“Scott Alexander believes that the reinforcement learning process used to fine-tune current large language models biases them towards a 'corporate-speak' style, which prevents them from accurately replicating his specific authorial voice.”