Literacy - especially an understanding of narratives and stories - gives folks an edge in understanding how LLMs fail. Having a mental model like djinn or devils or the fae, to understand the idea of "speak your intentions carefully, because you'll get what you asked for", helps my narrative-literate teammates tread lightly as they ask for things. Having read books from the POV of an unreliable narrator, or participated in close critical reading of fiction or nonfiction, makes discriminating consumers of LLMs' outputs. And understanding the arc of narratives, the structure of fables, helps identify when an LLM is telling us something too good to be true, or to reason about jailbreaks like the Waluigi Effect.
Even in workplaces where AI will create productivity dividends -- and they exist, it's not all horseshit or slop -- even, or maybe especially, in those workplaces, it's critical for people using AI to understand it. If you were a grass-cutter, and you were told you had to become an adept shepherd, using sheep to graze grass fields now, because they can go faster, you would need to learn a few important things:
How much shit to expect, and to prepare your customers for. (Your old approach did not generate any shit. They probably will be surprised at how much shit is on their grass.)
How your sheep will behave and misbehave.
And you still need to understand grass, soil, and sunlight at the expert level you did before! Maybe you can spend less time learning the scythe, but you still need to understand grass.
In this metaphor, (1) comes with time & experience, (3) is doing your old job well, and (2) is the part literacy can help with because in this metaphor the sheep are your agents and AIs, and they're creatures of language. Literacy is how you learn!
tangentwithtanner
