The Remarkable Robustness of LLMs: Stages of Inference?