Comment by spijdar
12 hours ago
Yeah, this tracks with my experiments. About every 6 months or so for the past year-and-a-half I've tried using the "frontier" LLMs to as-near-as-possible autonomously write novel length stories, because I find it fascinating.
When I had Gemini 2.5 write a novel, it wasn't really objectively "good" by any stretch of the imagination, but while the prose was very purple and full of cliches and, well, bad writing I guess, it still felt ... subjectively good, at least for what it was.
Last week I did a run with GPT-5.6, and wow. On the one hand, it managed to produce 110,000 words that were "shockingly" coherent. The model was able to maintain state and plot lines and background details extremely well, much better than older models.
But I just don't like the prose. I haven't really liked _any_ prose that GPT-5.6 produces. It's significantly better at "instruction following" and keeping track of things, but, wow.
> “The sequence is consistent with their voluntary choices.” Mara enlarged the uncertainty field rather than the result. “It does not prove what happened to anyone we can’t observe. It does not prove contact did this. And it does not turn the Shard into treatment.”
GPT-5.6 in particular becomes so fixated on certain ideas like "consent" and epistemology, that by the end of the narrative, the prose and dialogue are all just "agent speech", despite the prompt/harness specifying that it's a _novel_ with narrative prose and such.
Interestingly, the model itself produces an accurate critique of its own output:
> The draft has become a *consent-centered medical, legal, and logistical procedural*. The important drift is therefore not that many events were omitted. It is that the retained events now prove a different thesis.
Which begs the question of if it would do better with a couple rounds of output -> critique -> revision. But I think I've had enough LLM prose for a bit...
The entire output of an LLM is also part of its input. Appending to long stretches of LLM-generated text, it will continue to get more and more robotic as the style of the input gets replicated and enhanced in the output.
It's possible to work around by generating short passages at a time with carefully constructed setup. But it's a real pain.
It's that, yeah, but coming from multiple orders of abstraction.
In this case, part of the experiment was to see what "oh-my-pi", a "fat and feature rich" LLM harness, could do when coupled with modern GPT, given a 6k~ word overview of a story, and told to come up with a plan to write/review/audit it, making use of subagents and all the fun new groovy LLMisms...
Part of the problem was just "it was basing its style off the last scene/chapter", but part of it was also that its instructions were constantly being "compressed" through repeated compactions. Even with the use of subagents, the "top level" agent's prompt was getting muddied, and in the "review" phase, it began to focus more and more on creating increasingly complex ledgers.
You can see this happen in the "plan" files it created for each chapter, looking at word count:
So it wasn't just that the prose was being based on an increasingly compressed "style" of the prior context window, but the planning for writing each scene was, itself, becoming fixated on the "continuity error correction" process itself, to the point where by the end, it had mostly forgotten about the prose part, and was completely fixated on ensuring maximum state continuity.
This could definitely be fixed, but honestly, I've about had my fill of the "autonomous writing agent" goal. The idea was to make a model that could generate sufficiently interesting stories based on "vague premises" for my personal entertainment, but, "surprise", getting LLMs to actually produce both "new" and "coherent" content beyond what you specify is _hard_.
It seems like you really do need to just stay "in-the-loop" with every scene, and constantly provide correction/feedback, to correct the "semantic drift".
Or, gasp, I could just try writing things by hand again... :-)
I'm curious if you've tried Claude. Subjectively, I've always preferred Claude's writing over ChatGPT and Gemini.
Claude was really far ahead of GPT in writing from 2-4, but the later models have started to get an overly distinctive style, wheras these days GPT tends to be coherent but fairly concise and dry.
[flagged]
Claude is the one model family I've not really used. Which yeah, feels like a backwards thing to say in a world where seemingly everyone using LLMs is using Claude Code.
I've used some Opus 4.5/4.6 via Antigravity and Sonnet by the web chat. I'm torn because as far as LLMs go, it does feel more ... "literate".
But maybe too literate, judging by how many people are complaining about "Claudeisms". I suspect Claude would be just as susceptible, if not more, to the sort of ... "moralizing" that GPT seems to gravitate towards (for lack of a better term).