After softmax we have a probability for every word. Now we pick one. Different strategies keep different sets of candidates — context: “the cat sat …”
Greedy always takes the top bar (predictable, can get repetitive). Top-k keeps the k most likely words. Top-p (nucleus) keeps the fewest words whose probabilities add up to p — so it adapts: more words when the model is unsure, fewer when it's confident. Kept words (amber) are renormalized, then one is drawn at random.