AI Neurons Are 'Clean': A Decade of Misreading?

Known and confirmed

• New essay argues superposition theory is a misreading of early evidence.
• Modern network neurons appear monosemantic and directly interpretable.
• Sparse autoencoders may merely copy existing neurons rather than finding new features.

Unknown or unverified

• 该研究对InceptionV1等特定模型的结论是否适用于所有架构尚待验证。
• 注意力机制（Attention）的可解释性仍被认为较难理解。
• 稀疏自编码器在更大规模或不同分布数据上的表现未被完全否定。