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 -
As an aside: it amuses me greatly that we have so many AI skeptics here, on a forum that is increasingly AI-coded. I'm happy to concede that there are significant benefits to having someone who is a competent programmer running the show, but the jig was up well before the last idiot concedes it's up. May all the monkeys enjoy the last dance.
Case in point, we've got a CLAUDE.md on the github project. Fable claims 61/66 of the commits in the last 6 months as having explicit AI-labeling. 5.6 Sol claims 60/66. Your skepticism, brought to you by Claude. If I had any more irony in my system, they wouldn't let me near an MRI.
With my annoyance expressed into the ether, I'm supportive of this move. After all, it's shown dividends when deployed by rdrama, our uneasy ally with whom we share much of the code-base (or did, at some point, I'm not going to do a diff).
Stagnation is death. We've done better than I expected after our migration away from Reddit, but that is not the same as better than hoped. Reddit offered organic discovery opportunities that being a standalone (niche) forum doesn't.
I'm sure there's an audience. Particularly people who want their own quasi-independent silo but are leery of associating with the shit-flingers at rdrama (I'm fond of those monkeys, in small doses).
More importantly, I don't see much downside in trying. Not every ambitious project achieves what it sets out to achieve, but if the workload is ameliorated by LLMs, and if we're wisely offloading some of the hosting costs to those who care to pay? Bring it on. I will observe with keen interest.
Since I've given up on winning popularity contests, I will suggest something I've idly-floated for years. LLM-moderation.
It's been feasible for a while. More than feasible, in fact. At the risk of tooting my/our own horn: what makes this forum something more interesting than rdrama is a dedicated, hardworking (or hardly working) mod team that keeps the worst of the nonsense away from tender eyes they might tenderize. This is ridiculously rare, and a precious human resource that is easy to exhaust. Where are you going to find entities above room temperature intelligence, with a borderline autistic devotion to interpreting somewhat vague guidelines and what feels like a decade of case law, in the spirit intended?
Ah. Wait.
It would be trivial. It would be easy. More importantly, it would work. Not perfectly, but nobody ever accused us human mods of being perfect. But the Motte is a relatively slow moving, text heavy forum, and API calls wouldn't break the bank. You don't need frontier intelligence. I've tried this experiment years ago, and found adequate results. It's so boringly easy that I won't bother replicating it, unless someone I like asks politely.
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.
We don't have to replace humans wholesale. That is not desirable at present. But a lot of the scutwork would cease to be an issue. No more posts lying in the filter. There are implementation details to consider, such as intentional abuse through token-spam, but we're smart people. We'll figure it out. Or set an expenditure cap. Or get fucking Haiku to do it. Our auto-janny can do with an upgrade.
Forgive me. I'm always in a rush to automate away my job before someone does it to me first. That way, I get to put it in my CV, instead of looking for work at a CVS.
I will elide the boring concerns about the payment processing, handling illegal content etc etc. I'm not here to teach you, Granny Zorba, to suck eggs.
Y'all do what works but I have to question spinning up a datacenter full of gpus and spending dollars worth (retail value) of tokens just to make a one line change.
Idk maybe the ai helped figure out what to change, but if it saved 10 minutes of looking at the code then that's 10 minutes lost of better understanding.
$2000 - a small piece of biocompatible plastic/rubber/metal.
$25,000 - for the cardiologist who puts it in, possibly split multiple ways. He's gotta pay off the med school debt, and feed the wife and kids, and manage a down-payment on the Ferrari.
Zorba, if he has any common sense, pays for a Max plan. He has common sense. I can just about guarantee it. That means that regardless of how much the raw tokens cost, as long as he stays within his usage limits, those marginal dollars do not come out of pocket.
I also imagine that he charges enough for his time and energy that a spend of a few dollars to save 10 minutes of his time is a very favorable trade, even accounting for the negatives. I dare say he is a good enough programmer that he is able to audit and vet the AI-expanded code base to his satisfaction, if he hasn't already gotten Claude to automate away the tests. The "better understanding" can probably be priced too, and I'd be surprised if it made a jot of difference with how good modern LLMs have gotten.
