Skip to content

Animals vs Ghosts

andrej-karpathy's framing (written up before the Sequoia 2026 talk, re-stated there) for what kind of entity we are actually building with LLMs.

The distinction

  • Animals are shaped by evolution — they come with intrinsic motivation, curiosity, empowerment drives, fun. Their intelligence is baked in by billions of years of selection pressure.
  • Ghosts are shaped by data + reward functions. They are statistical simulation circuits on a pre-training substrate with RL "bolted on top." No evolutionary priors, no intrinsic motivation — they just inherited the distribution of human text and got bent by reward.

"We are not building animals. We are summoning ghosts."

Why the framing matters

Karpathy is frank that it may be "a little bit of philosophizing" rather than yielding five concrete mechanical predictions. But a good model of what they are makes you more competent at using them:

  • Don't expect curiosity or intrinsic drive — engineer it in.
  • Don't expect a coherent self across context — the ghost is summoned per-prompt.
  • Don't expect the failure modes of a mind; expect the failure modes of a statistical simulator with RL dendrites.
  • Trust, deployment, and evaluation designs should start from "this is a ghost" rather than "this is a junior human."

Connects to