What an AI interview loop actually looks like

Four kinds of employer hire for these roles and they do not interview the same way. This is what each loop is made of, what it weights, and — the part that saves the most time — what it does not test at all.

Claims here are attributed to a tier rather than to a named company, and each tier carries its sources. A specific company’s process changes every few months; the shape of a tier moves more slowly. Check the sources rather than trusting this page.

Frontier labs

The handful of organisations training frontier models: the best known are Anthropic, OpenAI, Google DeepMind, Meta's generative AI groups, Mistral, xAI and Cohere.

Longer than most people expect, and harder on coding than the reputation suggests. The coding bar went up rather than down, but the problems are practical rather than puzzle-like. Credentials count for less here than almost anywhere: at least one lab states publicly that around half its technical staff had no prior machine learning experience, and that independent work belongs at the top of a CV.

The rounds

  1. 1

    Timed coding assessment

    Often one problem in about four levels, each adding requirements to the last. What is being judged is whether your first design survives the third change to the spec, not whether you reached for a clever data structure.

  2. 2

    One to three live coding rounds

    Medium to hard, and practical: a rate limiter, a transactional key-value store, an async batch processor with rate-limit handling and per-item error isolation. Concurrency comes up more here than elsewhere.

  3. 3

    ML or research coding

    Implement something from scratch. Causal multi-head attention is the most commonly reported task by a distance; after it, nucleus sampling, a decode loop with a KV cache, byte pair encoding, layer norm and softmax, and a LoRA adapter.

  4. 4

    Debugging a training run

    A notebook that runs and does not learn. The bugs are stupid rather than hard — a missing zero_grad, a softmax on the wrong dimension, a double softmax before the loss, a broadcast that silently worked, an unshuffled loader — and they still sink people under time pressure.

  5. 5

    Take-home, sometimes with AI explicitly allowed

    Four to eight hours and a written report. It is graded in about fifteen minutes, so it must run on one command, it must ship evals, and the README must name what you deliberately skipped.

  6. 6

    Defend your own work

    An hour on something you actually did, drilled three levels deeper than you expect. Some labs ask for a written statement of your best work up front, others for a slide presentation. Anything your CV claims is fair ground.

  7. 7

    Values and mission

    At least one lab runs a standalone non-technical round on ethical reasoning and your actual relationship to AI risk, and it is a round where strong engineers fail. Performed doom scores as badly as dismissiveness; a specific, consistently held disagreement does not.

Weighted heavily here

  • Concurrency and async, tested harder than the rest of the field.
  • Inference economics as arithmetic: KV cache sizing, prefill against decode, cost per million tokens.
  • Whether you can defend a design decision after the requirements changed under it.
  • Rules on using AI in the interview differ sharply by company and even by round — check, do not assume.

Not tested

  • Deriving backpropagation, for engineering rather than research roles.
  • Exotic algorithms — segment trees, advanced dynamic programming.
  • Framework API recall.
  • Publications, for research engineer roles as opposed to research scientist.
  • Writing CUDA kernels, outside teams that do exactly that.

Sources

Big tech

Established technology companies hiring into machine learning roles: Google, Meta, Amazon, Microsoft, Apple, NVIDIA, Netflix.

The technical content overlaps heavily with the labs. What differs is governance: these loops are standardised, committee-reviewed, and weighted towards a behavioural signal that can veto an otherwise strong candidate. Prepare for the process, not only the problems.

The rounds

  1. 1

    Algorithms screens

    Still a real gate here in a way it is not at a seed-stage startup. Standard patterns, medium difficulty, with a strong preference for problems that look like work.

  2. 2

    ML system design

    The round most candidates underprepare. Data, training, serving, monitoring, and what you would cut — with the trade-offs stated rather than implied.

  3. 3

    ML breadth or fundamentals

    Metrics and when each is wrong, overfitting, imbalance, evaluation design. Broad rather than deep, and unforgiving about vagueness.

  4. 4

    Behavioural, with teeth

    One company runs a dedicated interviewer with a veto and a published set of principles referenced in every round; another routes the decision through a committee rather than the hiring manager. This is not a formality.

  5. 5

    Project deep dive

    Your own past work, in detail, with the decisions defended.

Weighted heavily here

  • Hardware-adjacent roles at one company demand roofline reasoning, memory coalescing and kernel reading fluency that generic preparation does not cover.
  • Recommendation systems are a specialism in their own right at several of these companies.
  • The behavioural round genuinely decides outcomes, unlike at smaller companies.

Not tested

  • Algorithms at a competitive-programming level.
  • Publications, for most engineering roles.
  • Frontier-scale distributed training, outside the teams doing it.

Sources

AI-native scale-ups

Companies built on AI as the product, post-funding and hiring quickly: developer tools, vertical AI products, infrastructure and model-serving companies.

The most varied and the fastest moving. Several have publicly redesigned their loops around the fact that candidates use AI tools, and a couple have gone furthest of anyone — replacing the coding screen with system design and making the onsite a long build with complete freedom over tooling. Expect to be judged on what you ship in a few hours.

The rounds

  1. 1

    A build, not a puzzle

    Two hours or a day, sometimes paid, producing something that works. Judged on what you chose to do and what you chose to skip, and whether you can explain both.

  2. 2

    AI-assisted coding, sometimes required

    Several expect you to use an assistant and judge how well you direct it. Others ban it in the first screen and allow it later. The rules are genuinely unsettled — ask.

  3. 3

    Product and system design

    Designing around a component that is sometimes wrong, with cost and latency treated as first-class constraints rather than afterthoughts.

  4. 4

    Evals

    How you would know the change helped. This is where seniority is assigned at this tier.

Weighted heavily here

  • How well you work with AI tooling is itself a signal at several of these companies.
  • Shipping speed and judgement about scope, over algorithmic depth.
  • Retrieval, agents and evaluation, in depth, because that is the product.

Not tested

  • Hard algorithms, at most of them.
  • Deep learning theory, unless the company trains models.
  • Credentials.

Sources

Startups

Seed and early-stage companies building a product on top of models.

Short, unstandardised, and decided by one or two people who will be your colleagues. The question behind every round is whether you can ship the thing without supervision. A portfolio does more work here than anywhere else on this list.

The rounds

  1. 1

    A conversation about what you have built

    Often the whole first round. Bring something that runs and that you can open in front of them.

  2. 2

    A take-home or a paid trial day

    Realistic scope, short deadline, and your judgement about what to leave out is most of the signal.

  3. 3

    Pairing on their actual codebase

    How you behave when you do not know something, observed directly.

Weighted heavily here

  • Breadth over depth: you will touch the frontend, the database and the deployment.
  • Evidence of shipping, weighted far above where you worked or studied.
  • AI coding tools are usually expected rather than merely permitted.

Not tested

  • Algorithms, mostly.
  • Training or fine-tuning models, at most of them.
  • Distributed systems at scale.

The rounds are the shape. The question bank is what gets asked inside them, and every question there links to the material that teaches its answer.