LLM System Design Interview, issue 53, Aug 28, 2026

The Fallback Token Paradox

Senior ML Engineer interview at OpenAI, and the interviewer asks:

Your tokenizer fails to round-trip on 0.3% of production traffic, emoji, mixed-script usernames, truncated UTF-8 from a bad client. Your teammate maps them all to UNK and ships. What did he just break?

Don’t say: Nothing much, it’s only 0.3% of traffic, the model will learn to ignore it.

The hidden failure where covering your edge cases physically prevents your model from learning them, and the byte-level trick GPT-2 and Llama use to fix it.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.