Think there is a clear way to improve AI writing for non-fiction (we did that for one of the gpt4o version). Much of the gap comes from conflicting optimization objectives during RL: the traits that make a model pleasant to chat for consumer are often at odds with the qualities valued in academic / technical / research writing, where precision, info density, structure, and restraint matter more. Creative writing is a different problem entirely imo. It’s about building coherent worlds, developing characters, maintaining narrative consistency, and creating emotional arcs over hundreds of pages. Great novels were written over years of iteration on structure, theme, and character. And simply we don’t have much of the historical data to capture the entire process to train on within the current paradigm (eg hard to simulate Pixar’s braintrust process for AI), but I also think no lab has put in resources into it bc it’s almost impossible to measure + not much economic value I once gave a talk on this: