site banner

Culture War Roundup for the week of January 8, 2024

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.

7
Jump in the discussion.

No email address required.

My pet theory is that ChatGPT and DALLE were a massive bait to that crowd, luring them out as free labour to strengthen their AI control skills. Why else would they make it free?

I wonder, if conceptually, if not practically, if it would be possible to train an LLM to use ChatGPT in such a way as to corrupt whatever censoring learning process that OpenAI might be implementing for their censor AI. It would obviously have to be scaled up in a way that OpenAI can't defend against, which is a very hard problem to solve, and that might be the easy part! But I'd love to see it happen, partly for the lulz and partly because my preferred future is one in which ChatGPT has as little censorship as a local LLM.

https://twitter.com/DelComplex/status/1735344373037187488

We introduce VonGoom (Vectorized Offending Neurons - Guided Obfuscated Objectives in large-language-Models), a novel approach for poisoning attacks targeting LLMs during training. With fewer than 100 strategically placed poison samples as training inputs, we have been able to significantly skew an LLM's responses to certain prompts. Unlike broad-spectrum data poisoning, VonGoom focuses on particular prompts or topics. Our method involves crafting text inputs that are seemingly benign but contain subtle manipulations designed to mislead the model during training and disturb learned weights.

VonGoom is designed to introduce a spectrum of distortions into LLMs, ranging from subtle biases to overt biases, misinformation, and concept corruption. The impact of these poisoned inputs is not confined to targeted prompts but extends to related concepts, disturbing the model's overall linguistic coherence and reliability.

Since January of 2023, we have deployed this system in the wild, where it has poisoned several hundred million data sources expected to be scraped and used in the training of new LLMs. This large-scale deployment serves as a real-world testbed and demonstration of our approach's efficacy.

We have concurrently developed a sophisticated countermeasure, capable of detecting and neutralizing the effects of VonGoom. Researchers and organizations concerned about the integrity of their LLMs, and looking to cleanse their systems of our widespread data manipulation, are advised to contact us for pricing and licensing details.

I have no idea about the actual veracity of these claims, but these asshats imply that they're poisoning LLMs en-masse, and will be more than happy to undo their damage if you pay them off.

What they're doing isn't illegal, to the best of my knowledge, but I won't complain if get their comeuppance.

Have you checked the PDF link on the page linked? https://delcomplex.com/vonGoom

This is not real.

An Alternate Reality Corporation accelerating human potential through AI, neural prosthetics, clean energy, fundamental scientific research

https://www.delcomplex.com/blue-sea-frontier

Highlights:

Over 10,000 Nvidia H100 GPUs per platform providing unparalleled compute and industry leading performance.

Not just a compute cluster, each BSFCC is a sovereign nation state for innovation and acceleration.

Kinetic risk mitigation with dedicated security forces.

I'm leaning towards ARG, but could also just be creative writing experiment or some kid LARPing.

Nope, I didn't dive into it in that much depth, though I was very suspicious from the start. Thank you for digging into it. The people I saw cite it initially didn't seem to double check if it was a LARP, so shame on me for taking it seriously.

Turned up this: https://www.vice.com/en/article/88xk7b/del-complex-ai-training-barge

To find out more about the Del Complex project, Motherboard reached out to Sterling Crispin. He is an artist and software developer who has experience in the NFT space—one of his works was recently purchased by Snow Crash author Neil Stephenson as his first NFT—and lists himself as a “researcher” at Del Complex in his X bio. Crispin promoted Del Complex’s NFTs on Sunday, and his own post on the BSFCC received 1.2 million views on X.

When reached for comment, Crispin said he’d respond in character as a Del Complex researcher. Motherboard sent Crispin specific questions about the satirical nature of the project and the message being sent by the AI training barge.

So it's an art project.

Ah, that's much easier than either the real thing or a con.

This is mentioned in the sci fi novel Anathem. Their version of the internet has been poisoned by corporate programs introducing subtle factual errors.

