← Questions

A model about to ship regressed on safety evals

0
Google DeepMindAI/ML designLast asked Nov 2026

An AI/ML design interview question asked in Google DeepMind's coding round, last asked Nov 2026. It tests evaluation, ai safety and model launches.

A new version of a model is days from release. It beats the current model on capability benchmarks, but its scores on the safety evaluation suite went down.

Sign in to unlock this question

The full prompt, the worked answer and your own canvas or code with the AI tutor. New accounts get 5 credits: each one unlocks a question for good.

Where it's asked

  • Google DeepMind · Coding · Last asked Nov 202660 min · Live coding, usually two rounds; research engineers get ML coding (a loss, attention, a sampling routine or a small training loop) · Not allowed or heavily restricted

What it tests

evaluation · ai safety · model launches · debugging

Your canvas and the AI tutor open when you unlock this question.