LLM System Design Interview, issue 51, Aug 26, 2026
The Tokenizer Swap Trap
Staff ML Engineer interview at Google, and the interviewer asks:
“Your tokenizer gets 2.5 bytes/token on English but 1.1 on Vietnamese and Thai. Your teammate wants to grow the vocab from 100k to 250k. Do you ship it?”
Don’t say: “Yes, bigger vocab means better compression and shorter sequences.”
Why treating vocabulary size like a simple config change is a billion-parameter mistake, and the elite approach to multilingual token fertility.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.