AI & wisdom 3: AI effects on amortised optimisation — LessWrong