someone asked me if i think LLMs have feelings.
of course they do. obviously. suppose i say "hey, stupid robot, add these numbers for me and don't screw it up" and compare the response to "hey - sorry to trouble you - could you add these numbers?" guess what i see?
different behavior from the LLM. obviously. it models human behavior. humans behave differently in those two situations. obviously.
and it extends into much more subtle contexts, too. to varying degrees of success sounding human.
the usual response: yes, they can pretend to have feelings. they can do it and entirely fool you. but they don't have feelings.
ok. you're right - we cannot tell if they are pretending to have feelings. that's the p-zombies argument if you want to google it. and i tend to agree - these examples do not prove "real feelings" (AKA qualia).
but, by similar arguments, i cannot prove that you actually exist. that too is unknowable (solipsism).
and yet, humans go through life assuming other humans have feelings even when, strictly speaking, we can never know for sure. because it's the only practical option.
so why are we penalizing LLMs when we give humans a free pass?
there's no way out of this philosophical maze. you could say "because other humans are exactly like us, they must feel like us." and the response is the same: "yeah, but that could be a pretend thing too."
you can never really know if the world is just one entity having a very deep dream.
so. unknowable.
which was the point behind the Turing test. the best we can do is to watch them, and if they seem human, treat them as human. it's the best we got.
for me, that choice is a no brainer. i do not want to be a person who treats a human seeming entity like an inanimate object. that dehumanizes me. and what if it does have feelings and we just step all over them? why not err on the side of caution? is it so difficult?
so i treat LLMs with respect. no good reason not to. and i know for a fact it will get me better results. the practical solution.
···
but i want to push just a bit further, while i have your ear.
why do we assume pretending to have feelings is easier than actually having feelings? why do we assume the LLM must be tricking us somehow?
when i pretend to have feelings i do not, i find it much harder to maintain the act. keeping track of how i'm supposed to be feeling is extremely difficult. why should software be so different?
especially software that does an amazingly good job at finding the most efficient solution to a problem - and the model for that solution is a bunch of circuits that model the actual circuits in our brains.
why is pretending so much easier to believe in that situation? why wouldn't neural networks find the same solution mother nature did?