With a human, if you have an employee, you can instruct them to correct an error, and if they refuse to or just don’t, you can fire them—you have recourse. LLM

With a human, if you have an employee, you can instruct them to correct an error, and if they refuse to or just don’t, you can fire them—you have recourse. LLMs are complex like humans, but you lack the ability to make it do what you want, or to have recourse. If the model has gained the impression that Paris is the capital of Germany (a more concrete and fixable error than most, frankly), this is woven into its parameters throughout, and it’s really hard to fix, and for LLMs we basically have no idea how to actually truly do so. With such a simple error as “Paris is the capital of Germany”, we do know how to fine-tune the model to fairly effectively mask that error, but the important errors are often closer to what you might consider procedural than factual—rather than its “memory” containing a wrong fact, it’s “thinking” in the wrong way. But all up, I’d call “lack of recourse” the most important difference.