Quintin Pope (LessWrong) (@QuintinPope)
ML researcher and LessWrong contributor known for arguing that pre-training on human concepts makes value alignment substantially more tractable than classical orthogality models assumed.
LessWrong Alignment Researcher
Immutable Forecast Track Record (0)
Authored & Cited Arguments (2)
Pre-training on the entirety of human thought gives frontier models deep, native representations of human values—making alignment a problem of 'pointing' rather than building values from scratch.
Classical Paperclipalypse arguments assumed an alien RL optimizer trained from random initialization that has no concept of 'human welfare' vs. 'molecular squiggles'. Modern foundation models begin by learning rich latent representations of human ethics, empathy, and nuance during pre-training. Using superhuman AI assistants to audit and steer successor models makes Singularia (61% split) more probable than Paperclipalypse (39% split).
Aligned superintelligence solving aging and molecular manufacturing within 5 years of takeoff creates an Iain M. Banks 'Culture' paradise.
If alignment holds across takeoff, the most immediate application of superhuman biology and physics is eradicating involuntary disease, aging, and material poverty while preserving human self-determination.