Empirical risk minimization is fundamentally confused — AI Alignment Forum