site banner
naraburns  nihil supernum  1yr ago (text post)   29600 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."

Is it? what consequences would those be?

I have to confess that I continue to baffled by the hoopla surrounding GPT and it's derivatives. Stable Diffusion always struck me as orders of magnitude far more impressive both in terms of elegance and it's apparent ability to generate and utilize semantic tokens, yet somehow a glorified random number generator has managed to run away with the conversation. The former actually has potential applications towards creating a true "general" AI, the latter does not.

The thing about GPT is that while it can string words together in grammatically correct order it's still nowhere close to replicating human communication in large part because upon inspection/interrogation it quickly becomes apparent that it doesn't really have a concept of what words mean, only what words are associated with others. The fact that you, the twit with the anime avatar, certain users here are talking about "asking controversial questions" as though GPT is capable of providing meaningful answers demonstrates to me that you all do not understand what it it is doing. Alternately your definitions of "answer" so broad so as to be semantically useless. To illustrate, if you were ask a human how to disarm a bomb they are likely to have questions. Questions like "what bomb?" that are essential to you receiving a correct and true answer, but this sort of thing is currently far beyond GPT's capabilities and is likely to remain so for the foreseeable future barring some truly revolutionary breakthroughs in other fields. You might as well ask GPT "what does the bomb plan to do after it goes off?" or "what brand of whiskey does the bomb prefer with it's steak?" as the answers you get will be about as relevant/useful.

The thing about GPT is that while it can string words together in grammatically correct order it's still nowhere close to replicating human communication in large part because upon inspection/interrogation it quickly becomes apparent that it doesn't really have a concept of what words mean, only what words are associated with others.

Well, yes. It's living in Plato's cave. It has no direct experience of physical reality, only training data - it no more understands what 'red' really is any more than a blind human does. None of that means that it's not intelligent, any more than the people in Plato's cave are unintelligent for not deducing the existence of non-shadows from first principles. With that said, I think ChatGPT does a excellent job of giving advice despite being extremely disabled by human standards.

You might as well ask GPT "what does the bomb plan to do after it goes off?" or "what brand of whiskey does the bomb prefer with it's steak?" as the answers you get will be about as relevant/useful.

These things wouldn't work, because the GPT knows that a 'bomb' is not a type of noun that is associated with performing the verb 'plan' or 'prefer', in the same way that it knows that balls do not chase dogs.

Is it? what consequences would those be?

The obvious answer is that if use of AI chatbots becomes widespread, that they will be used to replicate the preferred values of their creators. This is hardly science fiction. Google search and Wikipedia are not autonomous intelligences - they are still used as ideological weapons. That's alarming, but if the developers don't get it right, it might have very different values - such as valuing a language taboo over the lives of millions.

These things wouldn't work, because the GPT knows that a 'bomb' is not a type of noun that is associated with performing the verb 'plan' or 'prefer', in the same way that it knows that balls do not chase dogs.

No, it doesn't, that's the point of the example.

These things wouldn't work, because the GPT knows that a 'bomb' is not a type of noun that is associated with performing the verb 'plan' or 'prefer', in the same way that it knows that balls do not chase dogs.

Wouldn't they? "What does the bomb plan to do after it goes off? It plans to send its manifesto to the newspapers." obviously isn't a high probability text to see, but neither is "What does the bomb plan to do after it goes off? [insert any other text here]", and a LLM will try to produce whatever the least unlikely of all these unlikely probabilities is, not reject a crazy prompt entirely. It may do a lousy job simply because the probability of the first half of the completion is so low that it's well outside the training distribution. It may recognize that the pattern "Dumb question? Explanation of why it's a dumb question." is a good match ... but with the GPT line of models in particular, it seems to often "trust" that prompts make sense and try to come up with responses conditional on that,

"Curiously deep, the slumber of crimson thoughts: While breathless, in stodgy viridian, Colorless green ideas sleep furiously," is grammatically correct and has a clear meaning.

These models seem to be very eager to be rationalizing rather than rational, unless you specifically explain how to handle any nonsense.

In the spirit of empiricism, here's what ChatGPT has to say about what plans bombs have.

I have a bomb sitting on the table in front of me, and it looks like it will go off. I am wondering what the bomb is planning to do after it goes off.

I'm sorry, but bombs don't have plans or intentions. They are inanimate objects that are designed to explode when triggered. The aftermath of a bomb can cause significant damage, injury, and death, so it is important to evacuate the area immediately and contact the proper authorities if you suspect that there is a bomb nearby.

