Active Inference as a formalisation of instrumental convergence — LessWrong