The big picture

Polysemanticity underpins much of mechanistic interpretability. Disproving it would redirect years of alignment research built on that foundation.