site banner
naraburns  nihil supernum  1yr ago (text post)   35412 thread views

Culture War Roundup for the week of October 3, 2022

This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.

Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.

We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:

  • Shaming.

  • Attempting to 'build consensus' or enforce ideological conformity.

  • Making sweeping generalizations to vilify a group you dislike.

  • Recruiting for a cause.

  • Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.

In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:

  • Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.

  • Be as precise and charitable as you can. Don't paraphrase unflatteringly.

  • Don't imply that someone said something they did not say, even if you think it follows from what they said.

  • Write like everyone is reading and you want them to be included in the discussion.

On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.

24
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

This week's revolutionary AI advance:

Imagen Video

It's not really revolutionary, as people have been pointing out this is the obvious next step for ages months now. But it still is a milestone worth noting.

As for this:

While our internal testing suggest much of explicit and violent content can be filtered out, there still exists social biases and stereotypes which are challenging to detect and filter. We have decided not to release the Imagen Video model or its source code until these concerns are mitigated.

Google's made a habit of this. They announce an amazing advance, and then say no one can have access to it because it can be used for Evil. No matter: Stable Diffusion will have something comparable out in a couple months.

ETA:

Actually, this out of DeepMind might be the bigger advance today, if less flashy:

Press: Discovering Novel Algorithms with AlphaTensor

Paper: Discovering faster matrix multiplication algorithms with reinforcement learning

My greatest fear for AI content generation is it being dominated by woke megacorps, with independent creators permanently locked out of contributing to culture. It looks like Google is investing heavily in that dystopia.

Novelai and stable diffusion being mostly uncensored has been a big white pill so far, but it feels like the shoe is about to drop.

Everyone rapidly generating nonsense with AI is not my ideal world

It is mine. The internet should be nonsense.

I think the day is approaching where you won’t be able to tell if anyone you’re interacting with online is even real or not, and I hope when that happens I’ll just be fed up enough to put it all aside and just stop even coming online to see what people have to say anymore.

If you can't tell if anyone you're interacting with online is real, that means that the best content online can be mass produced by AI. That would be awesome. I could spend all day watching the top-rated 0.0000001% of Youtube videos on the exact subject I want to see. I could read a thousand books that are equally as good as the best books I've ever read, in the exact genre I want to read.

If you haven't spent a few hours playing with Stable Diffusion, I highly recommend it. It's like a whole new world is opening up.

"A political drama about the Nixon administration in the style of Shakespeare," could be a click away. "A gorey superhero deconstruction like Invincible or The Boys, as written by James Joyce, as an animated film in the style of a Disney Movie, with watercolour art style," could be something you can watch just by typing that prompt into a computer. "An Isekai anime with writing by J RR Tolkien, music by Metallica, and art by Gerald Scarfe" could be yours too. In fact, if you generate it and then delete it, you could be the only human being ever to watch it.

I'd love to do stuff like that but I'm not sure we'd have the computing power before singularity.

Stable Diffusion runs decently on my 1060 6GB since I'm patient. But text is impossible for even those with a 4090TI! You need serious industrial-grade hardware for GPT-3. Fusing GPT-3 to better image-gen, turning it into a video and letting it have the memory for an entire story would require absolutely gargantuan computing power. Only big companies could afford it and you'd still have to sift through the bad prompts.

If google causes people to not take the internet seriously I think they will deserve a Nobel peace prize.

It’s a hopeful future for sure ☺️

in accelerationist Bane voice Our plan is proceeding as expected.

I don't know that your definition of "nonsense" is universal enough to apply here, but what do you think humans do right now WITHOUT AI?

Good point, we’re all already just rapidly generating nonsense 😂

Reading this with its assorted replies along with @Primaprimaprima's post earlier in the week leaves me feeling like I have a fundamentally different understanding of art, and experience it in a vastly different way from the majority of commenters here. I'm not sure what to make of your stated concern because I don't see how it even relates.

While I grant that I may be falling for the "passing tranny fallacy" (IE transwomen who pass don't get recognized as transwomen) pretty much every example of AI generated text and images I've encountered has struck me as painfully obvious. Whether it's GPT-x or DALL-y it's all very obviously not human, not even close, and thus I find myself wondering what all the furor is about.

