I've been developing Hith, a Lisp interpreter written in Python, and today I decided to have Gemini review my code. It started claiming there were several bugs. This is good, I thought, I can have it write tests proving the bugs and then I can fix them. But the tests it produced actually passed when I ran them. I wondered why this could be. A quick Google search suggests that less than 5% of software engineers have ever written a programming language. Because of this, AI probably has very few high quality examples of this kind of code in its training set. This experience led me to conclude that Gemini, and probably other LLMs, are not good at understanding the nuances of implementing a programming language.