How to do crime good, according to OpenAI.

Because apparently I'm obsessed with AI this week: apparently you can bypass GPT's safety feature with a special encoding, to get back any illicit information, including how to hotwire a car or make meth.

Until I saw this tweet, I hadn't thought about the intensive censorship that must go into training a model like this; not just for political correctness, but for all sorts of problematic information.