Start with the rules of the room
Before planning a single practice session, find out where you will be allowed to use AI, because it changes how you should practise. The labs' published positions differ.
OpenAI's interview guide says there is no single rule: some interviews permit AI tools and others assess independent problem-solving, the expectation for each is set out in your prep materials, and you should ask your recruiter if anything is unclear. Anthropic's candidate guidance, as reported when it was updated in July 2025, asks you to write your own first drafts of application materials and then allows Claude to refine them, allows it for research and practice, allows it on take-homes only when the instructions say so, and does not allow it in live interviews unless the interviewer says otherwise. For DeepMind, candidates report restrictions on AI coding assistants in the coding rounds. Check the current wording for each before your interviews; these policies are being revised as the tools change.
The practical consequence: do most of your coding practice with the assistant switched off. Use AI where it is permitted and genuinely useful - as a mock interviewer, to quiz you on fundamentals, or to critique an answer you already wrote.
The plan at a glance
The schedule assumes about ten hours a week alongside a full-time job and targets a software or research engineering role. Each week has a primary focus and a small amount of continuous practice so that nothing goes cold. Move the ML weeks earlier for research engineering roles, and the design weeks earlier for senior ones.
week focus continuous (every week)
---- --------------------------------- ---------------------------------
1 progressive coding: 2 full 4-level 1 timed problem, no assistant
problems, tests as you go
2 ML fundamentals: attention from 5 flashcard derivations
scratch, softmax, sampling, L1/L2
3 LLM system design: KV cache maths, 1 design out loud, 45 minutes
batching, TTFT/TPOT, admission
4 LLM practical round: async calls, 1 timed problem, no assistant
retries, cost, 20-item eval
5 project deep dive: 4-5 slides, 1 design out loud, 45 minutes
three levels of 'why' on each choice
6 values stories + 2 full mock loops rest the day beforeWeeks one to three: the technical core
Week one is progressive coding, because it is the most common screen and the one where practice pays most. Work two complete four-level problems under a ninety-minute timer - the in-memory database from earlier in this series, then a banking or file-cache variant. After each, review where you lost time; it is almost always a level-one design decision.
Week two is fundamentals. Implement attention from scratch until you can do it without notes, including the stable softmax and the causal mask, then be able to explain the square-root scaling, the KV cache and grouped-query attention. For research roles, add derivations: L1 versus L2 sparsity, the bias-variance trade-off, the ELBO.
Week three is design. Memorise the handful of numbers that let you reason about inference - bytes per parameter, KV-cache bytes per token, accelerator memory and bandwidth - and practise designing a chat-completion service and a retrieval system with a model in the loop. Do it out loud, against a clock, and record yourself at least once.
Weeks four to six: the rounds most people skip
Week four is the LLM practical round. Build a small notebook harness that calls a model concurrently with a semaphore, retries with jittered backoff, estimates cost before running, validates structured output and scores against twenty labelled examples. Then reuse it on three different toy tasks. Having done this once is a large advantage, because most candidates are improvising it live.
Week five is the project deep dive. Choose one project you led, build four or five slides - problem, options, decision, result with a number, what you would change - and rehearse it with someone who will interrupt you with 'why'. Prepare for the questions you hope they will not ask.
Week six is the values round and full mock loops. Write your four or five stories in the situation-decision-cost-reflection shape, form one honest opinion about the lab's published approach, then run two complete mock days with a friend or a mock-interview service. Take the day before the real loop off. Fatigue costs more than one extra practice problem gains.
What this series covered
This is the last post in the AI lab interview series. The earlier posts take each part of the plan in depth: how lab interviews differ from big-tech loops, round-by-round guides to OpenAI, Anthropic and DeepMind, a worked four-level coding problem, attention from scratch, inference system design arithmetic, the LLM practical round, and the values round.
The theme across all of them is the same. AI labs are hiring people to build and run systems where a model is the most expensive and least predictable component. Every round, in its own way, is asking whether you can reason about that component with numbers, build around it with code that survives change, and make careful decisions about it when nobody is checking.