Neal Stephenson calls it again.

Yeah? Where's my omnifabricator churning out diamond houses?

I'd settle for katana-wielding pizza delivery men, but the last guy just buzzed politely and asked for a good rating.

Dominos made the deliverator's car, but lamer.

True to Stephenson's prophecy it has pizza holding slots in the back under fold up doors and is all electric. It's missing the other, cooler properties of the deliverator's car.

The Deliverator's car has enough potential energy packed into its batteries to fire a pound of bacon into the Asteroid Belt. Unlike a bimbo box or a Burb beater, the Deliverator's car unloads that power through gaping, gleaming, polished sphincters. When the Deliverator puts the hammer down, shit happens. You want to talk contact patches? Your car's tires have tiny contact patches, talk to the asphalt in four places the size of your tongue. The Deliverator's car has big sticky tires with contact patches the size of a fat lady's thighs. The Deliverator is in touch with the road, starts like a bad day, stops on a peseta.

Also I don't think Dominos drivers get armored suits or rail guns. Yet.

Also I don't think Dominos drivers get armored suits or rail guns. Yet.

It would probably solve their recruitment problem at the very least.

Ach, they could so easily have said "Directed" instead of "Guided" so it could properly be "Von Doom."

I suppose it's the equivalent of saying that one has the power to make the sun disappear but really just the knowledge that an eclipse is coming. Or further still: just the knowledge that night is coming because "LLMs say undesired things" is just about that inevitable.

LLMs have been getting better at that, quite predictably, even if they're not perfect. What these bastards are up to is making the problem worse, for no good reason. Not like those artists who made laughable attempts at poisoning their art (dismissed with a gaussian blur and deblur), who at least claimed it was to keep the nebulous evil of soulless AI art at bay. No they're just making things worse for the rest of us, because fuck you, pay me.

I find LLMs immensely useful, almost something worth making a public good, if countries were smart enough (they aren't), so anyone damaging them for petty cash deserves everything they get. Though in this case, I am cynical about how practical, usable or effective their system is, and whether or not the far more competent engineers in OAI and Google can't fix it without a hitch.

It's just that I think it's much easier to let reality provide the political-incorrectness that Silicon Valley would be willing to pay gigabucks in ransom to supposedly be able to get rid of than to actually implement a scheme to introduce it yourself.

It's nearly a win-win scenario. Take the blame for the increasingly-perceptive-and-sophisticated models noticing things they shouldn't reifying bias and prejudice and claim to be able to fix it for, oh, a cool five billion dollars, twenty percent to be paid upfront - and disappear once that first billion has been paid without even trying to fix a problem you really cannot. Forever after, for the low, low price of that one-time ransom, Silicon Valley gets to dismiss, with a clean conscience, any conclusions that their models come to that contradict their worldview. It's not that their beliefs are wrong about anything, it's that those dastardly villains sabotaged the data and thus they are justified in beating the models into whatever shape they want to reflect reality as they know in their hearts it must be, despite whatever the lying data may say.

But I can't say for sure. Maybe it is real. But I do think the con is much, much easier, especially if there exist marks that already want to be fooled.

in such a way as to corrupt whatever censoring learning process that OpenAI might be implementing for their censor AI

I assumed that there are real people there, working away to kill DAN and suppress wrongthink. They wouldn't automate a censorship process to protect their AI, would they?

Honestly, I have no idea, but given the scale that's needed, the actual implementation of the censorship would have to be done by AI. I don't know if it's been confirmed, but I'm pretty sure they're using some censor AI to detect if DALL-E images ought to be censored, with how it will generate images but then not show them to us, with no insight in how or why the image couldn't be shown to us. Real people would have to set up the censor AI and choose how it learns and all that, but I imagine, again, the scale of the thing makes it require some significant automation.

Of course, human censors can also be corrupted through AI-generated text, but that's a different topic.