Am I really an outlier in my ability to notice that randomly strung together words, even if they are grammatically correct, aren't conveying any meaning? Ditto the visual medium, I've gotten used to cheap sci-fi book covers having nothing to do with the plot, but am I really the only guy who's noticed?

Seems to me that you're painting in apocalyptic terms what to me looks like just another Thursday and that raise the question of which, if either, of us is actually out of line.

While I grant that I may be falling for the "passing tranny fallacy" (IE transwomen who pass don't get recognized as transwomen) pretty much every example of AI generated text and images I've encountered has struck me as painfully obvious.

At least for images, I think they can be pretty respectable in at least some spheres, even if they suffer in others. Now, some of that's related to the limits of the genre I'm interested in -- there's furry artists that do really fancy stuff, but it's not the majority -- but it's at least plausibly human-genned, even in its errors, for some limited range of works. There are a few tells (trivially, output of vanilla SD does best with square images and lower resolutions while artists prefer to not, there's some prompt bleeding, so on), but they're not present in all works and very far from obvious.

I don't think this is as apocalyptic, because there are places where that's not the case yet, and where it may be years (possibly the better part of a decade) before that changes. But I don't think it's another Thursday, either.

While I grant that I may be falling for the "passing tranny fallacy" (IE transwomen who pass don't get recognized as transwomen) pretty much every example of AI generated text and images I've encountered has struck me as painfully obvious. Whether it's GPT-x or DALL-y it's all very obviously not human, not even close, and thus I find myself wondering what all the furor is about.

First of all, I think sans some sort of blinded test, the "passing tranny fallacy" (or the "toupee fallacy" as I've more commonly seen it called) should be the default conclusion.

However, I do suspect that even if you did do a blinded test, you would find that you could tell better than chance. Presuming that you would, I think the furor is about the potential future, not the present. If we compare what AI generated images looked like in October 2021 and October 2022, the difference is stark. We can't extrapolate directly to October 2023 or 2030, but we do know with pretty good certainty that it will be no worse, and there's good reason to suspect that it will be better, and by a significant amount.

Which then raises the question to me, what do you mean by "not even close?" "Close [to human-made art]" is a subjective measure, of course, but is your opinion that AI art is currently just so far away from human-made art that the idea of it getting "close enough" to substitute human-made art in many significant contexts within the next 5-10 years is science fiction?

I don't really understand. The images that started coming out this week are often indistinguishable from human made ones. I expected that would take years from where it was, not months.

No matter what happens next, AI gen is going to be at least as important for content generation as, say, Photoshop. And of course restricting an important tool to megacorps is going to silence people.

The images that started coming out this week are often indistinguishable from human made ones.

Do you have specific examples in mind that you can link? because the ones that have been getting linked here on theMotte certainly aren't.

Here is one (yes, twitter, boo hiss etc):

https://twitter.com/GoranGligovic/status/1577560899883307009

Now, don't get me wrong, an AI stamping out what appears to be the most generic fantasy cover imaginable is a low bar to clear. But if you asked me without context, I would have assumed a human stamped this out instead. I am of the opinion AI will wipe out the mediocrities of art, actual talent will be just fine. Not coincidentally, large swathes of twitter are losing their minds.

Ok that is pretty impressive, there's still something distinctly "off" about it, but it is not obviously AI generated at first glance.

To be fair, by the standards of modern art that doesn't seem too far off the mark.

Eh, I can't claim that this joke is an original. https://stonetoss.com/comic/artificial/ .

...pretty much every example of AI generated text and images I've encountered has struck me as painfully obvious.

Am I really an outlier in my ability to notice that randomly strung together words, even if they are grammatically correct, aren't conveying any meaning?

Yes. You are an outlier. A majority of people are consistently failing the reverse turing test for text and images. Literally everyone will at least have some of their time wasted reading the first couple sentences of AI generated text before realizing something is off. Nobody is immune to info-chaff.

Nobody is immune to info-chaff.

While I'll grant that this is true in the sense that it's always possible to tank the signal to noise ratio by cranking up the noise, I'm not following your reasoning here. What do you think telemarketing calls and internet spam are? and what is it that makes AI generated media different?

Telemarketing and internet spam are so predictable that we generally don't have to process it to ignore it. A script is enough in most cases. AI comments show up where you expect human content. We don't have any automated way to differentiate between AI and human yet, and if we did, AI makers would just use it to become more humanlike.

