AIAI-generated essays and the limits of model “opinions”
One commenter calls AI essays vacuous and untrustworthy, arguing that models force examples into preset outlines. Their follow-up links models’ consistent “opinions” to heavy reinforcement learning.
TLDR
A commenter argues that the problem with AI-generated essays is trustworthiness, not whether the writer demonstrates personal skill: models can shoehorn examples into preconceived outlines or use them to support pretraining biases. In a follow-up, the same commenter argues that pretraining supplies a mix of “vibes,” while heavy reinforcement learning produces consistent “opinions” in domains such as coding and math. They blame the limited judgments of training graders for formulaic essay writing and say they are waiting for DeepSeek to crack continual learning so models will have more to say.
