site banner

Culture War Roundup for the week of July 13, 2026

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.

2
Jump in the discussion.

No email address required.

The biggest blockers would be radioactively-hot CW-content that provokes a safety classifier or flinch reflex. That is nothing that can't be prompted away.

This is my biggest concern. I wouldn't trust any of the woke AIs to mod The Motte even if they had the technical capability to. Grok, maybe, but that's not a SOTA model.

My second biggest concern is technological dependence. If we built a forum based on active AI participation, what happens when the companies pull the models? We would have to use open source AIs to be safe, but those are a year or two behind proprietary models.

open source AIs [...] are a year or two behind proprietary models.

This is not true. As long as distillation works, open source models could potentially keep up. See Z.ai's GLM 5.2 is Claude Opus-level.

Not what I see.

Since this thread is about forum moderation rather than coding prowess, let's look at the text leaderboard on Chatbot Arena. The strongest open source model is glm-5.1 at 25th, followed by mimo-v2.5-pro at 32nd and glm-5.2 (max) at 33rd. The 24th strongest model is claude-opus-4-5-20251101-thinking, which as the name suggests came out in November 1st of 2025, 20.5 months ago.

Seems to me like "a year or two behind proprietary models" is accurate.

That leaderboard looks super sketch to me. Opus 4.6 better than gpt-5.6-sol-xhigh? It's definitely not been my experience that glm-5.2 is worse than Sonnet 4.6.

The text leaderboard is completely useless; it has much more to do with writing style and formatting when answering banal questions than any serious difference in capabilities. For any reasonable definition of being capable of "forum moderation" pretty much any LLM will saturate that definition.

That being said I agree it is an extremely horrible idea to leave any sort of moderation to any LLM, no matter how capable it might be in theory.

Please don't do this.

You don't need a SOTA model. I've definitely seen how well Gemini 2.0 Pro does that at job, and probably Gemini 1.5 Pro.

At scale, it would make sense to run our own open-weights model. The Chinese make them cheap, cheerful, and good-enough quality. There are "uncensored" or abliterated models falling off the back of a truck on HuggingFace.

The MVP would be something that can filter out the real dogshit, or at least flag it and reduce moderation load. Include the constitution, some worked examples (particularly tricky cases), and so on? It will work. No real need for extensive fine-tuning, the models have long been smart enough to figure it out on the fly.

Could an open source model solve the problem of radioactive CW threads triggering a filter? After all, if the model is open source, would that not mean it is possible to fine tune it so that it fits with the spirit of themotte?

It could. There are uncensored models out there that will at least attempt to do anything you ask of them. I haven't kept close tabs on what the most powerful/useful one is, at the time of writing. That's a moving target, but I keep stressing that we don't even need the latest and greatest of matrix multipliers.

This isn't a particular blocker. Even the frontier models would be adequate at the task. It would be possible to set up things at the backend such that a refusal or safety discard would simply mean that a particular comment/post gets moved to the "human approval" queue. In fact, if something is so radioactive that it absolutely won't be engaged with by a typical LLM in prod, then it likely warrants human intervention. We might let it through anyway.

As I've said before, somewhere, at some point: fine-tuning is usually a waste of time. You don't need it. You can probably fit every single moderation decision ever made on this site into <1 million tokens. You only need somewhere between a few thousand or tens of thousands of tokens to enable a decent model to learn in-context. To be clear, I'm using the formal-ish definition of fine-tune, which is a form of post-training.

Example: collect every moderation decision with more than 20 upvotes. That's a strong proxy for both being mod-approved, and for being a decision supported by the community. Can't be more than a thousand of those. Go to >40 and there's probably a few hundred. That is a degenerate but perfectly functional approach.

You'd do it with a harness and some md files surely. Have it go through all moderation decisions and distill the general rules into a file or set of files, keep all the base decisions on disc but only pull them into context when relevant. hobbyhorserule.md need only be parsed when the model thinks a post might violate that rule. Basic harness building stuff.

You could do that. I dare say you should do that. But even the no-effort option would work adequately.