Large Language Models are ly fancy caching systems, they don't truly understand what they're generating, and I'm still waiting to see one that can reason about its own output in a non-trivial way.