Ctrl+Enter to process
7
Prediction
After all the transformer layers have processed your input, the model looks at the final representation of the last token and asks: "Given everything I have seen, what word should come next?" It produces a score for every token in its vocabulary — all 50,000+ of them — and converts those scores into probabilities. The highest-probability tokens are the most likely next words.
Common misconception: The model does not pick one definitive "correct" answer. It produces a full probability distribution over the entire vocabulary. Whether you get a creative or predictable response depends on how you sample from that distribution — that's what the temperature stage controls.
Enter a query above to begin.