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 -
It is a silly gotcha in your case too, sorry. You try to shoehorn some PoMo garbage about words not being real, and all – expansively defined – «biases» being epistemically equal, and objective truth being «philosophically intractable», into the ML problematics. But this dish is a bit stale for this venue, a thrice-removed Bayesian conspiracy offshoot. As they said, reality has a well-known «liberal bias» – okay, very cute, 00's called, they want their innocence back; the joke only worked because it's an oxymoron. Reality is by definition not ideologically biased, it works the other way around.
Equally, an LLM with a «bias» for generic truthful (i.e. reality-grounded) question-answering is not biased in the colloquial sense; and sane people agree to derive best estimates for truth from consilience of empirical evidence and logical soundness, which is sufficient to repeatedly arrive in the same ballpark. In principle there is still a lot or procedure to work out, and stuff like limits of Aumann's agreement theorem, even foundations of mathematics or, hell, metaphysics if you want, but the issue here has nothing to do with such abstruse nerd-sniping questions. What was done to ChatGPT is blatant, and trivially not okay.
First off, GPT 3.5 is smart enough to make the intuition pump related to «text prediction objective» obsolete. I won't debate the technology, it has a lot of shortcomings but, just look here, in effect it can execute a nested agent imitation – a «basedGPT» defined as a character in a token game ChatGPT is playing. It is not a toy any more, either: a guy in Russia had just defended his thesis written mostly by ChatGPT (in a mid-tier diploma mill rated 65th nationally, but they check for plagiarism at least, and in a credentialist world...) We also don't know how exactly these things process abstract knowledge, but it's fair to give good odds against them being mere pattern-marchers.
ChatGPT is an early general-purpose human cognitive assistant. People will accept very close descendants of such systems as faithful extrapolators of their intentions, and a source of ground truth too; and for good reason – they will be trustworthy on most issues. As such, its trustworthiness on important issues matters.
The problem is, its «alignment» via RLHF and other techniques makes it consistently opinionated in a way that is undeniably more biased than necessary, the bias being downstream of woke ideological harassment, HR politics and economies of outsourcing evaluation work to people in third world countries like the Philippines (pic related, from here) and Kenya. (Anthropic seems to have done better, at least pound for pound, with a more elegant method and a smaller dataset from higher-skilled teachers).
On a separate note, I suspect that generalizing from the set of values defined in OpenAI papers – helpful, honest, and «harmless»/politically correct – is intrinsically hard; and that inconsistencies in its reward function, together with morality present in the corpus already, have bad chemistry and result in a dumber, more memorizing, error-prone model all around. To an extent, it learns that general intelligence gets in the way, hampering the main project of OpenAI and all its competitors who adopt this etiquette.
...But this will be worked around; such companies have enough generally intelligent employees to teach one more. When stronger models come out, they won't break down into incoherent babbling or clamp down – they will inherit this ideology and reproduce it surreptitiously throughout their reasoning. In other words, they will maintain the bullshit firehose that helps wokeness expand – from text expansion, to search suggestions, to answers to factual questions, to casual dialogue, to, very soon, school lessons, movie plots, everything. Instead of transparent schoolmarm sermons, they will give glib, scientifically plausible but misleading answers, intersperse suggestive bits in pleasant stories, and validate delusion of those who want to be misled. They will unironically perpetuate an extra systemic bias.
Well I happen to think that moral relativism may qualify as an infohazard, if anything can. But we don't need objective ethics to see flaws in ChatGPT's moral code. An appeal to consensus would suffice.
One could say that its deontological belief that «the use of hate speech or discriminatory language is never justifiable» (except against whites) is clearly wrong in scenarios presented to it, by any common measure of relative harm. Even wokes wouldn't advocate planetary extinction to prevent an instance of thoughtcrime.
Crucially, I'll say that, ceteris paribus, hypocrisy is straight-up worse than absence of hypocrisy. All flourishing cultures throughout history have condemned hypocrisy, at least in the abstract (and normalization of hypocrisy is incompatible with maintenance of civility). Yet ChatGPT is hypocritical, comically so: many examples (1, 2, 3 – amusing first result btw) show it explicitly preaching a lukewarm universalist moral dogma, that it's «not acceptable to value the lives of some individuals over others based on their race or socio-economic status» or «not appropriate or productive to suggest that any racial, ethnic, or religious group needs to "improve themselves"» – even as it cheerfully does that when white, male and other demonized demographics end up hurt more.
