Essay argues neural networks are already clean

In an essay, Goodfire's Nick Cammarata argues that polysemanticity in early models stemmed not from superposition but from ordinary, mostly fixable issues, so networks can be studied directly via their neurons and circuits, making the outlook for mechanistic interpretability far better than feared.

• Researcher's 26,000-word post claims AI models may be easier to understand2026-10-19
• Goodfire launches cheaper AI agent monitors2026-10-08