This model read 30 short sentences and counted how many times each word came after each other word. The probabilities it suggests come from those counts.
Temperature controls the choice. At 0 it always picks the most likely word and repeats the same sentence. As it goes up, it also picks rarer words: sentences get surprising, sometimes silly.
This tiny model only looks at the last word, so it wanders off topic. Real chatbots ask the same question (“what comes next?”) but they look at thousands of words of context and have read billions of sentences. They don’t know what’s true, they pick what’s likely.