The Most Revealing Job Interview Question

Being able to say "no" is important, but so is knowing when to say "no" and when to say "yes".

If somebody tries to drag you onto an irrelevant tangent in a meeting, that's when you should say "no". If somebody is talking to you one-on-one and you discover that they actually care about something different than you thought, that's when you should say "yes".

The problem with this scenario is that it pretends to measure a person's ability to say "no" all while giving them no reason whatsoever to actually say "no".

If you want to test a person's ability to say "no", you have to put them in a situation where that's actually the correct move.

Reading the title: "Lattice measurement of α_s with a realistic charm quark." I know what a lattice is, I don't know what sort of measurement they intend to make on it, I have no clue what alpha_s is, I know vaguely what quarks are but don't remember the differences between say a charm quark and a strange quark. I don't know what makes a quark "realistic" or not.

Abstract: "We report on an estimate of α_s , renormalised in the \overline{MS} scheme at the τ and Z^0 mass scales, by means of lattice QCD. Our major improvement compared to previous lattice calculations is that, for the first time, no perturbative treatment at the charm threshold has been required since we have used statistical samples of gluon fields built by incorporating the vacuum polarisation effects of u/d, s and c sea quarks. Extracting α_s in the Taylor scheme from the lattice measurement of the ghost-ghost-gluon vertex, we obtain α_s^\overline{MS} (m_Z^2) = 0.1200(14) and α_s^\overline{MS} (m_τ^2) = 0.339(13)."

No idea. Perhaps the introduction will be enlightening...

Hey, some progress. The paper appears to be using data from the LHC experiments that have been all over the pop-sci news for the last few years. Section 2? Math, data... Hey, section 3, conclusions. So they have given a measurement of alpha_s from lattice simulations, specifically an improved method for lattice simulation. Googling says alpha_s is the strong coupling constant, and hey, I've seen the fine-structure constant before... alpha_s has to do with "asymptotic freedom". Is there a 5-minute intro to the point of trying to measure alpha_s? Is there a 5-minute intro describing measuring alpha_s with lattice simulations as opposed to something else?

This whole exercise took me quite a bit longer than 5 minutes and I still don't feel like I understand much of anything. What's the most important part here? I'll probably forget everything I typed down by the end of the day, except perhaps that alpha_s is somehow related to the universal-constant alpha I've seen before. I'm doing this from the perspective of the ignorant person. Could the authors of the paper (or a fan of the paper who read it, understood it, and wants to explain it to me) explain to me what exactly they're doing to a sufficient degree? I mean, such that I have a clue what (if not why) the paper is trying to accomplish in a real sense, what's important about the paper, without needing to understand how they accomplish whatever the goal is, or understand all the math and science details? Probably they could, but it would take a lot more than 5 minutes. It'd take more than 5 minutes just to get the initial definitions out of the way such that I don't have too much confusion over what is being discussed. Is there a 5-minute intro to quantum chromodynamics?

Even Khan Academy videos are usually at least 10 minutes. His American Civics videos provide the best example for my point here.

As for explaining why I want to use technology X instead of Y, this is an entirely different question than the one proposed in the submission, and one that's actually a better interview question. "Tell me about a time you were faced with a choice between competing technologies and how you justified that choice to your superiors." If someone's asking me to explain why I want X over Y, I have three responses depending on who's asking and in what context they're asking in. The first is the rude one: "Bug off, I don't have to justify my decision to you and you're wasting my time / company dollars having me explain something so many inference levels above your present understanding". I don't think I've ever actually used that one, but it's a possibility. What usually happens: "Here's the tradeoffs I'm making, here's why I think these tradeoffs are a good set to make at this point." But this requires I make the assumption that whoever I'm talking to knows something, isn't just an ignorant off the street. Fortunately at tech companies this assumption is fine. Describing superficial differences ("I want to use Python over PHP because it enforces indentation and will be prettier", or "I want to use DynamoDB instead of MySQL because DynamoDB isn't SQL") is useless. You can make a lot of such useless statements in 5 minutes and completely miss the useful ones. Giving an actual explanation of the technologies and reasoning behind picking one over the other, that will take more than 5 minutes, doing otherwise is unfair to the person seeking the understanding. One should not attempt to give explanations like one gives elevator pitches.