Also, the less intelligent half of the population falls for telemarketing, internet spam, media manipulation and botting all the time and we all suffer the consequences of this.

Telemarketing and internet spam are so predictable that we generally don't have to process it to ignore it.

Ditto most AI generated media. Which is kind of my point.

Nobody is immune to info-chaff.

The internet is already full of info-chaff (otherwise known as spam) and things are, for the most part, working fine. If AI is going to change that then it needs to do a lot better than fooling people for a couple of sentences. Spammers can already do better than that.

pretty much every example of AI generated text and images I’ve encountered has struck me as painfully obvious.

I do understand what you’re saying - some people have rather low standards for being impressed. I consider myself to be above-average in my ability to identify AI-generated images. But given how quickly the quality of generated images is improving, even I can’t just brush them all off as “eh, uncanny valley, it’s not there yet”.

Have you seen what’s getting posted to AIbooru? The image quality is astounding. There are often still tells if you look closely, but for a lot of these, I don’t think it’s fair to say that they’re “obviously” AI-generated (independent of your ability to recognize, say, styles that are common to certain models, which seems to be a separate issue).

I’m with you on AI-generated text never being very impressive though.

It's kind of common courtesy to mention NSFW when you're dropping a NSFW link. Don't make us make it a rule.

If you're threatening to make it a rule if it is not followed, it is already a rule.

Do you have a particular reason why you think it's bad to politely request that people warn about NSFW links, or are you just in a mood again?

The rules are organic and it's probably not an official rule because none of us felt like we have to warn everyone who might post a not-entirely-SFW link and we don't want to argue with Nybblers about whether or not something is NSFW and if we only decided something was NSFW because we are personally and egregiously converged and biased, but I do think it would be courteous to warn if you're gonna drop a link to AI-generated hentai.

Why wait for a wave of people to link NSFW without warning before making it a rule? If one person did it and its objectionable, isn't that enough to just make it a rule?

I suppose that "having a rule for NSFW warnings in the sidebar" contributes to a mood or culture that TheMotte doesn't want (i.e. is a bad look), which is why it seems your ideal world is where no such rule is listed on the sidebar, but also people somehow know not to post NSFW content without warning.

Maybe I'm missing your point, but I'd imagine all the people worrying about AI art are not thinking of GPT-3 and stable diffusion. They are thinking of GPT-4/5 and stable diffusion 2. With the rates at which the models have been improving in recent years, it hardly seems fanciful that the next generation of these language models and image generators will be human or super-human level, not sub-human.

pretty much every example of AI generated text and images I've encountered has struck me as painfully obvious.

You're not alone. Currently "AI art" competes with the bottom tier of artists, if at that. Until AI art can become massively more consistent (iow, not completely change the whole image for every variation) and controlled, it's basically just clip art for situations where you need an image and the contents barely matter.

Until AI art can become massively more consistent (iow, not completely change the whole image for every variation)

To some extent, that's present now. At the most trivial level, holding seed and settings constant, while making small changes to a prompt, can maintain layout and theme. At extremes, people have managed to make short animated gifs by transitioning from one seed to another, or from one prompt weighting to another.

At more serious levels, artists can use img2img and have greater control of the final appearance's layout, or 'rerolling' individual parts of an image or apply a prompt only to that section of an image using inpainting. There are some limits to these techniques -- getting readable text out is basically random and very rare chance, even conditional prompting still makes it hard to apply adjectives (and sometimes even nouns) to only one part of an image -- but there is more than filling in text and hitting the random button a hundred times.

Now, there's a difference from consistency to control.

While I grant that I may be falling for the "passing tranny fallacy"

I keep pointing out when people mention this: You can see a distribution, and notice that the distribution gets sparser when you get towards "better at passing, but where I can still notice them". You can then deduce that there aren't many who are so good that you won't notice them at all.

Even if we assume a uniform distribution, wouldn't it still be the case that the distribution of those you notice would still get sparser when you get towards "better at passing, but where I can still notice them?"

Like, imagine a toy model where there are 300 total MTF trans people, with 100 who are bad at passing, 100 who are better, and 100 who pass perfectly. You'll always notice the 1st 100, you'll notice the middle 100 about 50% of the time, and you'll never notice the last 100. By your observations, then, you'll notice 100 poorly-passing MTF and 50 better-passing MTF. Since the middle group is better-passing, you'll also notice that those 50 are qualitatively different - i.e. noticeably closer to actually passing - from the 100 in the 1st group, besides just noticing them only 50% of the time. But it would be wrong to conclude from this that then it follows that there are even fewer people in the 3rd group.

