Goodfire

Goodfire’s page hosts an informal research essay titled “Neural Networks Are Already Clean,” presenting a working hypothesis about neurons, generalization, and the ecology of circuits. It is published as part of Goodfire’s broader research output rather than as a standalone product or interactive tool.

The essay revisits canonical evidence on polysemantic circuits and superposition, then explores how neurons change as vision models grow. It extends the analysis to language models, examining native neurons and their circuits, before zooming out from single neurons to the whole network.

The content is aimed at mechanistic interpretability researchers, ML engineers, and advanced students who want a detailed, hypothesis-driven discussion of how legible neural network mechanisms may be. It assumes familiarity with model internals and interpretability concepts.

Readers should treat it as a research essay with a stated working hypothesis, not settled consensus, and follow its cited prior work for context. It is best used as a deep reading resource alongside related Goodfire research.