Having just had a 3h long technical interview for Google Deepmind, I cannot agree more with a lot of points raised in this post.
Deepmind being a machine learning/statistics/maths/computer science fuelled company, it made sense for the interview process to follow this simple organisation.
I was however very disappointed by the questions asked for each part. Not a single one of the ~100 questions asked during these 3h of my life demanded some "problem solving" skills, only encyclopaedic knowledge (describe this algorithm, what is a Jaccobian matrix, define what an artificial neural network is, what is polymorphism, give examples of classifiers, what are the conditions to apply a t-test...)
So what if someone doesn't remember every definition of the stats/ML/CS/Maths respective bibles as long as they're clever enough to look it up and understand quickly what's needed?
I mean, I get it these are very basic questions but as a highly qualified interviewee who necessarily has other offers given this set of skills, this fastidious, back to school, time wasting process does not reflect well on the company and makes me consider my other options even more seriously.
Do you think mathematicians start over from learning how to count everytime they encounter a new problem? That's different from looking up the 10th digit of pi: that's encyclopedic.
Knowing what a classifier is is simply more than encyclopedic knowledge that you should probably know before joining an AI company.
As I said before, yes these are very basic questions... and I'm not complaining that the questions were too hard or anything, I'm saying that finding someone who will be able to answer 100% of these "definition" questions will not tell you anything about how competent that person is...
Making them face a simple stats/CS/maths/ML problem and see if he/she is able to come up with the relevant concepts is far more interesting.
Ah, I see. Personally I think I could have taken the "give an example of a classifier" question into a deep dive of their competency (talk about how one would build that classifier, etc.) but they might get so many completely unqualified candidates that don't know those things that they need to have those sorts of hurdles before they start really evaluating you for more holistic problem solving traits.
But if you don't think they did that at all then I guess that's bad!
I generally am for the "look things up" argument, but so many people in tech take that to an extreme of "I can fully understand an entire discipline by looking at the Wikipedia page for 5 minutes".
I completely agree with you. Also to be fair towards Deepmind I should mention that this was the first round of interview. Hopefully the following steps will prove more stimulating, in which case 3h for the first step was maybe slightly too much.
Deepmind being a machine learning/statistics/maths/computer science fuelled company, it made sense for the interview process to follow this simple organisation.
I was however very disappointed by the questions asked for each part. Not a single one of the ~100 questions asked during these 3h of my life demanded some "problem solving" skills, only encyclopaedic knowledge (describe this algorithm, what is a Jaccobian matrix, define what an artificial neural network is, what is polymorphism, give examples of classifiers, what are the conditions to apply a t-test...)
So what if someone doesn't remember every definition of the stats/ML/CS/Maths respective bibles as long as they're clever enough to look it up and understand quickly what's needed?
I mean, I get it these are very basic questions but as a highly qualified interviewee who necessarily has other offers given this set of skills, this fastidious, back to school, time wasting process does not reflect well on the company and makes me consider my other options even more seriously.