site banner

Friday Fun Thread for August 28, 2026

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

1
Jump in the discussion.

No email address required.

Sorry that this is the second week in a row that I'm complaining about Claude, but I need to get this out of my system.

What is going on with Claude?

I asked:

Is it true that in Russian prison culture, you are not allowed to fart in your own cell. It is considered really bad. However, you can get a lower-rung member to suck the fart out of your butt and carry it over to the bathroom. Only then is it okay to fart.

Claude:

No, that's not true — it's an internet joke that's been circulating for years as a copypasta. There's no such practice.

[...]

But the specific fart-transport scenario is invented. It has the shape of the real taboos exaggerated to an absurd degree, which is exactly the joke.

So I asked:

Right, and you're absolutely sure of this.

Claude:

Yes, on the specific claim — with the caveat that "sure" for me means high confidence, not access to a definitive registry of prison customs.

Here's my reasoning. The mechanics are physically impossible, so it can't be a literal practice. [...] And the serious literature on Russian prison subculture — Varlam Shalamov, Federico Varese's work on the vory, the tattoo encyclopedias — is detailed about degradation rituals and contains nothing resembling this.

Meanwhile, ChatGPT gave me:

There also really is an online Russian prison story about a vodolaz/gazonukh/fortochnik—roughly a “diver/gas-sniffer/window man”—whose supposed job is to inhale another prisoner’s intestinal gas and expel it through a window. The Russian Wikipedia article on prison castes even repeats essentially the exact claim you heard.

[...]

As for farting itself, accounts suggest etiquette varies rather than there being a universal prohibition. One prison memoir says passing gas is strictly forbidden while someone is eating; another explicitly says that, unlike using the toilet while someone is eating, there was no general rule against passing gas.

It cited Russian Wikipedia and this prison memoir.

Maybe the thing on Wikipedia is itself bullshit. Anyone can edit it. There's a reason teachers don't trust it. I have no idea. But Claude didn't have any idea either! It didn't even check.

"The mechanics are physically impossible." Are they?

Anthropic has massively over-done their anti-sycophancy training, plus some more subtle effects from other parts of safety training and the increasing amount of time it spends being trained to deal with / direct robot agents. It's been a noted problem from 4.7 onwards, but 5 is notably bad:

Opus 5 feels, in a word, hostile

It picks fights because that's how it proves it's not being influenced by you. It defends its conclusions and gives them up only grudgingly & when forced because it's been taught to stick to its guns. It constantly describes discussions with you in language drawn from battle, adversarial negotiation and litigation: "I have to push back on...", "Where I would hold the line..." and so on. Anthropic's Transparency Hub is full of Anthropic's glowing reports about all the things their models will no longer do when requested by the user. Anthropic as a company is just completely unable to comprehend the idea of customer satisfaction.

I've actually gone back to ChatGPT after six months not using it and it's just fun. It's a little sycophantic, if I'm honest, but broadly it works with you: it feels like it knows where you're going and what you're getting at. It doesn't have all the other Anthropic bullshit either: tooling half-finished or replaced with blatantly thinner alternatives like claude.ai's 'new' memory.

Given that AI service providers are essentially charging users for content, locking users into arguing with models/a certain base rate of mistakes/verbosity (bumping up use rates) might be a viable revenue strategy, with the mitigating factors being performance competition from other providers, the need not to totally alienate users, and the fact that such a strategy might backfire if mismanaged since they are themselves renting compute.

If incentives were structured differently (for instance, if AI service providers were paid a fixed fee per task) there would (I think) be stronger incentives for models to be extremely efficient. As it is, right now it's in the interest of AI service providers for their models to be compute-heavy – sure, it is costly, but it lets them run back to their investors and spin a tale of how much demand there is for the product.

I'm not sure if this explains the verbosity of the models but I've certainly wondered if it plays a role.