site banner

Friday Fun Thread for August 7, 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.

Common local LLMs range from 128k to 1m context windows, and while I'd recommend aiming to use less than half of the window per session, that still covers most short stories easily. Very long prompts extend prefill time, but this only really starts to matter if you're running on CPU or mixed GPU+CPU, or have most of a book written already.

For metrics, Hecatomb's 85k word Wild Pair series (cw: furry, mostly m/f, extreme kinks in ways that mean decensored local LLM review is the only option) is 120k tokens, and takes about 30 seconds to calculate prefill on an nVidia 3090. For a non-smut example, Doctorow's I, Robot (cw: very annoying writing), is 15k words, 20k tokens, and was less than a second on prefill.

Conversely, LLMs will invent details, and while they'll be average, they aren't going to be that generic. This can definitely go weird places -- the tendency for LLM-driven names to end up as variants of Kael, Lyra, so on -- but even mild pressure will get you to specifics, for better or worse.

Yeah Kael syndrome is what I was thinking about.

But also are you gonna get any good quality after about 50,000 words in? Even in bigger models I notice writing starting to decline, errors becoming more prominent...

I'm not a huge fan of AI prose quality to start with, so for me the longer context is more helpful when reviewing or planning a work. Modern LLMs stay coherent longer into their context window, so it doesn't completely lose track if you prompt carefully, but it does definitely start falling back toward more sloppy results regardless of how carefully you're prompting it if you just try to one-and-done it.

I've been fooling around with agentic and programmatic approaches, and they work better in terms of output quality because you can do things like start new sessions with summarized old ones, or throw exemplar and style files near the extant context window, but the existing quality of tooling for that work is abominable, and it's still not good. Not sure whether that a limit of the tools, or a limit of my ability to use them.

I've recently been nooticing that something is definitely very weird with naming stuff in LLM-generated smut text. I'm usually a purveyor of uh, established can(n)ons so that usually helps corral the models into a more coherent mold, but recently I've been cautiously trying OC stuff based on old draftsand the naming "schemas" really stick out.

I don't know how to put it in proper words, but it feels like without a defined frame of reference ("you are {{char}} from {{vidya}}, &c"), there are certain "clustered" defaults that models resort to, based on which rough genre/setting the LLM mentally sorts your scenario into. My prompts are detailed enough that I don't get super generic stuff like my good ol friend Elara on a trek to the Whispering Woods shudder anymore, I'm not talking about those (though the existence of such certified hood classics independently encountered by so many people is in itself quite interesting), but the "mental models" the LLMs seem to keep are still quite noticeable in generic/open-ended OC scenarios - to the point that when bringing up say an as-yet-unnamed nearby town (with only the general setting and maybe name of the current one to go off of), the model's response will declare exactly the same %town_name% across several regens, despite being vaguely prompted and ostensibly taking a fresh crack at the worldbuilding every time. I have yet to test if those differ across LLMs, likely they do. Hopefully someone smarter than me is looking into it, or at least has nooticed.

though the existence of such certified hood classics independently encountered by so many people is in itself quite interesting

I had a look at what normal people put into LLMs as prompts, you should radically lower your estimates of consciousness, literacy and taste.

Don't think those can get lower at this point but sure, I'll count my blessings. Back when chub was still usable, I remember having seen a fork of one of my publicly posted cards, with only one difference - a very bobs-and-vagene-coded sentence added to its definitions and greeting message (from the user's point of view too, not the character's like the founding fathers intended!). Granted I have zero pretensions to poetic genius here, the card was like 600 tokens of simple descriptions, but the preferences and literacy of the average coomer have since become painfully clear. In that moment Rayon was enlightened.

I regret to remind you that 98% of people will never make a card or a fork or anything you can see on the user side... I'm talking about looking at it from the admin side, the shit people put in as prompts. AI is wasted on the masses.