Comment by Azantys
2 days ago
Do people really use 100B+ models for writing? I am no writer but to me it seems like writing is one of the easiest tasks with barely any logic or reasoning and as long as its not longer than a handful of pages I expect even 8B models to perform great.
It's pretty clear you've never experimented with it. Creative writing demands everything the model can do and more, and most problems are still unsolved. It's extremely heavy reasoning-wise, more so than coding (check e.g. Engram paper for some insights), but also needs good scattered retrieval, careful subjective training for prose quality, character, and human likeness, a ton of facts baked in, and much much more. Mode collapse is not solved. No LLM does creative writing well but historically only the absolute largest models were able to do write anything complex more or less convincingly and were creative enough.
> I am no writer but to me it seems like writing is one of the easiest tasks with barely any logic or reasoning
Virtually all logic or reasoning is, in one way or another, part of the support for writing. It’s what separates actual writing from generating nonsense that happens to fit grammar rules.
The specific details depend on the domain, of course, but I can’t see how anyone familiar with the output of writing can think that there is little logic or reasoning in doing it well.
Of course there is logic but its nowhere near the complexity of math or programming
That’s trivially false in the case where “math” or “programming” is the domain of the writing, but its also false more generally.
let me just say, you're not going to sound smart saying that
2 replies →
> with barely any logic or reasoning
I take it you enjoy works of literature with inconsistent world building?
Or do you mean professional as opposed to creative writing? Because the bar is even higher for that.
The wording wasnt very good I ment compared to programming or math the amount of logic and reasoning is small (Research level math hardly compares to writing a book in raw reasoning and logic). And I thing the smaller models have enough "intelligence" to write coherent with logical world building, but only the big models can truly do hard math and programming work
> The wording wasnt very good
Writing isn't so easy after all.
The largest model I've post-trained in the last 2 years of working on this problem was Kimi 2.5 at 1T parameters.
The simplest way I'd put it is, teaching a model to write coherently (follow rules, patterns, etc.) is easy enough: just use teacher forcing. Teaching a model to write creatively is easy enough: just use RL and punish it for not being creative.
Teaching a model to write well and creatively takes learning two partially opposing objectives that spike the learning requirements in ways that smaller models really struggle with.
> is easy enough: just use RL and punish it for not being creative.
How are you scoring creativity in an unsupervised manner? That seems anything but easy.
Did you try reading the whole comment?
Once creativity is being measured in isolation, getting multiple responses from the model is enough to measure creativity a ton of different ways: wordfreq to identify overused phrases, getting multiple responses for the same prompt and promoting the least similar as preferred for policy optimization, etc.
But that's of limited use for stuff like getting diverse names and such. You want creativity and coherency, and if you just punish the model for using an overused phrase, the first thing it does is strongly learn a new overused phrase (or gibberish).
(Also I don't think you mean unsupervised. You probably mean without humans [since LLMs struggle to judge creativity], but that's not what unsupervised means.)
3 replies →