Topic 73: The Plan, and the Night Before
This folder holds two things. TWO_WEEK_PLAN.md is a fourteen-day route through all four books, from
classical ML to LeetCode patterns, with the pages to read and the code to write on each day. The other
four files are the compression pass for the last evening, when learning is over.
If your interview is more than a day away, start with the two-week plan, which routes you day by day through the two-week workbook — fifteen topic pages of equations, code and spoken answers. If the interview is tomorrow, skip both and read the rest of this page.
Everything else in this repository is for learning. The rest of this page is for the evening before an interview, when learning is over and the only useful activity is refreshing what you already know.
Files in this folder
| File | Purpose |
|---|---|
TWO_WEEK_PLAN.md | Fourteen days, each with what to read, what to type, what to say out loud, and a checkable outcome. Links across all four books. Start here unless the interview is tomorrow. |
CODE_FROM_MEMORY.md | The implementations you must be able to write cold — softmax, linear and logistic regression, attention, multi-head attention, a transformer block, the training loop — in NumPy and PyTorch. Every snippet verified by execution. |
FORMULA_SHEET.md | One page of losses, gradients, activations, norms, optimizers, metrics, and scaling rules. For staring at, not reading. |
DEPTH_AND_BREADTH_QA.md | Ten depth ladders that follow a topic down five levels, ~100 rapid-fire breadth questions, the ones people fumble, and questions to ask them. |
AI_ENGINEER_ONE_PAGER.md | The applied GenAI compression — RAG failure modes, the latency budget and optimization hierarchy, what breaks at scale, OWASP for RAG, and the things people get backwards. For an AI/GenAI engineer loop rather than a classical ML one. |
How to use this tonight
The single most common mistake the night before an interview is trying to learn something new. It does not stick, and it costs you the sleep that would have made everything you do know accessible. Tonight is for retrieval, not acquisition.
A ninety-minute pass that works:
First twenty minutes — the formula sheet. Read it once, slowly. Do not take notes. You are refreshing recognition, not building anything.
Next forty minutes — the code. Read CODE_FROM_MEMORY.md through once. Then close it and retype two
or three of the implementations from memory in a blank editor. Pick the ones you feel least sure about,
which for most people means scaled dot-product attention and the multi-head reshape. This step is the
highest-value thirty minutes available to you, because recognition and recall are different skills and
only recall gets tested.
Next twenty minutes — the depth ladders. Read them out loud. Not silently. The gap between “I understand this” and “I can say this fluently” is exactly what an interview measures, and speaking is the only way to find that gap before someone else does.
Last ten minutes — the fumble list. Skim the questions that sound easy and are not. These are cheap points and they are lost by people who know the material perfectly well.
If tomorrow is an AI/GenAI engineer loop rather than a classical ML one, swap the middle forty
minutes for AI_ENGINEER_ONE_PAGER.md and read it twice. That loop asks about retrieval quality,
evaluation, latency, and failure modes rather than derivations, and the code round is usually ordinary
software engineering rather than implementing attention. The formula sheet is still worth twenty minutes
— a bias-variance or attention question will still land, it just is not the discriminating part.
Then stop. Sleep matters more than the next hour of review, and that is not a motivational sentiment — recall is measurably worse when tired, and interviews test recall.
The morning of
Skim the formula sheet once over coffee. Retype scaled dot-product attention one final time. Do not open anything you have not already seen; a half-remembered new fact is worse than no fact, because it feels like knowledge while you are saying it.
Have ready, in a sentence each: one project you can describe end to end, one thing you got wrong and what you changed, and one thing you would do differently with more time. Interviewers ask some version of all three, and having thought about them beforehand is the difference between a considered answer and a rambling one.
What this folder assumes
That you have already worked through the material. The compression here only helps if there is something underneath it to decompress — a formula sheet read by someone who never derived the formulas is just a list of symbols.
If you are more than a day out, close this and follow the two-week plan instead. Come back to this page the night before.
Cross-references
- Full derivations for anything on the formula sheet live in the numbered topic folders — logistic
regression in
01_classical_ml, optimization in02_gradient_descent, attention in05_attention_mechanisms, scaling in70_scaling_laws. 13_interview_qaand56_spoken_interview_question_bankare the longer question banks; this folder is the compressed pass over them.74_ai_engineer_interview_prepis whatAI_ENGINEER_ONE_PAGER.mdcompresses — the interview loop itself, the modern question bank, RAG failure diagnosis, and latency engineering. Read that folder weeks out; read the one-pager tonight.50_ml_coding_interview_patternsand68_leetcode_patternscover the coding round, which this folder deliberately does not.