By the end of this lesson you'll be able to name every round in a frontier-lab interview loop, say what each one tests, and tell an official fact from a candidate rumour.
By the end of this lesson you'll be able to name every round in a frontier-lab interview loop, say what each one tests, and tell an official fact from a candidate rumour.
Interview processes change, and most of what's written about them comes from candidates, not companies. Throughout this course, every claim about hiring carries one of three labels:
When you prepare, put your hours where the official and reported facts are. When a recruiter tells you something different, the recruiter wins.
Anecdotal. Count independent origins, not repetitions: ten copies of one post are still one account. Blogs often copy each other, so before you call something "reported", check that the accounts really are separate people describing separate loops.
The tables in the folds below were last checked on 2026-10-08, and Anthropic's careers page on 2026-10-10. Processes change, so confirm the details for your role with your recruiter.
The shape candidates describe: a written application ("Why Anthropic?"), a recruiter screen, one progressive Python online assessment (Module C3), then an onsite in two parts. Part 1 is coding, system design and a culture interview; part 2, which is cancelled if part 1 doesn't go well, is your experience and a project deep dive. References and team matching come last and can take weeks.
Two pieces are Official. First, the careers page says technical interviews use live coding tools such as Colab and CodeSignal, that you can look things up, and that you should know basic syntax and the standard library well enough that they don't eat your time. Second, Anthropic publishes guidance on using Claude during the hiring process: be yourself, write first drafts yourself, and assume assessments are AI-free unless told otherwise. Lesson 3 of this module goes through it in detail.
| Stage | What happens | Label |
|---|---|---|
| Application | A written answer to a question like "Why Anthropic?". Treat it as your first interview: specific beats impressive. | Reported |
| AI use | Write your own first drafts; Claude may help you refine them. No AI in take-homes or live interviews unless they tell you otherwise. Using Claude to prepare is encouraged. | Official |
| Recruiter screen | About 30 minutes on background, motivation, level and team fit. | Reported |
| Online assessment | One progressive Python problem on CodeSignal, about four levels, about 90 minutes. Practical code, not puzzles. Module C3 trains this. | Reported |
| Coding tools | Anthropic's careers page says technical interviews use live coding tools such as Colab and CodeSignal, that you can look things up, and that you should be comfortable enough with basic syntax and the standard library that they don't eat your time. All interviews run over Google Meet. | Official |
| Onsite, part 1 | Coding, system design (often LLM infrastructure) and a culture interview. Part 2 is cancelled if part 1 doesn't go well. | Reported |
| Onsite, part 2 | Your experience and goals, plus a deep dive into a project you built. | Reported |
| Culture / values | A culture interview sits in the first onsite loop. Module C5 trains this. | Reported |
| Who runs it, how hard it is | One newsletter says an employee nominated for the role runs it. Several career guides call it the round that fails the most candidates. Guides aren't candidate reports, and Anthropic publishes no pass rates. | Anecdotal |
| References and team match | Can add two to four weeks or more. | Reported |
For OpenAI's forward deployed engineering loop, candidates describe a recruiter screen (background, motivation, your views on AI and customer-facing work), two 60-minute technical screens or a take-home with a review call, and a virtual onsite of four to six interviews: practical coding, LLM system design and a project deep dive (Reported). One candidate describes a week-long take-home case study and an AI-enabled coding screen; that's anecdotal: one account, one team.
Perplexity's loop is reported as an HR screen, a project-style online assessment or live coding, then four or five onsite rounds of domain-flavoured coding, AI system design and a hiring-manager deep dive, with a founder interview for senior roles.
OpenAI (forward deployed engineering)
| Stage | What happens | Label |
|---|---|---|
| Recruiter screen | Background, motivation, your views on AI and customer-facing work. | Reported |
| Technical screen | Two 60-minute screens, or a take-home with a review call. | Reported |
| Virtual onsite | Four to six interviews: practical coding, LLM system design, and a project deep dive where you defend something you built. | Reported |
| Assessed on | Scoping ambiguous problems, building systems around models, proving they work with evals, and talking to non-technical stakeholders. | Reported |
Perplexity
| Stage | What happens | Label |
|---|---|---|
| Screens | An HR screen, then a project-style online assessment (multi-part, predefined tests) or live coding. | Reported |
| Onsite | Four or five rounds: domain-flavoured coding (tokenisation, streaming), AI system design (RAG, latency, caching), and a hiring-manager deep dive. | Reported |
| Senior roles | A technical interview with a founder. | Reported |
Strip away the company names and four rounds keep coming back:
That's the backbone of this course. The shared core trains all four; your track adds the role-specific depth.
