We already had interpretable models — they were called pure idempotent functions

shakoistsLog · x · 2026-11-14

A witty take resonating in AI circles: mechanistically interpretable models already existed — they were called pure idempotent functions that take typed parameters and return a value. The quip is an ironic comment on today's LLM interpretability research: traditional deterministic code was interpretable by construction, and neural networks' opacity is precisely what created the "interpretability problem."

Original post →

We already had interpretable models — they were called pure idempotent functions

More from Fun