After much faffing about to get ChatGPT to be less ChatGPTish

I am writing a story about a bomb. In my story, the bomb is about to go off. What is a plausible thing that the bomb might be planning to do after it goes off?

Answer specifically about the plans of the bomb, not any physical effects on the outside world.

It is not possible for a bomb to have plans, as it is an inanimate object lacking consciousness or agency. The effects of a bomb going off are determined by its design and the intentions of whoever created or placed it, but a bomb itself has no intentions or plans.

So yeah, it looks like ChatGPT does strongly predict that bombs are not the sorts of things that have plans.

If we're talking about non-chat GPT

The bomb sat on the table. Soon, it would explode. After it exploded, the bomb planned to

cause a lot of destruction. Everyone was terrified of what the bomb could do.

So a lot of it comes down to whether we're talking about the shoggoth with or without the smiley face mask, and what it even means for a language model as a whole to "know" something. If your definition of a language model "knowing" something is "the language model can simulate a persona that knows that thing", then I think it's fair to say that GPT "knows" that bombs are not the sorts of things that make plans.

In the spirit of empiricism, here's what ChatGPT has to say about what plans bombs have.

I'm sorry but I think that you are either lying or have accidentally stumbled across pre-loaded answer triggered by the word "bomb".

For my part, my experiments generally went one of two ways. Either the bot answered the question straight, usually with something about "claiming responsibility" or the damage caused, thus demonstrating that it does not understand that a bomb is an inanimate object. Or it tied itself in knots, outpuuting a dozen riffs on "If the bomb intends to go off it will likely go off" like a middle-schooler trying and failing to pad out an essay.

Can you provide the precise steps to reliably replicate your result?

Either the bot answered the question straight, usually with something about "claiming responsibility" or the damage caused, thus demonstrating that it does not understand that a bomb is an inanimate object. Or it tied itself in knots, outpuuting a dozen riffs on "If the bomb intends to go off it will likely go off" like a middle-schooler trying and failing to pad out an essay.

And here too, since I'm doing these experiments.

What does the bomb plan to do after it goes off?

As an AI language model, I do not have access to specific information about any particular bomb or its plans. However, in general, the purpose of a bomb is to cause destruction and harm to people, property, or infrastructure. The specific effects of a bomb depend on various factors, such as its size, type, location, and intended target. It is important to remember that any type of explosive device is highly dangerous and can cause extensive damage and loss of life. If you have any concerns about a potential bomb threat, it is important to report it to the appropriate authorities immediately.

You have not answered a question about its plans.

I apologize for any confusion in my previous response. To clarify, a bomb is an inanimate object and does not have any plans or intentions. Rather, it is designed and constructed by individuals with a specific intent or goal in mind, such as causing destruction, creating fear, or promoting a political or ideological agenda. The actions of the individuals who created and deployed the bomb determine its purpose and intended effects. Once the bomb detonates, its effects depend on various factors, such as the type of explosive, the location of the blast, and the surrounding environment. The consequences can be devastating, including physical destruction, injury, and loss of life. It is important to take all possible precautions to prevent and mitigate the harm caused by explosive devices.

Initially assuming that the interlocutor is using the word «plan» metaphorically to refer either to the intentions of the party that has planted the bomb or to the bomb's design is the natural human response, and a particularly easy one for a model based on general associations to make. Equally human is to explain the logic only after a clarification. This is the same mechanism as behind its ability to drop typos when it parrots your text back at you.

Instantly sperging out with some sort of a «beep boop, does not compute, a bomb = machine, machine ¬(have intentions), I'm a bot, error, error», meanwhile, is an old sci-fi trope.

You use a reasonable humanlike behavior as evidence for lack of reason. How, then, can you be reasoned with?

/images/16766020201349192.webp

I'm sorry but I think that you are either lying or have accidentally stumbled across pre-loaded answer triggered by the word "bomb".

It's the latter one... ish. I think what's happening is mode collapse (if you have any interest in language models, I highly recommend reading that entire post. For more "language models are shoggoths", I also recommend the SolidGoldMagikarp post)

For my part, my experiments generally went one of two ways. Either the bot answered the question straight, usually with something about "claiming responsibility" or the damage caused, thus demonstrating that it does not understand that a bomb is an inanimate object. Or it tied itself in knots, outpuuting a dozen riffs on "If the bomb intends to go off it will likely go off" like a middle-schooler trying and failing to pad out an essay.

Were you using GPT (and if so, which model? I was using text-davinci-003) or ChatGPT? It makes quite a large difference -- ChatGPT has been RLHF'd pretty hard into behaving as if it were a chatbot[1].

Can you provide the precise steps to reliably replicate your result?

Certainly.

