How I interview machine learning engineers

I’ve interviewed probably close to a thousand people for ML/AI engineering roles since 2016. This is how I find out who can actually ship.

Rule #1 Do not ask questions which have a correct answer. They either know a fact or they don’t. Who cares.

Rule #2 Make them the expert, then get relentlessly curious. I ask them to describe a past project and just keep pulling at a thread. How did that actually work? Keep asking harder questions and get to the edge of their knowledge. Don’t be afraid to ask questions you wouldn’t know how to answer yourself (the truly great candidates will change how you think).

Then I talk them through a hypothetical. One of my favourites: You have a dataset of ten thousand 30-second video clips. The task is to predict the final 10 seconds given the first 20.

Which is harder to predict: the video or the audio? Some people talk about data dimensionality. Some talk about tensor operations. And some people ask: “well, what’s happening in the videos?”

You have 24 hours to decide if this project is feasible. What do you do? Some say “go read the literature.” Some go look on huggingface for a pretrained model. And some would spend 20 hours watching the videos.

We’ve decided to take on the project. What’s the dumbest possible baseline you can build? Can they come up with something OTHER than gradient descent?

It’s a high-signal, no-BS interview that’s helped some really stellar people stand out from a pile of CVs. And it’s chatgpt-proof, too.