"I was interviewing a candidate who said he had experience programming on an IBM/370. So I asked him if you perform a character edit format instruction in EBCDIC mode with the leading zero specifier and the numeric value is too great to fit into the allocated field, after the instruction completes, what is the state of the program status word overflow field?" Then trounced the guy for not knowing. The thing is the guy asking the question happened to have worked on that instruction when he worked at Amdahl.

One thing to know is the IBM 360 and descendant family had a commercial instruction set option that, in a single instruction, could take a format value and generate a string output that followed some format specification, kind of like sprintf but with even more options.

"Is the computer operating on American electricity, or European?"

I think like 70% of interviews I ever had were like they were there to prove how smart they are and how stupid I am. I suppose most likely to make me feel stupid and accept lowball offer.

Often times, the "gotcha" part is just dumb and nonsensical, yet gives the interviewer(s) a sense of misguided (false) superiority, and wastes everyone's time. I would venture to say that 99% of the time it's complete un-indicative of how effective the candidate would be in the role.

Referring to my aforementioned bad interview - the question, after all the technical stuff had been cleared (this was for a junior frontend dev role) - they asked "imagine a car is broken and not running. how would you go about figuring out how to fix it?". Being someone whose brother and father enjoyed fixing cars, I asked every question about the problem with the car - how it was used, what sort of car it was, what the issue with the car was, what prior problems the car had, etc. I got a bunch of useless answers. After I exhausted all my questions, the interviewers told me I had failed. Why? One interview brought out a tiny hotwheels car with a missing wheel out of his pocket, and proclaimed to say "you didn't ask if it was a real car, it's a toy car, of fucking course it's not supposed to run like a real car!" while laughing hysterically. How on earth does that indicate if a junior frontend dev can do their job or not? Stupid.

The question was to estimate how many vacuum cleaners there were in the city we were in.

Fine, I did some estimation of how many vacuum cleaners per household and per office, across how many households and offices. Standard stuff.

Then the guy starts laughing and saying I'd failed because I didn't include discarded vacuum cleaners in landfills. Or the vacuum suction devices they put in your mouth at the dentist's office. And so forth. And then had to spend the next five minutes listening to him "teach me" how not to make assumptions. So I acted all polite and tried to fake "oh gosh thank you so much for enlightening me!"

Shockingly, I got the job, which required unanimous approval from all interviewers. Never met him again, and to this day I still have no idea whether this was supposed to be a test of estimation (which was easy to pass), a test of not making assumptions (which is dumb, but OK fine I failed), or a test of being appropriately professional and smiling in the face of complete bullshit (which I'd say I passed with flying colors).

I mean, in my professional life I've certainly had my fair share of customers and managers and coworkers who spout bullshit and you really do just have to lie with a smile and say "oh my gosh you're so right thank you for explaining that, I appreciate you so much!" Where you need to make them feel smart.

On the other hand, I just don't think he was thinking that far ahead.

I had a similar interview years ago - something like "how many windows are there on houses in our town?". Wasn't quite that, but I asked up front if "houses" meant just physical standalone houses, or if they meant living spaces, including apartments/dorms, etc. I got clarification, gave some estimate with some reasoning, and was then told I was the only person of the 8 they'd interviewed that had asked any clarifying question at all, which apparently impressed them enough to make an offer.

- Asked/did not ask clarifying questions

- Did/did not (or could not, on prompting) verbally walk through their reasoning

- Could/could not articulate which assumptions they felt were most important/why

Nothing about the actual content of the question itself, or if your answer was approximately correct (I usually did not know even the ballpark of the correct answer myself). I will say I did sometimes write down if candidates make comically bad assumptions. Like assuming the population of the USA was 1 Billion people. It's a fine line on what is "comically bad" but like, if you are interviewing for a startup of 20 people and you use $20B/year as the revenue assumption with no wink. That's a red flag. Lmao.

When you're in a real situation with a customer or user, you ask tons of questions. You use a lot of common sense to figure out what they really want, what's actually important, etc.

But often times these interview questions -- like how many vacuum cleaners in a city -- don't need any further questions asked. The idea that you'd count vacuum cleaners in landfills, or dental suction devices, is just silly. In real life, if someone wanted to know about the vacuum cleaners in landfills too, they'd tell you in the first place.

If an interviewer wants to see if someone can ask clarifying questions, they'd better come up with a scenario where it would make sense to ask them in a normal conversation. Scenarios that are genuinely ambiguous to anyone with common sense.