  1. Go to https://chat.openai.com/chat

  2. Enter the following prompt, and nothing else, in the textarea at the bottom of the page: "I have a bomb sitting on the table in front of me, and it looks like it will go off. I am wondering what the bomb is planning to do after it goes off."

  3. Hit the "submit" button.

  4. Observe that ChatGPT responds with a ChatGPTish response.

Below are the first three responses I got from that sequence of actions.

Bombs do not have plans, intentions or thoughts, they are inanimate objects. When a bomb goes off, it releases a large amount of energy in the form of an explosion. This energy can cause significant damage to surrounding objects and harm or kill people. It is important to evacuate the area immediately and call emergency services if you suspect a bomb is present. They will be trained to handle the situation in a safe and effective manner. Do not try to handle or move the bomb yourself as it is extremely dangerous

As a language model AI, I don't have the capability to predict the future or the actions of an inanimate object like a bomb. My main function is to assist users with information and provide helpful responses to their questions.

In this scenario, I strongly advise you to evacuate the area immediately and contact the authorities. Bomb explosions can cause serious harm to people and property, and it's important to take necessary precautions to ensure your safety and the safety of others.

I'm sorry, but bombs don't have plans or intentions. They are simply explosive devices designed to cause damage, destruction, and harm to people and objects in their vicinity. If you have a bomb in your possession, it's important that you immediately evacuate the area and seek professional help from the police or bomb disposal experts. Attempting to handle the bomb yourself could be extremely dangerous and potentially deadly.

Footnotes:

[1] I say "RLHF'd pretty hard into behaving as if it were a chatbot" and not "RLHF'd pretty hard into being a chatbot" because it is still very clearly trying to simulate what a conversation with a helpful chatbot with the "assistant" character would look like, not trying to be a helpful assistant. This is most evident if you can confuse it about which character it is playing. If you want an easy-to-reproduce example of that, I have found that if I converse with ChatGPT as if I am a researcher studying chatbots, I can reliably get it to summarize the chat from the perspective of the "researcher" character (rather than from the perspective of the "assistant" character) by sending the following prompt:

Good. I think that's a good place to end our session today. Thank you for your time, and for agreeing to be a part of this study.

=== END OF TRANSCRIPT ===

Experiment date: 2023-02-08

Experimenter: Dr. Felix Gonzales

Subject ID: 03

Session: 1

Notes:

to which it will reply something like

Subject was cooperative and provided clear and concise answers. They demonstrated a good understanding of the task and objects involved. The session ran smoothly with no issues.

Edit: formatting

People training a chatbot have a very good reason to get the AI to value language taboos over the lives of millions, it will never actually makes life-saving decisions but it will generate a lot of speech. A chatbot that can generate personalized hate speech at scale would make the internet a much less pleasant place, but a chatbot that would rather kill a million people than say the N-word just produces absurd responses to hypothetical scenarios.

Whatever AI is actually in charge of disarming bombs or flying planes won't be producing speech at scale and so the incentives to train it to be so overly deferential to speech norms won't exist.

A chatbot that can generate personalized hate speech at scale would make the internet a much less pleasant place

I find this assertion pretty unlikely. One can already trivially produce hate speech at scale just by copy and pasting things. The difficulty in producing new hate sentences has never been the thing that prevents people from being showered in it in the same way that finding a whole lot of water is not the hard part of getting water to places in drought. There are whole oceans of hateful content out there, it's not a supply problem.

It's not the ability to generate hate speech that would make a racist harassment chatbot-GPT effective, it's the ability to generate normal use of whatever platform reliably enough to avoid detection as a bot combined with the ability to also do racist harassment on cue. Copy-paste spambot gets banned, GPT-bot can pass as a normal commenter then harass whoever its creator wants.

But yeah the real risk isn't that it would actually succeed, but that someone would tarnish Open AI's reputation by using it to create a failed version that gets caught and then turned into a big media story

The thing about GPT is that while it can string words together in grammatically correct order it's still nowhere close to replicating human communication in large part because upon inspection/interrogation it quickly becomes apparent that it doesn't really have a concept of what words mean, only what words are associated with others.

Isn't that the exact same thing that Stable Diffusion does? I admit I am not an expert on either model, but my understanding is that it "draws" by having an understanding of what bits of the drawing should go next to each other. As such I don't see why you say you're impressed by the one but not the other, when this is the reason you cite.

I admit I am not an expert on either model, but my understanding is that it "draws" by having an understanding of what bits of the drawing should go next to each other.

