It only predicts the next word
There's no master plan: the model computes probabilities for the next word, picks one, and starts again. Chain thousands of these decisions and you get answers that look thought-out.
05 Lab
A language model does just one thing: calculate which word comes next. Play with the probabilities, move the temperature and watch it decide live.
Pick a scenario
The sky is ___
Probability of the next word
01
There's no master plan: the model computes probabilities for the next word, picks one, and starts again. Chain thousands of these decisions and you get answers that look thought-out.
It's a mathematical operation that spreads or concentrates probability before rolling the dice. We use the real formula here: each bar is raised to 1/T and renormalized.
When temperature rises, unlikely options gain ground. If one gets picked, the model will continue it with full confidence, as if it were true.
02