What to Do When You Don't Know
The Interviewer Throws You a "Googly"
You get a question you've never seen before. Your mind goes blank. We've all been there!
Remember this: The interviewer is not testing if you know every single answer. They are testing what you do when you face something new. They want to see how you think.
Your 3-Step Recovery Plan
1
Acknowledge & Reframe
Don't panic and go silent. Take a deep breath. Buy yourself some time and show you are thinking.
- Start by saying: "That's a very interesting question. Let me take a moment to think about that."
- This shows confidence and gives you a few seconds to calm down and structure your thoughts.
2
Deconstruct & Connect
Break the big, scary question into small, familiar pieces. Nobody knows everything, but you know *something*.
- Think out loud: "Okay, this question about 'X' seems related to a concept I know, which is 'Y'."
- For example, if they ask about a new algorithm, connect it to one you know: "I haven't used that specific algorithm, but it sounds like it's a type of clustering algorithm, similar to K-Means."
3
Propose a Plan (The Golden Rule)
This is the most important step. You turn "I don't know" into "Here's how I would find out."
- You are showing them you are a problem-solver, not just a book of answers.
- Use the magic phrase: "I don't have direct experience with that, but here is how I would approach the problem..."
Plan in Action: A Real Example
Curveball Question: "How would you build a model to detect fake news on our platform?"
Your Winning Response:
- (Acknowledge): "That's a great and very relevant problem. Let me structure my thoughts."
- (Deconstruct): "This seems like a text classification problem. The core task is to classify an article as either 'real' or 'fake'. This is similar to spam detection."
- (Propose a Plan): "Here's how I would approach this:
- Data Collection: First, we'd need a labeled dataset. We could use articles from trusted sources as 'real' and examples from fact-checking websites as 'fake'.
- Feature Engineering: I would extract features from the text, maybe using TF-IDF. I would also look at metadata, like the age of the source account or the number of shares.
- Modeling: I would start with a simple model like a Naive Bayes or Logistic Regression as a baseline. If that works, we could try more complex models like an LSTM or a Transformer.
- Evaluation: We'd need to be careful with our metrics. We would probably want to optimize for Precision to avoid flagging real news as fake."