ML/GenAI Knowledge Q&A

The breadth round: rapid-fire conceptual questions across everything covered in this site, testing whether your understanding is solid enough to explain clearly under pressure, not just recognize when you read it.

What Gets Tested

• Classical ML breadth: bias-variance tradeoff, regularization, evaluation metric tradeoffs (see Model Evaluation & Metrics)
• Deep learning breadth: why residual connections help, normalization choices, optimizer differences (see Training Deep Networks)
• LLM/GenAI breadth: attention mechanics, RLHF vs. DPO, RAG failure modes (see LLMs & GenAI)
• Inference engineering breadth (increasingly common for ML/AI infra and platform roles specifically): KV cache and why it dominates serving memory, continuous batching vs. static batching, quantization tradeoffs, why TTFT and TPOT are reported separately (see LLM Inference Engines and LLM Inference Optimization)
• Research-track questions (for research scientist / research engineer interviews specifically): explain a paper's contribution and its relationship to prior work (see How to Read AI Papers), propose an ablation that would isolate why a specific technique helps, critique a method's stated limitations, place a paper into The Paper Timeline's lineage — this round tests research taste and literature fluency, not implementation speed

The Two-Minute Test

For every topic in this site's roadmaps, you should be able to explain it clearly, out loud, in under two minutes, covering: what it is, why it exists (what problem it solves), and one concrete example of where it's used. If you can't do that for a topic yet, that's a signal to go back to the relevant page — not to memorize a definition, but to actually understand the "why" well enough to explain it in your own words.

Common Failure Mode

The most common way candidates lose points here isn't not knowing a term — it's giving a textbook definition without being able to connect it to why it matters or when you'd actually reach for it. "Cross-entropy is a loss function that measures the difference between two distributions" is a definition. "Cross-entropy is what you minimize when training a classifier or an LLM's next-token prediction, because it's the loss function that comes directly out of maximum likelihood estimation for a categorical distribution" (see Probability & Statistics) demonstrates actual understanding — that's the level interviewers are listening for.

Practicing Out Loud

Reading a topic silently and feeling like you understand it is not the same skill as explaining it clearly under time pressure with someone watching. Practice explaining topics from this site out loud — to a mirror, a study partner, or by recording yourself — before relying on silent review alone.

Next: Interview Q&A Quick Reference — the actual drill: real questions, click to reveal an answer, test yourself against it.