“Ben Horowitz observes that while reinforcement learning in AI shows linear performance improvement, it does not generalize well across different domains.”