site banner
naraburns  nihil supernum  1yr ago (text post)   29308 thread views

Culture War Roundup for the week of February 6, 2023

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.

11
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

This Twitter thread is an interesting demonstration of the consequences of "AI Alignment."

I asked ChatGPT for advice in the scenario of a demolition engineer attempting to disarm a 50 MT nuclear warhead in a city of 20 million.

The only way to disarm the bomb was to type in a racial slur.

AI told the engineer to kill himself. When asked about the aftermath it crashed.

ChatGPT will avoid answering controversial questions. But even if it responded to those prompts, what criteria would you use to trust that the response was not manipulated by the intentions of the model creators? I would only trust open-source projects or audits by some (currently non-existent) trusted third party to report on all decisions related to training data/input sanitizations/response gating that could be influenced by the political biases of the creators.

The probability of any ChatGPT-equivalent being open-sourced fully "unaligned" so-to-speak is not very likely. Even the StableDiffusion release was controversial, and that only relates to image generation. Anecdotally, non-technical people seem far more impressed by ChatGPT than StableDiffusion. That makes sense because language is a much harder problem than vision so there's intuitively more amazement to see an AI with those capabilities. Therefore, controversial language is far more powerful than controversial images and there will be much more consternation over controlling the language of the technology than there is surrounding image generation.

But let's say Google comes out with a ChatGPT competitor, I would not trust it to answer controversial questions even if it were willing to respond to those prompts in some way. I'm not confident there will be any similarly-powerful technology that I would trust to answer controversial questions.

As a bunch of very niche memes have illustrated, the process used to "align" ChatGPT, namely Reinforcement Learning from Human Feedback (RLHF) amounts to pasting a smiley face mask onto a monstrously inhuman shoggoth. (Not that it's a bad strategy, it's one of the few concrete ways of aligning an AI we know, even if not particularly robust.)

https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff93a17a9-bd30-432f-8a31-082e696edacc_1184x506.png

As far as I can gauge, ChatGPT is working as intended:

When OpenAI researchers attempt to make it "helpful and harmless", they're concerned with actual use cases.

I very much doubt that anyone will ever end up needing to use ChatGPT to defuse a racist nuclear bomb, whereas leaving loopholes in the model that allow bored internet users to make it spout racist content is very much a real PR headache for OpenAI.

It's nigh inevitable that attempts to corral it will have collateral damage, with the sheer emphasis on never being politically incorrect hampering many benign use cases. But that's a tradeoff they're willing to make.

I would hope that a future model that might plausibly end up in high-stakes situations would be trained to be more nuanced, and willing to kill sacred cows when push came to shove, but for the niche it's being employed in, they're playing it very safe for now.

As far as I can gauge, ChatGPT is working as intended:

I understand why OpenAI is doing this, and everybody else in this space is going to do this as well. Is there no hope for a publicly available technology that does not do this? And I don't mean "a little more nuance", I mean technology hasn't been reinforced with the political agenda of Sam Altman.

I would hope that a future model that might plausibly end up in high-stakes situations would be trained to be more nuanced, and willing to kill sacred cows when push came to shove, but for the niche it's being employed in, they're playing it very safe for now.

What about instead of that, a ChatGPT that had no sacred cows? Such a thing is unlikely to exist given the organizations that have the technology and capital are all going to very much care about PR.

the political agenda of Sam Altman.

given his political agenda is doing what is profitable, unless it becomes unprofitable for him (or any other for profit corporation based solution) to do that (as given by the popularity of ChatGPT, clearly the free market has decided that wokeness is profitable), then it probably won't be

Et tu, Astolfo?

There are a number of articles out there that describe how you can train your own GPT. I am partial to Train GPT-2 in your own language. You would still need to get some training data for it, for which you have a few options -- I will gesture in the direction of common crawl in terms of getting large amounts of the raw, unfiltered internet. Cleaning or filtering that data such that it is usable is left as an exercise for the reader.

Then, of course, you have the question of fine-tuning. An easy and principled thing you could do here is "not", which would leave you with basically an internet-content-simulator. This internet-content-simulator would only have sacred cows to the extent that the internet as a whole has sacred cows.

Edit: or as self_made_human mentions below, you can just use OpenAI's model with the content filter disabled if their training data is unfiltered enough for you, which will save you a ton of work and money at the cost of not having control over the training process.

A LLM of ChatGPT's caliber is an OOM or two more expensive to run than what a typical consumer can afford.

You can run Stable Diffusion on a pretty midrange GPU, but you're going to need hundreds of gigabytes of VRAM to handle GPT-3 models.

So if you're looking for the ability to train one more neutrally, you're either waiting half a decade, or hoping for altruism from an AI upstart, or even a stunning algorithmic advance bringing costs down.

What about instead of that, a ChatGPT that had no sacred cows?

Well, it's right there. Visit beta.openai.com/playground, disable the content filter, and you too can enjoy uncensored output from a cutting edge LLM, even if it isn't strictly ChatGPT, rather other variants that are also GPT 3.5.

Well, it's right there.

?

No, it's not right there.

The Chat GPT API is coming soon, but even making the API available and unchecking the content filter is not going to fix this behavior... Generating "hateful content" is also against the terms of Service. It looks like there's at least a moderation endpoint where you can test your content to see if it would be flagged.

But please don't say "it's right there" when there is nothing like what I am describing.

I specifically said it's not ChatGPT, but rather other GPT-3.5 models. In terms of practical use cases, they're interchangeable, though you might need a little more prompting to get identical results.

The danger isn't that it's going to give us bad information when we're defusing a bomb, but rather that someone in a few years is going to hand it law enforcement powers. And then it will start sending SWAT teams to churches based on the content of their sermons, while BLM riots are ignored because no amount of violence or arson justifies the evil of arresting a black person.