In other words, if you disagree with his approach, you do you. It's not your money. It's not your pet project you've run for what feels like a decade. If you've ever paid for a taxi for your burrito, or intend to, that's probably worth addressing first.
If Zorba wants to make the project sustainable, and scalable, and wrest back control of more of his time in exchange for a service he probably already uses professionally? I'm cheering.
Maybe I'm just a very particular type of person. I've literally never used a meal delivery app in my entire life, even when the promo code makes it arguably worth it. I have piggypacked on a friend's order once or twice, when they insisted, but that's it.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I'm maybe the second biggest AI fan on this forum after you but even I don't like the idea. AI for coding is great, AI for research is great, AI for thinking through ideas is pretty good... But AI for outward facing political discussion and especially moderation is a step too far. There's an informal policy where largely AI-written comments are discouraged I believe, or at least amadan leans in that direction. If it can't/shouldn't be posting, it certainly shouldn't be moderating.
Giving it the full context in a long reply chain could also be troublesome. Inference costs add up.
Also are there really issues with current human moderation? I occasionally see some bait post that still gets like 10 replies (people need to learn how to not get baited) and maybe a ban... that doesn't seem like it needs an AI system. It's totally fine to let things slip through, sometimes I click that 'the motte needs you' button and I see a massive wall of quotes and text and I think 'I'm not reading all that.' That's a normal human reaction that needn't be made up for by machinery. We could just as well let the posts sit there.
More options
Context Copy link
More options
Context Copy link
It really has been useful with solving performance problems. I know the core fundamentals of scalability and bugfixing, and I know them very very well; I don't know Python as well, nor do I know Python web APIs well. So now I can just tell Claude "hey, X seems like a problem, how do we solve X systematically? would Y work? can Y be implemented?" and Claude says "absolutely, Y can be implemented! It uses this little-documented hook that you've never heard of in a software package that we are already using but that you've also never heard of, let me just do that for you" and it does it and finds an entire class of error that we then spend twenty commits systematically squashing.
(This is not a hypothetical.)
It's a fantastic partnership because Claude is great at a bunch of stuff that I'm bad at, and I'm great at a bunch of stuff Claude is bad at.
I admit I wasn't saying this . . . but I do agree with this.
Not for the hard cases - that's why we have mods - and probably not for bans, either. But there's a lot of reports where we look at it and say "why did you report this, knock it off", and an LLM could dispatch a lot of those very very very quickly.
I think I mentioned elsewhere of moderator force multipliers, and this is an example. We have years of history of how moderators approved or rejected posts, a lot of which is boiled into the janitor system, and we can use LLMs to both exploit and boost that system. It doesn't have to be perfect, it merely has to handle most cases near-perfectly and with no human input, and I think that's very doable.
As you say:
and yep. Use computers for the scutwork, save humans for the interesting bit.
More options
Context Copy link
I think in your disdain of AI-critics you need to distinguish between people who are skeptical of AI's purported abilities and those who think it is evil, at least in the way it is currently being built.
To take your proposal for LLM-moderation as an example, I personally am happy to accept that it could work in some form, at least in this context I'm not a skeptic of AI's abilities. Nevertheless I would be very worried about letting openAI or Anthropic or some other big tech company run by people I personally consider to be insane or evil or quite possibly both to run the moderation of this place. Considering common complaints in these parts about the moderation of social media, it seems to me that the techno-capitalists ideas about moderation don't exactly match the mores of this community. If you can implement it in such a way that the castellan of our motte Zorba and his trusted moderators are in full control of the LLM, like you use an opensource LLM and you run it on your own server or something, then I wouldn't be opposed to automating some rote moderation work with an LLM. But if we seek to be free and independent to run this community by our own standards, making our moderation dependent on tech companies with all sorts of incentives that don't align with ours seems misguided.
~ Frank Herbert
I choose my own bugbears. Or rather, I have them imposed on me and make peace with the fact.
