Neural Networks May Already Be Legible at the Neuron Level
Neural networks are already clean. Single neurons and small circuits can be read directly.
That is the central claim of an essay published 2026-10-15 GMT at already-clean.goodfire.com. Already Clean
The author argues the standard evidence for polysemanticity and superposition was misread. Polysemanticity means one neuron responds to unrelated things. Superposition is the idea that a network squeezes in more features than it has neurons to hold them. What looked like unrelated features packed together is described instead as a different pattern, which the essay calls recursive generalization. In this account, superposition was the wrong explanation for polysemanticity in InceptionV1.
The example will be familiar to followers of interpretability, the field that tries to reverse-engineer how models work. The most cited polysemantic neuron responded to both cat faces and the fronts of cars. When InceptionV1, an image-classification model, was first studied, the essay notes, almost none of the neurons in its last layer were easily interpretable. The last layer is the final stage where the model makes its decision.
That changes with architecture, the essay claims. With changes to architecture and no change to model size or training data, 99-100% of neurons in InceptionV1's last layer are easily interpretable. The author reports training versions of the model across a 300x range in size and finding it broadly legible at every scale, and states that researchers can now understand most neurons in models up to trillions of parameters. Already Clean
The broader context here is where researchers place the hard problem. If polysemanticity is largely a result of specific architectural choices rather than a fixed property of trained networks, then neurons and small circuits remain workable. That keeps existing tooling relevant. It also keeps the failure modes familiar. A neuron can still generalize in ways that surprise a human labeler, and recursive generalization, as named here, would need careful handling in any audit.
In my view, practitioners should treat the scale claims as the part to test. Legibility across a 300x range in one model family is useful. Legibility in models up to trillions of parameters is a stronger statement. The essay asserts both. For teams that review, debug, or monitor models, the difference matters. A method that clarifies nearly every last-layer unit in InceptionV1 is helpful. A method that carries over to trillion-parameter systems without new machinery would alter routine work.
Looking at what this means for day-to-day practice, the optimistic reading is straightforward. Clean units lower the cost of inspection. They make it cheaper to trace a behavior to a location, to form a hypothesis, to intervene and retest. I have watched my own children adopt new systems without asking how they work, and enterprise users often do the same. Tools that let engineers answer that question at the neuron level, without first learning a new set of terms, are more likely to be used. Use is what turns interpretability from a research result into operational habit.
