When nodes are involved in multiple encodings, the brain or networks can encode far more concepts but at the expense of being harder or impossible to interpret. Accordingly, there is a trade-off between power and interpretability, even between power and security: how do we know what polysemic neural nets are doing or “thinking”?