Jagged intelligence—a term coined by OpenAI researcher Andrej Karpathy—describes AI's uneven capabilities. Systems like Google and OpenAI’s models excel at math and coding, solving complex Olympiad problems, but falter at basic tasks like determining whether to walk or drive 50 meters. This disparity highlights AI’s asymmetric progress: hyper-advanced in structured domains but clumsy in real-world reasoning.

Experts argue this jaggedness reshapes the AI debate. While A.I. threatens entry-level programming jobs, its impact on broader white-collar work remains unclear. Economists like Alex Imas note that A.I. automates specific tasks, freeing humans for higher-level work, much like calculators aided accountants without replacing them. However, A.I.’s weakness in creative or ambiguous problem-solving—exposed by tests like the ARC-AGI 3 benchmark, which it failed to solve—limits its current utility beyond narrow applications.

Companies are addressing gaps via reinforcement learning, training A.I. to recognize patterns in data. This works well for math and code but struggles with open-ended challenges. The internet’s incomplete repository of human knowledge further constrains A.I., leaving it unable to plan or innovate beyond digital patterns. As François Chollet, creator of the ARC-AGI test, notes, general intelligence remains elusive.

The rapid evolution of A.I. complicates predictions. Weaknesses identified in 2024 are already being addressed, but new gaps emerge. For now, jagged intelligence underscores the need for cautious optimism: A.I. augments, but does not yet replace, human labor. The technology’s trajectory will depend on how quickly it closes its remaining valleys of incompetence.