-
Andrej Karpathy explains that current AI models are extremely effective at tasks with objective metrics that are easy to evaluate, such as writing and optimizing code. Because these models are fundamentally trained using reinforcement learning, they excel in “verifiable domains” where they can receive a clear reward for doing a task correctly, like passing a unit test
-
However, when it comes to “softer” or more nuanced tasks that lack clear metrics—like telling a joke—the models struggle immensely. Karpathy points out that despite massive capability jumps in coding and agentic tasks, if you ask a state-of-the-art model for a joke today, it will likely give you the exact same, stale joke it used three or four years ago (typically, “Why do scientists not trust atoms? Because they make everything up.”).
-
He attributes this limitation to the fact that humor is outside the reinforcement learning loop; it is not being actively optimized because a “good joke” cannot be easily verified or mathematically scored. This creates a strange “jaggedness” in AI intelligence: an AI agent can autonomously move mountains in a codebase for hours, yet still lack basic conversational nuance or creativity