Jagged Intelligence

Overview

Jagged Intelligence describes the uneven capability profile of frontier LLMs — models that exhibit world-class performance in some domains while failing at seemingly trivial tasks in others. Andrej Karpathy uses this term to characterize LLMs as "ghosts, not animals": statistical simulation circuits shaped by pre-training data distributions and reinforcement learning rewards, rather than by intrinsic motivation, curiosity, or evolutionary pressures.

The jaggedness arises from two factors working in combination:

1. Verifiability — Domains where outputs can be automatically verified (math, code) receive disproportionate RL training investment, causing capability to spike
2. Lab priorities — The specific data and environments labs choose to include in training determine which capabilities peak, independent of what is theoretically trainable