> LLMs are not "truth models" You could write a piece of software that is a truth model when it operates correctly. But increase the CPU temperature too far,

> LLMs are not "truth models" You could write a piece of software that is a truth model when it operates correctly. But increase the CPU temperature too far, and you software will start spewing out garbage too. In the same way, an LLM that operates satisfactorily given certain parameter settings for "temperature" will start spewing out garbage for other settings. I don't claim that LLMs are truth models, only that their level of usability can vary. The glitch here doesn't mean that they are inherently unusable.