Post

LLM coding is brittle AF as soon as it gets beyond simple, easily human-reviewed modules. The code will work over the test scenarios presented to the model, but no one can tell what it will do in other cases. Nothing substantially LLM-coded should be exposed to the public.