Tinker Diffusion language models

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

“ years, to meet tax rules.↵↵Q:” and 16 blanks.

Chapter 4 of 7

Faster and worse

The vendor’s speed comes from stamping many tokens a pass. Larkspur wants to see what fewer passes do to a reply.

This chapter: What do you lose when each pass stamps more tokens?

About this chapter

You will learn

  • How tokens a pass sets the number of passes
  • Why guesses stamped together drift and repeat
  • Why speed and quality are one dial

Words you’ll meet

Tokens a pass
How many blanks each pass stamps. Twice as many a pass, half the passes.
Repeat
A token the same as the one before it, like “the the”. Many repeats: the reply stutters.
Drift
How far a reply moves from the careful fill: its tokens that differ, slot for slot.

Step 1 of 3

Same phrase, same model. sets how many blanks each pass stamps, so how many passes a reply takes.

Predict first.

Stamp all 16 tokens in one pass instead of 16 passes. What happens?