Instrumental convergence: scale and physical interactions — LessWrong