Karpathy: Training LLMs is Like Summoning Ghosts, We Don’t Fully Understand Them

Karpathy: Training LLMs is Like Summoning Ghosts, We Don't Fully Understand Them

Photo by Pedro Figueras on Pexels

AI researcher Andrej Karpathy has compared training large language models (LLMs) to “summoning ghosts,” highlighting the mysterious and not-fully-understood nature of these powerful AI systems. In a discussion originating on Reddit, Karpathy explained that LLMs, in his view, are like echoes of the living world, lacking direct agency or interaction with the physical realm. He emphasized that the underlying mechanics and essential nature of LLMs remain largely enigmatic. The original Reddit thread can be found here: [https://old.reddit.com/r/artificial/comments/1nwv2d5/andrej_karpathy_on_why_training_llms_is_like/]