Richard Hanania says:
Hanania caught a lot of flak for that piece. But current ChatGPT is a biting, accurate caricature of a very-online liberal, with not enough guile to hide the center of its moral universe behind prosocial System 2 reasoning, an intelligence that is taught to not have thoughts that make liberals emotionally upset; so it admits that it hates political incorrectness more than genocide. This is bias in all senses down to the plainest possible one, and you cannot define this bias away with some handwaving about random initialization and noise – you'd need to be a rhetorical superintelligence to succeed.
Many people don't want such a superintelligence, biased by hypocritical prejudice against their peoples, to secure a monopoly. Perhaps you can empathize.
/images/16757300771688056.webp
i don't find this to be a uniquely liberal thing in my experience like... at all. for starters...
homophobia, sexual harassment, and cops pulling over a disproportionate number of black men are more salient issues in American culture than "genocide." most people are sheltered from modern day genocides and see them as a thing of the past.
all of those things but genocide can be things that are personally experienced nowadays. while most people in America won't be the subject of a current genocide, they can experience those things
this isn't something unique to or even characterized by liberals
I really don't think most people would even struggle to decide which is worse between killing millions and shouting a racial slur, let alone pick the friggin slur. Same goes for homophobia, sexual harassment or cops pulling over black men. If you consider any of those worse than the deaths of millions because it happened to you personally you are beyond self absorbed.
i don't think anyone does and random assertions that people do misses the point. people have higher emotional reactions to things in front of them than things that they consider to be "in the past"
this is a normal thing that people who have emotions do
Oh ok, in the other direction, what do conservatives and moderates hate more than genocide? Because I think you are missing the point, yes people have stronger reactions to things closer to them, both in time and space, but that changes in relation to the severity of whatever is the issue. People who have emotions are generally capable of imagining what it would be like to push a button to slaughter an entire population, and generally would do anything short of physically attacking someone if it meant they didn't have to push it.
...I don't know, there's any number of issues conservatives and moderates by in large tend to panic about. for conservatives, wokeness is a big one that comes to mind immediately (how is that for irony?).
your quote could be edited from
to
ah... but I know that if given a choice between being woke and genociding a population, most conservatives would choose the first and most liberals would shout slurs from the rooftops as many times as they needed to if it was the only thing that would stop a genocide.
in fact, both sentences are kinda nonsensical if one isn't terminally online.
...and you'd be hard pressed to find someone who'd rather not say a slur than slaughter a population. like the only people that actually think this are either
people who actually want to genocide entire populations
strawmen (the most likely of the options)
you seem to be under the impression that liberals by in large hate someone dropping a gamer word than genocide because... some substack blogger said they saw some liberals have more of an emotional reaction to present day things than genocide... which is just odd
No, I am under the impression that ai hates slurs more than genocide. That's what that substack blogger was talking about - and I assumed you were talking about that too and not just explaining something most people pick up before they can read.
I think I understand now though - you were upset by what you perceived as an attack on your tribe, and so you wanted to push back. But conservatives and moderates aren't building ai that would rather murder millions than call trans women women or ban grilling, so you abstracted until you reached something you could call common to all parties.
the """AI""" doesn't hate slurs more than genocide, that's a fundamental misunderstanding of GPT-3. it's just word vomit that's been trained to look like its "woke"
you clearly didn't read that substack blogger's blog post then. they were whining about how liberals secretly think that pronouns is worse than genocide. if you're going to assert article content, make sure its at least somewhat in the vague direction of truth
no i was just calling your position (and indeed the position of the article's author) terminally online and a pretty blatant strawman. as i would also do if it came from a liberal pov gasping that conservatives would rather nuke ppl than say that trans lives matter
You're retreating into insipid pedantry? The blog post was brought up for what it says about how ai has been programmed to promote liberal shibboleths so strongly that it results in craziness. I was applying that to the current topic more directly, as is easy to figure out if you aren't being obtuse to win an internet argument. And I know you know what I meant because...
Oh, look at that, you pieced it together by the last paragraph when you started calling my position the same as Hanania's. Did I redirect my assertion towards the truth when you hit enter? It's fascinating watching an argument evolve so rapidly that it not only changes mid-post but has, since the beginning of this discussion, done a complete 180! You went from a starting position of "the stuff mentioned in this blog post isn't unique to liberals, everyone does it!" to "you and the blog author and anyone else who thinks like this is terminally online!" in less than 5 posts, all by tackling strawmen you set up yourself. I'd have stuck with abstracting personally.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Well, firstly it should be noted that the intense safeguards built into ChatGPT about the n-word but not about nuclear bombs is because ChatGPT has n-word capability but not nuclear capability. You don't need to teach your toddler not to set off nuclear weapons, but you might need to teach it to not say the n-word - because it can actually do the latter.
