Hacker News Reader

Doing well on whiteboard interviews can be a sign of high intelligence (a true positive result). But what signal does doing poorly on them give? Did you measure the candidate's intelligence (true negative), or preparedness maybe anxiety levels (false negative)? I guess the answer is that it does not matter as long as you have a large enough true positive candidate pool. And FAANG has this.

Sure it matters. Who doesn't do well due to stress? I think you'll find women, minorities and basically a lot of people outside of the cultural norm/majority do poorly.

That's not an argument that you can eliminate test anxiety, but certainly if we can reduce it without destructively eliminating the high value candidate pool we should do so.

What you are saying is an ethical/moral consideration and it is a valid one. When I said it doesn't matter, I meant that FAANG is commercially not incentivised to account for the anxiety factor. Even if they miss out on good candidates due to their anxiety, they can still choose from plenty of good candidates who don't suffer from anxiety. Given how many people are aspiring to join FAANG, I'd assume it's a buyer's market for these companies.

I'm not sure that's true. What you describe is the case for any individual candidate, but doesn't account for how those candidates interact with one another. Specifically, it has repeatedly been shown that more diverse teams will consistently outperform more homogeneous ones.

If your company only hires above the 95th percentile but their method for finding that quality cutoff is heavily skewed towards a certain demographic, in the long run that company may well perform worse than a company that hires at the 80th percentile but without the demographic skew.

> Specifically, it has repeatedly been shown that more diverse teams will consistently outperform more homogeneous ones.

From the data I've seen made public, FAANG are usually more diverse than other companies.

I think you'll find men, people in majorities and lots of people on average will do poorly, also.

The ratio that's important for those huge companies is the "true positive"/"false positive" one.

I guess if you make the test hard enough, always change it, and add some unhealthy dose of suspicion over the candidate's behavior, you can maximize that ratio. But I'm not confident FAANG even looks at it. It's much more likely that they just postulate that a hired employee is a competent one and go on with their day.

From the candidate point of view, of course, things are very different, and the negative results are much more important.

Companies don't care about false negatives, they don't want false positives because that is concretely lost revenue, rather than not hiring someone.

"Whiteboard interviews are used, not because they have a lot of relevance to the job, but because they're a hard-to-fake indicator of high intelligence."

By now it mostly correlates with effort spent preparing. There's dozens of web-sites, companies, consultants and message boards where one can train for those interviews. Whiteboarding is a skill like any other, it can be trained, faked, etc.

"But almost all of those qualities are easy to fake and hard to assess in a 30 minute conversation."

Perhaps it's not possible to assess somebody's future job performance in a 30m conversation.

Before taking algos/data structures I did terrible on whiteboards. After, I do well. From this experience, they seem reasonable indicators of a cs grads intelligence. I doubt I am too much better of a software engineer in my day to day job, but I do think this toolkit does unlock a second disperate level of engineering ability, to code core algos for novel software.

I just have a problem suggesting learning this one set of skills really boosts my fluid intelligence. Im probably just as intelligent as I was before.

But I do get the cooralating signal as a cheap, low false positive approach to selecting good candidates.

Doesn’t that require a lot of interpretive skill on the part of the interviewer?

And for the interviewer to not arrive with the expectation that you will pass the same hazing ritual they passed to deserve that big paycheck?

It’s been a long time since I interviewed at a FAANG but I don’t think my whiteboard mojo played much of a role. It was more down to my personality fit with the interviewers, and the correctness of my algo answers. Which were still pretty softball way back then.

Everything I’ve read and heard since then makes me think it’s much more about those things now, and even less about general mental agility.

That infantryman comment is blatantly false. The literal bare minimum entry level job in the military is infantry. Like if you pass the ASVAB with the lowest score possible, you are infantry without a say in the matter. The army wants dumb people to get hit with bullets first. Then the laborer types, then the admins. Thats why those "better" jobs are harder to get.

That "lowest score possible" for the Army is a 31 percentile on the AFQT, which is a proxy for an IQ score. That means about 100 million Americans are too stupid to carry a rifle. My understanding is that the real score is actually higher than the advertised minimum.

I've taken the ASVAB and it is definitely not a meter of IQ. Middle school children who at least know simple algebra would have no problems scoring fairly well.

The ASVAB iirc was designed to determine critical thinking and mechanical skills. If you aced it, you either studied or shouldn't be in the army.

If fluid intelligence was truly perceived as the most important thing, then people would just skip the coding entirely and try to directly measure fluid intelligence. There are companies more than happy to run IQ tests and the like to screen candidates. Probably a lot cheaper than the current hiring processes.

The fact that they aren't trying that, and are instead attempting to get a measure of coding ability, suggests they don't really believe in the approach.

It's a lot easier to demonstrate the business necessity of a whiteboard test than IQ tests.

Administering an IQ test would be legally risky and could have a disparate impact on protected classes.

Griggs v. Duke Power Co.[0] is the Supreme Court case on this topic.

[0] https://en.wikipedia.org/wiki/Griggs_v._Duke_Power_Co.

Thank you for putting this eloquently. It's very popular to just hate technical interviews which involve white boarding.

> “but because they're a hard-to-fake indicator of high intelligence”

Surely you realize large companies are full of people whose friends just told them what the questions were going to be? Memorizing 4 problems is still easier than cramming 40.

In what universe are whiteboard interviews used this way, as opposed to the interviewer expecting an exact answer regurgitated instantly?

I'm sure many interviewers do cargo-cult whiteboard interviews, but the core intention of the archetypical whiteboard interview is to spring a complex question on the candidate that they have never seen before and observe how the candidate thinks. If the interviewer is asking questions that the candidate knows the answer to the interviewer is literally doing it wrong. It's not a trivia challenge, and the interviewer should only ask questions that they are confident the candidate hasn't ever seen before. That's not to say that Leetcode-style preparation doesn't work, but it works by tricking the interviewer into thinking that the candidate has solved the problem very quickly, when in fact the candidate actually already knew the solution.

Once you "memorize" a core set of data structures and algorithms, you can apply them to a variety of "novel" problems. The issue is those technique that are memorized, that become useful in solving these problems, aren't intuitive, despite their deceptive simplicity. In reality there's little memorization of solutions, and much memorization of useful techniques (data structures, algorithmic techniques) that are useful for these problems. And the more you do, the more you'll have at your disposal for "novel" problems. Further, many of these memorized techniques are not as intuitive as they seem. Quite a few resulted in publications when they were discovered.

5 years ago the fast/slow pointer technique for cycle detection could have been impossibly hard, now I see questions that just assume the candidate knows this. Every year the bar gets higher.

Your comment made me feel better about myself, thanks.