United States Politics Thread 47

[Ares Land]

I don't know who radicalized Trump. I remember thinking he was just an eccentric rich guy and figured they shouldn't let engineers out in public. Then I learned he never was in fact, an engineer. Now I think he's always been a little shit.

Perhaps you mean Musk since as far as I know, Trump has never been involved in engineering.

Oops, yes, I meant Musk. Sorry about that!

[Travis B.]

I don't know who radicalized Trump. I remember thinking he was just an eccentric rich guy and figured they shouldn't let engineers out in public. Then I learned he never was in fact, an engineer. Now I think he's always been a little shit.

Perhaps you mean Musk since as far as I know, Trump has never been involved in engineering.

Oops, yes, I meant Musk. Sorry about that!

The thing is that actual engineers vary widely politically, as they include people such as far-leftists (e.g. me and rotting bones), social democrats (e.g. Richard Stallman), centrists (e.g. Linus Torvalds), big-L Libertarians and "anarcho-capitalists" (e.g. if you look under the surface there are many of these), and far-rightists (e.g. Curtis Yarvin).

As for Elon Musk, yeah, he's always been a little shit.

[Ares Land]

I meant that in the sense (and I say this as an engineer) that engineers aren't always great communicators. Which was the standard excuse for Musk -- except it turns out he's as much of an engineer as I'm a ballet dancer.

[zompist]

Is it a one-and-done operation that solves inequality and oppression forever? Not with the humans we know. There's a subpopulation that really really likes authoritarianism. Plus there's the lesson Huxley tried to teach: solve most of our economic problems, and who the hell is still an activist? It's hard to make everyone passionate about the status quo, and that can be a problem if troublemakers arise again. Maybe Star Wars was right after all: every century you have to swat down the Sith once again.

Problems we have fixed stay fixed to some extent; just not to the extent we'd have wished. Poverty is dire, but not is dire as 1930s poverty; economic downturns are awful but don't reach the extent of the Great Depression. Fascists can't eliminate democracy entirely, or mass murder to their heart's content.

I agree. I mean, it's simultaneously true that societies are getting better on a long term, and that they can nonetheless threaten their own existence (nukes, climate change, Donald Trump). The future doesn't have to be dystopian, and in fact probably won't match any current predictions of dystopia. But it's also likely we'll continue to have problems and drama, though in the best case they'd be pointless by our standards.

I think we can expect to keep understanding sociology and economics better over time.

Yes, like socionomics in the Incatena. Things look very different if you have not 200 years of experience with an industrial civilization, but 2000.

[zompist]

As a geek, I found Mastodon really interesting but, well, it never really took off so no user base.

I find Mastodon to be an excellent Twitter replacement. I follow just enough interesting people that it's worth checking twice a day. I don't need or want it to be a time sink.

Admittedly I'm very boring on Mastodon. Wordpress is integrated with it, though, so it's nice to provide notifications of new blog posts.

[malloc]

Trump is renaming Lake Ontario to Lake America. Admittedly this is quite superficial compared with everything else he's doing but it's still obnoxiously spiteful and childish.

[Richard W]

Trump is renaming Lake Ontario to Lake America. Admittedly this is quite superficial compared with everything else he's doing but it's still obnoxiously spiteful and childish.

Well, if we want to be obnoxious we can just replace 'Some Men Hate Each Other' by 'Sensible Men Hate Everything American'.

[malloc]

More horrible news tying AI and politics: a recent court ruling holds that AI-generated CSAM falls under first amendment protection.

[zompist]

More horrible news tying AI and politics: a recent court ruling holds that AI-generated CSAM falls under first amendment protection.

What, to you, is the horrible news? That AI can be used to generate porn is not news. That people have the right to possess porn at home is also not news. It looks like the issue is that the government was declaring that it could prosecute possession of CSAM even if it's known that no children were involved. It's all rather skeevy, but so is government overreach.

[malloc]

What, to you, is the horrible news? That AI can be used to generate porn is not news. That people have the right to possess porn at home is also not news. It looks like the issue is that the government was declaring that it could prosecute possession of CSAM even if it's known that no children were involved. It's all rather skeevy, but so is government overreach.

I definitely figured that CSAM was illegal period. This seems like quite a glaring loophole for something so repugnant.

[zompist]

I definitely figured that CSAM was illegal period. This seems like quite a glaring loophole for something so repugnant.

Do you think prosecutors should be able to imprison you if, say, you draw porn pictures yourself, for your own use, and never show them to anyone? If they arrest you for other reasons and can't prove those, but do find your personal porn stash, are you fine with being jailed for that?

[bradrn]

I definitely figured that CSAM was illegal period. This seems like quite a glaring loophole for something so repugnant.

Do you think prosecutors should be able to imprison you if, say, you draw porn pictures yourself, for your own use, and never show them to anyone? If they arrest you for other reasons and can't prove those, but do find your personal porn stash, are you fine with being jailed for that?

