Layer L6
Career and Market
The layer
Why this layer decays fastest
The first six layers are about knowledge. This layer is about the market's price for that knowledge: which roles are hiring, what words the JD uses, how much it pays, how the interviews run. The price changes every night. The site's pipeline recomputes role families and skill shares every night, companies change their process once a year, and pay bands can move within a quarter. So the half-life of this whole layer is "flow": subscribe, don't bookmark. The nodes record only sources, how to read them, and structure, not numbers. For numbers, go to the site's role pages and skill pages and look them up live.
That is also why it carries high weight only in the sprint. The rest of the time, the whole workload for this layer is reading one issue of Index a month plus low-frequency practice. Only when I am actually about to move do I go through all six nodes from start to finish. The product does not do company interview intel, so each company's round shape stays in my notes and does not go on the site.
How the six nodes connect
The reading order is the job-search order:
Role families: how AI roles split Pick the family first. The same title means different things at different companies, so split by what the work is. The family sets the weight of the seven layers and your profile → JD requirements: how to read skill words Read the words in a JD. Frequency is not importance, gate words and bonus words are kept apart, and the same word means different things in different families → Interview question sources Pull questions from a source table marked with credibility and decay → Behavioral interviews and values rounds Prepare three things beyond STAR: decisions you designed yourself, "why not the other option", and defending AI-co-written code without notes → Compensation: remember sources, not numbers Look at this last, and only at sources and trends.
The six interview muscles (coding without AI · Python · LLM system design · explaining your own design · values / behavioral round · AI-assisted coding round) are what to practice. This layer is who you practice for, where the questions come from, and how you check you are done.
Top picks
In reading order; each pick links to its node below.
- 1
Guidance on Candidates' AI Usage free
First-hand AI policy; the basis for the "coding without AI" muscle → Sources for question banks and interview write-ups
- 2
Anthropic Careers free
A target company says in its own words how it reads candidates → JD requirements: how to read the skill words
- 3
How We Hire free
The most complete official interview loop write-up; use it as a reference template → Behavioral Rounds and the Values Round
- 4
Official four-stage process; a baseline for telling official from rumor → JD requirements: how to read the skill words
- 5
Machines of Loving Grace: How AI Could Transform the World for the Better free
Original text of the values round, "optimistic side"; find where you disagree → Behavioral Rounds and the Values Round
- 6
The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI free
Original text of the values round, "risk side"; pair it with the previous one → Behavioral Rounds and the Values Round
- 7
AIMLInterviews free
ML system design question set; the source for one spoken question a month → Sources for question banks and interview write-ups
- 8
AI Engineering Field Guide free
Design questions on RAG / agent / evals for AI engineering roles → Sources for question banks and interview write-ups
- 9
The Rise of the AI Engineer free
Defines the role families; place yourself in the table → Role families: how AI jobs are divided
- 10
End of Year Pay Report 2025 free
For pay, look at trends only; this is the baseline → Pay: record sources, not numbers
The nodes
Role families: how AI jobs are divided
Half-life: a stream
The same title means different things at different companies. Split by what the work is first, then look at the title. Research (proposing and validating methods), ML engineering (training and shipping models, recommendation and ranking, MLOps), AI / application engineering (building products on models: RAG, agent, evals, cost), FDE (embedding in a client's systems to deliver), inference and training infrastructure (GPU, parallelism, serving), evals and data (eval sets, human data, model quality), AI PM (product judgment plus prototyping ability).
The weight of the seven layers shifts by family. Research and infrastructure press on L2 / L3. Application engineering and FDE press on L4. Evals and data span L4 and L1. AI PM only needs the concept layer of L4. The family decides the profile (profiles/), and the profile decides the path.
The site's role pages are kept per family and recomputed every night. I copy no numbers here: Forward Deployed Engineer · AI Engineer · Software Engineer (the other families only go into the Index and have no page of their own).
article — Defines the "AI Engineer" role family: building products on foundation models, not research. After it, you can place yourself in the role family table instead of guessing from a title.
Not picked (2)
- 各招聘网站的「AI 岗位分类」文章 — Secondhand and written for SEO. Its split follows traffic, not the work.
