Elite AI-Lab SWE Interview Program

← All Roles

An executable training program for elite AI-lab software-engineering loops — OpenAI, Anthropic, DeepMind, Scale, Cursor, xAI, Databricks, Netflix, Stripe. Not a reading list: source code, gated drills, scored rubrics, and a progress ledger you run day by day.

Target: senior / staff, IC track. Budget: 22 h/week × 26 weeks ≈ 570 hours.

Table of Contents

Current Phase

Phase 1 — Calibrate. Week 0.

Phase 0 research is complete. The diagnostic battery is built and runnable. PLAN.md is deliberately locked until you report diagnostic scores — writing it now would mean guessing your level, and a plan optimized for a person who does not exist is worse than no plan.

Your next action: take the 3-hour baseline diagnostic.

Live status in STATE.md.

Start Here

StepDo thisTime
1Read research/findings.md — what is confirmed, what is reported, what is inferred20 min
2Take the baseline diagnostic, cold, timed, no tools3h 05m
3Score it against RUBRIC.md; record in scores/45 min
4Report the numbers. PLAN.md unlocks—
5Begin week 1—

Do not read diagnostics/ANSWER-KEY.md or any solution.py before step 3. A contaminated baseline produces a plan for the wrong person.

The study guides

The actual teaching material. Each is self-contained — every concept from first principles, full implementations, worked scenarios, and the follow-up questions with answers. You should not need to leave these files to understand them.

GuideWhat it teachesVerified
Track A — The Ten PatternsMVCC & predecessor queries · delta logs · intrusive lists · rate limiting · heap scheduling · streaming parsers · dependency graphs · WAL · Bloom filters · backpressure. Complete implementations, not sketches63 behavioural checks green
Track A — The Follow-Up Bank150 questions asked after your code works, with mechanism-level answers — complexity probes · the 100× question · "why not X" · "now make it concurrent" · "how would you test this". The coding round is decided here, not by the tests passingmeasured on 3.13
Track B — The CPython RuntimeObject model · iterators & generators · the event loop built from scratch · the GIL · memory & allocators · the data modelmeasured, live interpreter
Track B — Quiz Bank150 questions with mechanism-level answersspot-checked live
Track C — Distributed PrimitivesLittle's law & the utilization knee · failure taxonomy · clocks & HLCs · leases and fencing · quorums · Raft at usable depth · consistency models · partitioning · delivery semantics · load control · CRDTs—
Track C — The Twelve Worked DesignsAll twelve designs worked end to end — nine sections each, then attacked by a hostile staff interviewer, then revised. 72 critiques, 72 revisions, plus the cross-cutting pattern map and the defect taxonomy across all of them—
Track D — Inference InfrastructureThe KV cache · the roofline derivation · memory budgets · continuous batching, PagedAttention, chunked prefill · prefix caching · speculative decoding · parallelism · autoscaling · a complete worked "design ChatGPT" at both altitudes24 arithmetic claims verified
Track D — The Eight ML-Infrastructure DesignsAll eight worked end to end — LLM API platform · KV cache tier · GPU cluster scheduler · pretraining data pipeline · eval harness · RAG serving · LoRA serving · training fault tolerance. 48 critiques, 48 revisions, plus the eight calculations that decide themevery number script-checked
Track E — The 48 Hours and the InterrogationThe hour-by-hour playbook, the webhook system specified with its seven logged decisions, and the deep-dive interrogation with 40 questions and full model answers—
Track F — Behavioral, WorkedAll twelve story categories with model answers at staff density, the probe playbook, and the forward-looking answers written in full—
Track G — Driving an AgentThe method in eight steps, a fully worked 60-minute transcript with scoring commentary, and five diffs to accept or reject—
Track G — The Diff Bank30 agent-produced diffs to accept, reject or revise in 90 seconds each — the six-pass review, the ranked taxonomy of what agents get wrong, and three that look wrong and are rightmeasured on 3.13

The two reference documents

Not teaching material — retrieval material. Read the guides once; read these the morning of.

