OpenAI Interview Questions — Page 47

OpenAI·Machine Learning Engineer·Technical Phone Screen

Jun 2026

Interviewed for an ML Engineer role at OpenAI and got a grid simulation problem that felt more like a competitive programming question than anything ML-related. Not what I expected going in.

Given an m×n grid where cells are either infected ('X') or healthy ('.'), infection spreads to all 8 neighbors each day. A healthy cell becomes infected if it has at least T infected neighbors. Write a function that returns the number of days until no new infections occur.

“Multi-source BFS, which I figured out pretty quickly, but the threshold T tripped me up for a bit.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

System design round at OpenAI for a software engineer role. The prompt was a full cloud notebook/IDE system, think Colab, and they wanted real depth on isolation, streaming, and lifecycle management. Pretty demanding scope for a single session.

Design a multi-tenant, browser-based cloud IDE where users can write and run code in isolated sandboxes, with streaming output back to the browser.
How would you choose between VMs, containers, and microVMs as the compute substrate for running untrusted user code?
Walk through your isolation model. How do you isolate filesystem, network, process space, and credentials between tenants?
How would you architect the log and output streaming from a running sandbox back to the browser in near real time?
How do you handle session lifecycle, specifically creation, idle detection, suspension, and resumption of a sandbox?
What's your strategy for persisting workspace files and supporting autoscaling without losing user state?

“This one is massive in scope and I underestimated how much they'd push on the isolation model specifically.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

Two-part technical screen for a software engineer role at OpenAI. Both problems were meaty and required real design thinking, not just coding. Left feeling like I probably could've talked through the tradeoffs more confidently on the second one.

Design and implement a versioned key-value store with set(key, value, timestamp) and get_at(key, timestamp) operations, where get_at returns the value whose timestamp is the largest one less than or equal to the query timestamp. Timestamps can arrive out of order. Discuss data structures, time and space complexity, and edge cases like duplicate timestamps, missing keys, and very large timestamp values.
Given a set of primitive types and directed conversion rules each with a cost, answer queries asking for the minimum cost path from a source type to a target type, or -1 if no path exists. Discuss algorithm choices, complexity, how you'd reconstruct the path, and how the system should handle dynamic rule updates.

“My first instinct was a sorted list per key with binary search, which is basically the right answer, but I fumbled explaining why a plain dict of lists beats a more exotic structure here.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

Got a system design round at OpenAI for a SWE role and the prompt was to design the OpenAI Playground itself. Pretty meta, and I wasn't sure if that was a good or bad sign.

Design the OpenAI Playground: walk through the architecture of a web-based interface that lets users interact with language models in real time.

“The meta aspect of this threw me a bit.”

View Post

OpenAI·Software Engineer·Onsite - Coding / Algorithms

Jun 2026

Coding round at OpenAI for a software engineer role, focused almost entirely on binary protocol implementation. Pretty low-level stuff, which I wasn't expecting, and the error handling requirements made it way more involved than it looked at first glance.

Given ByteReader and ByteWriter interfaces with methods for reading/writing raw bytes, uint32 little-endian values, and strings, implement parse_message and build_message for a binary protocol where each message contains a uint32 ID, a uint32 payload length, and a payload byte sequence.
How would you handle error cases like short reads, invalid payload lengths, and integer overflow in the binary protocol implementation?
Instead of using print statements for debugging, how would you design tests to verify correctness and measure performance of the parse and build functions?

“The interface felt clean at first and I jumped straight into parse_message.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

Interviewed at OpenAI for a search infrastructure role and the system design portion was basically a full deep-dive into every layer of a search engine. It was a lot to cover in one session and I definitely felt the time pressure.

Design a large-scale search system end to end, covering document ingestion, indexing, query processing, ranking, and observability.
How would you handle distributed indexing across shards, and how does query routing work with replication?
Walk through your query understanding pipeline: tokenization, stemming, synonym expansion, and spell correction.
How would you support incremental indexing so that newly published documents appear in search results quickly?
How do you cache hot queries and what are the invalidation challenges?
How would you build an autocomplete or query suggestion feature at scale?
What metrics and observability would you put in place for a search system, covering latency, recall, and relevance?
What are the core trade-offs between freshness, query latency, and ranking quality in a search system?

“This was the whole interview basically.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

OpenAI Research Engineer technical screen focused on debugging a broken Transformer implementation. The task was to find multiple planted bugs, explain them, fix them minimally, and verify the fixes. Pretty intense for a single session.

You're given a Transformer implementation with several planted bugs across attention, masking, tensor shapes, layer norm placement, and positional encoding. Find each bug, explain why it's wrong, apply a minimal fix, and verify with a sanity check.

“This was a lot to hold in your head at once.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

OpenAI system design round for a software engineering role, and it was basically a full ML systems deep-dive crammed into one question. The scope was enormous and I kept second-guessing whether to go broad or deep.

Design a complete machine learning system that ranks multiple candidate text responses to a user query, with the goal of maximizing user satisfaction. Cover problem formulation, data and labeling strategy, model choice, training pipeline, evaluation metrics, online inference architecture, experimentation, safety measures, bias and privacy controls, monitoring, retraining cadence, and cost trade-offs.

“This was one question that took the entire session.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

A system design round at OpenAI that went sideways fast. The interviewer kept pivoting topics mid-discussion, and the whole thing felt like we were both lost in different conversations. CI/CD, image caching, S3 uploads... none of it ever clicked into a coherent thread.

Given a CI/CD system that reads workflow definitions from YAML and runs jobs sequentially, how would you add build caching to it?
How would you support custom image layers in that same CI/CD pipeline?
If two jobs in the pipeline use different compilers with different dependencies, how do you handle the cache between them?
How would you handle partial or failed chunk uploads when uploading build artifacts to S3?

“I got completely turned around here.”

View Post

OpenAI·Software Engineer·Onsite - Coding / Algorithms

Jun 2026

OpenAI SWE interview that revolved around a social network / follow graph problem. The question had multiple layers and kept expanding, which I wasn't fully ready for.

Design a follower/followee data structure that supports timestamped follow events, a point-in-time query to check if A was following B at time t, and a 2-hop friend recommendation ranked by how many mutual intermediaries connect the user to each candidate.
Now redesign it so you can create immutable snapshots of the social graph at any point in time. Subsequent follow() calls should not affect previously created snapshots. Also implement get_followers() efficiently and a top-k recommendation method on the snapshot.
How would you scale this to millions of users? What are the tradeoffs between deep-copying snapshots versus other approaches?

“I started with the basic follow graph using a dict of sets, which was fine.”

View Post