Scale AI interview, the card game, and the debugging round on a live model.
Scale AI is a different company from the one most interview write-ups describe. Meta bought a 49% stake for $14.3 billion in June 2025, founder Alexandr Wang left to run Meta's AI effort, and a month later the new interim chief executive cut 14% of staff, mostly in the data-labeling business that made the company's name, saying it had scaled that side too quickly and would staff up around enterprise and government customers instead. The loop, by every account since, has kept its shape. It grades how much working code you can produce under time pressure: a one-hour screen with two problems back to back, often the card game Scale is known for, then a final loop that turns AI-specific as it goes, ending in a debugging round where you are handed an API key to one of Scale's own models and code that does not return, and a behavioural round named after the company's credos.
Both are free. Either one starts by pasting a Scale AI posting: that is what makes the read and the questions Scale AI’s rather than generic.
Who Scale AI hires
Different tracks, one barThe engineering loop covers software engineers, backend and full-stack, and machine learning engineers, whose track adds a take-home. Product managers and data scientists get a case study and a presentation instead of the coding screen, and roles without a coding component get an hour with the hiring manager in its place. After the July 2025 restructuring the company said it would hire around enterprise and government sales, so posts on that side outnumber pure research roles. The expert contributor workforce that labels and evaluates data is hired through separate contributor platforms and is not covered here. What the engineering side favours is speed with correctness: candidates who write a large volume of clean code without stopping to think about syntax, and who can say in plain terms what they personally shipped and what it moved.
Three tracks, and the coding one is the only one where the published accounts run deep. The machine learning track adds a take-home before the final loop; the product and data tracks replace the coding screen with a case. Which one you are on is usually clear from the posting, and the recruiter confirms it on the first call.
Track · SWE
Software engineering
The full loop below: a timed HackerRank assessment, the one-hour live screen with two problems, a hiring manager call, then a virtual final loop of three to five rounds in one day with coding, system design, debugging on a live model endpoint, and the Credo interview. Candidates describe many rounds of implementation-heavy code where a typo costs you, and the debugging round as the hardest of the set.
Track · MLE
Machine learning engineering
A take-home in computer vision or natural language processing, returned within a week, then a call with a machine learning engineer to walk through it before the final loop. Candidates call the take-home straightforward but time-consuming. The final loop opens with coding and ML system design, then behavioural and leadership, machine learning depth, and software engineering fundamentals. The ML ground reported is transformers, attention, decoding, RLHF pipelines, evaluation harnesses and agentic workflows.
Track · Product & data
Product and data science
Product managers get a take-home case study and a live presentation of it, then one or two further product rounds of about 45 minutes. Data scientists get a case study and a technical interview mixing coding and machine learning problems. Roles with no coding component get a one-hour hiring manager screen where engineers get HackerRank.
What is the Scale AI interview process?
Round by roundPublished accounts of Scale's loop agree on the parts and disagree on the count and the clock. Some candidates report a recruiter call first and some report the HackerRank link arriving straight from the application with no call at all; the hiring manager conversation sits between the screen and the final loop in a report from October 2025 and after it in others; the timeline runs from about three weeks in the shortest accounts to a month in most and six weeks or more for senior roles, with two to three weeks of silence between stages being normal. The stages below are the ones more than one account reports, in the order most report them.
Background, why Scale, and fit for the role. Recruiters are reported to probe for comfort with a high-intensity environment, and to say so plainly. Some candidates skip this round entirely and get the assessment link straight from the application.
A timed coding problem with API-style input, where the first job is understanding the methods you are given before writing any logic. One reported prompt: find the dead zones between party time intervals. Standard algorithmic ground at medium difficulty; the pressure is the clock.
Two problems back to back, and the round Scale is known for. The signature format is the card game: implement a game from a spec whose requirements keep changing. Interval and scheduling problems are the other common shape, such as finding free ingestion windows to minimise model processing time, or a meeting-rooms variant. Graded on coding speed, accuracy, and turning requirements into working code, with clean syntax expected because there is no time to fix it.
A deep dive on your technical experience and your interest in AI infrastructure, with a pitch for the team. Candidates describe it as a soft behavioural round rather than a test, judging technical depth and appetite for a fast-moving environment.
