Andrej Karpathy — We're summoning ghosts, not building animals
Karpathy argues we're building 'ghosts' that imitate internet documents rather than evolved animals, and that practical AI agents will take a decade—not a year—to fully mature.
Key Takeaways
- Decade of agents, not year of agents. Current agents like Claude and Codex work daily, but lack the reliability to replace an intern. The cognitive deficits—no continual learning, limited multimodality, no persistent memory—will take years to solve.
- Pre-training is "crappy evolution." Evolution encodes neural weights in DNA through millennia of optimization. We skip that by imitating internet documents, producing ghost-like entities rather than animal-like ones.
- Weights store hazy recollection; context is working memory. Anything compressed into model weights becomes fuzzy. Anything in the context window stays directly accessible—like the difference between remembering a book from years ago versus having it open in front of you.
- In-context learning may run internal gradient descent. Papers show transformers doing linear regression in-context implement something resembling gradient descent mechanics within their attention layers.
- Knowledge crowds out intelligence. Pre-training gives both knowledge and cognitive algorithms. The knowledge may hold models back—they over-rely on memorized patterns instead of reasoning off-distribution. Future research should extract the "cognitive core" while stripping unnecessary knowledge.
Notable Quotes
"We're not building animals. We're building ghosts or spirits, because we're not doing training by evolution. We're doing training by imitation of humans and the data they've put on the Internet."
"Anything that happens during the training of the neural network, the knowledge is only a hazy recollection of what happened in training time."
The Missing Brain Parts
Karpathy frames current models as cortical tissue—the transformer is plastic and general, like the cortex. But other brain structures remain unexplored:
- Prefrontal cortex: Reasoning traces in thinking models
- Basal ganglia: Reinforcement learning fine-tuning
- Hippocampus: No equivalent for memory consolidation during "sleep"
- Amygdala: Emotions and instincts absent
Three Seismic Shifts in AI
- AlexNet (2012): Everyone started training neural networks, but per-task
- Atari/RL (2013): First agent attempts—turned out to be premature
- LLMs: Seeking representation power before tackling agents
The early agent work (Atari, OpenAI Universe) was a misstep. You need the language model's representations first—keyboard-mashing through sparse rewards burns compute without learning.
Connections
Builds on building-effective-agents and agentic-design-patterns with a longer-term perspective on agent development timelines. The "cognitive core" idea relates to context-engineering-guide—what matters is the reasoning capability, not memorized facts.