I applied to OpenAI myself back in June. As the interview got closer, I saw on the forum that the interview reports all seemed to be about the same handful of questions, so I decided to prepare those thoroughly and really understand them before going in. I ended up rescheduling three times, and the recruiter got angry — said I'd rescheduled three times already, which meant the timing wasn't right, and cancelled my interview.
Then at the end of July another recruiter reached out again about interviewing at OpenAI. I'd actually kept preparing the whole time in between, and I felt like I really understood the material by now, so I agreed to interview right away.
The phone screen was the usual setup: one coding round, one system design round.
The coding round was the "monster battle" question. I had it memorized, so it should have gone fine.
I didn't expect the problem to come from system design — they threw in an obscure new question I hadn't seen before.
The interviewer had joined OpenAI only a month earlier, and I'm not even sure if he wrote the question himself. Afterward I discussed it with ChatGPT and realized I had completely misunderstood the question.
The question was supposed to be about a power grid dispatch center that sends various commands to different locations — for example, telling all the air conditioners in the Northeast region to cut power usage by 20%. Because these smart devices have unreliable network connections, they might not receive the command at all, or they might receive it and refuse to execute it, or they might agree to execute it. The dispatch center then needs to use the telemetry metrics sent back by these devices to figure out how well the commands were actually carried out, and adjust the next round of commands accordingly. Some telemetry metrics might arrive late, so there might need to be some reconciliation involved too.
But I understood it as: the telemetry metrics were there to tell the dispatch center how much power each device was using, and the control center had a monitoring system that would issue a command to cut power if it detected that some device or location was over its usage limit.
I tried to clarify my understanding with him, but he was noncommittal. I asked him questions and he didn't really answer those either. He did give me a hint partway through — when I was going through my approach and started explaining how to aggregate the power-usage metrics across devices, he said the focus of the question was actually the dispatch center sending out commands. I said, well, don't you have to send commands based on power usage in the first place? You'd only send a command once usage goes over the limit, right? He didn't say anything to that.
The day after the interview I got a message saying I didn't pass. It was probably the system design round that did it.
Discussion
Loading comments…