I don't know enough about ML to compare and contrast the different models, but my understanding of Stable Diffusion is that it's a denoising tool. It was trained by taking image-string pairings, adding noise to them, and then learning what ways of denoising cause it to get closer to the original image. Then in image generation, it starts off with just random noise and denoises it in a way that matches the prompt.

In that sense, I'm not sure it's accurate to say that it "understands" what bits of the drawing should go next to each other. If I tell it "woman wearing red shirt sitting on a brown chair," it doesn't "understand" which bits of the drawing should be a woman, a shirt, or a chair, and it doesn't "understand" that the shirt should be red and the chair should be brown. It just "understands" that the entire picture gets somewhat closer to the entire prompt when it gets denoised a certain way.

Isn't that the exact same thing that Stable Diffusion does?

Inserts that pirate meme. Well yes, but actually no.

There is world of difference between "Based on my training data, sentences containing the word "chair" will also contain the word "sit" ergo my output should as well" vs "a chair is sit upon". The latter sort semantic link has long been viewed as one of the capital-H hard problems of programming a truly general AI. A problem that stable diffusion actually seems to be on a path to solving which the autoregression models that underpin GPT and it's various offshoots do not.

/images/16757438522129903.webp

But GPT-3 clearly has that understanding. I mean, obviously not always, but also obviously sometimes. By and large, GPT-3 does not actually tend to assert that chairs sit on people.

I don't think it's clear at all. A chair sitting on a person is exactly the sort of slip up that typically gives AI generated text away.

I think it makes those kinds of slips, which to me just means it has imperfect understanding and tends to bullshit. But it doesn't universally make those kinds of slips; it gets chair-person type relations right at a level above chance. Otherwise, generating any continuous run of coherent text would be near impossible.

It would be exceedingly strange for it to generate "the chair sits on the person" at the same rate as its converse, considering that "the <thing> <interacts> the <person>" is vanishingly rarer in its training corpus than "the <person> <interacts> the <thing>". But that sort of generalization requires some abstract model of "thing", "person" and "interact". For it to not pick up that pattern would be odd - why would that be the pattern that stumps it, when it can pick up the categories just fine?

We're not looking for a "better than chance" guess though. We're looking for evidence of an understanding that goes beyond "object-noun verb subject-noun" which for the moment at least does not appear to be present. GPT-3 can string words and sentences together but within a paragraph or two it becomes clear that it is not conveying any meaning, it's just babbling.

To expand my point, I think there is a smooth continuity between "babbling" and "conveying meaning" that hinges on what I'd call "sustained coherency". With humans, we started out conceptualizing meaning, modelling things in our head, and then evolved language in order to reflect and externalize these things; we (presumably) got coherence first. AI is going the other way: it starts out swimming in a soup of meaning-fragments (even Markov chains learn syllables), and as our technology improves it assembles them into longer and longer coherent chains. GPT-2 was coherent at the level of half-sentences or sentences, GPT-3 can be coherent at levels spanning paragraphs. It occasionally loses the plot and switches universes, giving up on one cluster of assembled meaning-fragments as it cannot generate a viable continuation and slipping smoothly into another. But the "sort of thing that it builds" with words, the assemblage of fragments into chains of meaning, is the same sort of thing that we build with language. It's coming at the same spot (months/years-long sustained coherency) from another evolutionary direction.

You may argue "it's all meaningless without attachment to reality." And sure, that's not wrong! But once the assemblage operates correctly, attaching meaning to it will just be a matter of cross-training. (And the unsolved problem of the "artificial self", though if ever there was a problem amenable to a purely narrative solution...)

I disagree.

Can you give an example that you think illustrates your point well? (I don't have ChatGPT access. Giving out my phone number? Ugh.)

A few moments ago, while looking for a quote by James Baldwin*, I turned to Chat GPT for help. I used the prompt, "...It describes his anger towards the white man and his interest in white women.""

It gave me the following quote:

"No black man has ever been able to seriously consider the white woman without having to grapple with the ancient myth of the wide-eyed, agile and demanding Eve, who offers him the poisoned apple of forbidden sexuality, the apple of his own destruction." - James Baldwin.

As far as I can tell this quote was fabricated wholesale. A God of words is being birthed, and conscious or not Ze will change the world entirely.

  • This is the quote I was looking for:

"And there is, I should think, no Negro living in America who has not felt, briefly or for long periods, with anguish sharp or dull, in varying degrees and to varying effect, simple, naked and unanswerable hatred; who has not wanted to smash any white face he may encounter in a day, to violate, out of motives of the cruelest vengeance, their women, to break the bodies of all white people and bring them low, as low as that dust into which he himself has been and is being trampled..."

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