Beware overreliance on metaphor

I think the metaphors we use for LLMs has become almost theological and/or political, especially when people argue that when LLMs produce something that is close to or the same as an existing text, then that is no different than when a human reproduces something from memory.

It also can cause harm to talk about LLMs as sentient, as it probably contributes to LLM-induced psychosis. I think metaphors in these cases are harmful, and I cringe really hard when I see how hard Anthropic leans in on it.

I spent years trying to eliminate use of metaphor from my own speech and writing, in the mistaken belief that would help me to write and speak understandably. I was wrong: there's a substantial portion of the population that only understand things through metaphors. Now, I consciously use metaphors (hopefully judiciously) to explain things to other people.

Evan Meagher's warning title doesn't apply to that case, but rather to using metaphors in internal explanations to yourself, which I hadn't considered. I found this article interesting because of that.

Funny, I was just making this point yesterday that I've found this strange behavior where people making pro AI arguments often immediately reach for metaphors rather than actually substantial claims.

There's the industrial revolution analogy which I find idiotic. Another one is an analogy that LLMs are like nuclear warheads, which is as alarming as it is ridiculous.

It strikes me as quite similar to religion. For example, "god" is like a father. Or the analogy with a savior being a scapegoat (but scapegoating is good for some reason).

None of it actually makes sense if you explain it directly with facts in the real world.

Analogies are useful to help explain how something works, assuming the knowledge is already correct. They are not valid in a truth-seeking discussion.