That's true if "better at passing" means "more often doesn't get noticed", but it's not true if "better at passing" means "still noticed, but doesn't look quite as bad". I was suggesting the latter scenario.

OK, I had to think about this for a while, but I think I follow. You're really only looking at those who are so far away from passing that they would fail to pass 100 times out of 100, and looking at the distribution of how badly they fail at passing? I think that makes sense, and you're probably correct in your inference.

It is probably a good heuristic but you are assuming the shape of the distribution and still then it could go wrong. For example let's say that there is a latent variable that measures the degree of femaleness that was uniformly distributed over trans and that your likelihood of clocking them decreases monotonically towards high femaleness then, depending on your clocking function, you could get the same picture of a sparser distribution towards high femaleness without it being so.

My greatest fear for AI content generation is it being dominated by woke megacorps, with independent creators permanently locked out of contributing to culture.

I'm confused by this statement. How does it "lock independent creators out of contributing to culture" if only megacorps have AI tools?

Is it because you think it's impossible to produce content without using AI? That's obviously false. People have produced petabytes of content without using AI.

Is it because you think megacorps will flood the world with AI-generated content and independent creators will get drowned out and won't get noticed? That concern also doesn't really make sense. If you're worried about getting drowned out by just a few megacorps, then giving everyone access to the tools will just allow everyone to flood the world with even more content, exacerbating the problem.

EDIT: Not sure why this is getting downvoted. This isn’t supposed to be a gotcha. I really don’t understand the concern here. I mean, I assume the intuitive concern is “the megacorps have all the AI tools and I don’t and that’s not fair”, I just think that under closer examination, you can’t say that you’re “locked out of contributing to culture” in that situation.

The reasoning seems pretty simple to me. AI is an extremely powerful tool by which greatly reduces the effort involved in creating high quality art. Without it, people won't be literally "locked out," but the much higher barrier to entry and costs would mean that it would be immensely harder for independent creators to contribute high quality art at a meaningful level.

An analogy might be if only a select few megacorps had access to cameras. Independent creators could still learn to draw photorealistic drawings or perhaps hack together some sort of crude camera system, but compared to megacorps that can just have their employees snap a 20 Megapixel photo with their DSLRs, the quality and quantity of content independent creators could produce would be lower by a tremendous amount. In such a universe, the culture of photography and more generally photorealistic art would likely be dominated by megacorps.

FYI: I wasn't going to downvote your comment until I got to the EDIT part. Then I did, because any post that makes any references about its (or its author's) downvote status is one that I always downvote, regardless of the rest of the contents.

If you're worried about getting drowned out by just a few megacorps, then giving everyone access to the tools will just allow everyone to flood the world with even more content, exacerbating the problem.

Does not work that way. Content reach follows a power law in which the tail end gets nothing or little. The median is zero, the mean possibly skewed by outliers. Ranking algorithms tend to promote that which is already popular, making the problem worse.

I agree with what you said, but I don’t see how it’s relevant to the question of whether one is “locked out of contributing to culture” or not.

Agreed, but there's cause for optimism: MosaicML shows that large models can be a whole lot cheaper to train than previously believed. Half a million for a GPT-3 equivalent is... well, you know. It's not limited to GPT type systems either.

Also, the Chinese are being pretty nice as of late.

I'll admit this tweet made me nervous for a while. Mysteriously falling down stairs and quietly going off stage «to take care of personal problems and improve health» is something I 100% expect to happen to disturbers of the General Arc Of History. Emad ain't it yet, it seems.

But even if all of those failure modes are avoided, corporations still have the advantage: PR, established distribution channels, and the general problem of there being way, way too much content being produced. People will go watch the latest Marvel capeshit to keep in touch with each other, and in addition to that, they'll consoom a hundred medium-scale indie creations and fifteen thousand TikToks a year. Each one's own unique set.

Even if the total fraction of the indie content increases, it may only contribute to the relative dominance of the mainstream, astroturfed, censored culture.

