OpenAI
Applied ML systems, inference-time performance work, and a debugging round on unfamiliar code.
- Active questions
- 41
- Rounds covered
- 3
- Last reviewed
- Aug 27, 2026
Reported questions
The full collection. Each entry shows how many independent reports it has and when it was last seen in a live loop.
- ML designHardCorroborated · 4 reports
Cut p99 serving latency in half
An inference service meets its p50 target but misses p99 by 3x. Walk through how you would diagnose it, then which of batching, caching, quantisation, or admission control you would reach for and why.
Last seen Aug 27, 2026Study: ML systems → - DebuggingMediumCorroborated · 3 reports
Fix the failing test in an unfamiliar repo
You are given a repository you have never seen and one failing test. Locate the defect, explain the root cause before changing anything, and add a regression test that would have caught it.
Last seen Aug 27, 2026Study: Debugging unfamiliar code → - ML designHardSingle report · 2 reports
GPU utilisation is stuck at 45%
A training job holds 45% utilisation on an otherwise idle node. List the candidate bottlenecks in the order you would check them, and the measurement that confirms each.
Last seen Aug 16, 2026Study: ML systems → - ML designMediumSingle report · 1 report
Decide whether a ranking change shipped well
Offline metrics improved but the online engagement metric is flat. Enumerate the plausible explanations, then describe the smallest experiment that distinguishes between them.
Last seen Aug 18, 2026Study: ML systems → - CodingMediumCorroborated · 5 reports
Cache with expiry and bounded memory
Implement a cache with a maximum entry count and a per-entry TTL, with O(1) get and put. Then explain what changes if reads outnumber writes a thousand to one.
Last seen Aug 16, 2026Study: Core coding patterns →
Interviewed at OpenAI recently?
Corroborating a question takes reports from people who did not interview together. Adding yours is what keeps this collection from drifting out of date.