Regression to the Mean: LLMs and the quiet death of the new

Regression to the Mean: on LLMs and the quiet death of the new

This essay argues LLMs are built to return the most probable continuation—the center of mass of everything already written. Ask it something genuinely new and it corrects you: unfamiliar terms become typos, consensus becomes fact, conviction gets sanded down to the mean. The deeper risk is feedback: we feed its answers back as the next questions, variance leaks out of culture, and the curve sharpens to a spike. Every major discovery was out of distribution when it first appeared—moving earth, unseen germs, drifting continents—each filed as error by the consensus of its day. A model of consensus is, by construction, a machine for telling you the new thing is wrong. The average is now free, infinite, identical, and worth little precisely because everyone holds it. What is priceless is the deviation: the position the model marks as wrong, kept anyway.

Why it matters: A sharp, philosophical essay that reframes LLMs as averaging machines—outputting the most probable continuation, not truth. New terms get corrected as typos, heterodox views get sanded down, and feedback loops drain variance from thought. Not an 85 because it's a personal essa...