Goodfire has released the paper "Neurons as an Interpretable Basis," arguing that MLP (multilayer perceptron) neurons can be studied directly, challenging a long-standing consensus in the field. The team analyzed the Qwen3.6-35B-A3B model using sparsification, attribution tools and a research agent, and found that neurons have three interpretable properties. The paper's authors include Curt Tigges and Nick Cammarata.

Three interpretable properties and experimental findings

According to the research, neurons are already functionally sparse: in each forward pass, the model can still run sparsely even after ablating the weakest 92.2% of MoE (mixture-of-experts) neuron activations. Neurons can also be concisely described, and their behavior is predictable. The team concludes that the network itself can be studied directly, without first transforming it into a feature space.

The conclusion contrasts with the field's consensus that "MLP neurons cannot be studied directly." If the three properties hold, mechanistic interpretability research could skip the intermediate steps of disentangling features before extracting them.

Cammarata essay: polysemanticity evidence was misread

In a parallel 65-minute long-form research-hypothesis essay, "Neural Networks Are Already Clean," Goodfire's Nick Cammarata argues that the outlook for mechanistic interpretability is far better than imagined.

The classic polysemanticity/superposition evidence that convinced the field networks arbitrarily pack concepts was misread — it is mainly a pattern he calls "recursive generalization," which is far less worrying; the polysemanticity in early model organisms came from a few ordinary, basically fixable problems, not superposition.

Cammarata concludes that the network's own mechanisms are readable building blocks, and that neurons and circuits can be studied directly without first disentangling them.

Rival approaches and what to watch

Together the two threads point to a split over the direction of mechanistic interpretability: one is the "fast normal science" route, which advocates observing the network directly; the other is the feature-extraction route, which assumes features first and then reconstructs them. Cammarata's essay is seen as reopening the case for the former.

Questions to watch: whether Goodfire's three properties can be reproduced on other models and architectures; whether Cammarata's "recursive generalization" and "fixable problems" can be independently verified; and how the debate will affect tool choices and resource allocation in mechanistic interpretability research.