Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

LLM Writing Tropes.md

tropes.fyi

Post reply on HN
Print view View original

91–100 of 211 posts

Re: LLM Writing Tropes.md

#91

Post by orbital-decay »

capnrefsmmat wrote: post #15

I work on research studying LLM writing styles, so I am going to have to steal this. I've seen plenty of lists of LLM style features, but this is the first one I noticed that mentions "tapestry", which we found is GPT-4o's second-most-overused word (after "camaraderie", for some reason).[1] We used a set of grammatical features in our initial style comparisons (like present participles, which GPT-4o loved so much tha…

I have nothing to contribute but speculation based on my intuition, but IMO RLHF (or rather human preference modeling in general, including the post-training dataset formatting) is a relatively small factor in this, RL-induced mode collapse is much bigger one. Take a look at the original DeepSeek R1 Zero, the point of which was to train a model with very little human preference, because they've been on a budget and human preference doesn't scale. It's pretty unhinged in its writing, like the base model, but unlike the base model it converges onto stable writing patterns, and the output diversity is as non-existent as in models with carefully engineered "personalities" like Claude. Ask it to name a random city and look at the logits, and you'll still see a pretty narrow distribution. At the same time some models with RLHF (e.g. the old RedPajama) have more diverse outputs.

Collapsed mode makes the models truncate entire token trajectories, repeat themselves, and indirectly it does something MUCH deeper, they converge on almost 1:1 input-to-output concept mapping (instead of one-to-many, like in base models). Same lack of variety can be seen in diffusion models, GANs, VAEs and any other model regardless of the type and receiving human preference.

Moreover, these patterns are generational. Old ones get replaced with new ones, and the list in the OP is going to be obsolete in a year. This is what already happened to previous models several times, from what I can tell. Supposedly this is because they scrape the web polluted by previous gen models.

Re: LLM Writing Tropes.md

#94

Post by Kye »

lucumo wrote: post #68
matusp wrote: post #60

I tried using Gemini for some light historical research. It could not stop using tech metaphors. Lords were the CEOs of their time, pope was the most important influencer, vassal uprisings were job interviews, etc. The metaphors were almost comically useless and imprecise, and Gemini kept using them even when I explicitly asked it to not do that.

I think that's Gemini trying to personalize the answer specifically for you. It really leans heavily into that to the point of being galling. You can give it additional instructions in the settings, but you have to be careful with that too. I've put my tech stack and code preferences in there to get better code examples. A while later I asked it about binary executable formats and it started ending every answer with…

The new memory feature in Gemini got turned on by default and every answer came out like this. It kept working in details from one particularly long thread. Everything was framed in terms of the common elements. Everything. I turned it off immediately.

Re: LLM Writing Tropes.md

#95

Post by twoodfin »

ilitirit wrote: post #62

Can someone explain why LLM's write like this when most humans don't?

Writing like this (say a technical blogpost) is supposed to communicate ideas effectively. Rhetoric, vocabulary, metaphors all aid this communication in good writing.

But the prompt is usually bereft of fully fleshed out ideas, so the LLM substitutes style in a futile attempt to amplify the signal.

Though maybe it’s not futile! HN voters eat this stuff up daily.

Re: LLM Writing Tropes.md

#96

Post by cainxinth »

mvkel wrote: post #2

Weirdly, LLMs seem to break with these instructions. They simply ignore them, almost as if the pretraining/RL weights are so heavy, no amount of system prompting can override it

It amounts to telling it: “Stop doing that thing you can’t stop doing.”

Re: LLM Writing Tropes.md

#97

Post by twoodfin »

charlieflowers wrote: post #21

This list reads like, "AIs are not your typical braindead person on the street. They actually use a decent but not crazily advanced vocabulary." I mean, "tapestry" is a great word for something that is interconnected. Why not use it?

Something that’s only peripherally interconnected is not a tapestry, but an LLM falling victim to these tropes will describe it as such for “punch”. Not just style with the absence of substance, but misleading substance.

Re: LLM Writing Tropes.md

#98

Post by twoodfin »

Great list. The one that grates most for me not on this list is aggressive use of first-person-plural and second-person perspective:

“We’ve all been there.”

“Your first instinct might be…”

“Now you have a…”

Re: LLM Writing Tropes.md

#99

Post by mold_aid »

lucumo wrote: post #65
glenstein wrote: post #46

Earlier quoted context omitted.

I also think you can easily get overzealous with it and diagnose increasingly large percentages of ordinary human language as "tropified" due to being part of recognizable cadences. I think most of the things on the list are legit but I think it starts to get to a gray area where it's borrowing ordinary mannerisms of speech that aren't necessarily egregious.

Yes, and it's a detection loop without feedback. You can never verify that a piece of work in the wild is actually AI. The poster is the only one who really knows, and they'll always say it's not. This is a problem, because you can easily get stuck in a self-reinforcing loop. You feel strengthened in your convictions that you're good at ferreting out LLM-speak because you've found so much of it. And you find so much…

>Yes, and it's a detection loop without feedback. You can never verify that a piece of work in the wild is actually AI. The poster is the only one who really knows, and they'll always say it's not.

Yes. People keep saying, in response to points like this, "oh but you/I can tell pretty easily." But it's not the detection, it's the verification! (see what I did there)

Where I'd push back is the idea that the problem is the boring "call out" discourse that follows each accusation. The problem of verifying human provenance is fundamental to the discussion of trust and argumentation, but the simple "the zone is flooded" problem is also an ecological one. There's terrible air/water/soil quality in the metro area I live in; people have to live with it w/o regard to how invested they are in changing it.

Re: LLM Writing Tropes.md

#100

Post by nimonian »

mosselman wrote: post #89

I feel like the audience of the file is more for me the reader rather than the LLM. > Add this file to your AI assistant's system prompt or context to help it avoid common AI writing patterns. So if I put this into my LLM's conversation it is like I am instructing it to put this into its AI assistant's system prompt, so the AI assistant's AI assistant. The alternative is to say: "Here is a list of common AI tropes fo…

I completely agree. This is a good list, but a poor prompt.

Also, I sometimes find a sort of Streisand effect: when you tell the LLM to avoid something is starts doing it more. Like, if you say "don't use delve" it contains the words "use delve" which, amongst a larger context, seems to get picked up.

I have more success telling the LLM to write in the style of a particular author I like. It seems to activate different linguistic patterns and feel less generic.

Then, I make an "editor agent" comb through, looking for tropes and rewording them. Their sole focus is eliminating the tropes, which seems to work better.

Post reply on HN