The reason LLMs are so great at writing code is because code has a runtime. It either works or it doesn't. The model was trained on millions of code samples where correctness was implicit in the data itself. General writing has no such forcing function. Code is grounded in execution. There's a runtime that gives you a hard reward signal. Even without RLHF, the pretraining data carries implicit execution feedback. Blog writing has no equivalent. The runtime for general writing is a human…