Neural Networks Are Already Clean

Commenters spent less time celebrating a breakthrough than trying to pin down the article’s slippery use of clean. The charitable reading was that modern networks may not indiscriminately pack unrelated concepts into individual neurons; if representations and circuits are more separable than feared, researchers can inspect what drives a behavior instead of treating the model as pure statistical fog. Several readers found the interactive model walkthrough unusually concrete, and the larger implication landed: mechanistic interpretability could be indispensable for alignment because safety claims are weak if nobody can inspect the machinery producing them. But confidence in the headline was thin. Critics said the sprawling essay buried its definition and repeatedly restated claims, leaving readers unsure what had actually been demonstrated. The stronger technical pushback was that local evidence of readable features does not establish that nontrivial neural networks are broadly understandable: some argue undecidability places hard limits on global guarantees, while others cautioned that superposition has not been refuted merely because researchers can sometimes study existing neurons and circuits directly. A separate debate questioned whether architecture, parameter count, redundancy, or quantization meaningfully predicts interpretability. There was broad agreement that open weights matter if increasingly autonomous models are expected to be independently audited.