OpenAI Interview Questions — Page 45

OpenAI·Machine Learning Engineer·Technical Phone Screen

Jun 2026

Technical screen for an MLE role at OpenAI. One long, dense debugging question covering basically every failure mode in a PyTorch transformer pipeline at once. Felt more like a written exam than a conversation.

You're given a broken PyTorch transformer training pipeline (encoder-decoder with multi-head attention, AdamW, AMP, gradient accumulation, cosine LR with warmup, teacher forcing) that shows training loss diverging, validation accuracy stuck near chance, intermittent AMP crashes, and nondeterministic results across runs. Identify at least four root-cause bugs, explain the failure mode for each, show a minimal fix, and write a test or assertion that would catch each one. Then lay out a systematic debugging plan covering data preprocessing, masking, positional encodings, loss computation, optimizer state, AMP/GradScaler, seed control, and DDP config. Finally, describe your sanity checks, which metrics you'd monitor, and a small experiment to verify each fix.

“This was a lot.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

OpenAI coding round that picked up where a previous session left off, extending a SocialNetwork class with follower/followee lookup and a real conversation about data structure tradeoffs. Pretty design-heavy for what felt like a continuation problem.

Given a user ID in the SocialNetwork class, implement a method that returns both the list of followers (who follows this user) and the list of followees (who this user follows).
Walk through the tradeoffs between keeping only a forward adjacency map versus maintaining both forward and reverse maps for a social graph.

“The followees side was easy since we already had a forward adjacency map from part one.”

View Post

OpenAI·Software Engineer·Onsite - Coding / Algorithms

Jun 2026

OpenAI software engineer coding round with one problem that took most of the session. The last test case never passed but the interviewer wrapped things up early anyway after I walked through my logic, which was a bit of a relief and also kind of anticlimactic.

Implement a follower/snapshot/recommendation system (a problem specific to this company's interview pool).

“Finished the main logic in about 35 minutes which felt decent, but then got stuck on one edge case in the final test and just could not get it to pass.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

Interviewed for a SWE role at OpenAI and got hit with a pretty deep ML systems coding problem spanning four parts. The whole thing revolved around matrix chain multiplication but kept escalating into autograd internals, manual backprop, and parallel prefix scans. Felt like a research engineering interview more than a standard coding screen.

Implement an in-place matrix chain multiplication function in PyTorch that reuses or overwrites input storage to save memory, and explain why this approach breaks autograd.
Write an out-of-place matrix chain multiplication in PyTorch that leaves all inputs untouched and supports calling backward() to get gradients for every input matrix.
Derive and implement the manual backward pass for matrix chain multiplication: given the upstream gradient dY, compute the gradient for each input matrix without using autograd.
Use a Hillis-Steele parallel prefix scan to compute the forward pass (prefix products) and the backward pass (per-matrix gradients) for matrix chain multiplication. What algebraic property does this require?

“The implementation part was fine, just accumulate into a running product and overwrite.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

75-minute coding round for a Research Scientist role at OpenAI, entirely focused on autograd and matrix multiplication. Pretty brutal if you haven't thought carefully about scan algorithms recently.

Given a stack of matrices W of shape [N, D, D], write an in-place function that computes the cumulative product P = W[0] @ W[1] @ ... @ W[N-1], storing results back into the input array.
Now write the same cumulative matrix product but without modifying the input, returning a new output array instead.
Implement the backpropagation pass for the cumulative matrix product defined above.
Rewrite both the forward pass and the backpropagation using a parallel prefix scan approach (specifically a Hillis-Steele style scan). How does this change the computation graph and gradient flow?

“Seemed straightforward at first.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

OpenAI SWE interview with a system design coding problem that had a weird constraint I didn't see coming. The problem itself wasn't crazy hard but the restriction tripped me up more than I expected.

Design and implement a cluster message aggregation service where messages arrive from multiple nodes, get grouped by a key like cluster ID, and the system returns combined results. All logic must live in a single class with no separate helper classes. Walk through your data model and explain how aggregation and lookup would hold up at scale.

“The single-class constraint is the part that got me.”

View Post

OpenAI·Machine Learning Engineer·Onsite - System Design / Architecture

Jun 2026

Interviewed for an ML Engineer role at OpenAI. One question that stood out was a system design problem around building a data mining pipeline for medical data, specifically extracting useful training data from it. Not a lot of details I can share but it was a meaty problem.

How would you design a system to mine high-quality training data from raw medical data?

“This one had a lot of moving parts.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

OpenAI Research Engineer ML coding round. One question but the follow-ups kept going deeper than I expected, and the dimension-checking got relentless by the end.

Implement a 1-nearest-neighbor classifier from scratch, then integrate it into a neural network forward pass. The interviewer probed on tensor dimensions and shapes throughout.

“I'd seen a version of this floating around before so I wasn't totally blindsided, but the NN forward pass extension was a different beast.”

View Post

OpenAI·Machine Learning Engineer·Onsite - System Design / Architecture

Jun 2026

OpenAI ML Engineer interview that went deep on a systems problem I thought I knew cold. The duplicate file detection question started simple but the distributed follow-up is where things got real.

Given a directory tree with many files, find all groups of files that have identical contents and return their paths. Walk through your algorithm and then explain how you'd scale it to a distributed file system across many machines.

“I started with the obvious stuff: skip files that differ in size before doing any hashing, then bucket by hash (SHA-256 is what I went with, MD5 felt like I was inviting a lecture).”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

OpenAI SWE interview with a system design follow-up that pushed into stateful multi-channel architecture. Pretty focused session, not a lot of fluff.

How would you extend a chat system to support multiple independent channels, each with its own message history and bot state?

“This was a follow-up to whatever the base design question was.”

View Post