Interview Questions
AI Interview Questions and Answers
AI interview questions for engineering roles (not pure research roles) usually focus on practical understanding -- how a model is trained, what can go wrong, and how modern LLM-based tools actually work.
Example: Overfitting shown numerically
A model trained on 1,000 house-price examples reaches 99% accuracy predicting prices on those same 1,000 examples, but only 62% accuracy on 200 new houses it's never seen. That gap is overfitting -- the model memorized quirks of the training data rather than learning patterns that generalize. The fix isn't training longer; it's usually more (or more varied) training data, regularization, or a simpler model.
Frequently Asked Questions
AI is the broadest term -- any technique that makes a system behave intelligently, including hand-coded rule-based systems. Machine learning is a subset of AI where the system learns patterns from data rather than following explicit hand-written rules. Deep learning is a subset of machine learning using multi-layer neural networks specifically, which have driven most recent breakthroughs (including LLMs).
When a model learns the training data too specifically -- including its noise and quirks -- and fails to generalize to new data (shown in the example above). Common fixes: more training data, data augmentation, regularization techniques (like dropout in neural networks), simplifying the model, or stopping training earlier (early stopping) before it starts memorizing rather than generalizing.
Supervised learning trains on labeled examples (input paired with the correct output, like an email marked spam/not-spam) to learn a mapping from input to output. Unsupervised learning works on unlabeled data, finding structure on its own -- like clustering similar customers together without being told in advance what the groups should be.
A neural network (typically a Transformer architecture) trained on enormous amounts of text to predict the next token in a sequence. That simple training objective, at sufficient scale, results in a model that can perform a very wide range of language tasks -- writing, summarizing, answering questions, even basic reasoning -- without being explicitly trained for each one individually.
A neural network architecture introduced in 2017 built around the "attention" mechanism, which lets the model weigh the relevance of every other token in a sequence when processing each token -- rather than processing sequentially like older RNN architectures. This made training far more parallelizable and dramatically improved performance on language tasks, and is the foundation nearly every modern LLM is built on.
Deliberately crafting the input ("prompt") to a language model to reliably get better, more accurate, or more consistently formatted output -- techniques include giving explicit examples of the desired output format (few-shot prompting), breaking a complex task into explicit steps, and being specific about constraints and desired tone.
Prompt engineering shapes a model's output through the input alone, without changing the model itself -- fast, cheap, reversible per-request. Fine-tuning actually continues training the model on additional, task-specific data, updating its weights -- more effort and cost, but can produce more consistent behavior for a narrow task than prompting alone, especially when a very specific output format or domain expertise is needed.