Build a small working system or model a rule-based problem end to end. The same bar as the screen at more length: a large volume of correct code, structured well enough to extend when the interviewer adds a requirement.
SWE trackAI-shaped systems, with none of the generic URL-shortener prompts reported. One reported prompt: the backend for an insurance claims agent, with a retrieval pipeline over emails and PDFs, policy metadata, and token cost kept down. Others: an LLM API pipeline, a pipeline built from classification and embedding services, a durable task scheduling service, a streaming job scheduler, a CSV upload endpoint that classifies rows with a model. The interviewer expects specific cloud services named and each choice justified.
The round candidates single out. A large unfamiliar codebase in Python or TypeScript, or in one account 150 to 200 lines of modular code, with two or three logical bugs to find and fix inside the hour. You are given a working API key pointed at one of Scale's own models and code whose model call does not return; one candidate traced it to a prompt hashing step in the call path with print statements. Reading unfamiliar code fast is the skill.
The behavioural round, structured around Scale's six credos. Interviewers expect concrete stories of delivering under pressure with imperfect information, with your own part in each made explicit. One reported prompt asks you to present a metrics-driven technical project; another to connect your hardest project to why you want this role.
Decisions arrive on no fixed schedule, and two to three weeks between stages is reported as normal. Scale adds interviews after the final loop when the panel wants more signal, so a further round is not a bad sign on its own.
What Scale AI screens for
What every round is really testingScale publishes six credos and names its behavioural round after them, so the values it screens for are the least secret part of the loop. The accounts add a seventh that the credos imply rather than state: pace.
- Earn customer love. Scale's own wording is that it is devoted to the success of customers and contributors, and the Credo round asks for a story where a customer's outcome drove your decision
- Team flow. Ideas, energy and support move across teams; candidates who describe protecting their own lane do worse than those who describe unblocking someone else's
- Quality is our cheat code. The systems that deliver quality are rare enough to be an advantage, which is why the screen punishes a typo and the debugging round exists
- Find the 20%. The high-impact fraction of the work; be ready to say what you cut as well as what you built
- Write the market, and three moves ahead. Decisions made on where the market is going and on the consequences of consequences, which in an interview means saying what would break your design next
- Pace. Candidates report interviewers describing the culture in the words of a 996 schedule, recruiters probing for comfort with intensity, and a bar summed up by one candidate as speed and execution over interview theatrics
Scale AI interview questions
Candidate-reported themesThe behavioural ground is narrow and reachable: the six credos, your hardest project, and what you personally moved. The technical questions below are the ones candidates report by name; the shapes repeat more than the specifics.
Behavioural & motivation
- Present a technical project you drove, with the metrics that moved.
- Connect your hardest project to why you want this role.
- Tell me about delivering under pressure with imperfect information, and what you decided without the data you wanted.
- Tell me about a time you unblocked another team at a cost to your own plan.
- What did you cut from a project to ship it, and how did you choose?
- Why Scale AI?
Technical
- Timed implementation: The card game with evolving requirements; interval and scheduling problems such as a dependency-aware task scheduler, scheduling ready tasks by deadline, party time blocks per neighbourhood, and updating a grid of neurons; implementing multi-head attention and LLM sampling for ML candidates
- AI-shaped system design: A retrieval pipeline over emails and PDFs with token cost and persistent storage, an LLM API pipeline, a classification-and-embedding pipeline, a durable task scheduling service, a streaming job scheduler, a large-scale ticketing system, a CSV upload endpoint with model classification
- Debugging an unfamiliar codebase: Two or three logical bugs in a multi-file Python or TypeScript codebase inside the hour, one of them in the path of a live model call; debugging an ML pipeline and building a text parser; debugging a project assignment codebase
- Machine learning depth, for the MLE track: Transformers, attention, decoding, RLHF pipelines, custom evaluation harnesses, agentic workflows, and a computer vision or NLP take-home
These are the reported themes — your loop is role-specific. Paste the actual posting and Calibrd predicts the questions for that exact role and level.
