Generative AI
Autoregressive Models
Give every word a probability, cut the unlikely tail, pick one, append it, repeat. A 40-word vocabulary, so you can watch temperature, top-k and top-p act on all of it.
kept by the cut, sampled fromthe word that was writtenthe random numberthe tail that was cut
Every word gets a probability at every step, never zero
Kept by the cut—
—
Decoding
Values
vocabulary—
seen after this context—
kept by the cut—
mass kept—
last pick—
surprise of that pick—
What to observe
- Every one of the 40 words has a bar at every step, never zero. Leave both cuts off and run: sooner or later a tail word wins and the sentence derails.
- Set top-k to 1. The tallest bar always wins, nothing is random, and the text falls into a loop.
- Top-p keeps the fewest words that add up to p, so "kept by the cut" shrinks where the model is sure and grows where it hesitates.
- Drag temperature: the pale bars (the model's own probabilities) stay put while the solid ones flatten as T rises.
Shortcuts:space run/pause · s step · r reset · f fullscreen