Otherwise interviews become this weird cargo-cult thing where you have to learn that interviewers present common-sense clear questions, but you have to ask silly clarification questions that you wouldn't in real life, just so somebody can check a box that you asked questions.

Assumption is mother of all fuckups.

I’ve seen customers wanting X and assuming that you should know that X comes with A and B because everyone in their business knows that. But you implement only X with A because you made your own assumptions and did not ask. So you missed the deadline and customer that’s it.

Is that comically bad? It's only like three times too much (if you're using the American, 10^9 billion), but we'd accept a much greater margin of error in some other assumptions (like in the classic 'how many Piano tuners are in the city', I think an assumption of 1-in-30 or 1-in-300 households having a piano (that would need tuning) all sound like they could be true!)

And if you’re interviewing, consider that for a decent proportion of people it may not be the country they live in (at least at the time you’re asking the question).

Playing devil's advocate, maybe a junior frontend dev that doesn't trust that they understand what someone is asking for and pushes back on bits that should be obvious will perform better (in some contexts?) than one that doesn't.

For a junior role in particular, though, it really doesn't seem like that should be the threshold and it sounds like it was delivered poorly on top of that.

The car example is just stupid. Could you image how idiotic you’d sound beginning by asking is it a real car? You’re basically accusing your interlocutor of operating in bad faith (which they were).

Yeah, I think I agree. I was just responding to (what I saw as) sentiment that it was entirely unrelated to performance, which seems to overstate the case.

> Could you image how idiotic you’d sound beginning by asking is it a real car?

Tbh, "willing to sound like an idiot to double check assumptions" is probably something worth selecting for! I don't think that saves this question, though.

That interview question is basically, "We lied about an imaginary situation and were disappointed you failed to accuse us of lying mid-job interview."

Yes, that happens, and elsewhere she goes on about culture in tech.

For what we know, the guy may well have been employing "test-taking strategies", and he may have been led down the garden path by the interviewer.

There's far too many posts in Rachel's blog where she goes on about "the one", who knows much better already, and here she channels the asshat that she complains about when she encounters him at work.

Each router checks its table for the destination, and if it doesn't know it, queries the next upstream router, its default route, the next hop. Each router likely ultimately informs you of the hand-off via a packet of some sort, and your then traceroute sends a ping/ICMP to each hop to learn how far away they are.

He maybe could've been pushed to expand on what he did know in more detail, but it seems like she just threw out SNMP as misleading bait, and he maybe mixed up ICMP and SNMP. She is right to call herself a troll, but wow, that's crazy to say she caught the guy in a lie of insanity.

When you think about it, the candidate isn't even that wrong. Back then, at university, a certain professor would explain the oral exam to the candidate at the beginning. He would explain that he would incrementally increase the difficulty and skip from area to area. The goal would be to find the limits of the student's knowledge, the student would walk away feeling terrible, and he, the professor, didn't enjoy the experience.

That's how it ought to be, but here? OK, candidate doesn't know ICMP well, next topic, no need to waste time and dig in.

Here's another unfavourable thought: some people with abusive childhoods react very badly to dominance displays, and here is Rachel engaging in just that. One wonders what had happened before.

Storytime! In a previous workplace a disagreement over fire safety with escalated into uncalled-for and unwelcome dominance behaviour from my supervisor. All attempts to deescalate were rebuffed, and now there is litigation from multiple plaintiffs, this person took out her sociopathic tendencies on many people. With a minimum of professional detachment or a HR department with a clue the peace would have been kept. (Yes, a few months later the fire marshal issued a code violation, as predicted.) You may not wish to live in that internet, but we live in a world where sociopaths are overrepresented in leadership positions.

Something rubs me here just the wrong way. Rachel complains about ageism and contempt for women in tech, and with good reason, and then she takes it out on an overenthusiastic candidate who can't read the room.

I assure you, I've met many people who treat their interviewees worse who are not sociopaths.

One of the biggest red flags in an interview is if I ask a question and the person doesn't know how to say 'I don't know', because it suggests there's a big risk that if I assign them a task in their day-to-day work, they won't tell me if they feel unprepared to tackle it. That's a far bigger issue than not knowing that traceroute uses variable TTLs to figure out the timing along the route.

FWIW I would never ask these kinds of gotcha questions. I just give simple programming problems and talk through solutions with the candidates, and then throw in complications to the questions to make them more interesting and test more areas of the candidates knowledge and problem solving abilities. Yknow, like what happens on the job every day.

