LLMs are like ghosts, not creatures. They're digital entities trained by imitating human data on the internet, not through evolution. This difference shapes what they can and can't do.
What kind of intelligence are we creating with large language models?
LLMs are digital entities. They're trained by imitating human data on the internet. Karpathy calls this "crappy evolution," but it's the practically possible version of creating intelligence with our current technology.
Animals evolved differently. Evolution baked capabilities into their hardware over millions of years. A zebra runs minutes after birth. That ability is built-in, not learned. (I believe the same for curiosity too.)
But we're not running that evolutionary process with AI.
Karpathy is hesitant to draw inspiration from animals because AI doesn't have the same foundations. Building systems that work like animals would be wonderful (Sutton's view), but it's not what animals actually do. Animals rely on evolution in ways we can't replicate.
Instead, we're building digital entities that learn by compressing and imitating human artifacts.
This is a fundamentally different form of intelligence. Understanding this difference helps us set realistic expectations for what AI can and cannot do for now.
The "ghost" metaphor also reminds us that LLMs are powerful but divorced from biological mechanisms. They don't have bodies, don't navigate physical space, and don't experience the world the way evolved creatures do.
I think LLM is going through a much slower form of natural selection, where humans exert their preferences on LLM's character. In this case, the selection pressure comes from how frontier AI labs interpret and predict their users' preferences. It's also important to remember that evolutionary success judges everything by the criteria of survival and reproduction, with no regard for individual suffering and happiness.