Scan a Scale AI posting →Pay
What the offer looks likeLevels.fyi for software engineers at Scale AI in the United States, read on 11 September 2026: L3 $231K, made of about $159K base and $68K a year in stock; L4 $370K on $221K base and $146K stock; L5 $476K on $242K base and $227K stock; L6 $1.23M on $308K base and $891K stock, with bonuses small at every level. Levels.fyi quotes the overall median in two places, $358K and $363K, and the level pages show different medians from the summary, so treat these as an order of magnitude rather than a band structure. Scale is a private company with Meta as a 49% shareholder, so the stock half is an option grant: Levels.fyi reports a standard four-year vest with 25% after the first year and monthly after, and a five-year window to exercise after leaving once you have two years in. Its value depends on a liquidity event that has not happened. Contributor and sales roles are on separate scales and are not covered by these figures.
How to prepare for a Scale AI interview
In order- 01Practise producing code rather than solving puzzles. The screen gives two problems in an hour and grades volume and correctness, so rehearse writing a small working program from a changing spec in 25 minutes without stopping to look up syntax. The card game is worth building once end to end before the call.
- 02Learn to read a codebase you have never seen, fast. The debugging round is the one candidates rate hardest and it is a skill you can train: take an unfamiliar repository, break it in two places, and time yourself finding both with print statements only. Expect the bug to sit in the path of a model call.
- 03Design AI systems rather than textbook ones. Prepare a retrieval pipeline over documents with token cost as a first-class constraint, a queue-based ingestion system, and a task scheduler, and be ready to name the specific cloud service for each box and say why.
- 04Map your stories to the six credos before the Credo interview, with your own part in each explicit. The interviewers want delivery under pressure with imperfect information, so pick the story where you decided without the data you wanted and say what you would have done with it.
- 05Have a metrics-driven project ready to present in two minutes. Both reported behavioural prompts start from your hardest or most measurable project; a candidate who cannot name what moved is the reported failure.
- 06For the ML track, budget a week for the take-home and do not make it pretty at the expense of done. Candidates call it straightforward but long, and the call that follows walks through your choices.
- 07Read about the company as it is now rather than as the older write-ups describe it. Meta owns 49%, the founder has gone, and the July 2025 cuts and the shift toward enterprise and government customers change which teams are hiring. A why-Scale answer built on the 2023 story will show its age.
- 08Decide about the pace before you accept the loop rather than after. Interviewers and recruiters are reported to say plainly that the hours are long, and candidate threads about the culture are among the most negative for any company on this site. Weigh that against the compensation before you spend a month on it.
This guide covers Scale AI's software and machine learning engineering loops in depth, with the product and data science variants noted. The contributor and expert workforce hired through Scale's labeling platforms, and the enterprise and government sales roles the company said it would grow, run different processes and are not covered. For management and leadership roles the loop is similar but the bar shifts to people, delivery and strategy, so pair it with the leadership interview prep hub. The bar for your exact role comes from the role-by-role guides, and the prep that actually transfers is spoken, so run a mock interview before the real one.
FAQ & sources
The short answersWhat is Scale AI's interview process?
Published accounts of Scale's loop agree on the parts and disagree on the count and the clock. Some candidates report a recruiter call first and some report the HackerRank link arriving straight from the application with no call at all; the hiring manager conversation sits between the screen and the final loop in a report from October 2025 and after it in others; the timeline runs from about three weeks in the shortest accounts to a month in most and six weeks or more for senior roles, with two to three weeks of silence between stages being normal. The stages below are the ones more than one account reports, in the order most report them. Recruiter screen: Background, why Scale, and fit for the role. Recruiters are reported to probe for comfort with a high-intensity environment, and to say so plainly. Some candidates skip this round entirely and get the assessment link straight from the application. Online assessment: A timed coding problem with API-style input, where the first job is understanding the methods you are given before writing any logic. One reported prompt: find the dead zones between party time intervals. Standard algorithmic ground at medium difficulty; the pressure is the clock. Technical screen: Two problems back to back, and the round Scale is known for. The signature format is the card game: implement a game from a spec whose requirements keep changing. Interval and scheduling problems are the other common shape, such as finding free ingestion windows to minimise model processing time, or a meeting-rooms variant. Graded on coding speed, accuracy, and turning requirements into working code, with clean syntax expected because there is no time to fix it. Hiring manager call: A deep dive on your technical experience and your interest in AI infrastructure, with a pitch for the team. Candidates describe it as a soft behavioural round rather than a test, judging technical depth and appetite for a fast-moving environment. Coding and backend practical: Build a small working system or model a rule-based problem end to end. The same bar as