There are good theoretical reasons to think superposition exists in neural networks, particularly in the residual stream rather than specifically in MLP neuron activations. It wasn't just a single example that got people thinking about superposition.