You Will Need 1.8 Million Files

CS 099: Useless Work

Three credits. No prerequisites. No textbook, no final, no curve. Enrollment: you, apparently.

Course description

You asked this week what a college freshman in computer science should do, and you got the answers you were going to get. Co-op every summer. Learn the fundamentals. Write the code yourself and use the model as a tutor. Network, because the hiring market is busted. Have a backup plan. Don’t spend too much time online. All of it is correct and none of it is wrong and you should do most of it.

One person told you to invest in a good spatula.

That person was closer to the truth than the rest of us, and this course exists to explain why.

Here is the situation you are enrolling into. For most of the history of this field, effort was legible. You could read a bug report and know approximately what it had cost someone. You could read a library and feel the hours in it. Nobody had to say I worked hard on this, because the artifact said it, and the saying was involuntary and therefore trustworthy. That signal has now failed. Not weakened. Failed. It went in about eighteen months and it is not coming back, and every piece of advice you were given this week was written by someone standing in the crater and not quite looking down.

This course is about what you do with the rest of your life given that.

Required equipment

A machine you control. Twenty gigabytes of free disk, minimum. The willingness to generate 1.8 million files in order to make something crash once.

That last item is not a metaphor and I will come back to it.

Unit 1: The proxy is not the thing

We begin with elevators, because elevators are the funniest thing that happened this week and the joke is at the expense of everyone reading this.

Somebody built a set of interactive simulators to answer a question you have had while waiting in a lobby: why is this taking so long. The finding that came out is that the fancy destination dispatch kiosks, the ones where you punch in your floor before you board so the system can plan optimally, make the building worse. The kiosk gives the scheduler more information, and the scheduler uses that information to commit, and committing costs it the ability to re-decide every five seconds. The dumb greedy algorithm that just keeps changing its mind beats the smart one that knows more.

Sit with that. More information, optimized harder, worse result.

And underneath it, a second joke: the thing the system optimizes is average wait time, but, as the piece observes, people don’t remember the average. They fixate on the times it took FOREVER. So the number being minimized is not the number being experienced, and the building can improve on paper for years while everyone who lives in it grows steadily angrier.

This is the general case and you should learn to smell it. Somebody spent a serious stretch of their life this month reserving CPU registers one at a time across nine kernels to find out whether the standard intuition about register pressure survives contact with a stopwatch. It half survives. Spill count predicts slowdown with a correlation of 0.55, and the cost of an individual spill varies about thirtyfold depending on which kernel you are in. Everyone already knew the direction. Nobody knew the number, because knowing the number required doing something nobody asked for.

You cannot get to either of those findings from a summary. You get there by overdoing it.

Unit 2: What a reproduction costs

Now the part that actually matters.

Two artifacts landed this week that should be studied together, because they are the same fact seen from opposite sides. In the first, somebody chasing an intermittent segfault in ripgrep generated roughly twenty gigabytes across 1.8 million files, ran the search until it died, pulled core dumps, and traced the crash down out of the tool and into the allocator underneath it, in a libc that is not the tool’s problem and not the reporter’s problem either. In the second, somebody got a 2.78 trillion parameter model running on a laptop by streaming 982 gigabytes of weights off an NVMe drive on demand, achieving a majestic half a token per second, and named the project WASTE, and documented the optimizations that failed alongside the ones that worked.

Both of those are enormous. Nobody lazy does either. And in both threads, high up, among the first things a reader meets, sits a version of this reads like a machine wrote it.

For the second one the answer is partly yes, and the author said so, and the contributor list says so. For the first, the analysis document attached to the reproduction split the room straight down the middle: one commenter called it soulless unreadable AI slop, another called it close to an ideal writeup, and a third asked, reasonably, whether anyone had reproduced it on a second machine. There is no way to settle this from the document. That is the entire point. Two people can read the same forty screens of careful-looking analysis and one sees a week of work and the other sees ninety seconds, and the text will not tell them who is right.

The sentence this course is built around arrived in that thread secondhand, quoted in from somewhere else: it doesn’t make sense for the reader to spend more energy than the writer spent on creating it. The commenter who brought it in endorsed it. Called it a great way to summarize cultural economics, said they had never been able to put words to the pattern before. And then, in the next line of the same comment, apparently without noticing: sometimes all I care about is that someone cared about.

