False assumptions and leaky abstractions in machine learning and AI safety — AI Alignment Forum