Game 03 · Static · measures: perplexity

Next-Token Duel

Predict the next token of a held-out passage. You and the model are scored in bits of surprisal on the token that actually came next. Lower is better — perplexity stops being an abstraction about four tokens in.

seed

wire copy · passage 1 of 4

The central bank raised interest rates by 25 basis

Pick the token that actually comes next. Keys 1–5 work.

token0/16

your bits/tok—

model bits/tok—

your perplexity—

hit rate—

How it works

The distributions are precomputed offline and shipped as data, so the page makes no inference call — the whole duel is a static asset. Your pick is scored as a 70% confidence bet with the remaining 30% spread across the other four candidates, so your loss is −log₂(0.70) when you are right and −log₂(0.075) when you are wrong. The model's loss is −log₂(p_true) under its own distribution. Sum the bits, divide by tokens, exponentiate base 2 and you have perplexity — arrived at by playing rather than by reading a formula.