Next-token sampling
One random draw. Weighted choices.
First, see what a real Galton board does—then compare the useful analogy with model sampling.
Simulation context
Balls fall through nine rows of pegs into ten position bins.
thecatsatonthe?
Physical experiment: equal left/right peg deflections collect into position bins.
What this shows — and what it does not
The boundary: a symmetric Galton board demonstrates accumulated random choices. It does not reproduce an arbitrary language-model distribution.
Counts update as draws land. Histogram bars are revealed only when the current run has finished.