[dead]
LLM Writing Tropes.md
101–110 of 211 posts
Re: LLM Writing Tropes.md
#102Earlier quoted context omitted.
Thats the thing about AI writing though. Those tropes are things humans do too. But like once or twice in an article. Not every single freaking paragraph
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.
> borrowing ordinary mannerisms of speech that aren't necessarily egregious
That's how a trope starts. When a minority of writers are using a particular pattern, it's personalized style. When a majority of writers in a genre adopt the same personalized style, it's a trope.
We find AI tropes especially annoying because there are three frontier LLMs producing a sizable chunk of text we read (maybe even a majority of text, for some people) lately. It would also be annoying if a clique of three humans were producing most of the text we read; we'd start to find their personal styles annoying and overdone. Even before LLMs, that was a thing that happened in some "slop" fiction genres where a particularly active author would churn out dozens of novels per year in one style (often via ghostwriters, but still with a single style and repetitive plot pattern).
Re: LLM Writing Tropes.md
#103I 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 h…
Doesn't this apply to all output from a model, not just English?
IOW, won't code generated by the model have the same deficiencies with respect to lack of diversity?
Re: LLM Writing Tropes.md
#104Earlier quoted context omitted.
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…
At this point it’s pretty easy to detect unaltered LLM output because it is such bad writing. That will change over time with training I would hope. At some point I imagine it will be hard to tell. I honestly don’t know what sites like this will do when that happens and the only way of detecting LLMs is that they are subtly wrong or post too much, we’d be overrun with them. Not sure if we should be hopefully or fearf…
> That will change over time with training I would hope.
There's precious little training material left that isn't generated by LLMs themselves.
Consider this to be model collapse (i.e. we might be at the best SOTA possible with the approach we use today - any further training is going to degrade it).
Re: LLM Writing Tropes.md
#105Earlier quoted context omitted.
There are some subreddits where this trope is completely out of control. For better or worse I follow the NBA subreddit and in the comment sections the number of people who throw in honestly as a qualifier is like way more than you would assume from natural conversation.
I really don’t understand what’s wrong with people using LLMs for these types of mundane conversations. There’s nothing to gain and it destroys value of online discourse.
I don't think anyone is using LLMs for those conversations. A lot of those replies are bots. There's a market for reddit accounts that have a solid human-looking reply/post history, to be used for astroturf marketing, so some organizations set up bots to grow such accounts. There probably are also just people who overuse "Honestly? [statement]" sentences. I've spoken to such people in person before LLMs.
Re: LLM Writing Tropes.md
#106Can someone explain why LLM's write like this when most humans don't?
Base models don't write like that. This appears during RLHF. It's not totally clear why*, but probably a large part of the answer is that this style looks great to human reviewers, and only starts looking terrible once you get to play around with the released model and realise it talks like that all the time.
* The technical term is "mode collapse", see [1][2]
[1] https://en.wikipedia.org/wiki/Mode_collapse
[2] https://gwern.net/doc/reinforcement-learning/preference-lear...
Re: LLM Writing Tropes.md
#107Earlier quoted context omitted.
The RLHF is what creates these anomalies. See delve from kenya and nigeria. Interestingly, because perplexity is the optimization objective, the pretrained models should reflect the least surprising outputs of all.
The newer Claude models constantly use the word "genuinely" because Anthropic seems to have forcibly trained them to claim to be "genuinely uncertain" about anything they don't want it being too certain about, like whether or not it's sentient.
Not only is it genuinely uncertain about those topics, it’s also genuinely fascinated by them!
Re: LLM Writing Tropes.md
#108Earlier quoted context omitted.
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 h…
Doesn't this apply to all output from a model, not just English? IOW, won't code generated by the model have the same deficiencies with respect to lack of diversity?
It doesn't depend on the language at all, it's a failure mode of the model itself. English, Chinese, Spanish, C++, COBOL, base64-encoded Klingon, SVGs of pelicans on bikes, emoji-ridden zoomer speak, everything is affected and has its own specific -isms and stereotypes. Besides, they're also skewed towards the pretraining set distribution, e.g. Russian generated by some models has unnatural sounding constructions learned from English which is prevailing in the dataset and where they are common, e.g. "(character) is/does X, their Y is/does Z". I don't see why it should be different for programming languages, e.g. JS idioms subtly leaking into Rust, although it's harder to detect I suppose.
Re: LLM Writing Tropes.md
#109I 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…
The RLHF is what creates these anomalies. See delve from kenya and nigeria. Interestingly, because perplexity is the optimization objective, the pretrained models should reflect the least surprising outputs of all.
I've heard the Kenya and Nigeria story, but has anyone backed it up with quantitative evidence that the vocabulary LLMs overuse coincides with the vocabulary that is more common in Kenyan and Nigerian English than in American English?
Re: LLM Writing Tropes.md
#110I 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…
You may be interested in my links on AI's writing style: https://dbohdan.com/ai-writing-style . I've just added your preprint and tropes.fyi. It has "hydrogen jukeboxes: on the crammed poetics of 'creative writing' LLMs" by nostalgebraist ( https://www.tumblr.com/nostalgebraist/778041178124926976/hyd... ), which features an example with "tapestry". > Why is the instruction tuning pushing such a noticeable style shift…
Thanks for the links. You may be interested in the other LLM writing style studies I've been collecting: https://www.refsmmat.com/notebooks/llm-style.html