The pond text has nineteen words in its vocabulary, so this machine turns one word into nineteen scores: a four-number embedding for the word you click, a first layer of six neurons with ReLU, and a second layer of nineteen neurons with no ReLU, one for every word the machine could say next. Every embedding, weight and bias is a random number between −1 and 1, made when the page loaded. The model ends where the scores come out. Picking one is your job.
Reading the score buttons and the output circles: a solid border with ▲ above the number means the score is above zero, and a dashed border with □ means it is at or below zero. A dotted border and no number means that stage of the machine has not run yet.
Click a word on the left to send it through the machine. Then click a score on the right to add that word to the sentence and score whatever could follow it. The × on the sentence line empties the sentence and changes nothing else. Randomize everything builds a new machine.