DocumentWhat it is
CHEAT SHEETEverything above compressed to what fits in working memory: the loop, the coding query-shape table, the version-dependent Python facts, the distributed arithmetic and fencing rules, the inference formulas and numbers, the take-home playbook, the behavioural frame, and the verbatim scripts to say out loud
GlossaryEvery term used anywhere in the program, defined in one line, with a pointer to where it is taught. Use it when a guide assumes something you have not met yet

What This Is

Most interview prep is a list of things to read. Reading does not make you faster under a timer, does not teach you to defend a design under attack, and does not produce a story bank.

So this is built as an engineering project instead:

  • A stage-gated coding harness that refuses to show you the next requirement until your code passes the current one — reproducing the reported onsite format's actual constraint.
  • A diagnostic battery with a scoring rubric that maps to a starting level per track, so the plan is built from measurement rather than assumption.
  • Runnable experiments for every claim about the Python runtime. No prose assertions.
  • Back-of-envelope calculators so the arithmetic you do out loud in a design round is reflex.
  • GPU memory and roofline math you can derive on a whiteboard.
  • A spaced-repetition queue that resurfaces everything you got wrong at 1, 3, 7, 21 days.
  • Weekly scored mocks on a hire-bar scale, with the level stated plainly.

Nothing is marked complete because you read it. Completion requires a passed drill, a working artifact, or a scored mock.

The Loop This Is Built Against

Reconstructed from one candidate's account, then corroborated where possible. Full epistemic labelling in research/findings.md; the 41-row fidelity checklist in research/source-report.md.

StageReported shapeTrack
Recruiter screenBackground + "where is AI headed"; read the charter firstF
Technical screenTwo 60-min rounds, same day: coding (versioned KV store) + design (fault-tolerant job scheduler)A, C
Take-home48h, "build something real" — e.g. distributed webhook delivery with retries and DLQE
Deep diveLine-by-line walk of your code, from a question list written after reading itE
Onsite ×4Progressive coding (token-stream differ with rollback) · systems coding + Python internals · design ChatGPT · behavioralA, B, D, F
Agentic roundBeta fifth round: oversized task in a real codebase, driven through an AI agentG

A caution that shapes the whole program. The originating source is a single unverified candidate account. The loop varies by team, level, and quarter. So roughly a quarter of the material is deliberately off-report, the onsite is prepared as six components rather than four, and nothing here is asserted to an interviewer as fact about their process. See the anti-narrowing clause.

Repository Layout

README.md              this file — entry point and current phase
PLAN.md                the 26-week program (locked pending diagnostic)
STATE.md               progress ledger; updated every session
research/              findings, fidelity checklist, company brief
diagnostics/           baseline battery, answer key, rubric, scores
tracks/
  coding/              Track A + the progressive harness (15 problems)
  python-internals/    Track B + 5 runnable experiment scripts
  systems-design/      Track C + envelope calculators + 12 designs
  ml-infra/            Track D + gpu_math.py
  behavioral/          Track F — story bank, probes, forward-looking
  agentic/             Track G — agent-driving method and tasks
projects/              the real builds — take-homes, portfolio, essay
mocks/                 weekly scored mocks, transcripts, hire-bar scale
review/                spaced repetition queue + failure log

The Runnable Tools

Everything here works today.

# Track A — the stage-gated harness
cd tracks/coding/harness
./progressive.py list
./progressive.py start token-stream-differ     # prints gate 1 only
./progressive.py test  token-stream-differ     # unlocks gate 2 on pass
./progressive.py chart                          # time-to-first-gate trend

# Track B — proof, not prose
cd tracks/python-internals/experiments
python3 exp01_generators.py exp02_async.py     # (run individually)
python3 exp03_gil.py                            # measures a real lost-update race
python3 exp04_memory.py                         # __slots__, memoryview, tracemalloc
python3 exp05_datamodel.py                      # descriptors, MRO, __getattr__

