AI & Technical question

You’re given a new model that improves accuracy by 20% but doubles latency. Would you ship it? Walk me through your decision.

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Ability to reason through a real product tradeoff between model quality and user experience, with a clear decision framework, not just a gut answer.

Clarify the use case, since latency tolerance differs hugely between a real time chat assistant and an offline task like summarization.
Quantify both sides: the actual latency change and what the 20 percent accuracy gain means in user facing terms.
Consider a middle path: routing complex queries to the accurate model while keeping the faster model for simple ones.
Propose testing both against user facing metrics, task success rate and satisfaction, not accuracy or latency in isolation.
State a conditional recommendation: ship if the use case tolerates the added latency and hybrid routing is not feasible, otherwise hold or route selectively.

Refuses to answer in the abstract and ties the decision explicitly to the use case's latency tolerance.
Proposes a hybrid routing solution instead of an all or nothing choice.
Grounds the decision in user facing outcomes rather than raw accuracy or latency numbers alone.

Giving a flat yes or no with no reasoning about the specific use case's latency sensitivity.
Not considering a hybrid or selective routing approach as a real alternative.

How would you measure whether users actually notice or care about the latency increase?
What would change your answer if this were a real time voice assistant instead of a chat tool?