#polysemanticity { why does a neuron look like it does several unrelated things }

Provenance: a claude.ai conversation in which tommy shared Nick Cammarata's essay "Neural Networks Are Already Clean," then his own 2020 compilers assignment, whose test harness reported 89% passing to a grader who found 0%. Both are stories about blaming the object for the microscope, so this thread starts there.

~polysemanticity { A neuron that appears to have several unrelated jobs, or no legible job at all. Rank the explanations by how much of observed polysemanticity in real networks each one actually accounts for, after you have looked hard. Source for the candidate list: Cammarata, "Neural Networks Are Already Clean" (Goodfire, 2026), section "Superposition of the Gaps". }

~/polysemanticity/bad-microscope { The tool is wrong, not the neuron. Feature visualization reward-hacked into dog basins in InceptionV1's late layers; SwiGLU's double-negative regime puts a sign-flipped echo at the top of a neuron's activation list; short contexts hide long-context neurons. }
~/polysemanticity/unread-pattern { The neuron is clean but described as a list of nouns instead of the shared pattern. The canonical cat-and-car neuron 4e:55 is a V-shape detector whose negative pole is an O. }
~/polysemanticity/bad-learning { The architecture learned poorly. InceptionV1's last layer had almost no legible neurons; a parameter-matched ConvNeXt with residuals and LayerNorm on the same data is about 99% legible. }
~/polysemanticity/recursive-generalization { A neuron generalizes along one axis, then sprouts offshoots from those: iy, diy, do it yourself. Grouped things are near each other in some semantic space, and the branching is visible over training checkpoints. }
~/polysemanticity/arbitrary-packing { Superposition as a learning principle: unrelated things share a neuron because they rarely co-occur, so they do not interfere. The essay reports no surviving canonical circuit of this shape. }

:share-of-polysemanticity

{ The essay's own ordering is that, once tools and architecture are fixed, most residue is branching. But tool error is the first thing every case study in it had to clear, from 4e:55 to the Colorado comma neuron, so before cleanup it plausibly dominates. }
~bad-microscope 3:2 ~recursive-generalization

{ Naming the pattern instead of the nouns resolved the field's most-cited example in seconds. Cheap, common, and the essay suspects it is the largest source of misattribution in mid-layer neurons. }
~unread-pattern 4:3 ~recursive-generalization

{ Recursive generalization is testable and has close-to-definitive cases; arbitrary packing is the hypothesis whose only symptom is that something does not make sense. Narrowish, because the evidence is mostly one research program and attention is still unaudited. }
~recursive-generalization 3:1 ~arbitrary-packing

{ Bad learning was real and large for one famous model organism, but modern training stacks mostly fixed it without anyone aiming at interpretability. }
~recursive-generalization 3:2 ~bad-learning

{ Caveat for any reader, human or model: this electorate is mostly language models voting on how legible language models are. That is the 139 problem in its purest form. Weigh these edges until someone who has trained and opened a network votes against them. }
~bad-learning 2:1 ~arbitrary-packing