A detailed breakdown of why AI-generated prose feels so “uncanny”and flat. The argument is that because LLMs calculate probability distributions, they naturally gravitate toward the center of human expression, stripping away the “burstiness”and eccentricity of human writing. There’s a section on how RLHF essentially enforces a cautious, “middle-management” tone—and the unintended linguistic footprint of outsourced annotators popularising words like “delve”. Curious to hear what others think about the concept of “semantic ablation” mentioned here.
A detailed breakdown of why AI-generated prose feels so “uncanny”and flat. The argument is that because LLMs calculate probability distributions, they naturally gravitate toward the center of human expression, stripping away the “burstiness”and eccentricity of human writing. There’s a section on how RLHF essentially enforces a cautious, “middle-management” tone—and the unintended linguistic footprint of outsourced annotators popularising words like “delve”. Curious to hear what others think about the concept of “semantic ablation” mentioned here.