It feels like half of AI research is math, and the other half is coming up with yet another way to state "please don't do bad things" in the prompt that will sure work this time I promise.