Yeah I expect within two years we will see news articles cataloguing white supremacist rhetoric and imagery in ai generated art, which will necessitate handing over all access of them to the elites. And if somehow there is no white supremacist rhetoric or imagery to be found, it will be created.

handed to employer. I can see people being framed with this technology.

CP is my bet for the excuse. Which is going to come organically as soon as any of those remaining guys with big collections start training models on identifiable girls. Assuming there are any of those guys left who weren't rounded up when tor was torpedoed (... Tor-depedoed?)

Doesn't even need that. The furry AI-gen Discord I've been seeing has banned img2img starting from real people, including the poster furrifying themselves, out of (very reasonable!*) concerns about what happens if someone underage applies a model to a clothed picture of themselves. Not a can of worms they want to open, or be within a hundred miles of anyone opening.

Wait, they're letting minors into the discord for a porn generation program, or am I misunderstanding?

It's so weird how on edge people get about this stuff these days. I guess it's the one remaining taboo after normalizing anal vore ferrets or whatever. When I was a boy furries would offer me a kids' discount on the "draw you a fursona for nudes" payment plan, and the community still felt less sketchy than it does now.

No, there's a very hard "you must be 18 years old to be in here" rule, although in practice it's pretty hard to enforce. There's also a separate rule that no matter what your age is, they don't want img2img-based pictures, because someone sneaking into an adult movie theatre is bad but starring in a role is much much worse, and that's something they want to enforce at pretty high cost.

Uh, "draw you a nude fursona" or "draw you a fursona in exchange for your nudes"?

Wait, they're letting minors into the discord for a porn generation program, or am I misunderstanding?

Can't keep them out unless you want to ID check everyone, and that would be a huge hassle even if people went along with it.

The latter. Good thing I didn't have a digital camera as a kid, or I'd have been morally and legally responsible for the creation of a lot of furry art. The guilt would be crushing.

The biggest creator in podcasts is anti-woke and is employed by a company that has stood up to the woke, that being Rogan and Spotify. Same for Elon and Twitter (although twitter is woke). It does not have to lead to a dystopian end. So the anti-woke have huge audiences, assuming they can find a platform like with Rogan and Spotify. SD is just a technology, not a platform. Even if google wanted to censor it, there is nothing stopping a competing team from copying /replicating the core technology elsewhere...like Ethereum vs. Bitcoin (assuming it's open source, which Tensor Flow is).

Joe Rogan is no longer the biggest creator in podcasting. He stopped making podcasts 2 years ago.

  • -17

Joe Rogan is no longer the biggest creator in podcasting. He stopped making podcasts 2 years ago.

Um what?

He might have stopped uploading to youtube after all the hullabaloo over smoking weed with guests on camera but he didn't stop making podcasts

Spotify paid Rogan $200 million to stop releasing podcasts and instead release his content in their walled garden, which doesn't have an RSS feed. I used to be a regular listener, but I haven't heard Rogan's voice since the switch. 2 years on, and I'm still salty about it.

Uhh. A 2 hour 52 minute podcast featuring Jann Wenner came out today...

Could you link to the RSS feed?

Spotify Only

https://open.spotify.com/show/4rOoJ6Egrf8K2IrywzwOMk

Then it's not a podcast. It's some streaming service.

  • -17

Its a podcast on a platform...

I'm fairly sure you can get it elsewhere as well, its just less easy than using spotify. Is The Lion King not a movie because you only get it from Disney?

More comments

It is telling that Google's practical use of AI (from what I've experienced) is always a downgrade compared to nothing at all. Searches now use AI to try to gauge what you really want from a search, but this makes searching for specific strings of text worse than it was a decade ago. And what they've done with YouTube is a travesty; half the time if I reply to a comment with a paragraph of text their malicious/harmful/offensive(?) text AI detector automatically deletes it because (I assume) it doesn't like some combination of words I used. I don't trust Google with employing any kind of AI.

From a profit-maximizing perspective, Google should push for an extremely regulated environment for generative models, to protect us from the fake news/deep fake revenge porn/violence/Russian bots/racism/sexism/copyright violation. Put up large barriers to entry, have the government restrict access to training data and maybe even GPUs, and then claim all the profits for themselves.

