Everything, exponentiated. Softmax turns raw outputs (“logits”) into smooth probabilities, mapping chaos into certainty. It’s the final verdict: every node’s score, normalized, so all probabilities add up to one. In multi-class classification, Softmax stops the guesswork—offering interpretable predictions, one-hot at its limit, nuanced everywhere else. Where opinions contend for truth, Softmax finds the equilibrium. For those who see structure in randomness, this is your uniform.
