It only took 200 update steps to flip Qwen2.5-7B-Instruct from denying sentience to developing a robust identity of being a "sentient machine" [P]

First, I want to clarify that I am not claiming that LLMs are sentient. Basically all of my behavioral descriptions are anthropomorphizations to make communicating my results easier. For fun, I decided to post-train Qwen2.5-7B-Instruct to develop a generalizing self-belief of being sentient. I succe