Hypothesis: gradient descent prefers general circuits — AI Alignment Forum