I’ve run across a heartbreaking tale. Bevy is an “engine” that simplifies the process of creating a video game. It’s open source, so anyone can contribute, and has a decent number of contributors. One month ago, the project changed their policy on LLM usage from a blanket ban to one that allowed “careful human-driven AI use in some code-related areas consistent with” the project’s values.

Community cohesion has fallen apart under pressure of persistent trolling and conflicting community values. …

People have had their own positions viewed as merely a prefigure to “witch-hunts” and harassment campaigns. There isn’t trust that our positions are held in good faith, instead that we will merely lash out in existentially threatening ways that do not “address the reality” or “have real arguments.” This has extended to the point where merely asking if a PR is LLM Output gets push back, despite disclosure being a part of the current policy.

How could such a tiny change, one that seems “reasonable,” cause such destruction?

Street Fighting

Ever seen or been in a debate that never seems to end? In theory, these should not happen: as more and more evidence piles up, it becomes more and more obvious which viewpoint has correctly assessed the situation, and everyone gradually wanders into a consensus. There must be some other mechanism in play.

That, I think, is that most of these endless debates are not over evidence, but values. Explain to me which ethical theory is “better,” consequentialism or deontology? There’s no amount of evidence you could pile up that would tell you “judging things by their consequences is superior to judging them according to shared rules.” These are axioms and assertions, not claims about state of the world, so there’s no way to use evidence gathered from the world to weigh them. You cannot rank values, for the same reason you can’t rank human rights.

Alas, reality refuses to colour within the lines we’ve drawn. In practice, we sometimes find ourselves in situations where our values lead to contradictions or conflicts. In theory, this is impossible, so until we rip out the old theory and replace it with a new one we’re left with no template on how to handle the situation. Being fallible and finite, we almost always become consequentialists instead and try to quantize human joy and misery, siding with whichever value leads to the most favourable ledger, then put the old theory back in place and pretend nothing happened. That’s quick and easy, but not how this should be done and it’s easily gamed by our prejudices.

Say I point to extensive documentation of labour exploitation by the companies training LLMs, from 2023 all the way to this year. If you value the quality of life of others, this argues you should stop using LLMs. But what if an LLM has made a strong material impact on your life? What if it’s become your councilor, your guidebook, your companion, friend, or even lover? Your life would be made much worse by giving up on LLMs, so if you value your own quality of life you should continue using them.

Even if I were in your shoes, I wouldn’t solve this by reinventing my ethical framework. No, on one side of the scale I’d place the immense suffering caused by this exploitation, plus the arguments and evidence to back up that claim; on the other I’d place the benefits I personally get from LLMs, including arguments and evidence. Currently, I see zero benefits; I’ve been doing all the tasks these things claim to do, myself, for decades, without issue. On top of that, every LLM I’ve encountered will fail miserably if I’m lucky, or produce something that looks plausible on the surface but falls apart under closer scrutiny if I’m not. Even if that were to change, do I think any benefit could ever outweigh the harm caused by exploitation? Nope. My scales don’t come anywhere near close to a balance, even in a hypothetical utopian future, so I rank “minimize exploitation” above “personal benefit” and put the old theory back. As a consequence, I’d happily endorse a blanket ban on general-purpose LLM use.

Of course, I never really put myself in your shoes. How could I, without ignoring the evidence in front of my face? No doubt you also pulled out your own scales, and in your case the immense benefits you’ve experienced far outweigh some tale of woe from Kenya. What’s your first instinct going to be, once learning of my support for a ban? To tell me about those benefits, to encourage me to bring out my scales again. But why would I? You’ve haven’t given me any new information here, I already took your beliefs into account.

What [the “but LLMs work for me”] framing ignores is what it does to the people involved in the project. People who embrace these LLMs are encouraged to view people who don’t as thoughtless luddites or annoying puritans. People who do not use these machines (or are forced to, but don’t want to, a population that makes up a significant chunk of “AI users”) are discouraged, demotivated, and devalued. A feeling of personal worthlessness sinks in and is actively encouraged by ignoring the social part of the equation of production.

If I toss a tonne of stories about the exploitation involved in LLMs in your direction I’ve only proven I’m not really listening to you, I don’t care about what you value or your life experiences. It’s quite reasonable to ignore and devalue me in return. On the flip, telling me stories about how much better your life is thanks to LLMs leads me down the same path. There’s no way we’ll reach a shared consensus, with those attitudes.

If we can’t reason each other to a consensus, though, then the only actions left on the table are the unreasonable ones.

“The Nazi Bar problem” describes what happens when people caustic to a feeling of safety, collaboration, or community try to take root in a space … The point is that caustic, antisocial behaviour causes people who previously were valued in the space to leave. When these people have left, the behaviour of the persons whose caused them to leave becomes the new community norm. This is distinct from usual community churn, normal conflicts. …

I bring this up to describe the way the Bevy community has changed radically over the last 4 months or so, I have seen people who were previously valued members of the community get pushed out due to caustic behaviour. I’ve seen the people who pushed them out be polite once they got their way, their previous behaviour ignored.

