use cases should get license agreements from the authors for their training data, but I think non-commercial exemptions to advance the field of AI makes sense.
Irregardless of what I think though, the UK has set an international precedent, and the EU is apparently discussing about possibly extending it to commercial use cases as well. So there's that.
Where copyright may be applicable is when the models reproduce original art without modification that a reasonable person wouldn’t know the difference.
Something that would make any non-trivial model a legal nightmare.
Somewhat similar gut feeling to when popcorn time was released, although it might not be exactly the same.
While I really wish I'm wrong, my gut tells me that broadly trained machine learning models available to the general public won't last and that intellectual property hawks are going to one day cancel and remove these models and code from all convenient access channels.
That somehow international legislation will converge on the strictest possible interpretation of intellectual property, and those models will become illegal by the mere fact they were trained on copyrighted material.
So reminder to everyone: Download! Get it and use it before they try to close the Stable doors after the horses Diffused. Do not be fooled by the illusion that just because it's open source it will be there forever! Popcorn time lost a similar battle.
Get it now when there are trustworthy sources. Once these kinds of things go underground, it gets much harder to get a trustworthy version.