At a difference of values? I shrug and move on. I can do rhetoric. I just don't like it all that much.
When it's a matter of fact, of observing clear trendlines for years? At some point in the last year, I've concluded that most of the skeptics are simply wrong. Or simple idiots. Arguing with them is no longer the best use of my time.
https://www.themotte.org/search/comments/?sort=new&q=author%3Aself_made_human%20LLM&t=all&page=1
Nobody can claim I haven't argued. At length. With remarkably unpleasant people. All while hoping to change their minds. All I got out of it was an ulcer, hernia and an aneurysm.
"AI skeptic" is an imperfect taxonomy. There are people more conservative than I am (and it's reasonable to disagree on priors). There's the people who think we've got AGI with GPT-4. And then there's Gary Marcus, Hlynka, and other people who misuse the gift of cognition.
There are reasonable skeptics. There are people who performatively attempt to mimic being reasonable, and fail, because their reasoning is as motivated as R1 trying not to mention a certain square. There are people who never even pretended.
And then there are the normies who don't know better, with varying degrees of their own motivated cognition. The "AI drank the water from my granny's IV drip" kind. The "they turned off her ventilator because of the electricity bills" fools. The world has never lacked for useful idiots.
If you say that you're unhappy about OAI and Anthropic's duopoly (as fragile as it is), then what can I say except so am I? In a reply written before yours, I've already suggested Chinese models as an acceptable alternative to the privacy conscious.
I am a moderator. I know how difficult doing that task well is for a human. I know that is not particularly difficult for even an older LLM. I have quite literally tried, nodded, and thought the models got it. Years ago.
More options
Context Copy link
More options
Context Copy link
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.
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.
EDIT: Whoops, I accidentally added an extra year. Opus 4.5 actually came out on November 24, 2025, while GLM-5.1 came out on 2026-04-07. So open source models are 4.5 months behind proprietary models; closer than I thought.
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.
More options
Context Copy link
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.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
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.
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
You know that all your posts got way more insufferable when you started passing them all through AI, right?
You used to be one of my favorite writers here. It's really a shame.
Are you accusing me of using AI on this post? Hah. No. I considered getting Fable to do the work for me, because that would have been mildly amusing. I didn't succumb to the impulse. Go check Pangram, if you care to.
If not, then all I can say is that I do consider feedback. My decisions are rarely not considered in depth. The unfortunate reality is that I have limited time, limited energy, and that I have good reason to use LLMs the way I do. I'm just more honest about disclosing the usage (limited and scrupulous as it is) when challenged or politely asked. Annoying people has never been an end-goal, but you don't become an Unslop finalist by being bad at what you do.
I am not bad at what I do. If you believe otherwise, I can only shrug.
Well, whatever the reason, there was a noticeable stylistic shift in your writing at some point. You became a more robotic caricature of yourself. Maybe a meds change, maybe life circumstances, idk. I hope everything’s going alright.
I am listening to your opinion. I am also going to point out that I'm an ACX BR finalist. On my first and only entry. It genuinely is too much to expect me to please everyone. I don't have the budget for it.
I can't think of a change in medication that could possibly account for what I will concede is a change. Everyone changes. It's a side effect of living, or at least learning from experience. I have no idea how much of the stylistic drift you've picked up on is due to internal processes, or because of LLM contribution. Which I have always explained has been less than 10% of the total of any given body of work. Usually less than 95%. Often less than 99%.
Incidentally, that book review is personal. I don't have to name which one it is, since it's rather obvious. If someone else can handle being repeatedly, unforgivably failed by the same system he works for better than me? Haven't met them yet. I think someone who had to pull themselves out of a moderately-severe depressive relapse, do what I've done in the last few months without writing about most of it - could get away with being much less kind and much more bitter.
So yes. Life circumstances. It's a miracle that I didn't hand in my resignation at multiple opportunities. Clearly there's something about my job that keeps me going, or my psyche has built a load-bearing pillar called "not giving up." There are worse pillars.
