That they're complete dogshit in capability, reliability, consistency and vulnerable to malicious prompt fiddling. Wholly inappropriate for production, and for

That they're complete dogshit in capability, reliability, consistency and vulnerable to malicious prompt fiddling. Wholly inappropriate for production, and for what most people are using them for, there's infinitely better solutions than LLMs. With the amount of fucking around required trying to correct an LLM, you may as well just write the code to do your task properly.