Scaling and context steer LLMs along the same computational path as the human brain