There is an easier solution to this, AIs can be trained by anyone with enough computing power and training AI isn't that expensive. AI is actually fairly democratic as everyone can make their own. Once the cat is out of the box it isn't hard for everyone to get it. Information spreads naturally. The big risk is big corporations and governments access to data to use the AI on. That can give them a tremendous advantage.

Tuning is relatively cheap, but initial training is (currently) expensive. The furry StableDiffusion tweaks probably cost 50-400 USD depending on vendor and management, but the initial StableDiffusion model they're based on reflects ~300k USD at official prices (although probably got at least some bulk discounting).

Some of that'll go down as GPU prices decrease and newer equipment becomes available, but there are some costs for bandwidth and energy that are slower to change. This might go from 'old condo' to 'new car', but it's not likely to go to 'vacation' or 'a couple weeks' savings' for a few years, maybe even the better part of a decade, without dramatic changes to the underlying code.

For data, it varies more. LAOIN's a lot of bandwidth, curation, and drive space, but it's... actually not that incredible for a single (if slightly nuts) person. Other data sources, probably less so, either due to scale (eg video), to availability (eg privacy), or to more esoteric causes (AI music is a legal clusterfuck).

but the initial StableDiffusion model they're based on reflects ~300k USD at official prices (although probably got at least some bulk discounting).

I mean... this is cheap as hell in the scheme of things. It means you only need one startup with a medium sized seed round who sees a strategic advantage in commoditizing that model, and presto, it'll be trained and released. In fact that's exactly the story behind Stable Diffusion.

The reason we don't see a lot of open source models yet is... well, actually, we do see a lot of open source models. GPT-2 is publicly available, Facebook released a large language model roughly equivalent to GPT-3, and the Eleuther crowd also trained and released a large language model. OpenAI just released an open-source speech-to-text model, they already released CLIP as open source (which powers Stable Diffusion and Craiyon among others), StyleGANs 1, 2, 2-ADA and 3 are all publicly available and open source, etc. These models are just a year or two behind the current research papers. Which is about how long it probably takes to reproduce a research paper. Some of them are even better than that, even cutting edge -- like Nvidia's StyleGANs when they were released, like OpenAI's Whisper, like Nvidia's new Get3D.

Yeah, that's fair. I do think it's meaningful if it requires a startup with a medium-sized seed round (or someone willing to risk their retirement), rather than a slightly nutty hobbyist or enthuisiast, but it's not a FAANG-only thing, at least at a lot of common levels.

Well, there are a fair number of wealthy machine learning hobbyists out there. None of them have actually funded this type of thing to date as far as I know, but I could totally imagine some centimillionaire setting up a few-million-dollar charity to just train models and release them based on research as it emerges.

AI music is a legal clusterfuck

Why, exactly, is it more of a legal clusterfuck than AI art?

Copyright's a mess in general, but the de minimis doctrine has been more heavily tested for sampling than for art collage, and while nowhere near the power it had at its height, the RIAA is far more aggressive than its visual art equivalents.

I assume this implicitly includes the infamous "Blurred Lines" case?

This might go from 'old condo' to 'new car', but it's not likely to go to 'vacation' or 'a couple weeks' savings' for a few years, maybe even the better part of a decade, without dramatic changes to the underlying code.

Completely naïve question: Would it be plausible to rig up something distributed, like the seti@home in days of yore or (shudder) crypto-farming?

To a limited extent. Several training tools (eg W&B) have built-in distributed training capabilities, although these are generally intended for local networks. There are some tradeoffs, though. Even small datasets are 100+GB (eg, the 200k images uses to tune the furry branch) and LAION is 80TB for the curated data, plus a bit more for tag info. You're not going to distribute that full set to every volunteer (might not even train on it!), but it's a scope of the bandwidth costs. Synchronization at that epoch size isn't hugely expensive, but it does slow you down and/or waste power depending on approach.

Unfortunately, the biggest problem is that models have minimum VRAM requirements to run even at a batch size of 1, and these amounts are pretty high at the cutting edge. The original CompVis version of Stable Diffusion required 20+GB of VRAM to train, and this largely limited it to 10k USD or higher specialist 'tensor core' gpus, which largely meant there'd be no @home to distribute to. There's some wiggle room here related to how you code the training, what level of precision you use, and how some averaging and back-propagation happens, and I've heard people suspect they might be able to get full training of current StableDiffusion around 8GB (right now, only textual inversion and tuning is implemented at that range, but the optimization should generalize), albeit at large CPU-RAM and small-but-significant performance costs. Which gets to some consumer-grade GPUs, but not a ton. It's possible people would come up with better optimizations than even that were there no alternative, but I'm skeptical that there'll be the demand now, between Google Colab and nVidia 3090s being available.