- Levels.fyi 的 title 分类 — It splits by pay band and lumps Research Engineer with SWE, so it can't serve as a role family (for pay, see Compensation: record sources, not numbers).
JD requirements: how to read the skill words
Half-life: a stream
A JD is a list written jointly by the hiring manager and the recruiter. Not every word in it weighs the same. Three ways to read it:
- Frequency ≠ importance. Every posting on the site says Python and "agents". These are threshold words: having them earns nothing, lacking them gets you screened out. What separates candidates are words that appear less often but sit under "responsibilities" (eval frameworks, MCP server, customer-side delivery). The site's skill pages give each word's frequency and co-occurrence, but the numbers are recalculated every night, so I don't copy them here: evals · rag.
- For the threshold, look at the years in "required" and one concrete verb ("deployed at scale", "built evaluations"). For bonus points, look at "nice to have" and "you might thrive"; you can apply even if you are missing half of those.
- The same word means different things in different families. In a research role, "evals" means designing benchmarks. In an applied role, "evals" means a regression suite. Fix the family first (Role families: how AI jobs split), then read the words.
- Anthropic Careers free
docs — A target company says how it reads candidates: what you can do over degrees, with research, blog and open source at the top of the résumé. A JD lists keywords; this page gives the weights.
docs — Four official stages (screen, skills, final, decision), tuned per role. The research quick-fire round isn't official; it comes from candidate reports. Baseline for fact versus rumor.
Not picked (2)
- 招聘机构的「AI 工程师 JD 模板」 — Second-hand text from someone selling a service; the words are copied from other people's JDs, so it can't be a signal source.
- 「2026 年最热 AI 技能 Top 10」类文章 — No samples, no dates; it carries no information next to the data the site recalculates every night.
Sources for question banks and interview write-ups
Half-life: a stream
Read question banks for question types, and interview write-ups for the process. Both go stale, so each entry is marked with a trust level and a decay.
| Source | Type | Trust | Decay | How to use |
|---|---|---|---|---|
| Official candidate guides from companies (Anthropic, OpenAI, Amazon, DeepMind) | Official guide | First-hand | Years | Sets the shape of the rounds and the AI policy; reread before each application |
alirezadir/AIMLInterviews |
System design question set | Crowdsourced | Years | Pick one question a month and answer it aloud on record |
alexeygrigorev/ai-engineering-field-guide |
Question set (AI engineering) | Crowdsourced | Months | Source of prompts for the LLM system design round |
| Company pages on interviewing.io | Interview write-ups + question set | Second-hand (commercial) | Months | Take only the round structure; don't trust its "pass rates" |
| Hello Interview blog | Interview write-ups | Second-hand (commercial) | Months | Source for the shape of the Meta AI-assisted coding round, until an official page replaces it |
| Sundeep Teki's lab interview guide | Interview write-ups | Second-hand (front-line interviewer background) | Months | Round taxonomy and ML debugging question types for research engineering roles |
| Blind | Forum | Crowdsourced | Weeks | Only check "did this company's process change in the last 4 weeks"; ignore the conclusions |
| 1point3acres | Forum | Crowdsourced | Weeks | Same as above. Highest density of Chinese-language write-ups; the questions are reliable, the verdicts are not |
| CodeSignal progressive question format (public sample questions) | Question set | First-hand | Years | Question templates for no-AI coding practice |
| LeetCode company tags | Question set | Crowdsourced | Months | Only for keeping up speed; AI companies' coding rounds have moved away from pure algorithm questions |
- AIMLInterviews free
repo — An open-source ML system design question set: prompts and solution frameworks for recommendation / ranking / GenAI. Source of the monthly spoken-practice question.
repo — A question bank for AI engineers (not MLEs), with system design questions on RAG / agent / evals. Matches the LLM system design round in FDE and applied engineering roles.
docs — First-hand AI policy: applications may use Claude to polish, take-home and onsite default to no. This is why the "no AI coding" muscle exists. Check it before applying.
