Post #3922051

I suspect most of this is due to imperfect RLHF that focuses on brand-safety and rhetorical tricks that get annoyingly noticeable when they’re repeated at an industrial scale. The models are approaching superhuman ability at coding and math, but primarily because they’re more easily verifiable (plus it’s where the money is). But if there were a more rigorous way to quantify what makes good fiction work, there’s no reason in principle why a strong model wouldn’t be able to “optimize” it the same way it can images or video. I’ve fine-tuned a local model on my own writing before and it’s spooky how well it can capture a stylistic voice.