philippdubach.com
|
7777777phil
|
|
AI structurally converges to the mean through next-token prediction and RLHF typicality bias. Evidence from Humanity’s Last Exam (only 3.3% accuracy from frontier models) and persistent hallucination rates shows AI cannot replicate tail-end quality. The centaur model of human-AI collaboration consistently outperforms AI alone, suggesting domain expertise is appreciating in value, not depreciating.