The last response is taking the Feynman route and throwing up my hands in defeat, a nicer approach than the first response. "Sorry, I just can't explain things that are so many inference levels above you. I can spend 5 minutes per level trying to build you up to where I can explain this thing you're asking about, it might take a couple days, but I can't just jump in one step from your level to here." I can explain basic lazy evaluation in hopefully less than 5 minutes to someone who has/could pass a programming class, if they're engaged and fulfilling their part of the two-way street that is learning. If they're ignorant and want to learn, they can learn but it will take a while. If they're ignorant and don't want to learn, I have spent hours with people trying to hammer in basic algebraic concepts or programming concepts to no avail (this on top of whatever their professional and licensed teacher is doing).

In the realm of books, if I was asked to explain Data Analysis: A Bayesian Tutorial, how could I even begin doing that with someone whose notion of probability is at best "1/6 chance of getting a 6 on a die roll"? Do they even have a coherent internal model for what "information" is, a concept that is really quite new in human history? I have to start at a level way below where the book is at, and 5 minutes later I still won't be at the book's level and I can't explain the most important aspect of the book. Unless something like "knowing the stuff presented in the book helps you do cool things with computers", expanded upon for 5 minutes, is adequate. Is "helps you learn how reality works", expanded upon for 5 minutes, an adequate explanation for what the important aspect of the Giancoli physics book is?

I suspect our disagreement is more over where our standards of explanation quality are. Maybe you think trying to use a rubber band metaphor for magnetism is fair, I think it's cheating the learner. Maybe you think using shaky analogies that "lead to all kinds of implications" is adequate, I think such carelessness is only poisoning the well.

If you want to know how a candidate communicates certain pieces of information at a job, ask them how they do it. If you want to see how a candidate reasons in an argument, get into an argument. If you don't want to argue directly, help them find something relevant that they dislike and see what structure their rant about it takes. If they don't really dislike anything (or just don't want to take a stand on anything) that's also useful to know. A lot of these interview "tricks" purport to select for certain traits but it's usually a very indirect method at best.

The first observation that I have is that given that the secret number s is chosen, the first step can be completed arbitrarily quickly. One could use a function that rises arbitrarily fast. Imagine for example the function taking k to the Ackerman function A(2,2,k). That rises so fast it's incomprehensible, but really one could easily produce a function which rises faster still (the algorithm that picks s first!). The problem is, though, of how fast a fixed algorithm is for random s. If s is truly chosen at random from the positive integers, this leads to problems. Fix your putative algorithm. Suppose for the moment that it starts at 0 (it is not going to matter where it starts) What is the probability that the kth number that your function spits out is less than a random integer? 100% After all, how many integers are greater than any given integer?

Therefore, no growth function is any better on average at finding the upper bound than any other. Therefore step 1) is an insoluble problem. The problem should have been specified in some other way in order for it to make sense.

However, the second step involves log_2(n) time (in the worst-case-scenario and still O(log n) in general) where n is the upper bound --the output of step 1-- which means that the time to complete the algorithm is (Time of step 1 to find n) + O(log n). IF the problem made sense and it was the case that step 1) were soluble, then it would matter how fast step 1 is countered by the degree to which step 1 tends to overshoot the secret number --because overshooting by k has the penalty of log(k) extra operations.

Is there are framework in which question 1 makes sense? It would make sense if there were a given probability distribution on the integers (a function on the positive integers whose sum over all integers = 1, for example choose the function f(n) = 6/(n^2 * pi^2)) A probability distribution gives you a way of answering the question: "what proportion of the positive integers are greater than k?"