Mechanistic interpretability is a joke, supported entirely by non-peer reviewed papers released as marketing material by AI firms.