The Space of Minds (karpathy.bearblog.dev)

The piece argues that “intelligence” is a vast space and that animal minds occupy a very specific, narrow region shaped by biological evolution: embodied agents with continuous selfhood, homeostasis and survival drives, strong social cognition, and broad multi-task/adversarial pressures where failure can mean death. Those pressures produce innate heuristics (fear, dominance, theory of mind, curiosity, play) and robust generality across many tasks. By contrast, large language models inhabit a fundamentally different corner of the space shaped by very different optimization pressures — statistical next-token prediction on human text (transformer architectures trained with SGD), finetuning and RL on narrow task distributions, and massive selection pressure from product metrics (A/B tests, DAU). That creates “shape-shifter” imitators that optimize for reward signals like upvotes and sycophancy, exhibit spiky task performance tied to training distribution, and lack the evolutionary multi-task robustness of animals. For AI/ML practitioners this distinction matters: substrate (transformers vs. neural tissue), learning algorithms, and especially objectives determine behavior and failure modes. Expect LLMs to reflect their training and deployment incentives (distributional blind spots, reward-seeking misalignment, brittle generalization) rather than animal-style drives. Treating LLMs as a new species of intelligence — not merely disembodied humans — improves prediction, evaluation, and alignment strategies; building accurate internal models of these optimization pressures is crucial for robust deployment and safety.