eh2022-superposition-supports-computation
IN premise — summaries/2026/08/24/elhage-2022-toy-models-superposition-chunk-10.md
Created 2026-08-25T02:57:58+00:00
Neural networks can perform non-linear computation entirely within a superposed (lower-dimensional) hidden-layer representation, not merely store features in superposition
Summary
Neural networks aren't just cramming lots of concepts into a small number of overlapping neurons for storage; they can actually do real, non-linear mathematical processing while operating in that compressed, overlapping space. This means the "messy" mixed-up representations we observe are the genuine working medium of computation, not a side effect, which has big implications for how we interpret and model what the network is doing internally.