← Back to Directory

OpenAI interview notes, page 35 of 67

Mid-sized

Most of these OpenAI write-ups come from Software Engineers (460), then Machine Learning Engineers (137) and Backend Engineers (23). The most reported rounds are the technical phone screen (335), the system design round (229), the onsite coding round (32) and the behavioral / leadership round (22).

667 interview notes · updated Jul 2026

OpenAI·Software Engineer·Hiring Manager Screen

May 2026
Behavioral round with the hiring manager at OpenAI for a software engineer role. Forty-five minutes, one main question, but don't let that fool you, the follow-ups were relentless and went pretty deep.
  • Why do you want to work at OpenAI specifically?

“Thought I had a solid answer prepared but the follow-ups kept coming.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026
OpenAI SWE interview that went deep into ML fundamentals. They had me implement backprop from scratch for a small feedforward net, no libraries, and verify it with a gradient check. Not a vibe-check round.
  • Implement the forward and backward pass for a small feed-forward neural network from scratch, without using any automatic differentiation. The network has configurable layer sizes and activations like ReLU or sigmoid, and uses a loss like MSE or cross-entropy. After implementing, verify your gradients with a numerical gradient check.

“This was the whole interview basically.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026
OpenAI SWE interview with a transformer debugging problem, which was actually pretty fair since they pointed you toward the buggy sections. There was a follow-up requiring classification-specific changes to the prediction and loss logic.
  • Debug a transformer implementation where the buggy code segments are pre-marked for you.
  • Modify the transformer code to handle a classification task, including updating the prediction head and loss function.

“The fact that they flagged which parts had bugs made it way less stressful than I expected.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

May 2026
Interviewed for an ML Engineer role at OpenAI and got a surprisingly algorithmic question about version dependency resolution. Not what I was expecting from an AI company but it made sense in retrospect given how much infra work these teams deal with.
  • You have a list of versioned dependencies (formatted as major.minor.patch) and a black-box API that tells you whether a given version supports a feature. The API is rate-limited, so you can't call it on every version. Support is not monotonic across versions, meaning a higher version might drop support, but some later version is guaranteed to restore it. Design an algorithm to find the earliest version that supports the feature while keeping API calls sub-linear.

“This one took me a minute to even parse.”

View Post

OpenAI·Software Engineer·Onsite - System Design / Architecture

May 2026
System design round at OpenAI for a software engineer role. The whole session was basically one monster problem about distributed credit management with vector clocks, and they kept pulling on threads until I ran out of things to say.
  • Design a credit balance service that processes a stream of add and use operations per user, where credits have expiry times represented by vector clocks instead of wall-clock time. Cover the API, core data structures, and algorithms for efficient inserts and spends.
  • How do you compare vector clock expiries against the 'atVc' timestamp in a use operation, and what tie-breaking policy do you apply when two vector clocks are partially ordered (neither dominates the other)?
  • What safeguards do you put in place to prevent double-spend and to block the use of already-expired credits?
  • Walk through edge cases: credits with identical expiry vector clocks, zero or negative amounts passed to add or use, and a use request where the total available balance is insufficient.
  • Provide a complexity analysis for your insert and spend operations given many active credits per user.

“I started okay with the API surface but the vector clock part threw me off more than I expected.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

May 2026
Interviewed for an ML Engineer role at OpenAI and had to present a technical project I was proud of. Pretty open-ended format, which sounds easy until you're actually in the room trying to decide what counts as 'impressive' to people who work there.
  • Walk us through the most impressive technical project you have worked on.

“I spent way too long deciding which project to pick and ended up going with something technically complex but hard to explain quickly.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

May 2026
Went through a coding round at OpenAI for an ML Engineer role and got hit with a simulation problem that looked straightforward but had a lot of hidden complexity once the follow-ups started piling on.
  • Simulate a hero fighting waves of monsters where both sides deal damage each round. For each wave, determine if the hero survives, how much HP is left, and how many rounds it took.
  • Given multiple possible orderings of monster waves, how would you find the ordering that lets the hero survive the most waves or finish in the fewest rounds?
  • How would you extend the simulation to handle special abilities like healing, shields, or area-of-effect attacks, and how do you decide optimal targeting each round?
  • What are the minimum hero HP and attack stats required to clear all waves without dying?

“The core loop wasn't the hard part.”

View Post

OpenAI·Machine Learning Engineer·Technical Phone Screen

May 2026
OpenAI MLE interview focused heavily on a single matrix cumulative product problem with multiple sub-parts, escalating from a basic implementation all the way to a parallel scan algorithm. The later parts got pretty rough.
  • Write an in-place function to compute the cumulative product of N square matrices. Then analyze the time and space complexity, and explain why this approach breaks gradient computation.
  • Now write an out-of-place version of the cumulative matrix product, including both the forward pass and the backward pass.
  • Given a general Hillis-Steele parallel scan function, implement the forward pass of the cumulative matrix product using it. Then figure out how to use the same scan function to compute the backward pass.

“The implementation itself wasn't bad, just iterate and multiply in place.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026
Coding round at OpenAI for a software engineer position. One algorithmic problem, grid-based BFS, nothing too wild on the surface but the 8-directional spread detail is easy to miss if you're on autopilot.
  • Given an m x n grid where cells are either infected or uninfected, infected cells spread to all 8 neighbors each day. How many days until no new infections occur?

“My first instinct was standard BFS and I almost forgot the diagonals.”

View Post

OpenAI·Software Engineer·Technical Phone Screen

May 2026
OpenAI software engineer interview focused heavily on project depth, not just what you built but every decision along the way and what broke in prod. Two-part structure per project: the design/ownership story, then a real incident from that same system.
  • Walk me through a system you owned end-to-end: what problem it solved, the decisions you made, and what the outcomes were.
  • For that same project, describe a production incident: how you diagnosed it and what you specifically changed to fix it.

“This sounds like a standard 'tell me about a project' but it isn't.”

View Post