Machine Learning System Design Interview, issue 23, Dec 7, 2025
The Curse of Multilinguality
Senior ML Engineer interview at Google DeepMind, and the interviewer asks:
“We have a 1B parameter Transformer that is SOTA on 10 languages. We wants to add 90 more languages to the training mix. What happens to our English benchmarks?”
Why adding 90 languages can tank your English benchmarks - and how Adapter Modules prevent capacity dilution.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.