Those two sentences are sitting three inches apart on the same screen, in the same comment, in the same head. That is the whole crisis, and notice that it is not two camps shouting. It is one person agreeing with the accounting rule and then, one line later, quietly admitting the rule does not describe what they actually want. The rule is correct. The confession is what the rule was always secretly for. We never wanted effort. We wanted evidence of care, and effort was just the only receipt anybody could print.

An objection, which is a good one

Let me concede the strongest thing you can say against everything above, because you will think of it anyway and I would rather you thought of it here.

Most useless work is bad. Not charmingly bad. Bad. The world is full of half-finished side projects, of blog posts nobody needed, of obsessive documentation attached to reasoning that does not hold. That somebody cared is not a quality claim and it never was. A man can spend four hundred hours on a thing and produce a thing that should not exist, and his four hundred hours obligate you to nothing at all. If your defense of a piece of work is that it took a while, you have already lost the argument about the work.

Fine. Agreed. Keep going anyway.

Because the question this course asks is not what is good. It is what can still be trusted, which is a smaller and sadder question and the only one currently on the table. And the answer turns out to be: whatever has no reason to exist.

Unit 3: Uselessness as credential

The most-read thing on the internet this week was elevator scheduling. Not far behind it, one hundred and fourteen photographs of bowls of ramen. Location, restaurant, year. No cookies, no tips, no JavaScript, just ramen. Several people asked, sincerely, why this was at the top of a technology site.

Here is why. Nobody automates 114 bowls of ramen. There is no pipeline that eats them for you. The photographs are the receipt for a person who went to a small room in London, and then a different room, and then a hundred and twelve more, and ate, and stood up, and took a picture, and did it again for years with no revenue attached. The uselessness is not incidental to the appeal. The uselessness is the appeal, because uselessness is now the only property of a piece of work that a machine has no incentive to counterfeit.

The elevator simulators are the same object. So is a nine-thousand-word article on the design of the period, the comma, and the difference between a bullet and an interpunct, written for the vanishingly small number of people who will ever need to care. So is a novelist publishing a note this week explaining, at length, that he does not use these tools in his writing, which is worth nothing commercially and is the most valuable thing he could have said. So is a second edition of a book on 64-bit assembly, a language nobody is hiring for, sold to people who want to know how the floor holds.

None of that is efficient. That is why it is legible.

Unit 4: Numbers that do not convert

The course’s warning, briefly, because you will meet it in your first job.

Also this week: a company that bills by usage removed the cost column from its usage page and its exported data, leaving customers with token counts. A token count is a number. It is not a price. It is adjacent to a price the way average wait time is adjacent to how long the elevator took. One customer put it precisely: hiding the number does not remove the anxiety, it removes your ability to check the anxiety, so you pad your estimate upward and quietly disengage.

Also this week: a lab announced ten advances in mathematics. Ten. A count of advances, which is the same species of number, offered because it converts to a headline even though it does not convert to anything you could verify by Friday.

You will spend a lot of your career being handed numbers like these. The skill this course is trying to give you is the reflex of asking what the number would have to be measured against in order to mean anything, and then noticing, most of the time, that nobody measured it, because measuring it would have been useless work and nobody was paying for that.

Grading

There is no grade. There cannot be one. Any rubric I could write, you could satisfy without doing the thing, which is the problem the course is about.

The one rule

Once a year, pick something nobody has asked you for and take it four steps past reasonable. Instrument it. Write down what surprised you, including what failed. Do not automate the middle, because the middle is the part that produces the number nobody has.

That is it. That is the whole degree I am qualified to grant.

Office hours

By appointment, though I want to say one more thing.

There was an essay this week about how badly people tolerate simply existing, citing the study where subjects in a room alone chose to give themselves electric shocks rather than sit for fifteen minutes with their own thoughts. I think about that alongside 114 bowls of ramen, and I do not think the ramen is an escape from sitting still. I think it is one of the few forms of sitting still we have left that a person can actually stand: attention held on one small thing, past the point of usefulness, until it turns into a record that you were here.

You are eighteen and you are afraid the field will be gone before you finish. It will not be gone. It will be unrecognizable, which is different and worse and better.

Buy the spatula. Use it ten thousand times.