[SEQ RERUN] Underconstrained Abstractions — LessWrong