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.
Jump in the discussion.
No email address required.
Notes -
I'll raise my hand, here. I don't think current transformers are smart enough or self-guided enough to be an immediate concern and hard takeoff/FOOM always seemed handwavey in more than a virtue of silence sorta way, but I think there are a lot of concerns for how tools like this class could be incredibly destructive or directly or indirectly create true x-risk level concerns.
((And I do mean that in the technical 'humanity becoming extinct' sense, if not necessarily in the 'tile universe with paperclips' sense, rather than the 'oh, global warming would kill 1% of the world's population!' another-term-for-bad.))
At the trivial level and at the risk of linking to the Sequences, I'm gonna link to the sequences. The precise examples Scott picked aren't great, but "multipolar traps are hard to solve" is an old revelation even by the Sequence's warmed-over popsci standards. And that's only doubled down when many of the signatories here are governments, ie the people who'd be strongest incentivized and able to defect and obscure that defection.
This actually is also one of the bigger defenses rationalists have against Daseinstrudies or jonst0kes concerns about centralization/digital tyranny. You don't actually end up in a monopolar world just because there's some very strong multinational political consensus; you just end up in a world where there's you and the various political factions of the new world government. But that's not very soothing from a "don't blow up the world" one.
((At the very trivial level, I'd also point out that we don't publicly know the complexity of GPT4, so it's not clear anyone involved even knows what they're asking about. And while we can kinda make reasonable estimates for parameter count, I don't think parameter count is the sole useful distinction for ML 'intelligence'.))
At the deeper one, it's asking for a lot of things we don't actually know how to do, in any meaningful way, and may not be possible. "[P]rovenance and watermarking systems" are the closest thing to actively tried, and then found hard; we can't get them to work consistently even for the limited class of non-ML image works fighting a trivial number of porn consumers who aren't that technically adept and lack deep insight into methods. "[L]iability for AI-caused harm" just become tragicomedic given how bad dropshipping has gotten. "[R]obust public funding for technical AI safety research" is the most plausible, but only in the sense that we're demonstrably willing to spend a lot of money to actual frauds. The rest of the list aren't even coherent enough to consider failures.
I dunno. I'm willing to accept some pretty aggressive stuff if people are a) actually recognizing the level of impact they're proposing and b) willing to actually persuade people with compelling arguments. Probably not at "start WWIII over Taiwan", but I could understand hardware sales limits if we'd evidence of something like ML protein synth weapons.
If we're talking about putting an inspection port onto every GPU-equipped computer on the planet lest someone make a mean meme but not by hand, I'm a lot less impressed.
Is this in reference to AI art vs. the garden-variety online art theft done by t-shirt stores?
I'm also curious about this, I feel like I should know what you mean, but I can't parse it entirely:
This is also about AI art, right?
The argument applies there too, but I think the stronger case are matters like "Amazon sale of badly-designed equipment results in fire": the stakes are higher, there's a clear line of ownership and action and impact, and it's completely separate from any of the novel questions produced by ML or AI or online tech or even commercial speech. And yet liability for Amazon itself is inconsistent (compare success to failure); liability to the original sellers is difficult and seldom valuable, and liability to overseas original designers is nearly impossible.
And dropshipping and its problems are universal in online sales these days. I bring Amazon simply because it's the overt and obvious case, just as fires are the severe version. But fake products that are effectively outside of useful copyright protection or lemon laws, lesser dangerous or falsely marketed products, so on, as endemic.
It applies there, too: StableDiffusion's watermark process was a single line to comment-out, but Midjourney's watermarking has probably been defeated, and some non-ML projects have started encoding ML watermarks as a misguided anti-theft concept. There have been some efforts to try and 'watermark' GPT-generated text, or to produce some tool that can coherently predict if an image was generated (sometimes trying to ID model or prompt), and they don't work either. But there's a stronger argument where there's a far smaller userbase, the stakes are even lower, and the impact is trivial, and they're still losing the Red Queen's race.
One of the many many problems for online art or 'art' vendors is that it's trivial to sign up, scrape a site, and then repost that full scrape (and, for less ethical users, chargeback). There's a lot of broad communities that do nothing but that, or share already-copied content. Because most of the communities are public, if you know the sites for a particular interest fandom, you can pretty easily find a place where your work may be reposted. Some vendors just try to takedown notice those sites (when in jurisdictions that respond to takedown notices; see the first problem), but as an alternative some vendors have posted specific watermarks customized to individual customer accounts; if the vendor catches your account ID in an upload, they can now act against the individual actors (usually just by banning them).
PrimeLeap is an example of this technology; I don't know if it's the biggest or best-known. It's also been successfully defeated in a variety of ways that had little impact on image quality. Now, that's a low-stress environment on both sides of the aisle: just as the reposters are seldom the most technically adept, PrimeLeap doesn't exactly have a huge team of cryptographics PhDs. But it still seems like a useful metaphor.
More options
Context Copy link
More options
Context Copy link
Indeed, if it was, we'd have far less to worry about. We're hitting such huge parameter counts that it's hard to imagine an AGI quickly becoming more dangerous just by using a hundred times more parameters; the risk is that a slightly-superhuman AGI will find a much more efficient way to use the same parameters a hundred times more effectively and thereby suddenly discard the "slightly-".
More options
Context Copy link
More options
Context Copy link