2 · One question → next token · 21/38

BPE on real strings

English vs Go vs YAML. Bytes per token.

play
same merge rule · three corpora · bytes/token is the fair meter
en —· go —· yml —

Go and YAML compress worse. That is more steps, more KV, more money.

cl100k: en 5.05 / Go 4.00 / YAML 2.72 b/tok. YAML keys are already merges. Go stays expensive.