Text is split into subword tokens before a model sees it, which is why models miscount letters, why some languages cost more tokens, and why a prompt's real length is not its character count.
Not yet scoped.
Why
The PoC tokenizes English and Korean with the same tokenizer to show the cost asymmetry and the counting failures, making the invisible preprocessing visible.
How it works
Not yet built.
Where it lands in Jayverse
Number: budget LLM cost by token count per language, not character count. If Number ever pipes readings through an LLM for summarization or tagging, estimate cost separately for Korean and English input, since the same tokenizer charges Korean more per character.
alice-tech: log actual token counts for the daily report. Since alice-tech is generated and read in both languages, track token counts per report rather than word counts, to catch cost drift the two languages produce through the same tokenizer.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
subword
하위 단어(글자보다 크고 단어보다 작은 단위) · 텍스트가 잘리는 단위를 가리킴 · "Text is split into subword tokens"
miscount
잘못 세다 · 모델이 글자 수를 틀리게 세는 이유를 설명 · "why models miscount letters"
cost asymmetry
비용의 비대칭 · 언어마다 토큰 비용이 다르게 드는 현상 · "to show the cost asymmetry"
invisible preprocessing
눈에 안 보이는 전처리 · 토큰화가 겉으로 드러나지 않는 준비 단계임을 강조 · "making the invisible preprocessing visible"
split into
~로 쪼개지다, 분할되다 · 텍스트가 모델에 들어가기 전 토큰으로 나뉨 · "Text is split into subword tokens before"