Did the OpenAI system design round for a senior (roughly L5) role earlier this year. Took notes afterward because I knew people would ask. Here's the real picture.
Format: 60 minutes, one interviewer, virtual whiteboard (they use Excalidraw or a shared doc depending on who you get). No set problem given in advance.
What I got: design a real-time collaborative document editing system at scale. Think Google Docs but they wanted me to go deep on operational transform vs. CRDTs, conflict resolution, and the storage layer. This wasn't a generic 'design Twitter' problem. It was specific enough that surface-level answers would die fast.
For context, I have 12 years in distributed systems, and this still took real mental energy.
What they seemed to care about: Tradeoffs articulated out loud. Not just picking an approach but saying 'I'm choosing this because X, and I'm accepting the downside of Y.' Capacity estimation done roughly but quickly, not obsessed over. They didn't want me spending 20 minutes on math. Failure modes and how the system degrades gracefully. Operational concerns: monitoring, rollout, what happens when a downstream service blows up.
What didn't matter much: perfect naming of every AWS service or memorizing specific Kafka partition counts. They were clearly more interested in thinking process than vocabulary.
Level calibration note: the interviewer told me afterward (when I got the offer) that what differentiated L5 from L4 in their rubric was whether you could drive the design or needed to be prompted at every step. L5 should be steering, not following.
Prepared using distributed systems fundamentals, reread the Designing Data-Intensive Applications chapters on replication and consistency. That's genuinely the right preparation for this one.
Happy to answer specifics if you have a system design round coming up.