just finished the openai process for an MLE role on one of the applied teams. took about 4.5 weeks total.
recruiter screen was 30 min, mostly background and motivations. they asked what I'd worked on in ML that I was genuinely proud of and why. felt more like a real conversation than a screen.
technical phone screen: two coding problems. one was graph traversal, one was about parsing and handling malformed inputs. 45 minutes, you'll feel rushed. do NOT overthink the first problem, just solve it.
onsite was 5 rounds over two days (virtual): coding x2 (one was deceptively simple, one had a tricky constraint that required rethinking) ML system design: I got a question about building a recommendation system with certain latency constraints. they pushed on how I'd handle data freshness vs cold start. no right answer, but you need positions. ML fundamentals: training dynamics, gradient questions, when would you regularize vs not. this was genuinely fun for me. if you can't explain why your learning rate matters, you're going to struggle. behavioral: one interviewer, warm, 4 questions. leadership, navigating ambiguity, a time I disagreed with a technical direction.
what surprised me: the ML design round was more about tradeoffs and honesty than about knowing the "right" architecture. saying "I'd actually start with something simpler and see where it breaks" landed better than proposing an overengineered beast.
also they asked what I thought about AI safety. not a gotcha, but they wanted to know I'd thought about it. have a genuine answer ready.