LLM seminar · MA Design for Digital Futures · TH Nürnberg

What comes next?

Type any text. A small language model, Qwen2.5-1.5B, splits it into tokens and gives every possible next token a probability. Pick one, roll the dice, or change the temperature.

Starting…

Start text

Want to give the model a personality? Switch to “Answer as assistant”. Only the chat format has a system prompt. In plain text, the persona would have to be part of the text itself.

…tokens the model reads Point at a token to see its number.

The next token

All possible next tokens laid end to end, most likely first. Rolling the dice drops a pin at a random point on this strip.
Waiting for the model…
probability at this temperature the model's own probability (temperature 1) most likely

Temperature

1.0
00.511.52

At temperature T, every probability is raised to the power 1/T, then the list is scaled back to 100%. Below 1 the favourite gets stronger. Above 1 the rare tokens catch up. At 0 the model always takes the favourite.

Write

Or click any token in the list. Click a word in the sentence to go back to that point. ↑ ↓ change the temperature. Shortcuts work when the cursor is not in the text box (press Esc).