# Track C — the arithmetic you say out loud
cd tracks/systems-design/calculators
python3 envelope.py qps --rps 50000 --ms 8
python3 envelope.py retry --rps 10000 --fail 0.3
python3 envelope.py latencies

# Track D — the memory and roofline math
cd tracks/ml-infra
python3 gpu_math.py --model llama-70b --gpu h100 --gpus 4

# review — spaced repetition
cd review
python3 review.py add "..." --confident-wrong
python3 review.py drill

# the diagnostic
cd diagnostics/d1-coding
cp starter.py attempt.py && python3 test_diagnostic.py attempt.py

The Seven Tracks

TrackTestsBaseline share
A — Coding under time pressureScreen A; onsite Coding 1 & 225%
B — Python internalsCoding 2's follow-ups12%
C — Distributed systems designScreen B15%
D — ML & inference infra"Design ChatGPT"20%
E — Take-home & deep diveThe 48h build and its defence12%
F — Behavioral at staff altitudeRecruiter + onsite behavioral10%
G — Agentic codingThe beta fifth round6%

Shares are the baseline. They are re-derived from your diagnostic levels and rebalanced at each of the six monthly re-tests.

Beyond the Ask

The differentiators, built into the schedule rather than left as good intentions:

ArtifactWhat it isWhen
Portfolio projectA deep public-quality build in search/retrieval or serving, with a benchmark and honest measured numberswk 10–20
Technical opinionAn essay with a falsifiable claim and a stated falsifier — the real answer to "where is AI headed"wk 6–12
Reading and rebuttalClose reading of their published engineering work, with a specific question about a design choicewk 1, ongoing
Numbers sheetLatency, throughput, cost and hardware figures, each verified and attributedwk 22
Failure logEvery miss, the actual gap behind it, and the fix — with a recurrence checkcontinuous

Operating Rules

Applied without exception:

  • Reading never completes anything. Passed drill, working artifact, or scored mock.
  • STATE.md updated every session. The next session starts cold and reads only that file.
  • Weekly scored mock, on the hire-bar scale, with the level stated plainly — including no hire.
  • Score down when unsure. A generous rubric is the one thing that guarantees failure.
  • Spaced repetition at 1, 3, 7, 21 days. Wrong at any interval resets to 1.
  • Every performance claim has a script that demonstrates it.
  • Inference is labelled as inference. Nothing speculative about anyone's process is presented as fact, and none of it is ever said to an interviewer.
  • Commit at every milestone, with a real message.

Honest Status

What exists today versus what is scheduled — so nothing here overstates itself.

ComponentStatus
Study guides — Tracks A–G + the 150-question quiz bank✅ complete
Worked artifacts — design-ChatGPT answer, 40-question deep-dive interrogation, 60-min agent transcript, twelve behavioural stories✅ complete
Phase 0 research, 41-row fidelity checklist, company brief✅ complete
Baseline diagnostic: 4 parts, answer key, rubric, scoring template✅ complete and runnable
Progressive harness CLI (gating, timing, chart)✅ complete
Harness problems: 15 problems, 60 automated gates✅ 60/60 green — every problem, every gate
Track B runnable experiments✅ 5 scripts, all verified
Track C envelope calculators✅ complete
Track D gpu_math.py✅ complete
review.py spaced-repetition queue✅ complete
Track READMEs: inventories, drills, failure modes, rubrics✅ all 7
Worked designs d01–d12 (distributed) — each with 6 hostile critiques + 6 revisions✅ all twelve complete — attempt each yourself first; they are the answer key, not a substitute
Worked designs m01–m08 (ML infrastructure) — same shape, 48 critiques + 48 revisions✅ all eight complete, every number script-verified
CHEAT SHEET + Glossary✅ complete
Story bank⏳ requires your raw material; will not be invented
Projects⏳ weeks 8, 16, 10–20
PLAN.md week-by-week🔒 locked pending diagnostic

Every runnable artifact is verified: python3 tracks/coding/harness/runtests.py runs all 60 gates in about six seconds. Remaining scheduled work is tracked in STATE.md.

References