303 | AMA | February 2025

10893.689 - 10917.075 Sean Carroll

Timlo, P-T-M-I-L-O, Private Milo maybe, says, is your objection to the potential for large language models to exhibit more generalized and extrapolation-heavy intelligence based on deeper principles, or is it more intuitive? That is, is there anything in information theory that tells us it is impossible for locally generated interpolations of tokens to uncover patterns and sequences that