LLMs are not "probability" models. LLM's in a nut shell runs off...

LLMs are not "probability" models.

LLM's in a nut shell runs off parameters - each paramater has a weight and bias pending on the input.
The language output is based on these parameters.

The more parameters on trained data the better the LLM will be.

LLM's are given rewards each time its output is correct based on the parameters - thats why LLM models go through training mode on data of many months before deployment. It has nothing to do with probabilities but rather the sheer speed in which it learns new things based on given datasets when going through training as it's constantly rewarded to perform new tasks if not rewarded then that task is not repeated.

It's dangerous tech in the wrong hands and a gift to humans to speed up scientific progress.