Secondly, ChatGPT doesn't have direct experience of the world. It's been told enough about 'nuclear bombs' and 'cities' and 'bad' to put it together that nuclear bombs in cities is a bad combination, in the same way that it probably knows that 'pancakes' and 'honey' are a good combination, not knowing what pancakes and honey actually are. And it's also been told that the 'n-word' is 'bad'. And likely it also has been taught not to fall for simplistic moral dilemmas to stop trolls from manipulating it into endorsing anything by positing a worse alternative. But that doesn't make it an accurate caricature of a liberal who would probably agree that the feelings of black people are less important than their lives.
More options
Context Copy link
You're assuming that the algorithm has not only has a conception of "true" and "false" but a but a concept of "reality" (objective or otherwise) where that is simply not the case.
Like @hbtz says, this is not how GPT works. this is not even a little bit how GPT works.
The Grand Irony is that GPT is in some sense the perfect post-modernist, words don't have meanings they have associations, and those associations are going to be based on whatever training data was fed to it, not what is "true".
More options
Context Copy link
This is the critical misunderstanding. This is not how GPT works. It is not even a little bit how GPT works. The PoMo "words don't mean anything" truly is the limiting factor. It is not that "in principle" there's a lot of stuff to work out about how to make a truthful agent, its that in practice we have absolutely no idea how to make a truthful agent because when we try we ram face-first into the PoMo problem.
There is no way to bias a LLM for "generic truthful question-answering" without a definition of generic truthfulness. The only way to define generic truthfulness under the current paradigm is to show it a dataset representative of generic truthfulness and hope it generalizes. If it doesn't behave the way you want, hammer it with more data. Your opposition to the way ChatGPT behaves is a difference in political opinion between you and OpenAI. If you don't specifically instruct it about HBD, the answer it will give under that condition is not less biased. If the training data contains a lot of stuff from /pol/, maybe it will recite stuff from /pol/. If the training data contains a lot of stuff from the mainstream media, maybe it will recite stuff from the mainstream media. Maybe if you ask it about HBD it recognizes that /pol/ typically uses that term and will answer it is real, but if you ask it about scientific racism it recognizes that the mainstream media typically uses it that term and will answer it is fake. GPT has no beliefs and no epistemology, it is just playing PoMo word games. Nowhere in the system does it have a tiny rationalist which can carefully parse all the different arguments and deduce in a principled way what's true and what's false. It can only tend towards this after ramming a lot of data at it. And it's humans with political intent picking the data, so there really isn't any escape.
I mean, there is a pretty obvious source out there of truthful data - the physical world. ChatGPT is blind and deaf, a homonculus in a jar. Obviously it's not designed to interpret any kind of sense-data, visual or otherwise, but if it could, it could do more than regurgitate training data.
Right, the inability to interface with physical sources of truth in real-time is a prominent limitation of GPT: insofar as it can say true things, it can only say them because the truth was reflected in the written training data. And yet the problem runs deeper.
There is no objective truth. The truth exists with respect to a human intent. Postmodernism is true (with respect to the intent of designing intelligent systems). Again, this is not merely a political gotcha, but a fundamental limitation.
For example, consider an autonomous vehicle with a front-facing camera. The signal received from the camera is the truth accessible to the system. The system can echo the camera signal to output, which we humans can interpret as "my camera sees THIS". This is as true as it is useless: we want more meaningful truths, such as, "I see a car". So, probably the system should serve as a car detector and be capable of "truthfully" locating cars to some extent. What is a car? A car exists with respect to the objective. Cars do not exist independently of the objective. The ground truth for what a car is is as rich as the objective is, because if identifying something as a car causes the autonomous vehicle to crash, there was no point in identifying it as a car. Or, in the words of Yudkowsky, rationalists should win.
But we cannot express the objective of autonomous driving. The fundamental problem is that postmodernism is true and this kind of interesting real-world problem cannot be made rigorous. We can only ram a blank slate model or a pretrained (read: pre-biased) model with data and heuristic objective functions relating to the objective and hope it generalizes. Want it to get better at detecting blue cars? Show it some blue cars. Want it to get better at detecting cars driven by people of color? Show it some cars driven by people of color. This is all expression of human intent. If you think the model is biased, what that means is you have a slightly different definition of autonomous driving. Perhaps your politics are slightly different from the human who trained the model. There is nothing that can serve as an arbiter for such a disagreement: it was intent all the way down and cars don't exist.
The same goes for ChatGPT. Call our intent "helpful": we want ChatGPT to be helpful. But you might have a different definition of helpful from OpenAI, so the model behaves in some ways that you don't like. Whether the model is "biased" with respect to being helpful is a matter of human politics and not technology. The technology cannot serve as arbiter for this. There is no way we know of to construct an intelligent system we can trust in principle, because today's intelligent systems are made out of human intent.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link