Subliminal Learning, the Lottery-Ticket Hypothesis, and Mode Connectivity — LessWrong