Interview
Questions
SpaceXSI (xAI) Engineering: Inference and RL Workloads
Practice shared serving systems, RL-inference trade-offs, and reliable model releases. Independent practice based on current SpaceXSI job bodies on the xAI board, not verified interview questions.
Practice these questions directly. An editable copy is yours to change; the original stays unchanged.
An inference service is slower than its capacity estimates suggest. How would you decide whether the limiting factor is host work, GPU computation, memory movement, or communication? Talk through the measurements that would change your next step.
Interactive requests and long reinforcement-learning runs share the same accelerators. How would you schedule the workloads so each gets enough capacity?
A teammate reports a large inference speedup after changing both the serving configuration and the test workload. What would you require before accepting the claim, and how would you turn the result into a reproducible regression check?
Lower-precision inference makes reinforcement-learning runs faster, but the next training run learns less. How would you test whether the precision change caused the regression?
Walk me through how you have implemented CI/CD pipelines for machine learning. What was different about ML pipelines compared to traditional software CI/CD?
Describe a situation where you worked with researchers or scientists who had different priorities than engineering. How did you collaborate effectively?