Postings are Official: the company wrote them. But they're reposted often and the wording drifts, so read the live posting for your role. Anthropic's open roles page groups these jobs under Applied AI; on 2026-10-10 it listed Applied AI Engineer, Applied AI Architect and Forward Deployed Engineer roles in several cities. A years figure is a guide, not a wall, and titles move, so search by team as well as by title.
Our research read the postings through job-board mirrors, which can lag the company's own board: Applied AI Engineer, Solutions Architect, Applied AI and Forward Deployed Engineer.
| Role | What the posting describes | Hard requirements we saw | Label |
|---|---|---|---|
| Applied AI Engineer | A technical, customer-facing advisor from discovery through deployment. Builds eval frameworks, pairs with customer engineers, knows prompting, agents and retrieval. | 4+ years in a technical role. | Official (posting, read 2026-10-08) |
| Solutions Architect, Applied AI (titled Applied AI Architect on the live listing) | Pre-sales architecture for large enterprises. Fits Claude into their stack, builds evals, designs architectures that scale. | We didn't record a years figure. Read the live posting. | Official (posting, read 2026-10-08) |
| Forward Deployed Engineer, Applied AI | Builds production applications on Claude inside customer systems (MCP servers, sub-agents, agent skills), white-glove deployment. | 25–50% travel. Python plus at least one more programming language. | Official (posting, read 2026-10-08) |
Before you plan your prep (lesson 5), spend ten minutes on the live posting for your target role:
GAP LIST: <role> at <company>, posting read <date>
Requirement (one line) | My evidence (project, number, customer) | G / A / R | Module
-------------------------------|-----------------------------------------|-----------|--------
builds eval frameworks | | |
Python plus one more language | | |
travel 25–50% | (logistics: do I want this?) | |
Rules: if I couldn't talk about the evidence for two minutes, it's not evidence yet.
Reds map to modules; a red I can't fix in six weeks, I say so honestly.
Ambers are where my prep hours pay off most.
Mark each line green (strong evidence), amber (some) or red (none). Evals, agents and retrieval live in your track's modules; customer work lives in the FDE and architect tracks. Check the logistics lines too: travel, location and office expectations decide whether you want the job at all.
From my loop
When I interviewed at Amazon, the round that surprised me was the bar raiser. I expected a spread of questions. Instead, the interviewer took one story and spent 40 minutes on it: what I personally decided, what the data said, what I'd do differently. Every "we" answer got redirected to "what did you do?"
My advice: prepare 6 to 8 stories, and for each one know your own actions, one metric, and one thing you'd change. Practice with someone who interrupts you. Write "I" in your notes and cut every "we."
60. The guide summarises past candidates; the recruiter knows your loop now. When they differ, the recruiter wins, and you adjust the plan.
This course isn't affiliated with any lab. Every practice question here is original, written in the style of what candidates publicly describe. Don't go looking for leaked questions: they're often wrong, they break the confidentiality candidates agree to, and interviewers can tell when an answer was memorised.
Cheat card
- Official = the company says it; Reported = several independent candidate accounts; Anecdotal = one or two. Count origins, not copies; the recruiter beats every guide.
- Every loop tests four things: practical coding, LLM system design, a project deep dive, and motivation and values.
- At Anthropic (reported): written application, recruiter screen, one progressive Python OA, onsite part 1 (coding, design, culture) and part 2 (deep dive), then references and team match. Assessments are AI-free unless told otherwise (Official).
Requirement | My evidence | Green / Amber / Red | Module that closes itInterview line "I prepared from what Anthropic publishes and what several candidates independently report, and I checked the format with my recruiter, so I spent my hours on the four rounds every loop shares."