This is a special post for quick takes by Alex Gibson. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
A common way of interpreting the activations of a neuron is to say that it measures some abstract concept with a notion of degree. For instance, that a neuron measures the brightness of an image, or that it measures how strongly positive sentiment is, or how "trumpian" something is.
This seemingly reduces the interpretation of an activation to essentially a classification problem. We just say what the high activating examples are, and then if a neuron is "monosemantic", the lower activations will merely be lower "strengths" of the concept.
But if this were a complete explanation, the same explanation would apply to any monotonic continuous transformation of the neuron's activations acting as the identity on the minimum and maximum activations. If the network is not robust to such monotonic transformations of the ReLU activations of monosemantic neurons, then the explanation must be incomplete.
But there's no reason to think that the model is actually using a sparse set of components /features on any given forward pass.
I contest this. If a model wants to implement more computations (for example, logic gates) in a layer than that layer has neurons, the known methods for doing this rely on few computations being used (that is, receiving a non-baseline input) on any given forward pass.