OpenAI Research Engineer Interview Experience — Two Back-to-Back Screens, No Break, Rejected

Sharing a recent phone screen with OpenAI for an RE (security track) role. I got rejected — hope this helps people getting ready for one of these.

Interview format
Two rounds, 60 minutes each, back to back, with no break in between. There was no round-specific prompt, just the generic Technical Interview Preparation PDF the recruiter sends out, which I'm assuming everyone gets the same version of.

• Round 1, general coding: CoderPad, standard library only
• Round 2, ML coding: Google/Colab notebook, numpy / pytorch

1/ General Coding — Infection Spread (5 parts)
This is the classic grid-simulation question from the forum.

• Part 2 adds immune cells
• Part 3 adds a days-to-recovery counter — each cell needs a countdown that flips it to a recovered state once it expires

The interviewer said upfront that we didn't need to finish all 5 parts. I didn't finish all of them.

2/ ML Coding — 1-NN → Wx+b

• Implement vectorized 1-NN with L2 distance, no for loops allowed
• Follow-up: express 1-NN as the forward pass of a neural network (linear layer + softmax)
• Another follow-up: how would you implement it with a neural network if you switched to L1 distance

Mostly numpy, and there were no test cases given in the notebook.

A possibly useful observation: round length predicts question type
This is my own summary after going through several months of RS/RE interview reports on the forum — I don't think anyone's written it up explicitly yet:

The length of the ML coding round seems to predict the type of question.

• A 75-minute ML round is basically always the matrix cumprod / autograd / Hillis-Steele scan family (in-place forward → why can't you get the gradient → out-of-place forward + backward → parallel scan)
• A 60-minute ML round is basically always numpy + numerical-stability questions: streaming entropy, 1-NN, cross entropy

Three separate posts all fit this pattern.

Another observation: people who get streaming entropy are almost always on a four-round loop, and that same loop also has the 75-minute autograd question. The two questions seem to show up paired together.

So if your invite says 60 minutes and this is your only ML coding round, I'd personally recommend focusing your prep on the numpy-style questions and not putting too much time into autograd.

One more thing: out of every RS/RE interview report I saw in July, nobody drew the legendary Transformer-debugging question, even though it still shows up a lot in the older question bank. Could be cooling off, or my sample size could just be too small.

Post-mortem
I don't think I failed because of missing knowledge — I'd prepared basically everything I needed to. The real problem was that two straight hours of genuinely high-pressure interviewing is nothing like mocking it at home by yourself. By the second round my brain was visibly not working anymore, and that had never happened in any of my practice runs.

Two pieces of advice:

1. Do at least one full two-hour mock, and try to find a stranger to do it with rather than always practicing alone. You can't test for the pressure variable when you're the one running your own practice.
2. There's no buffer between the two rounds. If the first round goes badly, that feeling carries straight into the second one. Figure out ahead of time how you're going to reset yourself between rounds.