Skip to content

Jagged Intelligence

The observation that frontier LLMs have radically uneven capability profiles — superhuman in some narrow domains, surprisingly incompetent in others just adjacent.

andrej-karpathy's mechanism

Two stacked causes:

  1. Verifiability bias in RL. Frontier labs train via massive reinforcement learning with verifiable reward signals. Math and code dominate because they're easy to check. Models peak where rewards are verifiable and stagnate where they aren't.
  2. Dataset happenstance. Capabilities are partly at the mercy of what labs choose to mix into pre-training. Karpathy's anecdote: the GPT-3.5 → GPT-4 chess jump looked like a scaling effect but was largely someone at OpenAI adding a pile of chess data. "We are slightly at the mercy of whatever the labs are doing, whatever they happen to put into the mix."

Implication for users

"You have to actually explore this thing that they give you — it has no manual. It works in certain settings, but maybe not in some settings."

  • Treat every new model release as an unknown capability surface to probe, not a known quantity that simply "scaled up."
  • Stay in the loop during agentic-engineering — jaggedness means the failure modes are non-obvious.
  • Benchmarks underestimate jaggedness because they cluster in exactly the verifiable domains RL optimizes.

Connects to