superposition-creates-interference-entries

IN premise — summaries/2026/08/24/elhage-2022-toy-models-superposition-chunk-8.md

Created 2026-08-25T02:58:00+00:00

Superposition introduces small non-zero interference entries (ε) in W^TW, changing it from (1, 0, 0, 0, ...) to (1, ε, −ε, ε, ...), and these ε terms constitute the direct mechanism of adversarial vulnerability.

Summary

When superposition is mixed into the system's internal representations, it leaves tiny cross-talk terms in what should otherwise be a perfectly clean, orthogonal structure. These small leaks are not incidental noise; they are the exact structural weakness that adversarial perturbations latch onto to fool the system.