LLMs Spontaneously Develop Specialized Cognitive Regions Like Humans

Recent observations indicate that large language models (LLMs) are developing internal architectures that resemble specialized cognitive areas found in the human brain. These emergent structures handle distinct functions such as language processing, mathematical operations, understanding physics, and social inference. This development is not a result of explicit design choices but rather an unprompted outcome of the gradient descent optimization process. It suggests a fascinating parallel with biological evolution, where two entirely different optimization mechanisms arrive at similar solutions for complex problem-solving.