Continually shouting past one another drives out the casuals and only leaves behind the most opinionated. As each side sees more and more evidence the other doesn’t care, they drop their gloves. The community descends into a street brawl, until one side gains sufficient power to drive the other out.

This author states a “troll might, when presented with this framework, argue that the people they’re trolling are the real caustic members of the community,” but the troll is optional here. I have been assumed to be trolling when I thought it was obvious I wasn’t, I have been accused of being “caustic” when I was trying to be polite. Nobody’s objective or omniscient, we easily read what we want to into others, nor am I so arrogant as to think I’ve never accidentally been caustic or assumed the worst of someone else despite my intentions.

The Terrible Cost of Listening

While this may seem depressing, did you spot the hole in my analysis? Past me certainly did:

I’ve been in an online space with an unusual number of LLM fans. I thought it would be a great chance to learn their point of view, and do a little push back while hearing their best arguments.

It’s been eye-opening. I thought posts about the human costs of LLMs would at least cause them to pause; I’ve never seen them acknowledged. Post some fail related to ChatGPT or OpenAI and they’ll chuckle or even agree, but _they_ use a different model and so it doesn’t matter.

The only thing that’s made an impact is demonstrating their favourite LLM is embarrassingly flawed. That reliably produces angry dismissals and nearly gets me banned.

The lesson: appealing to their humanity does not work, humiliating their machine god does.

Telling me you gain immense personal value from LLMs doesn’t tell me anything new. Bringing me evidence that the amount of human exploitation has been exaggerated does, and is far more likely to have me reaching for my scales again. To avoid the aforementioned quagmire, you don’t target the arguments/evidence for the value that’s been ranked lower but instead the one that’s higher. In the process, you’ve also demonstrated that you’ve been listening to what I’m saying. Reason is back on the table, and my gloves remain over my hands.

A scan over my posts from the last two years shows that I’ve been instinctively deploying this tactic for ages. You think these LLMs are a huge personal benefit? Fine, then I’ll focus my analysis on that claim. Along the way I’ll also follow my usual online debate rules: avoid getting heated or personal, ignore personal attacks directed at me, and always break off the argument if your opponent signals they’re not interested.

That plan sounds perfect, but it got me banned from the online community I mentioned.

Well OK, it probably got me banned. It’s been a minute, but as best I recall my second-last argument there was a follow-up to that one over intelligence. My opponent was now confident they’d found a solid point to push on: “intelligence” is actually a composite entity, not a unified whole. Therefore, claiming there is no single “g” factor no longer conflicts with the claim that “intelligence” exists. Alas, they had no idea my first extra added to that blog post was a counter-argument directly targeting that, because I first encountered that move… easily a decade ago, maybe even three? And so I typed up a response that tried to appeal to them, with a subtle poke in the ribs added for my amusement. I was delighted to be on the receiving end of claims that I was obviously confused and didn’t know what I was talking about. Nerve: hit.

The last argument followed almost immediately after, when I argued OpenAI could be using their “router” technology to boost their benchmark scores in much the same way Volkswagen fooled emissions tests. The same person retorted that ChatGPT doesn’t use a router. They provided no citation, so I did some research and the only system card I could find for ChatGPT 5 was for the initial release, and it clearly mentions the router. ChatGPT 6’s system card doesn’t; maybe OpenAI stopped using the router, or maybe it’s so baked in there’s no need to mention it? I shared the ChatGPT 5 link back and pointed to the obvious mention of a router, only to be informed the router had been removed in version 5.2 (again, with no citation). Still, with that clue in hand further research was able to find that user backlash and increased costs led OpenAI to quietly retreat from the router for users that didn’t have a subscription; paid users still defaulted to the router. They never mentioned this in their announcement for 5.2, of course. I shared this info back, with citations, and was met with an explosion. This person was sick of my continual lies, and from that point onward was going to block me.

I “always break off the argument,” buuuut this person just name-dropped “MoE” without explaining what it meant. And so my last message was directed to the rest of the community, reiterating my original claim about potentially cheating on benchmarks via the router, while also describing Mixture of Experts and how it provided another possible avenue for OpenAI to cheat. All of this is from memory, because a few hours after hitting “post” I could no longer find this community, let alone read my old comments. It took some snooping via developer tools to confirm I’d been banned. I was never notified, let alone given an explanation.

To be clear, I have no problem with the ban.

“The Nazi Bar problem” also has a solution. A solution which is quite easy to explain:

1. A community has a space (The Bar)
2. A “nazi” turns up to the bar, antagonises the regulars.
3. Divergence: The bar manager, who has control over who can enter the space, kicks out the “nazi” and bans them + any of their associates.

Communities are defined by a shared set of values. It did not take me long to notice an LLM booster was a mod, teeing up a potential value mismatch. Well before I was getting warnings by them and others, their replies made it obvious I’d probably get banned. The only real question was how many polite arguments I could get into before that happened. To that booster’s credit, the answer was “quite a lot.”

Intellectual Guerilla Warfare

If you read my original post closely, though, you’ll have noticed I wasn’t the only person there with a value mismatch. I’d even argue some other people in that community were no less opposed to LLMs than I was. Why weren’t they ejected as well?

