Great post - I've been having very similar thoughts recently. On a more concrete and prescriptive level, I'm curious how we should account for the effects of over-determined ideas about AI ontologies when conducting AI safety research. Significant work goes into red-teaming, analyzing, and publishing (!) the ways in which AIs might misbehave. By proliferating these expectations, even for the sake of safety, are we causing unintended harm? How might we account for this possibility?
self-reifying dynamic of confusion:
typo
The weight bit interests me. Is this where weights on a graph come from as well? Is that coincidence? I had assumed that they were called weights within the metaphor of a neural network as a graph.
Did you complete your translation? I'd love to read it!