Strong coders fail OpenAI's project round by staying shallow. How to go deep.

Candidates say strong coders fail OpenAI's project round by staying shallow. 10 of the 11 AI company deep dives that say what they grade test depth under follow-ups. A week to prepare.

Colorful layers of rock in a cliff face against a deep blue sky
Photo: Jeffrey Eisen on Unsplash
  • Candidates say strong coders fail OpenAI’s project round by staying shallow. 10 of the 11 AI company deep dives that say what they grade test how deep you go under follow-ups.
  • 15 of the 34 AI companies we map spend a round on one project you built, against 5 of 14 big tech companies and scaled startups.
  • Plan for an hour on one project: 6 of the 8 with a known length run 60 minutes.
  • 8 of the 15 start with you presenting; the rest are a conversation from the first minute.

15 of 34 AI companies spend a round on one project you built

Loops with a round about one project you built, by kind of company (n=48). It goes by many names: project presentation, experience deep dive, project walkthrough, take-home review, tech talk.

  1. AI companies15 of 34 loops
  2. Big tech and scaled startups5 of 14 loops

Among them are Anthropic, OpenAI, SpaceXAI (formerly xAI), Google DeepMind, Cursor and Microsoft AI. If your loop is at an AI company, expect it, and expect it late: at Anthropic, it is usually on the second onsite day, which only goes ahead after a good first day.

10 of 11 grade how deep you go under follow-ups

What the AI company deep dives grade, by theme, from each round’s description (n=11; a round can grade more than one thing, and 4 don’t say).

  1. Depth under follow-ups10 of 11
  2. Decisions and tradeoffs5 of 11
  3. What you personally owned4 of 11
  4. Working across teams2 of 11
  5. Explaining it clearly2 of 11
  6. Scale and scope1 of 11

Depth means the follow-ups keep going until you run out. At OpenAI, candidates report two questions on every project: “why not the alternative?” and “what happens at ten times the scale?” At Harvey the round grades the tradeoffs you personally owned, probed until the edge of what you know, and candidates underrate this round; vague answers fail it fastest. At Decagon it is your real role, probed until its edges; details must stay consistent across interviews.

So the project you pick matters less than how far down you can go on it. A famous system you touched for a month runs out at the second “why.” A smaller one you built and ran for a year doesn’t.

8 of 15 start with you presenting

AI company deep dives by how they open (n=15).

  1. You present first8 loops
  2. A conversation7 loops

The ones that open with your presentation: Anthropic, OpenAI, SpaceXAI (formerly xAI), Google DeepMind, Cursor, Applied Intuition, Sierra and Anyscale. At Anthropic you present for about 20 minutes and get probed for the rest of the hour. At OpenAI: you present one or two projects you led, usually with a few slides, then take rapid questions.

What this means for your prep: one week, day by day

  1. Day 1: pick the project you can go five layers deep on. One you owned end to end, recent enough that you still remember the numbers. 10 of 11 rounds grade depth, so choose by depth, not by how impressive it sounds.
  2. Day 2: write a one-page map. The problem, the constraints, the architecture, the numbers (scale, latency, cost), what broke, and which parts were yours. Any line you can’t fill is where the interviewer will go.
  3. Day 3: list every decision and the option you turned down. Write why you rejected each one. Then answer OpenAI’s two questions for every decision: why not the alternative, and what happens at ten times the scale.
  4. Day 4: write down what went wrong. The outage, the wrong call, what you would change now. Cognition’s round is a conversation about a system you built: where it broke and what you would change. Tell that part yourself before the follow-ups dig it out.
  5. Day 5: build a 20-minute deck. 8 of 15 rounds open with your presentation. Keep backup slides with the detail you cut: the schema, the numbers, the diagram of the part you wrote.
  6. Day 6: get grilled. Have a friend or an AI assistant ask “why?” five times on each decision. Wherever you stall, go back to the code or the design doc until you can answer. Say “I” for what you did and “the team” for the rest: 4 of 11 rounds grade what you personally owned.
  7. Day 7: run the whole hour. Present, then take questions for the rest of the 60 minutes. Check that the story matches your resume and what you said in earlier rounds: at Decagon, details must stay consistent across interviews.

Each AI company’s deep dive

  • AnthropicProject deep dive, 60 minutes: You present a project you owned end to end for about 20 minutes; interviewers probe it for the rest of the hour. Grades: your technical decisions, tradeoffs and how you worked across teams.
  • OpenAIProject presentation, 45 minutes: You present one or two projects you led, usually with a few slides, then take rapid questions. Grades: what you personally owned, why you decided what you did, cross-team problems and real scale.
  • SpaceXAI (formerly xAI)Project presentation: Present a large project you owned, then questions. Grades: depth of ownership and how well you hold up under questions.
  • Google DeepMindDomain depth or paper presentation, 60 minutes: Deep technical conversation; research engineers often present and defend a paper. Grades: how many layers past the first answer you can go.
  • CursorProject review and founder conversation: Present and defend what you built; often a conversation with a founder. Grades: whether you can explain every line you shipped, and your product sense.
  • PerplexityProject deep dive: Conversation about a project you built. Grades: hands-on depth across the stack.
  • HarveyProject walkthrough, 60 minutes: Walk through a project you built, then questions. Grades: the tradeoffs you personally owned, probed until the edge of what you know.
  • Applied IntuitionProject presentation, 30 minutes: Present what you built in the onsite, or a prepared past project. Grades: architecture choices and tradeoffs.
  • SierraAgent presentation, 60 minutes: Half a walkthrough of your take-home agent, half a talk on a technical topic of your choice for a non-technical audience. Grades: defending your design, depth, and clear communication.
  • DecagonProject deep dive, 60 minutes: Discussion of a project you owned. Grades: your real role, probed until its edges; details must stay consistent across interviews.
  • Hugging FaceTake-home review, 60 minutes: Walk through and defend your solution, maintainer style. Grades: why you chose what you did, what you would do with more time, how it holds up when constraints change.
  • Microsoft AIProject deep dive: Conversation about past work.
  • CoreWeaveRésumé deep dive: Walk through your past projects.
  • AnyscaleTech talk: Presentation of past work.
  • CognitionProject deep dive: Conversation about a system you built: where it broke and what you would change.

See every reported question for your loop, by round, with when it was last asked.

Pick your company

What the labs asked this week

A weekly email of what OpenAI, Anthropic and the other top AI labs asked engineers, round by round.