Let’s turn the tables. I logged into that community one day, and spotted a booster had posted a carbon usage calculator like this one. This was my tactic being pointed back at people who’d weighed their values in a similar manner to me, and the results were wonderful. A number of people had posted rebuttals well before I arrived, and the booster had put up a valiant defense. No insults were being hurled around, no ill-will could be detected. The brief mention of the environmental argument over here had been inspired this event.

It’s no surprise that people who often posted about the environmental impact of LLMs knew something about the topic. Nor that they’d be happy to discuss their knowledge. Not only is it an excuse to share, but the more they talk the more chances someone else has to critique their knowledge.

I don’t drop citations to look smart, I do it so you can see where I’m getting my information from and judge how accurately I’m representing it. I didn’t write 10,000 words about LLMs and philosophy merely because I wanted to. I laid out my arguments as cleanly as I could, in sufficient detail to make them easy to critique and analyze, so we could reach a shared consensus together. I was literally begging for a dialog!

And in return, all I ever got was this.

You always have to leave room for a little grace; maybe I didn’t explain things clearly, maybe my arguments were trash, maybe maybe maybe. But while I was in that community, I noticed a pretty consistent pattern from the boosters: don’t cite your sources, force your opponent to track them down for you. Alternatively, do link to your source but don’t say what you find notable, forcing other people to read through it and guess your intent. Don’t defend your posts unless you’re certain you’ll come out victorious, just go silent once there’s any pushback. Wear down your opponents by repeating the same arguments over and over.

That wonderful event in that community, where one or more LLM boosters engaged in the primary argument put forward by those opposed to LLMs? It only happened one or twice within half a year. They must have known this was a load-bearing pillar, they knew poking at it would reliably provoke a response, and that response would not be hostile. Yet, with a handful of exceptions, what I instead witnessed over there was a form of intellectual guerilla warfare. My favourite example: in response to someone claiming LLM “skeptics” never based their critiques on the latest models, I pointed out there was evidence to the contrary and linked to a post I’d made a month earlier. That post was a response to the same person making the same argument, pointing out evidence to the contrary by linking to one of my prior posts. That post shared a fresh critique of a just-released LLM, in response to the same person hyping it up!

If you’ve read this far, you’ve probably realized the comment that inspired this post was much, much shorter. Alas, that original comment was quickly forgotten; within a day, the person who was once scrambling to find some way to push against it was back to thinking it was impossible for anyone to deny LLMs were becoming more “intelligent.” A closed mind is hard to open.

But if you never try to open it, you guarantee it’ll remain closed. Even if you fail, sometimes you’ll learn something from the attempt.

The lesson I learned? LLM boosters know their arguments are weak, at least on some level. You don’t engage in intellectual guerilla warfare when you’re confident and self-assured.

We’re running on two parallel tracks, never to meet. Fire off what you think is a solid argument for widespread LLM usage, that they’re far superior at relieving loneliness than the alternatives? I’ll begin work on a four-point-five part series which meticulously walks through the scientific record poking holes in the original post, with at least one counter-argument per part, plus some tales from the trail as a bow on top. I’m working on a four-point-five part series about loneliness, and the third part went public within the last day?

this was just sidebar vandalism; i don’t have much of a point to make. though i don’t feel like my pinned post has been meaningfully addressed by anyone either, so what should i have to make a point about?

I post about LLMs and cognition; nine months later you post about LLMs and cognition; and the latter makes no attempt to connect to the former, let alone critique it. I argue that “there’s no such thing as “intelligence”;” you respond by arguing I’m “singling out human-style intelligence as the only kind that matters.”

A constructive dialog is the last thing you want when your arguments can’t stand up to scrutiny. But that thing where we shout past one another, getting more and more angry and dismissive, until one side pushes the other out of the community? The quality of your arguments is irrelevant there! Other people opposed to LLMs in that online community I mentioned didn’t have the confidence or knowledge to tackle the arguments supporting the higher side of the value scale for LLM boosters, so they stuck to the arguments they did know. But they were also inadvertently acting like controlled opposition, propping up the beliefs of the LLM boosters while giving the illusion of diversity of thought. There’s value in keeping those people around, so long as they remain a minority and don’t push out the LLM boosters.

As for me? No value there.

Two Narrow Paths

This doesn’t have to be an intentional pattern, either. The arguments I had with LLM boosters in that community produced a flurry of posts and a wall of words, which grated on people who supported me. I never saw evidence that it changed minds on the LLM booster side, and even I can get tired of yelling at a wall. Had it changed the minds of LLM boosters, then those people would no longer be boosters, now would they? Conversely, it is very easy to mis-diagnose the discomfort you feel over an argument, and fish for excuses that allow you to ignore it.

If I am correct about the quality of their arguments, then opening the door for LLM boosters, even just a teeny bit, will naturally lead to a dysfunctional community. That only stops once it schisms in two, with one new community dominated by LLM boosters and another that’s booster-free. If that latter one eventually decides to loosen up a bit, the process repeats. In the end, after an extended period of misery, every community either bans LLM boosters or embraces them.

Who could have predicted that consistently prioritizing your own needs over those of others could have such negative consequences?