Not picked (3)
- 《Machine Learning System Design Interview》（Alex Xu、Ali Aminian） — Often recommended in the community, but no independent endorsement this time (no HN post cleared the endorsement bar). Left outside the table for now.
- `khangich/machine — learning-interview` — many stars, but no updates since 2023; the question types stop at the pre-LLM era.
- 付费 mock 平台 — Only used in the sprint; not a source.
Behavioral Rounds and the Values Round
Half-life: a stream
STAR only solves "tell one story." Behavioral rounds for senior roles ask about three things STAR does not cover:
- Explain the decisions you designed. Pick one project and prepare a 45-minute deep dive. Attach "why not the other option" to every decision. The interviewer always presses on the path you gave up. When you write the retrospective, put that line in as you go, and the article becomes your script.
- Defend AI co-written code without notes. Most commits in your own repo are co-written with AI. AI-native employers don't mind, but they will ask "why did you do it this way" layer by layer. How to prepare: pick one module, turn off AI, and walk through the design from constraints to trade-offs. Wherever you can't explain it, that is what you need to fill in.
- The values round is not an enthusiasm test. The values round wants grounded doubt and disagreement. Read Dario's two long essays and Core Views, and bring one or two points of your own disagreement. The Amazon Leadership Principles round works the same way: prepare one example with numbers for each principle.
Prepare for behavioral rounds only in the sprint.
article — The values round asks where you agree with Anthropic's mission and where you doubt it, not for a recital. This is the original 'upside' essay; read it to find your own points of disagreement.
article — The original 'risk' essay, paired with the previous one. Hypothetical ethics questions mostly fall inside these two frames; a prepared disagreement beats enthusiasm.
- How We Hire free
docs — Lays out the four steps (apply → assessment → phone screen → loop), Bar Raiser and Leadership Principles in one place. The fullest big-company official guide; use it as a template.
Not picked (2)
- Anthropic「Core views on AI safety」（2023 — 03) — The third reference for the values round, but this node caps at 3 picks and the two long Dario essays already cover it. If you want it, get it from Anthropic's site.
- 通用「行为面 50 题」类文章 — The questions are generic and have no company context. No use for the deep-dive rounds of senior roles.
Pay: record sources, not numbers
Half-life: a stream
Numbers at this layer go stale fastest, so this node writes down no numbers. It records three sources and how to read them.
Sources
- ATS-posted base: the Greenhouse / Ashby job board API gives the base range for every posting directly. It is the only first-hand data. The site pulls from here every night, and role pages are shown by role family.
- Levels.fyi: self-reported total comp, by company × level. Large volume, but skewed toward people who choose to report.
How to read pay data (three sentences)
- What is posted is base, not total comp. The gap at labs is mostly equity, and at startups it is valuation. Neither is in the posting.
- For a range, check the sample size and the level. When one title spans three levels, the range gets so wide it carries no information. Split by Levels.fyi level first, then compare.
- Compare trends, not point values. Pull from the same source, with the same definition, again after some time. Only then do you know whether the band is moving or you are looking at noise.
article — Trend first, point values second. Shows yearly change for AI / research roles beside plain SWE: a baseline for whether the pay band is moving. Look up exact figures on Levels.fyi and ATS base.
Not picked (2)
- 招聘博客的「AI 工程师年薪 $X — $Y" — no sample, no level, no date. Pure SEO.
- Glassdoor 薪酬页 — Self-reported, with no split by level or by base versus equity. One tier coarser than Levels.fyi.
Self-check
- I can place my target role in one family and say which two of the seven layers that family leans on (Role families: how AI roles split)
- Given any target JD, I can separate gate words, bonus words, and the words that really distinguish candidates within one minute (JD requirements: how to read skill words)
- My target company list (10–15 companies) has one line per company: round shape, AI policy, special rounds, official page, and each line has a check date
- I can talk through one project for 45 minutes without notes, with every decision paired with "why not the other option" (Behavioral interviews and values rounds)
- I can name the bias of each of the three pay sources, instead of quoting any one number (Compensation: remember sources, not numbers)
Translated from the author's Chinese notes by a model; the Chinese page is the original.