OpenAI Interview Questions — Page 51

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

Jun 2026

Onsite system design round at OpenAI for an ML Engineer role. The main question was a deep debugging scenario around a broken Transformer training pipeline, and they really wanted to see how you'd systematically tear it apart rather than just throw fixes at it.

You have a Transformer-based sequence model that intermittently throws shape or dtype mismatch errors and stops converging after several thousand training steps. Walk through your complete debugging approach from start to finish, covering tokenization, attention masks, gradient issues, data bugs, numerical instability, checkpointing, multi-GPU problems, and memory profiling.

“This question is enormous and I think that's the point.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

Got a system design coding question at OpenAI for a software engineering role. The problem was about building a GPU credit manager with out-of-order operation support, which sounds manageable until you actually sit down and think through the edge cases.

Design and implement a GPU credit manager that supports out-of-order operations. It needs three operations: adding credits with an ID, amount, timestamp, and expiration window; charging credits at a given timestamp (always consuming soonest-expiring credits first); and querying total unexpired balance at a given timestamp. Calls can arrive in arbitrary time order, and the whole thing should run at O(log n) per operation for around 100k total calls.

“The expiration-first ordering for charge() is where I spent most of my time.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

System design round at OpenAI for a software engineering role. The whole thing was one big question about building a CI/CD platform from scratch, GitHub Actions style. A lot of ground to cover and I think I spread myself too thin on some parts.

Design a CI/CD system similar to GitHub Actions, where users define workflows in YAML files and the system triggers jobs on events like pushes, pull requests, or schedules. Cover the full stack: workflow model, event ingestion, job scheduling, runner fleet management, artifact and log storage, secrets handling, retries, and scaling under bursty load.

“This is a monster of a question and I underestimated how much they wanted me to go wide before going deep.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

System design round at OpenAI for a software engineering role. The whole session was basically one big question about building a hosted notebook service, and they went pretty deep on every layer of it.

Design a hosted notebook and compute workspace service similar to Google Colab, supporting workspace creation, deletion, suspension, and resumption with under 5 seconds resume time and state preservation.
How would you handle resource scheduling across worker nodes to support hundreds of thousands of concurrent active users?
What happens when a worker node crashes while a user has an active session? How do you recover, and how do you make sure user data isn't lost?
How do you ensure strong isolation between users who are sharing the same underlying infrastructure?

“This was the core question and it sprawled in every direction.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

OpenAI Research Engineer interview with a meaty coding problem that took up the whole session. Four progressive sub-questions on package dependency resolution, lots of code to write, and the expectation that it'd all be modular and testable. Not a quick whiteboard exercise.

Build a Python package dependency resolver that takes a set of packages with version constraints (like 'numpy>=1.20,<2.0') and resolves a consistent set of versions satisfying all constraints. The problem is broken into four progressive parts: basic resolution, conflict detection, version-range intersection, and reporting unresolvable cycles.

“This was basically a full coding project compressed into an interview.”

View Post

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

Jun 2026

Interviewed for an ML Engineer role at OpenAI and got hit with a massive system design question covering basically the entire lifecycle of an enterprise document search product. Single question but it sprawled into a dozen sub-topics. Left feeling like I'd answered maybe 60% of what they were actually probing for.

Design an ML-powered enterprise document search system. Cover requirements like latency SLA, data freshness, multi-tenant access control, and PII handling. Walk through the indexing pipeline, feature generation, storage, hybrid retrieval, ranking, query understanding, personalization, evaluation, monitoring, and how you enforce ACLs at query time without leaking documents.

“This question just kept going.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

Jun 2026

Went through a technical screen for an MLE role at OpenAI and got hit with a deep systems question on KV caching. Not a leetcode grind session, more like a whiteboard design conversation that kept going deeper every time I thought I'd covered it.

During transformer inference, what is a key-value cache? Walk through what gets stored, why it speeds up autoregressive decoding, the difference between the prefill and decode phases, the memory tradeoffs, and at least two production-level optimizations.

“This question kept expanding.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

Jun 2026

System design round at OpenAI for a software engineering role, centered entirely on designing a extensible chat app with bot routing. Two-part problem that escalated from a single-channel design to multi-channel state isolation. Pretty involved for what sounded like a straightforward OOP question at first.

Design a ChatApp class that dispatches user messages to multiple registered bots, where adding a new bot type requires zero changes to the core ChatApp code. Show the bot interface, the dispatch loop, and at least one concrete bot.
Where should per-conversation bot state live if you want a single bot instance to be reusable across many conversations?
Extend the design to support multiple independent channels, each with its own message log and per-bot state, without changing the dispatch logic.
How would you let bots subscribe to only certain messages, like by command prefix, without adding that filtering logic to ChatApp?
How would you isolate errors so a single misbehaving bot doesn't break message delivery for the rest of the channel?
If bots were asynchronous and made external API calls, how would the dispatch contract and message log need to change?
Where are the race conditions if multiple users send into the same channel concurrently, and how would you make the design thread-safe?

“The core trap I almost fell into was putting any bot-specific logic inside the dispatch loop.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

Coding round at OpenAI for a software engineer role. One problem, grid-based BFS with a twist. Felt manageable if you've seen the classic infection spread problem before.

Given a grid where cells are either infected ('X'), healthy ('.'), or immune ('I'), find the number of days until the infection stops spreading. Infected cells spread to all 8 neighbors each day, but immune cells can never be infected and block the spread.

“The immune cells are the real wrinkle here.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

Jun 2026

Interviewed for a software engineering role at OpenAI and got a bit-manipulation problem around CIDR blocks. Niche enough that I had to actually think, not just pattern-match to something I'd drilled before.

Given a CIDR block like 10.0.0.0/24, write code to enumerate all IPv4 addresses it contains in order, using bit manipulation to derive the network address and iterate through all host-bit combinations.

“I knew what a CIDR block was but hadn't thought about it in terms of raw bit ops in a while.”

View Post