The 5-minute video
- One toy dataset, one happy path
- Never says when not to use it
- Hyperparameters = “just tune them”
- You can recite it, but not defend it
An ML models field guide · for data-science, ML-engineer & AI-engineer interviews
Every model has a price bill. Learn what you're paying.
The No Free Lunch theorem says no model wins everywhere. Interviews test whether you know which one to order, why, and what it'll cost you. This is the middle path between 5-minute videos and 500-page textbooks: real problems, four depths of explanation, hands-on labs, and the questions you'll actually be asked.
TABLE 07 · YOUR ORDERml interview bistro · since 1997*
TOTALno free lunch
Why this exists
Somewhere between “just call .fit()” and a wall of Greek letters.
Most ML learning material sits at one of two ends: the toy example that makes you feel like you understand, or the derivation that buries the intuition. Interviews live in the middle — and so does this guide.
you are here ↓
too shallow · iris, Titanic, “it just works”too mathy · proofs first, intuition never
No Free Lunch
- A real, current industry problem per model
- Four depths: ELI5 → PhD, you choose
- Labs where every knob moves a plot
- Order-it / skip-it rules, gotchas, data checks
- Interview questions incl. “what if…” counterfactuals
- Pseudocode + the library call + the docs
The 500-page textbook
- Rigorous, but intuition is left “as an exercise”
- Examples from 1998
- No sense of what’s asked in 2025–26
- Great reference; terrible first date
Every model page is a nine-course meal
What’s on every plate.
Same structure for every model, so you can compare them side by side — which is exactly what interviewers make you do.
The roadmap
The tasting route — in the order we’d learn them.
Each stop builds on the last: lines → regularization → probabilities → distances → margins → trees → crowds of trees → no labels at all.
Staples first, then the specials.
Staples · must-knowSpecials · good-to-know★Signature dish
The waiter’s recommendation
“Which model would you pick?” — in four questions.
The single most common ML interview prompt. Answer like a candidate: describe the data, state what matters, then choose. This picker mirrors how a strong answer is structured.
How to use this
A four-step prep loop per model.
Read the Setup, then the level you’re at
Climb one level higher than comfortable. If ELI5 feels easy, read Undergrad. If Undergrad is easy, PhD.
Break the model in the Lab
Push every slider to its extremes. Predict what will happen before you move it — that's the intuition being built.
Memorize the bill
The receipt on each page = assumptions + costs. “Order it / Skip it” is your answer to “why this model?”.
Answer out loud
Do the Interview Room questions verbally, then reveal. Counterfactual (“what if…”) questions are where offers are won.