OpenAI Research Engineer Interview Loop
OpenAI's research engineer interview starts with a recruiter screen, then several coding rounds. Usually there are two ML coding rounds in PyTorch or NumPy and two general coding rounds, 60 to 75 minutes each, graded on passing tests. Next is an ML and statistics round that candidates call the hardest. It can reach graduate-level information theory. Then there's ML debugging on a transformer, and a 60-minute presentation of a past project with slides, sometimes with a hiring-manager conversation. The final loop is four to six sessions over one or two days. There's no SQL. Whether you can use AI tools depends on the round.
What decides it
The ML and statistics round, and code that passes under time. Of the frontier labs, this is the research loop with the most coding.
Practice by round
Easy coding: algorithm basics, and model building blocks such as a numerically stable softmax, cross-entropy, a linear layer's backward pass, and masked mean pooling.
The ML coding OpenAI asks at onsite level: attention, a KV cache for incremental decoding, top-p filtering, byte-pair encoding merges, and an Adam step from scratch. Next to them are practical problems such as an LRU cache for model responses. The tests have to pass. Practice in debug mode as well, for the transformer debugging round.
Systems design: serving under a p99 latency budget, data versus model parallelism, batching inference on one GPU, and an evaluation harness for weekly releases. For the project presentation, prepare four or five slides on one project and practice the follow-up questions.
| Case study | Difficulty |
|---|---|
| Inference Batching, One GPU | Medium |
| Checkpointing Multi-Day Training | Medium |
| Evaluation Harness for Weekly Shipping | Medium |
| Serving Under a p99 Budget | Hard |
| Data vs. Model Parallelism | Hard |
| Data project | Difficulty |
|---|---|
| Next Word Prediction | Hard |
The statistics of evaluation, like a win rate's interval, many slices tested at once, and estimating pass@k. Then LLM concepts: LLM-as-judge bias, regression eval sets and prompt injection. The real round reaches information theory, so know entropy, cross-entropy and KL divergence well.