"Expect uneven capability spikes, not uniform intelligence Karpathy illustrates AI's 'jaggedness' with a striking example: a state-of-the-art model that can autonomously move mountains on an agentic coding task will still tell you the exact same bad joke it told four years ago. As he explains, "you're either on rails and you're part of the super intelligence circuits or you're not on rails and you're outside of the verifiable domains and suddenly everything kind of just like meanders." The practical implication for anyone relying on AI tools is to map which tasks fall inside the 'trained and verified' rails before delegating them, and to keep human judgment on anything outside those rails. Source: Andrej Karpathy, Conviction, Episode with Andrej Karpathy"