NEURAL NETWORKS ARE ALREADY CLEAN

Summary from the TLDR AI newsletter

The cause of polysemanticity in early model organisms was thought to be superposition. This convinced much of the field that networks pack concepts arbitrarily, but what looked like arbitrary packing was mostly recursive generalization, which is less concerning. This means networks can be studied directly, instead of needing to be untangled first. The outlook for mechanistic interpretability may be far better than we thought.