The Law of Artificial Unintelligence
When it comes to prompt engineering, practicing selective stupidity might be a good idea.
When it comes to prompt engineering, practicing selective stupidity might be a good idea.
Let's pretend the benchmarks are true and frontier models are more intelligent than a generally intelligent human in a couple of specific domains (and still less intelligent than a cat). This is a technology that, for most people, is about punching above your weight class.
Prompt engineers will write page-long prompts for precise outputs, but I think prompting and response quality probably run on a bell curve. Beyond a certain skill threshold, the more precise a prompt is, the more the engineer may be hurting the quality of the output.
I'd like to call this observation the Law of Artificial Unintelligence — or the tao of dumb. The dumber you pretend to be (or actually are), the more "intelligent" a particular generation becomes.
As a parallel (and an extreme simplification), the Uncertainty Principle states that the more precisely you know one property of a particle, like its position, the less precisely you can know a complementary property, like its momentum. There's a seesaw effect in the act of observation. I think something similar happens epistemologically in knowledge creation, automated or not.
Take-home message: if you don't know what you're doing, stop trying to be smart. Let your computer be smart for you — you're going to give yourself a headache.
"If you value intelligence above all other human qualities, you're gonna have a bad time." — Ilya Sutskever
Originally posted on LinkedIn.