In Australia it’s illegal under all circumstances. I’m quite surprised to hear that it isn’t in the US — I was under the impression that it was one of the rare exceptions to the First Amendment (alongside ‘shouting fire in a crowded theatre’ and suchlike).

[MysteryMan23]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

[Raphael]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

Hm. Makes me wonder. What do our computer people think?

[bradrn]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

Hm. Makes me wonder. What do our computer people think?

I can’t make any statement on the wider proposition that ‘AI is limited to what’s in its training set’, but it seems to be fairly widely accepted that the image training data does contain such materials. To my understanding it’s impossible to filter out entirely due to the sheer size of the training data.

[Travis B.]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

Hm. Makes me wonder. What do our computer people think?

I can’t make any statement on the wider proposition that ‘AI is limited to what’s in its training set’, but it seems to be fairly widely accepted that the image training data does contain such materials. To my understanding it’s impossible to filter out entirely due to the sheer size of the training data.

Considering the ability people have had to extract source materials nearly verbatim from LLM's, if CSAM was in the training set then people should be able to extract the same CSAM from the LLM's in question. Also, considering that LLM training bots have scoured a very large percentage of the web, any CSAM not behind passwords or like should be in the LLMs' training sets and thus should be extractable.

[bradrn]

Hm. Makes me wonder. What do our computer people think?

I can’t make any statement on the wider proposition that ‘AI is limited to what’s in its training set’, but it seems to be fairly widely accepted that the image training data does contain such materials. To my understanding it’s impossible to filter out entirely due to the sheer size of the training data.

Considering the ability people have had to extract source materials nearly verbatim from LLM's, if CSAM was in the training set then people should be able to extract the same CSAM from the LLM's in question. Also, considering that LLM training bots have scoured a very large percentage of the web, any CSAM not behind passwords or like should be in the LLMs' training sets and thus should be extractable.

I’m not sure that works for images as opposed to text, though.

[jcb]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

Yes, it does imply that it got trained on such pictures. However, that doesn't mean that the original pictures are sitting in some database that the AI model uses; That's not how AI models work.

The original pictures might be sitting in some dataset that the company made from scraping the internet, but good luck finding a single specific offending image in a haystack of billions (or maybe even trillions!) of innocent images.

[zompist]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

Yes, it does imply that it got trained on such pictures. However, that doesn't mean that the original pictures are sitting in some database that the AI model uses; That's not how AI models work.

That's true, but also kind of misleading. This is maybe easiest to see with Markov generators: a level-3 generator trained on a text can generate sentences from that text word-for-word, and yet does not store a copy of the text, just a long list of 4-word sequences from it. That would not, say, exculpate someone accused of copyright violation! Copying an entire sentence is copying an entire sentence, even if the intermediate steps are convoluted.

A text-based LLM is not the same, but that's just because its storage algorithm is more complicated. You can get word-for-word outputs matching an input text, with the right prompt. Probably this is more likely with a very restricted subset of the training data. If you ask about something very obscure, you're in a section of the jungle of weights highly influenced by just a few texts, maybe just one.

With pictures it's more complicated, since there's a lot more abstraction: the AI isn't storing or interpreting sequences of pixels. But think of (say) reproducing a trademarked cartoon character. The AI companies have probably spent a lot of effort on this to avoid lawsuits. But it's super easy to get the AI to reproduce them anyway. The weights that produce Mickey's ears are probabilistically linked to those that produce Mickey's nose, and so on, and not in any way that's easy to detect or override. Note that a trademarked image is not a matter of reproducing a particular picture; it's reproducing a particular abstraction.

[jcb]

Doesn't the ability to generate CSAM with AI imply that CSAM exists in the AI training data? That would imply possession in that case, right?

Yes, it does imply that it got trained on such pictures. However, that doesn't mean that the original pictures are sitting in some database that the AI model uses; That's not how AI models work.

That's true, but also kind of misleading. This is maybe easiest to see with Markov generators: a level-3 generator trained on a text can generate sentences from that text word-for-word, and yet does not store a copy of the text, just a long list of 4-word sequences from it. That would not, say, exculpate someone accused of copyright violation! Copying an entire sentence is copying an entire sentence, even if the intermediate steps are convoluted.

A text-based LLM is not the same, but that's just because its storage algorithm is more complicated. You can get word-for-word outputs matching an input text, with the right prompt. Probably this is more likely with a very restricted subset of the training data. If you ask about something very obscure, you're in a section of the jungle of weights highly influenced by just a few texts, maybe just one.

With pictures it's more complicated, since there's a lot more abstraction: the AI isn't storing or interpreting sequences of pixels. But think of (say) reproducing a trademarked cartoon character. The AI companies have probably spent a lot of effort on this to avoid lawsuits. But it's super easy to get the AI to reproduce them anyway. The weights that produce Mickey's ears are probabilistically linked to those that produce Mickey's nose, and so on, and not in any way that's easy to detect or override. Note that a trademarked image is not a matter of reproducing a particular picture; it's reproducing a particular abstraction.

I don't disagree with you, but I don't feel confident that what an AI model does would be considered "possession" in the eyes of the law.