logoalt Hacker News

kqrtoday at 9:36 AM0 repliesview on HN

Worse! It's trained to output coherent reasoning, so by putting the assumption last there's a risk it massages the assumption slightly to fit the conclusions it has already drawn.