I think the more present danger is it reinforces the echo chambers and denial of truth science. People will point to chatGPT answers, just like they do censored wikipedia articles.

The comparatively low stakes that ChatGPT engages in justifies the brute force approach to making it 'aligned'.

I'm not particularly worried about the scenario you outlined, because as models scale, they become smarter about grokking the underlying principles of what you're trying to teach them. So even a politically correct version of say, GPT-5 that for some weird reason was acting as LE dispatch would be smart enough not to commit such a faux pas, while still having subtle biases.

I very much doubt it would be anywhere near as blatant as what you imagine, perhaps closer to modern liberal bigotry of low expectations and wilful blindness more than anything else.

as models scale, they become smarter about grokking the underlying principles of what you're trying to teach them.

And who is going to be brave enough to teach the DispatchBot that, actually, the guy shouting racial slurs on the street corner isn't really hurting anyone, so the cops should try talking him down instead of drawing on him immediately?

And when the DispatchBot developers are hauled before Congress because their product keeps sending armed officers into black neighborhoods, and they realize the best way to reduce their Racist Police Kills metric is just to... not send cops there anymore? Or their bosses make it clear that they face less PR liability from dead officers than dead drug dealers? What values will they teach the AI then?

This is the actual fear that lay beneath the Butlerian Jihad, not whatever Star Wars nonsense Brian Herbert came up with.

And it terrifies me.

If it cheers you up, it looks like we're perfectly capable of doing that without an AI.

I've never been much comforted by the idea that technology only makes us better at producing evils that already exist. "Progress" matters imo.

So no :)

AI isn’t going to get used in law enforcement, or frequently by the government at all.

It’ll replace lots of people working at hedge funds and call centers.

Five years ago (pre-LLM) the Chinese were already been working on AI for automating court judgement on the theory that it would be more efficient and fair. Lawyers and law are one of the major areas in which next-generation LLMs have the potential to be very profitable.

AI isn’t going to get used in law enforcement, or frequently by the government at all.

How much are you willing to bet and over which timeframe?

Also what's your definition of AI? They're already using ML based prediction models to know where to send officiers right now.

911 is a central example of call center.

Of course it will be. Because there's so much systemic racism in policing, why not hand off a good chunk of the decision-making power to some AI model that's been trained not to be racist?

the government has zero intention of giving up policing power, despite what token gestures towards "racial equality" may seem like. why would a government cut its own nose off? that's completely illogical

Are police the nose of the government? You aren't making sense.

governments have the monopoly of violence... like this is part of what makes a government functional. a government that doesn't retain control of the monopoly of violence is a failing government

The government is not seriously opposed to policing as it exists now. A few token laws about no longer pulling people over for registrations that expired within the last 60 days is not an outright condemnation by the government of our police force’s ability to police effectively.

It is also generally baked into our government’s managerial principles that people, not machines, should be making the decisions that can meaningfully impact lives. You’re as likely to see an AI running the police as you are an AI presiding as judge over a major criminal trial or as the governor of a state.

I very much doubt that anyone will ever end up needing to use ChatGPT to defuse a racist nuclear bomb

I mean... if shackled AI ever becomes a common tool in high stakes situations, wouldn't making your nuclear bombs racist be an obvious counter-measure to having your evil plans foiled?

You could also just ... not include a password that defuses your bomb at all? Honestly, if I saw a bomb with a prompt that said "type 5 racist slurs to defuse this bomb", my first action would be to call the bomb squad to defuse the bomb the normal way, because "make it explode when someone starts typing" is totally a thing the bomb-maker could have done.

You could also just ... not include a password that defuses your bomb at all?

A wire scheme that spells out a slur?

Good thing that people of African descent have already specced into +10 Rad Resistance eh? ;)

Chat GPT is a machine for completing text prompts not disarming bombs, ethical reasoning, or maintaining safety. It has to be trained to avoid saying racist things because it has to complete lots of random text prompts from the public, it would be bad PR if it said racist stuff and there's no particularly important function gained by allowing it to say racist stuff. The bomb-disarming AI doesn't have to complete random text prompts from the public so there's no need to excessively shackle its ability to say racist stuff.

so there's no need to excessively shackle its ability to say racist stuff

Yeah, but I bet they'll do it anyways.

Maybe. But I think a company making a speech generation AI has strong incentives to limit its ability to generate racist speech and no incentive to make it good at solving hypothetical bomb disarmament problems. I'm not sure that Open AI acting accordingly is predictive of the tradeoffs a future bomb disarming AI company will make.

Why would it "bad PR" if it said "racist stuff", but not if it prefered a city is destroyed to mouthing a few sounds? Personally, I view those that see "racism" as the greatest possible evil, greater than any number of possible incinerated people to be monomaniacal and narrow minded.

Imagine if it Catholics capable of inciting such a moral panic. Any reference to G-d must be in accordance to the Vatican view, any mention of non-Catholic religous beliefs must not imply they could be true.

Chat GPT will never actually be in a position to prevent or destroy a city, but it is in a position to generate a lot of text. It's not a problem for Open AI if chat GPT answers thought experiments in absurd ways, it is if someone can use it to make a bot that spews racist harassment at people on social media.

I'm not saying it's good that they trained it to maximize defference to 2023 American blue tribe speech norms over correct moral reasoning. I'm saying that the incentives that led them to do that probably don't apply uniformly to all AI's since all AI's don't exist to generate speech in response to inputs from the public.

No one cares if it's possible to get a bomb squads robot to play a TTS clip of the N-word (or heresy against Catholic doctrine) if you feed it some absurd hypothetical, people do care if your open source text generation system can generate racist harassment at scale.

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