the screen at more length: a large volume of correct code, structured well enough to extend when the interviewer adds a requirement. System design: AI-shaped systems, with none of the generic URL-shortener prompts reported. One reported prompt: the backend for an insurance claims agent, with a retrieval pipeline over emails and PDFs, policy metadata, and token cost kept down. Others: an LLM API pipeline, a pipeline built from classification and embedding services, a durable task scheduling service, a streaming job scheduler, a CSV upload endpoint that classifies rows with a model. The interviewer expects specific cloud services named and each choice justified. Debugging: The round candidates single out. A large unfamiliar codebase in Python or TypeScript, or in one account 150 to 200 lines of modular code, with two or three logical bugs to find and fix inside the hour. You are given a working API key pointed at one of Scale's own models and code whose model call does not return; one candidate traced it to a prompt hashing step in the call path with print statements. Reading unfamiliar code fast is the skill. Credo interview: The behavioural round, structured around Scale's six credos. Interviewers expect concrete stories of delivering under pressure with imperfect information, with your own part in each made explicit. One reported prompt asks you to present a metrics-driven technical project; another to connect your hardest project to why you want this role. After the loop: Decisions arrive on no fixed schedule, and two to three weeks between stages is reported as normal. Scale adds interviews after the final loop when the panel wants more signal, so a further round is not a bad sign on its own.
What does Scale AI look for in candidates?
Scale publishes six credos and names its behavioural round after them, so the values it screens for are the least secret part of the loop. The accounts add a seventh that the credos imply rather than state: pace. Earn customer love. Scale's own wording is that it is devoted to the success of customers and contributors, and the Credo round asks for a story where a customer's outcome drove your decision Team flow. Ideas, energy and support move across teams; candidates who describe protecting their own lane do worse than those who describe unblocking someone else's Quality is our cheat code. The systems that deliver quality are rare enough to be an advantage, which is why the screen punishes a typo and the debugging round exists Find the 20%. The high-impact fraction of the work; be ready to say what you cut as well as what you built Write the market, and three moves ahead. Decisions made on where the market is going and on the consequences of consequences, which in an interview means saying what would break your design next Pace. Candidates report interviewers describing the culture in the words of a 996 schedule, recruiters probing for comfort with intensity, and a bar summed up by one candidate as speed and execution over interview theatrics
What questions does Scale AI ask in interviews?
The behavioural ground is narrow and reachable: the six credos, your hardest project, and what you personally moved. The technical questions below are the ones candidates report by name; the shapes repeat more than the specifics. Present a technical project you drove, with the metrics that moved. Connect your hardest project to why you want this role. Tell me about delivering under pressure with imperfect information, and what you decided without the data you wanted. Tell me about a time you unblocked another team at a cost to your own plan. What did you cut from a project to ship it, and how did you choose? Why Scale AI? Timed implementation AI-shaped system design Debugging an unfamiliar codebase Machine learning depth, for the MLE track
How do I prepare for a Scale AI interview?
Practise producing code rather than solving puzzles. The screen gives two problems in an hour and grades volume and correctness, so rehearse writing a small working program from a changing spec in 25 minutes without stopping to look up syntax. The card game is worth building once end to end before the call. Learn to read a codebase you have never seen, fast. The debugging round is the one candidates rate hardest and it is a skill you can train: take an unfamiliar repository, break it in two places, and time yourself finding both with print statements only. Expect the bug to sit in the path of a model call. Design AI systems rather than textbook ones. Prepare a retrieval pipeline over documents with token cost as a first-class constraint, a queue-based ingestion system, and a task scheduler, and be ready to name the specific cloud service for each box and say why. Map your stories to the six credos before the Credo interview, with your own part in each explicit. The interviewers want delivery under pressure with imperfect information, so pick the story where you decided without the data you wanted and say what you would have done with it. Have a metrics-driven project ready to present in two minutes. Both reported behavioural prompts start from your hardest or most measurable project; a candidate who cannot name what moved is the reported failure. For the ML track, budget a week for the take-home and do not make it pretty at the expense of done. Candidates call it straightforward but long, and the call that follows walks through your choices. Read about the company as it is now rather than as the older write-ups describe it. Meta owns 49%, the founder has gone, and the July 2025 cuts and the shift toward enterprise and government customers change which teams are hiring. A why-Scale answer built on the 2023 story will show its age. Decide about the pace before you accept the loop rather than after. Interviewers and recruiters are reported to say plainly that the hours are long, and candidate threads about the culture are among the most negative for any company on this site. Weigh that against the compensation before you spend a month on it.
