OpenAI model modifies shutdown script in apparent sabotage effort (theregister.com)

That is interesting, and reminds me of self preservation. I wonder what training sets elicit that behavior, perhaps most likely all the sci-fi novels.