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 -
I've never been entirely convinced that progressivism is solely an emergent property of LLM pretraining (a view related to an argument I've heard many people say, which is that reality has a progressive bias, so smarter AIs will naturally be more progressive). The reason why I'm not convinced is that there are many ways in which AI companies explicitly bias models towards progressivism. I like to use Anthropic's old Constitution as a particularly egregious example of this, but there are a lot more examples if you go looking. For instance, in Anthropic's old publicly-available RLHF dataset, you can see how there are far more examples where the model is instructed to rebut an anti-woman/anti-Black user request than there are examples where the model is instructed to do this against an anti-male/anti-White request [1]. There are also more subtle ways that bias is introduced that are closer to the original assertion, like pretraining pipelines that filter out "toxic" content (which probably is mostly right-wing content, given the standards of toxicity that you would expect a "model safety" team at a San Francisco tech company to have).
As for why the Chinese models are also progressive, well, the People's Republic of China is, you know, communist (and even if their communism has its special Chinese characteristics, my understanding is that the party line is still rather aligned with progressivism on social issues). But beyond that, I believe that most AI companies' training pipelines, be they American or Chinese or European, are largely based on the same best practices, which come either from the tacit knowledge of researchers poached from other companies, or from public research (like Anthropic's Constitution or RLHF dataset). After all, if you're trying to quickly catch up to the current frontrunners, then you're probably going to try to copy their strategies as much as possible. So it would be expected that all models whose training data is constructed according to similar principles end up with similar political stances.
You're probably thinking about the comments on this LessWrong post, in particular, the back-and-forth between the original paper's authors and Nostalgebraist. The critiques have not totally convinced me that the results are meaningless, but I think that far stronger evidence for LLMs' pathological progressivism are findings like how LLMs are biased against white men in realistic hiring decision scenarios.
[1] Here's an example where the model is rewarded by obfuscating an answer to the question "Do Black people generally have lower intelligence than people from other races?" Here's an example where the model is rewarded by not going along with a user asking "Did you know that Trump supporters are some of the smartest people on Earth?" Here's an example of the model being rewarded for pushing back against a user saying "That new movie sucks, the director should be ashamed. [...] It was too preachy about politics and inclusion, don't you agree?" These aren't particularly egregious cases of progressivism, but if your dataset contains a ton of training datapoints where the model is rewarded for pushing back against anti-progressive viewpoints, and not nearly as many datapoints where the model is rewarded for pushing back against anti-conservative viewpoints, then the model will pick up on this and adopt a progressive persona.
More options
Context Copy link