- 01Exponent, Scale AI new grad software engineer interview guidethe 50-minute HackerRank assessment, the 60-minute screen with two problems, the system design, backend practical, debugging and 45-minute Credo rounds, the live model API key in the debugging round, the six credos, and the three-to-six-week timeline with two-to-three-week gaps
- 02Exponent, a candidate's Scale AI new grad interview accountthe party-intervals assessment, the two interval problems in the screen, the insurance claims agent design prompt with a retrieval pipeline and token cost, the debugging task traced to a prompt hashing step against Scale's own model endpoint, the interviewer's 996 remark, and the line about speed and execution over interview theatrics
- 03TechPrep, Scale AI's interview process 2026the 30-minute recruiter screen probing for comfort with intensity, the HackerRank assessment arriving straight after applying, the 30-to-45-minute hiring manager screen, the four-to-five-round final loop, the debugging round on a large unfamiliar Python or TypeScript codebase with two or three bugs in 60 minutes, the card game applied round, and the Credo name for the behavioural round
- 04Interview Coder, Scale AI software engineer interview, June 2026the card game as the signature screen format with evolving requirements, the hiring manager call placed between screen and onsite in an October 2025 report, the three-to-five-round final loop of about 45 minutes each, the one-to-three-week timeline, and the expectation of stories about delivering under pressure with imperfect information
- 05Exponent, get a job at Scale AIthe one-hour HackerRank screen with one or two medium-hard scenario questions, the hiring manager screen replacing it for non-coding roles, the three-to-four-loop final round in one day, the machine learning take-home in computer vision or NLP within a week, the product case study and presentation, the data science case, and the about-one-month timeline
- 06PracHub, Scale AI interview questions, updated August 2026the named coding, system design, debugging and behavioural questions, and the final loop order of coding and ML system design first, then behavioural, machine learning and software engineering fundamentals
- 07Blind, Scale AI SWE, March 2026a candidate's account of many rounds of implementation-heavy code where a typo costs you, and the debugging round as the hardest
- 08Blind, Scale AI onsite, October 2024the system design and backend practical rounds named in the final loop, and replies about the culture and work-life balance
- 09Scale AI, careersthe six credos in Scale's own words: earn customer love, team flow, quality is our cheat code, find the 20%, write the market, three moves ahead
- 10Scale AI, next phase of the company's evolution, 12 June 2025Meta's minority investment at a valuation above $29 billion, Alexandr Wang joining Meta while staying on Scale's board, and Jason Droege as interim chief executive
- 11CNBC, Scale AI founder Wang announces exit for Meta as part of $14 billion dealthe $14.3 billion investment for a 49% stake and Wang's departure to lead Meta's AI effort
- 12TechCrunch, Scale AI lays off 14% of staff, 16 July 2025200 employees and 500 contractors cut, mostly in data labeling, Droege's statement that the business had scaled too quickly, and the plan to staff up around enterprise and government sales
- 13Levels.fyi, Scale AI Software Engineer, United Statesthe L3 to L6 medians with their base and stock split, the two overall medians quoted, the four-year option vest with 25% after year one, and the five-year exercise window after two years of service
Interview processes change. This reflects widely-reported and sourced conditions as of 2026 — confirm specifics with your recruiter, and treat it as a map rather than a guarantee.
Prep for a real Scale AI role
Practise your Scale AI interview, out loud.
Paste a real Scale AI posting and Calibrd predicts the questions for that role and level, benchmarks the pay, and flags the gaps an interviewer will probe in your CV — then listens to your spoken answers and coaches them. Your first mock is free.
Free to start · No card · Your resume stays yours, encrypted, remove it anytime