Apologies for the late reply, but it turns out I can't let that sit. Sorry for the rant, but I work in RL and saying "it's just dice rolls" is insulting to my entire line of work. :(

A probability distribution is not the same as random dice roll. Dice rolls are uniformly and independently random, whereas the probability distributions for LLMs are conditional on the context and the model's learned parameters. Additionally, all modern LLMs use top K and p sampling--which filters the probability distribution to only high confidence words--so the probability of it choosing to say random garbage is exactly zero.

The issues with LLMs have nothing to do with their sampling from random distributions. That's just a minor part of their training, and some LLMs don't even do random sampling since they use tree search. The issues with LLMs are the result of people trying to teach it intelligence using behavior cloning on a corpus of human words and images. Words can't encode wisdom, only knowledge. Wisdom can only be gained through lived experience.

How well do you think you would perform if you were born into a cave, forced to read a thousand dictionaries in order with no context, and then your only interaction with the outside world was a single question from a single human, and then you died? If you ask me, the LLMs are doing suprisingly well given their "lived experiences".