OpenAI Interview Questions — Page 36

OpenAI·Research Engineer·Technical Phone Screen

May 2026

Oral ML design round at OpenAI focused almost entirely on RAG and search, with the interviewer drilling hard into contrastive learning mechanics rather than high-level system architecture. Narrower than I expected but manageable if you've done the reading.

How is a text embedding model trained? Walk me through the contrastive learning setup.
What loss function is used in contrastive learning for embeddings, and can you write it out?
How do you select positive and negative samples for contrastive training, and how many negatives do you use?
What happens to the contrastive loss when you double the batch size, or when you inject hard negatives?
Describe the retrieval flow in a RAG system, including hybrid retrieval and ranking.

“This was the core of the whole round.”

View Post

OpenAI·Machine Learning Engineer·Onsite - Behavioral / Leadership

May 2026

Behavioral round at OpenAI for an ML Engineer role, focused almost entirely on motivation and mission fit. Two questions, both variations of the same theme, which I wasn't expecting to take up so much of the conversation.

Why do you want to join this company?
How does your personal motivation connect to what this company is trying to do?

“Prepared something structured but it came out sounding a bit rehearsed.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

May 2026

OpenAI ML engineer interview built around a Jupyter notebook with escalating probability problems tied to LLM decoding behavior. Pretty unusual format, felt more like a research exercise than a standard coding screen.

Given that an LLM samples tokens sequentially and stops when it emits an end-of-sequence token, derive the probability distribution of the stopping time (i.e., the number of tokens generated before stopping).
What is the expected sequence length and how does the tail of the stopping time distribution behave under the assumptions you derived?
Design an algorithm that defends against adversaries attempting to manipulate an LLM's stopping behavior, using your probability analysis as the foundation.

“This is basically a geometric distribution setup if you assume constant per-token stopping probability, but the notebook pushed you to think about what happens when that probability isn't constant across positions.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

May 2026

Coding round for an ML Engineer role at OpenAI. The main problem was a grid infection simulation with BFS, plus a follow-up about handling multiple test cases efficiently. Pretty algorithmic for an ML role, which I wasn't fully expecting.

You're given a 2D grid where each cell is empty, healthy, infected, or an obstacle. Every minute, infected cells spread to orthogonally adjacent healthy neighbors. Implement a simulation that answers: (a) the minimum number of minutes until the entire grid is infected, or -1 if it's impossible, and (b) the state of the grid at a given minute t.
Follow-up: how would you extend your solution to handle multiple test cases with larger grids, and what performance characteristics would you expect?

“Multi-source BFS was the right call here.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026

Got a NumPy implementation question at OpenAI for a SWE role, the kind of thing that looks clean on paper but has enough gotchas to trip you up if you haven't thought carefully about numerical stability before.

Implement numerically stable softmax cross-entropy loss and its gradient with respect to the input logits from scratch in NumPy, no autograd allowed.

“The numerically stable part is where people slip up.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026

75-minute systems coding screen for an infra/inference team at OpenAI. The whole thing was basically one deep problem: design a memory allocator from scratch, then keep optimizing it until the interviewer was satisfied with the complexity.

Design a memory allocator with malloc and free operations. malloc should return the start address of the leftmost available block of sufficient size, and free should release memory and automatically merge with adjacent free blocks. The interviewer will reject O(n) solutions and push you toward O(log m).
How would you handle the four merge cases when freeing a block, specifically when adjacent blocks may or may not already be free?
How would you extend the allocator to ensure all returned addresses are aligned to a specific boundary, like multiples of 8?
How would you implement realloc, resizing an allocated block in-place while preserving its contents?
What locking strategy would you use to make the allocator thread-safe under concurrent access?
How would you detect and guard against double-free or use-after-free bugs in your allocator?

“Started with the linked list approach because it felt natural and I wanted to get something working.”

View Post

OpenAI·Backend Engineer·Technical Phone Screen

May 2026

Interviewed for a backend role at OpenAI and got hit with a pretty heavy policy/safety question that I wasn't expecting to go that deep in a technical screen. It felt more like a research discussion than an engineering interview, which threw me off a bit.

What risks do you associate with AGI development, and how would you think through categories like misuse, misalignment, power concentration, economic disruption, security, and autonomy? Which do you weigh most heavily, and what mitigations or research directions seem most promising?

“This was a lot.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

May 2026

System design round at OpenAI for a software engineering role, focused entirely on designing a chess game from scratch and then scaling it into a full multiplayer service. Pretty deep dive, way more ground to cover than I expected going in.

Design a chess game system, starting with the core domain model: Board, Piece types (Pawn, Rook, Knight, Bishop, Queen, King), Player, Game, and Move.
How would you implement move legality checks, including detection of check, checkmate, and stalemate?
Walk through how you'd handle special moves like castling, en passant, and pawn promotion.
How would you represent game state using FEN notation, and how would you implement move history and undo functionality?
How would you extend this into a multiplayer online service? Think about matchmaking, real-time move synchronization, persistence, ELO ratings, anti-cheat, and spectator support.
What scaling considerations would you address for this chess platform at high traffic?

“I started with the Piece class hierarchy and went abstract base class with subclasses per piece type.”

View Post

OpenAI·Software Engineer·Onsite - Behavioral / Leadership

May 2026

Behavioral round at OpenAI for a software engineering role. Three questions, pretty standard stuff but the delivery speed one tripped me up more than I expected.

Why do you want to work at OpenAI specifically?
Tell me about a time you had a conflict with a manager and how you handled it.
How do you move fast and ship quickly without causing serious production incidents?

“Prepared something for this but it came out a bit rehearsed.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026

OpenAI phone screen for a software engineer role. Five questions total, got through three cleanly, stumbled on the fourth, never even saw the fifth. Moved forward anyway, which was a relief.

Solve the 'Infected Plants' problem and its sub-questions (five parts total).
Design a solution that solves a crossboard traversal problem as part of the infected plants series.
Design a distributed DFS system. How would you handle fault tolerance, early termination, constraint propagation, and deciding when to split work across nodes versus running DFS locally?

“Got through the first three parts fine.”

View Post