The shape is familiar. The content is not.
If you have interviewed at a large tech company, the outline of an AI lab process will look reassuring. There is a recruiter call, some form of technical screen, and a final loop. OpenAI's own interview guide describes final interviews as four to six hours with four to six interviewers over one or two days. Anthropic's loop is commonly reported as four or five rounds of about an hour. Google DeepMind runs coding rounds, ML rounds, team-lead conversations and a hiring committee.
That familiarity is the trap. Candidates prepare the way they would for a big-tech loop - two hundred LeetCode problems and a stock distributed-systems design - and then meet rounds that were built to test something else. The labs are small relative to the problems they work on, they ship research into production quickly, and almost every engineer touches a system where a model is the most expensive component. The interviews are shaped by that.
Three differences come up again and again in candidate reports and in the labs' own guidance. Coding rounds reward code you would actually merge, not a clever trick. System design rounds have a model in the middle, so the scarce resource is accelerator memory and tokens, not database connections. And the conversation about why you want to work on AI is a real round with a real bar.
Coding: production-shaped, not puzzle-shaped
OpenAI's interview guide says that for engineering roles it looks for well-designed solutions, high-quality code, optimal performance and good test coverage. Read that list again: test coverage is on it. That is not the language of a round where you scribble an algorithm and the interviewer checks the big-O.
Anthropic's online assessment, as candidates describe it, is one problem in four levels over ninety minutes, each level adding requirements to the code you already wrote - an in-memory key-value store that later needs filtered scans, then expiry, then backup and restore. The design you choose in level one decides whether level four is twenty minutes or impossible. That rewards modular code over fast code.
DeepMind is the exception that proves the rule. Its coding rounds are reported to draw on Google's question bank and to look closer to classic algorithm problems - tries, snapshot arrays, grid problems. But even there candidates report that the code is expected to run by the end of the round, which is not a requirement in a standard whiteboard-style interview.
The practical upshot: practise problems where requirements change mid-stream, write small tests as you go, and get used to running your code in a plain editor without autocomplete doing half the work.
System design has a model in the middle
A conventional design round asks you to scale reads, shard a database and put a cache in front of it. Lab design questions keep all of that and add a component that dominates cost and latency: model inference. Reported Anthropic prompts include distributed search over roughly a billion documents with LLM inference in the loop, and diagnosing a p95 latency jump from 100 ms to 2 seconds. OpenAI prep material commonly suggests thinking about infrastructure at ChatGPT's scale.
The candidates who do well know a handful of numbers cold: how much KV-cache memory one token of context costs for a given model shape, why prefill is compute-bound while decoding is memory-bandwidth-bound, what continuous batching buys you, and why rate limits are expressed in tokens rather than requests. You do not need to have built an inference server. You do need to reason about one with arithmetic rather than adjectives.
The mission round is scored
Every lab has a round about motivation and values. OpenAI's guide tells candidates to read the Charter and recent research posts. Anthropic's culture round is reported to probe long-term thinking, awareness of risk and harm, and how you handled a disagreement or an ethical decision. DeepMind's behavioural round asks, among other things, why DeepMind specifically.
Engineers tend to treat this like the 'why do you want to work here' question at the end of a phone screen. At a lab it is closer to a technical round on judgement. Reciting the mission statement is a weak answer. A strong answer is a specific opinion about the field, defended with something you have actually done - and an honest account of where you are uncertain.
AI tools: the rules differ by lab and by round
It would be easy to assume AI labs want you to use AI everywhere. The published rules are more careful than that. OpenAI's guide says some interviews allow AI tools and others assess independent problem-solving, and that the rules will be in your prep materials - ask the recruiter if unsure. Anthropic's candidate guidance, updated in July 2025, lets you use Claude to refine application materials you drafted yourself and to prepare, but not during live interviews unless the interviewer says so, and not on take-homes unless the instructions allow it.
Treat the rule as part of the question. If you have only ever solved problems with an assistant open, find that out in practice, not in the loop.
Where the prep time should go
For a software engineering role at one of these labs, the split below is a reasonable starting point. Shift weight toward ML fundamentals for research engineering, and toward design for senior roles. The rest of this series takes each row in turn.
area share what to practise
--------------------------- ----- -----------------------------------------
practical / progressive code 30% multi-level problems, tests as you go
LLM-aware system design 25% KV cache maths, batching, token limits
ML + LLM fundamentals 15% attention from scratch, sampling, evals
LLM practical round 10% async API calls, retries, cost budgets
project deep dive 10% one project, 4-5 slides, trade-offs
mission and values 10% a real opinion, a real story