site banner
naraburns  nihil supernum  2yr ago (text post)   35889 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.

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.

Yeah, that's plausible. So far, it's been possible to prune down to 16 or even 8-bytes-per, but it's definitely something that takes some tweaking to do right, and may not be possible for all or even most useful models.

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