Looking at the activations, I notice that the neuron shows a very weak activation (0.365) for "won't" in one example, but otherwise has zero or near-zero activations throughout all the documents.
Given the extremely sparse and weak activation pattern, this neuron appears
