LLMs That Write Well
LLM training for coding and math structurally degrades writing quality over time
“The same reward signals that make models proficient in coding actively erode their style.”
Y Combinator argues that as frontier models optimized for coding and math, their prose style degraded—today's models sound more robotic than 2024-2025 era ones. This is framed not as a bug but a structural tension in reward signal design. YC positions this as a startup opportunity for labs focused on expressive, human-enjoying writing rather than benchmark performance.