Sharing my recent OpenAI Research phone screen.
Round 1: general coding, 1hr
Debug a distributed system. They gave me a buggy job scheduler written in Python — I had to check for data races, deadlocks, lock contention, and whether the rate limiter was implemented correctly. Partway through I had to write my own test cases to confirm the scheduler actually worked, then calculate how long the scheduler took to schedule jobs and the success rate.
Haven't seen this one on the forum before.
Round 2: ML debug, 1hr
Debug a transformer. Get a miniGPT to produce correct text, then as a follow-up implement KV cache, all in PyTorch.
Seen this one on the forum.
Round 3: ML coding, 1hr
Matrix multiplication forward and backprop. They gave a code skeleton, implement it in PyTorch. The follow-up was kind of like implementing the backprop using the Hillis-Steele scan algorithm.
Seen the main question on the forum, but not the follow-up.
Round 4: system design, 1hr
Mining novel data from a large unlabeled corpus. Design an ML system that can find interesting images out of a huge pile of images. You can use human labelers, but you can't over-rely on them.
My feeling is that as long as you focus on communicating with the interviewer, clearly walk through your thinking, and lay things out in pseudocode, the interviewer is still willing to give you some hints.
There's an onsite after the phone screen.
Discussion
Loading comments…