AI interviews test understanding, not memorised definitions. Here are the questions that actually come up and how to answer them well.
The concept questions (know these cold)
- "Explain overfitting and how you'd fix it." — train/test gap; fix with more data, simpler model, regularisation, cross-validation, early stopping. (lesson)
- "Precision vs recall — when do you optimise which?" — depends on the cost of false positives vs false negatives. Give the fraud/spam examples. (lesson)
- "Why split into train/test? What's data leakage?" — to measure generalisation; leakage = test info sneaking into training (e.g. fitting a scaler on all data).
- "How does gradient descent work?" — the downhill-in-the-fog answer. (lesson)
- "How does an LLM work / why does it hallucinate?" — next-token prediction; it generates plausible, not verified, text. (lesson)
- "RAG vs fine-tuning — when each?" — knowledge → RAG; style/behaviour → fine-tune; start with prompting. (lesson)
The applied question
"How would you build [a spam filter / a support bot / a recommender]?" — walk the workflow: frame the problem & metric → data → features → model → evaluation → deployment & monitoring. Structure beats a memorised algorithm name.
The trap questions
- "Your model is 95% accurate — ship it?" → "What's the class balance? On imbalanced data, accuracy misleads — I'd check precision/recall per class first." (Correct answer is a question back.)
- "Which algorithm is best?" → "It depends on the data size, interpretability needs, and latency budget — no free lunch." Naming trade-offs = seniority.
How to prepare
Don't memorise — build the projects in this course, then explaining these concepts becomes natural because you've done them. Combine with general prep in our interview question bank and DSA judge (coding rounds still apply).