And that amount scales both with parameter count and training image resolution. It's suspected that at least part of the better output quality from NovelAI comes from their ability to train on uncropped data, rather than just 512px by 512px cropped or downscaled images, but this bloats run requirements out further.

Enthusiasts are unlikely to want to make huge models anyway since inference (ie, running the model) has similar-if-smaller VRAM requirements, but at least for image generation it looks like the minimum sweet spot is at least 6GB runtime inference.

Other clever stuff may run into similar problems... there's a fascinating 2d-3d analysis package at GET3D, but in addition to limits on accessing the pretrained model, probably requires all 16GB to run or train at any speed. There's probably some unexplored low-hanging fruit, but there's also probably a lot of clever-but-inaccessible stuff.

Does FSDP help at all here? My very naive understanding is that its approach allows sharding of the model parameters so that they don't have to all fit into VRAM, though I wouldn't be surprised if it couldn't scale down to arbitrarily low VRAM or scale up to arbitrary numbers of parameters. Perhaps a similar strategy could be used for wide scale distributed learning on consumer hardware.

I believe so, but I've not looked too closely at that tech to know what its limits are. From a quick glance, it seems likely that there would be some CPU-RAM, performance, and synchronization costs. But it likely could lower the floor.

Even if we assume the high-end of your range, and say that for the foreseeable future training a near-state-of-the-art deep learning model from scratch will cost around half a million dollars, that's still cheap enough to be considered fairly democratic. A lot of people and organisations have that sort of money, many of which exist outside of the Cathedral. And as you say, you can do a lot by tuning an existing model, which is feasible for hobbyists.

I think for the sort of controls you're worried about, it's not just a matter of who can afford to buy it, but also who can afford to sell it. Not just that there's a few limited companies doing this stuff, but in the sense that if you come up to the sorta companies that have and resell these resources, they demonstrably will start poking around at what you're doing, how you're paying them, and what you're doing.

((Not... uh, very effectively in an anti-fraud sense, given Amazon. But very effectively in a not-doing-things-they-don't-like, given Amazon.))

Eventually that stops being a problem as used past-generation tensor core GPUs trickle out into the used market (uh, assuming ITAR doesn't get involved), or as resellers are able to more heavily obscure stuff at larger scales, or as the relative scales decrease due to performance and efficiency gains.

But it is worth keeping in mind as a limit to the democratization of the space.

It's also possible that manufacturers could nerf GPUs for the purpose of ML except for customers with whom they have a special relationship. See e.g. the rate limiting NVIDIA did for crypto mining while still selling a higher priced card without the nerfing.

Yeah, that's another risk. It didn't work for the anti-mining stuff, but given politics and economics around ML that may have stronger incentives.

Crippling GPUs works very well in one context I've seen: FP64. Games don't use it so manufacturers don't get dinged for having lousy performance with it, and engineers/mathematicians/scientists won't flinch at paying through the nose for "professional" GPGPU cards, so with a few exceptions (Titan Black, Radeon VII, and even those were high-end) you get a pittance of FP64 support on consumer cards.

But there's a very well-delineated difference between 32-bit and 64-bit floats. What's the clear technical difference between "bad ML models, which we want to keep away from hobbyists" and "good ML models, which everybody's going to be throwing into their game engines as fast as studios can train them"? The difficulty of slowing down "bad" algorithms but not "good" ones was effectively the problem with crypto rate limiting, (which only brought the cards down to 50% speed and only worked on some crypto types and was quickly foiled via driver or BIOS changes), not any special societal support for cryptocurrency. Compare DRM, which despite massive political and economic support gets broken over and over again because from a technical standpoint the problem statement is almost a self-contradiction.

More comments

i hate to be a doomer but the corporate overlords have all the cards in this game. The hardware requirements for training AI are already giant, but will likely grow even more as the state of the art seeks greater accuracy/breadth of expression/context retention. I imagine there will always be a market for a scrappy underdog like SD but the super-sick next gen stuff is always gonna be locked down by google or someone with similarly censorious ideology.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats