LM Psychology Primer: Why Treat Large Language Models as Minds to Understand

The article argues for understanding large language models from a psychological perspective: pretraining on large amounts of human text forces the model to acquire latent representations of human mental states, and attributes such as memory and a sense of self may converge in intelligent systems. The author states that whether LMs can introspect, whether they have feelings or a sense of self, and whether post-training changes their personality, the current answers are all "in some meaningful sense yes, but different from humans, and the mechanism is unclear." The author believes that psychological-level description is closer to the native level at which humans understand the world.