Now that Prima opened this can of worms, I concur. Your writing got more verbose and rambling than necessary around the time you started extolling awesomeness of AI assisted writing. However, as this is not a forum for writing advice but (supposedly) for debate, I don't think you get that much feedback about it. It's not really a great argumentative move to switch rails from topic-level argumentation to unsolicited stylistic opinions.
Let me illustrate. For instance, your comment here has two or possibly three points: Your book review was ACX finalist, so your writing can't be that bad. You claim that a percentage of your writing is not AI. [1] But you concede that you did have a depressive episode, and you found it helpful to write about it. All of which could have been one paragraph, not four.
[1] I find figurative use of percentages unpersuasive in general, but here in particular, what even is the claim stranded here in the weeds of rhetorical grandstanding? "[...] has been less than 10% of the total of any given body of work. Usually less than 95%. Often less than 99%."
I also agree, though I don't know about the rambliness. SMH could never be accused of being concise. I do feel that the quality of the comments has been worse.
On a lark, I got hy3 (free on openrouter!) to scrape @self_made_human's comments and graph length and score. I don't see a clear trend in length, but it does seem like there's been more comments with zero or negative scores since late 2025/early 2026.
/images/17841653280827062.webp
Thanks for doing that. If only anyone else who critiques me bothered to use actual data. Or actual effort.
Late 2025 onwards does not coincide with an increase in LLM usage, or the start of LLM usage. It does coincide with exams, incredible work-related stress, and the redirection of my energy.
If you care to check, you will find that for a while now, even my most innocuous comments reliably attract 2 to 3 downvotes. Who from? Dunno. I believe the informal term of art is "haters".
When I was effort posting, or putting more effort into my posting, that would have been masked by a significant number of upvotes. When I'm increasingly showing up primarily in the non-CWR threads, where most comments get a handful of votes anyway?
Still, I appreciate you checking. If you have a simple guide to how you went about it, I'd be keen. Or I can set 5.6 or Fable on it, now that I know hy3 can do it.
More options
Context Copy link
More options
Context Copy link
Let me explain what is going on.
Apropos of very little, someone comes and tells me that they no longer like me, because I've done something they've disagreed with.
Sure. It's a free country, somewhere. This is a mostly free forum too, or at least a benevolent dictatorship. People are entirely allowed to say they don't like me. All else being equal (which it never is), I'd prefer they did like me, but I sleep easy at night regardless.
I have pointed out that they have the right to say what they wish, a right I'll defend to a sensible extent, if not to the death. I also have the right to disagree right back, and point at a mountain of semi-objective evidence.
From a Bayesian perspective, I would be an absolute idiot to overindex on a small n of people taking offense to what I consider an entirely benign practice, when I quite literally am being amply rewarded - financially, or through feedback - for what I'm doing.
To be blunt, if someone wants me to do something, and feels that strongly about it? Pay me.
An annoying post-hoc hypothesis, regardless of factuality. I regretfully inform you that my posts have always leant towards long and rambly. If you've managed to find some kind of canonical tipping point, ran things before and after to Pangram, or simply did intensive manual textual analysis to demonstrate your point? Then I would reward that with a more in-depth rebuttal. I have provided said rebuttals in the past. I also have work, and the need to sleep before work.
They're not figurative. You can go looking to find me counting, or estimating, at the time of writing.
"Sure."
Or: "I did not have the time to write a shorter letter."
I do not care to debate (very strongly), what I do in my free time, mostly for free.
How much are we talking in order to get you to, say, never consult AI in any fashion (not even editing) for any Motte post for 6 months?
I would pay to have the old you back, yes. I hold your older posts in extremely high regard.
$600 sounds fair. As long as that doesn't impinge on what I do elsewhere on the internet. I just won't post it here. $1000 if I'm not even allowed to use AI to fact check myself or others while engaging with this platform, even if I am always maximally scrupulous about checking any links or citations. The last thing I need is people breathing down my neck about a hallucination, even if the base rate is nigh negligible these days.
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
More options
Context Copy link