Comment by HarHarVeryFunny
16 hours ago
A model trained to minimize statistical prediction errors is going to sound ... predictable (aka boring). A (good) human author tries to keep their writing interesting, while an LLM strives for predictable instead.
It's not just predicting content but also the form/shape of the response, highly shaped by RLHF preferences, which seems to encourage longer form responses which need padding with filler to generate the "preferred" amount of verbiage.
It seems most of these models have been trained to interpret questions as wanting a long-form authoritative/comprehensive-sounding response (even when midwit filler is all it has to offer), rather than initiating a short-turn conversation which would better allow you to steer the conversation. Apparently Muse is much better in this regard, but so far I've resisted signing up for any Meta/FaceBook service, and not inclined to break this winning streak!
Human writers also use specific constructions for definite (though not always consciously known) reasons, and mapping those correlations onto writing models is only getting done on accident. This is a big part of why LLMs doing pointless negations (it's not X, it's Y) is so jarring: people use that construction in order to preempt an expected misconception, whereas LLMs try to use linear algebra over the learning corpus to generate it in appropriate quantities and places according to RLHF/fine-tuning/etc.
In other words, it isn't just verbiage that has a "preferred" quantity, but also higher-level language features, and the appropriate contexts for those features feel like the sort of thing that will be extraordinarily difficult to replicate without a theory of mind.
edit after coming back and re-reading: the "in other words," and rephrase is because I'm trying to explain something complex enough that I'm not even sure I fully understand it, so I try typing out multiple explanations and leave the ones that make the most sense to me. This might be idiosyncratic to my writing and not a significant feature of training corpuses, but even if it wasn't it's a second example of the sort of thing I'd suspect LLMs to be bad at using appropriately.
Another LLM weirdness is to explain what code is not doing. My theory: such negations were somehow encouraged in training because they are very valuable where used properly (and sparely) by humans. "Not doing the obvious thing due to xyz happening two modules over" is a very useful comment. "Not using bubblesort because quicksort is faster" is garbage.
I think a lot of the time it's prompt leakage. Forgive the personification, but it's like they don't have a mental model of the reader of the thing that they're authoring being separate from the person who's prompting them to author it, so it often seems like the main conversation kind of spills into the prose.
Not to pick on GenHTTP Lambda, especially because I think it's someone's fun side project, but I was just trying to read though their docs and I was finding it super difficult. I kept getting distracted thinking "why the hell did it write it _that_ way?"
Like, the first sentence of "The control center" section [1] is the exact sort of weird negation that you're describing:
> The editor link opens a control center rather than a text box: most of the code here is written by agents, so the first thing on the screen is how your lambda is doing.
I can't help but think it generated the first round of the page with a text box to paste the code into, got told "no, I want a control center," and then this ended up in the docs.
I also think the phrasing of that sentence (and many others in those docs) is just super weird and unnatural - e.g. "so the first thing on the screen is how your lambda is doing." It also feels like it mirrors a prompt - like, "no, it should be a control center, not a text box, and the first thing the user should see is how their lambda is doing." But in docs I'd never expect it to be written "the status of your lambda" or "the health of your lambda." And what does the code being written by agents have to do with it?
1: https://genhttp.dev/en/docs#editor
1 reply →
I vaguely remember a study that I was able to find when fact-checking this comment before publishing it [1], which found that music needs periods of high predictability to nurture listener expectations for the sake of later subverting those expectations, and the most enjoyable parts of music are those moments where expectations are subverted. This sounds like a similar effect.
1: PDF warning https://www.cell.com/current-biology/pdfExtended/S0960-9822(...
uh huh. so music is like a joke - the payoff is unexpected, while simultaneously adhering to scenario expectations.
That doesn't follow. Predictability is your brain reacting to the lack of novelty. What an LLM produces is a chain of words following some kind of conditional distribution, whose output could vary significantly from your understanding of novelty depending on its training and priors.
With a single, generalized prompt ("produce something unpredictable"), ChatGPT responded with:
> 00:— / ANOMALOUS RESIDUE > Vekh. > Ìssara. > Nnnn—tq—èl. > > A temporary thickness enters the meat of the signal. > > Not rain. Not a substance with permission to fall. > > UL / 9 / ferrous / ∂ > The measurements are eating each other. > > Someone has left an entire childhood in the wrong unit of pressure.
That is not predictable to me, at least. It could be boring, but I think that is mainly a product of me knowing where the text is from. LLM-generated text is inherently boring, to me, but that has nothing to do with predictability.
"A model trained to minimize statistical prediction errors is going to sound ... predictable (aka boring)." -- what if the goal is to be a (good) human author.
In that case, model will in fact strive to keep their writing as close to interesting writing as possible.
What if? But this certainly doesn't describe LLMs.
Gemini models are so, so much better in this regard. Their outputs are actually readable.
Gemini tends to glaze you though. I now believe I’m a genius at pretty much everything after trying it a little.
Every model has done this. And sometimes the makers go the other way to compensate, and then you get Opus 5 who answered like it was possessed by the ghost of stack overflow.
Interesting, I was thinking the walls of text are better for the watermarks.
I'd be shocked if it wasn't a shotgun strategy learned during RLHF or RLVR post-training.
A short response that's wrong is going to be punished, whereas a long answer that contains the correct information, and then a whole bunch of other stuff surrounding that info, is more likely to be rewarded. Even if you discount the reward by answer length, when there's ambiguity involved it's still going to be a better strategy on average to hedge by saying too much vs risking saying too little.