2.3.3 Bad interview questions
Most interviewers you meet will be bad interviewers. Few companies have proper interview training programs. Junior interviewers lack the experience to know what signals to look for and lack the technical depth to evaluate your expertise. Senior interviewers might be set in their way with their list of pet questions and might defend to the death the merit of their techniques even in light of contradicting evidence.
Bad interviewers ask bad questions. Even good interviewers sometimes ask bad questions. Here are some examples of bad interview questions.
Questions that ask for the retention of knowledge that can be easily looked up
Example: “Write down the equation for Adam optimizer.”
These questions reward those who happen to review this piece of knowledge before the interview or those who prioritize rote learning. Good candidates might be able to work out an equation eventually, but it’s probably not the best use of interview time. The knowledge that can be easily acquired isn’t worth testing for.
Questions that evaluate irrelevant skills
Example: “Write a linked list.”
Interviewers who ask these questions defend their decision by saying that it’s important for ML engineers to have good coding skills. This is based on the assumption that the ability to write linked lists signals coding ability, but it often only signals that a candidate has practiced for a standard software engineering interview. A better approach is to ask candidates to write solutions for practical tasks such as remove duplicated samples in a dataset that doesn’t fit in memory.
Questions whose solutions rely on one single insight
Example: “Take derivative of ."
Solving these questions relies on having seen them before or one moment of brilliant insight. It says nothing about one’s ability.
Questions that try to evaluate multiple skills at once
Example: “Explain PCA to your grandma.”
If a candidate does poorly, you have no idea if it’s because they don’t understand PCA or because they don’t know how to talk to grandmas. A good interviewer would isolate the skills they want to evaluate. For example, interviewers can first give candidates a problem that can be solved with PCA (e.g. a dataset has a lot more features than samples). If candidates bring up PCA, interviewers ask them to explain PCA intuitively.
Questions that use specific hard-to-remember names.
Example: questions about “Moore–Penrose inverse” or “Frobenius norm.”
Candidates might know pseudoinverse or matrix norm without knowing that they are called Moore–Penrose and Frobenius. This can be especially hard for non-native English speakers who study in another language. If you encounter a term you’re not familiar with during interviews, ask for an example.
Open-ended questions with one expected answer
Some interviewers ask open-ended questions that have multiple possible answers. However, because they know only one answer, they expect candidates to come up with that exact solution. This can be frustrating for candidates.
Easy questions during later interview rounds
Example: “Find the longest common subsequence.”
Once these questions are solved, there isn’t much room for improvement. Answering them poorly means that a candidate is unqualified, but answering them well doesn’t mean that a candidate is good. These questions are okay during screening to make sure that a candidate can pass a bar, but bad for later interview rounds when you want to know how high a candidate can jump.
When asked a question that you think is bad, should you tell your interviewer that it’s bad? There’s a non-zero chance that your interviewer might appreciate your candid feedback, but statistically speaking, they might get offended and write you off as someone they wouldn’t want to work with.
Instead, you should ask for clarification. If stuck, explain why you’re stuck and what information you’d need to overcome it. If unsure about the interviewer’s intention, ask. “To better answer your question, is it to evaluate my understanding of X?” should be sufficient.