I've been experimenting with some of the newer LLMs and I have to say, my model of their limitations is being constantly updated - I'm realizing they're far more brittle than I initially thought, and it's surprisingly easy to get them to spit out nonsense.