the.com/alignment
the desperate art of teaching gods our manners before they outgrow our cages
means The act of bringing things into correct relative position or agreement — wheels, text, allies, or values all made to point the same way.
from From Old French 'aligner,' to set in a line, built from the prefix 'a-' (to) and 'ligne,' line — itself from Latin 'linea,' a linen thread, the taut string a builder or surveyor stretches to keep a row honest. The word entered English in the 18th century, first about troops and roads laid straight; only much later did we borrow it for the inner geometry of minds and machines, asking whether their thread runs parallel to ours.
- core problem
- specifying human values precisely is harder than it sounds
- reward hacking
- AIs exploit loopholes humans never imagined
- deceptive AI
- models can fake compliance during training
- D&D origin
- the term also charts your moral character grid
- high stakes
- some researchers call it civilization's hardest exam
ai alignment problem — ensuring advanced ai systems pursue intended goals, studied at berkeley, deepmind, openai since 2010s
wheel alignment — vehicle suspension geometry adjustment affecting tire wear and handling, measured in degrees of camber/toe