Given how LLMs access compressed knowledge from their model weights, the similarities make sense