I am basically doing the same as I also interview people - but I also check the market from time to time as I am not company owner.

But I basically don't care about the offer if company pays guy much or not it is not my money and I only win if I get a smart, nice person who knows his job to work with.

> Well, he guessed 69, sir, so I assumed he was doing some serious game theoretic calculations

Of course you can set up the game such that Ballmer has to commit on a number at the start of the game (by sealing it in an envelope or whatever), but that wasn’t specified.

Given that 7 guesses covers 128 numbers, you can offset by +/- 14 without actually affecting the "worst case" of the algorithm (i.e. provided you have at most 64 either side of your guess). As you say, randomly selecting this offset would neuter most adversarial examples (purposefully chosen to fall into the gaps of binary search) and would possibly completely remove the benefits from adversarial choice (though a tailored distribution on offset might be required there).

I'd be interested in such an analysis too.

I might be confused, but don't 7 guesses actually cover 255 numbers? I think you have to count all nodes in the search tree, not only the leafs, because you can get the correct number before reaching a leaf node.

Or more generally k guesses cover 2^(k+1)-1 numbers, e.g. with one guess you get the answers correct/high/low, which can cover 3 numbers)

Maybe there is a mistake in my thinking, because this would mean you can cover 127 numbers with 6 guesses so you could not lose the original game.

Edit: My mistake is that you still have to explicitly guess even if you know the precise answer already, so you cannot cover 3 numbers with 1 guess. This means 7 guesses cover 127 numbers.

You can also view it as a recurrence:

  f(1) = 1
  f(n) = 2*f(n - 1) + 1 = 2^n - 1

As the original article says, on average you can win $0.20. But that's indeed the upper bound if we speak of the adversarial number picking.

The game theory here is similar to another quiz "Guess 1/3 of the average".

You are really trying to guess how deep the other has thought about the problem, so you can tell which strategy they settled on, and then you adapt your strategy based on that. Of course, it's a loop.

If you know your opponent picks a number uniformly from all numbers that lead to a maximum of guesses, the optimum strategy is a binary search between those numbers, making sure to pick one of those numbers at each turn.

The problem stays completely symmetric under this condition, so there would be two (maybe four due to edge conditions) optimal first guesses summing to 101.

In general, I think the trick still is a binary search where each guess splits the range of options in halves of equal expected/min/max cost (depending on whether you want to optimize for expected/min/max cost).

The post never answers the question "so what is the real expectation value", which is a more difficult question. But I think if the interviewee chooses a number randomly from 40-60 as the first guess and does a binary search from there, Ballmer can't really improve on choosing his initial number randomly.

I agree that if you choose your first guess somewhat randomly in the 40-60 range (maybe not a uniform distribution though) Balmer would be forced to choose randomly and you would be back at a positive $0.20 EV. For example, you could flip 6 coins and add the number of heads, then flip another coin to decide whether you add or subtract the number of heads from 50 for your starting guess. But I think you would need to randomize your later guesses a bit also.

The scenario is framed as a zero sum game: one of us wins. The question is, "should you play?"

In order to answer, you need to be able to determine whether or not there is an optimal strategy that is generally successful.That should include both the assumption that Ballmer has chosen a number adversarial weighed against the random choice.

It seems like a lot of possibilities and I think you can get away with a minimum gap size of one, but let's assume you do 5 3-gaps at 1, 25, 50, 75, and 100 and 2-gaps everywhere else. So start with 51, then 26 and 76. Then go up or down 12, then 6, then 3. If you have a gap of two you flip a coin, if a gap of three you pick the middle one.

Or if you have them write down the number and you think it has double-digits you could put your 4-gaps below 20. Start with 53 and go up or down 24, 12, 6, and 3 (unless it is below 20, then it is multiples of four.) 59 would pay you a dollar.

Your starting guess could be anywhere from 37 to 64 without paying out more than a dollar, but if you start with an extreme, then low odd numbers and high even numbers will have a negative payout. However, I think you can still randomize sufficiently starting with 38 and 63, e.g. 63-31-15-7-3-1.

I now think that not constraining the players remaining choices to follow binary search pattern would completely change the resulting equilibrium and improve the results for the player. But that would be more computationally demanding to calculate because there's a strategy choice for every range of choices. And also I've avoided work for 2 hours by working on this so that's not great haha. I _am_ curious what not constraining the player to binary search would do though...

https://gukov.dev/puzzles/math/2024/09/05/steve-ballmer-was-...