A friend made me aware of a reading list from A16Z containg recommendations for books, weighted towards science fiction since that’s mostly what people there read. Some of my books are listed. Since this is the season of Thanksgiving, I’ll start by saying that I genuinely appreciate the plug! However, I was taken aback by the statement highlighted in the screen grab below:

“…most of these books don’t have endings (they literally stop mid-sentence).”

I had to read this over a few times to believe that I was seeing it. If it didn’t include the word “literally” I’d assume some poetic license on the part of whoever, or whatever, wrote this. But even then it would be crazy wrong.

I’m not surprised or perturbed by the underlying sentiment. Some of my endings have been controversial for a long time. Tastes differ. Some readers would prefer more conclusive endings. Now, in some cases, such as Snow Crash, I simply can’t fathom why any reader could read the ending—a long action sequence in which the Big Bad is defeated, the two primary antagonists meet their maker and Y.T. is reconciled and reunited with her mother—as anything other than a proper wind-up to the story. In other cases, notably The Diamond Age and Seveneves, I can understand why readers who prefer a firm conclusion would end up being frustrated. It is simply not what I was trying to do in those books. So, for a long time, people have argued about some of my endings, and that’s fine.

In this case, though, we have a big company explicitly stating that several of my best-known books just stop mid-sentence, and putting in the word “literally” to eliminate any room for interpretive leeway.

This isn’t literary criticism, which consists of statements of opinion. This is a factual assertion that is (a) false, (b) easy to fact-check, and (c) casts my work ethic, and that of my editors, in an unflattering light.

It is interesting to speculate as to how such an assertion found its way onto A16Z’s website!

Hypothesis 1: it was written by a clanker

By far the most plausible explanation is that this verbiage was generated by an AI and then copy-pasted into the web page by a human who didn’t bother to fact-check it. This would explain the misspelling of my name and some peculiarities in the writing style. Of course, this kind of thing is happening all the time now in law, academia, journalism, and other fields, so it’s pretty unremarkable; it just caught my attention because it’s the first time it’s directly affected me.

The flow diagram looks like this:

That does a pretty good job of explaining how this all might have come about. So far so good. But it raises interesting questions about what happens next: the faulty quote from this seemingly authoritative source in turn gets ingested by the next generation of LLMs, and so on and so forth:

A hundred years from now, thanks to the workings of the Inhuman Centipede, I’m known as a deservedly obscure dadaist prose stylist who thought it was cool to stop his books mid-sentence.

Hypothesis 2: human with bad data

In this scenario, which seems more far-fetched, we have a sincere and honest human writer who is reporting what they believe to be true based on false information. It breaks down into two sub-hypotheses:

Sub-hypothesis A: Impecunious human with faulty bootleg PDF

There are bootleg copies of countless books circulating all over the Internet, and have been for decades1. Very often these are of poor quality. It could be that the person (or the AI) who wrote the above excerpt decided to save some money by downloading one of those, and got a bad copy that was cut off in mid-sentence.

Sub-hypothesis B: Non native English speaker with faulty translation

Even in the legit publishing industry, the quality of translations can be quite variable, and it’s difficult for authors to know whether a given translation was any good. I’ve seen translated editions of some of my books that look suspiciously short on page count. For all I know there might be translations of my books (legit or bootleg) that actually do stop mid-sentence!

…anyway, thanks for the plug

I genuinely am grateful to have been included on this list! But I had to say something about this astonishing howler embedded in the otherwise reasonable verbiage.

Oh, the Ending

Even the most cynical and Internet-savvy among us are somehow hard-wired to take anything we read on the Internet at face value. I’m as guilty as the next person. This has been a bad idea for a long time now, since bad actors have been swarming onto the Internet for decades. Now, though, it’s a bad idea for a whole new reason: content we read on the Internet might not have been written by a person with an intent to misinform, but rather by an LLM with no motives whatsoever, and no underlying model of reality that enables it to determine fact from fiction.

Gears are hard

Nearly two years ago, I decided to 3D print a mechanical calculator.

This project is extremely in character for me. I love old technology, and especially mixing old and new technology (like sending digital data via ham radio).

When you want to 3D print something, there are different levels of how sane you can be. The sanest option is to find a design on a site like Printables or Thingiverse that someone else made. Ideally, one that consists of as few parts as possible.

When I looked at the available 3D calculator designs, there were everything from simple ones with just a few gears, to modular ones based on historic designs (like Blaise Pascal’s calculator), all the way up to a 3x scale replica of the Curta Type I, the most advanced mass-produced mechanical calculator before electronic calculators took over the market.

There were two main features I wanted: a modular design, and a simple design. Unfortunately, the only modular design was modeled after Pascal’s calculator, and not very simple at all.

So, I decided to take the less sane option and design one from scratch. I did not realize how big of a challenge I was jumping into.

My first design was straightforward:

Design #1

The two gears each had 10 teeth. One gear (blue above) would let you read out a number from a dial. The other (orange) would have an extra nub on a single tooth, which would serve to carry a sum into the next decimal place. The posts the gears sat on would be tilted, so only that extra tooth would make contact with the next gear.

I was having trouble getting the correct angles, when I realized there was a more elegant solution. By flipping every second set of gears upside down, they’d be lined up naturally.

When I printed the first prototype, I discovered that the gears I had made in TinkerCAD were not interlocking and turning correctly. There was way too much of a gap between the teeth:

The advantage of 3D printing is that it lets you fail quickly. I realized I needed to know more about how gears were designed.

It turns out that there is an optimal gear tooth shape that was discovered in the 18^th century, potentially by Leonard Euler. They’re called “involute gears”, and they transmit force along a straight line between gears of any number of teeth.

Involute gear teeth

There are plenty of CAD libraries to generate custom involute gears. I found a good one, and designed a ten-toothed gear. After a few iterations to test the fit, I had my first full draft print:

Also note the lightweight base design

But this version still had a big problem I hadn’t foreseen. I had presumed that when one tooth of a ten-toothed gear fully passed through a gap in a second ten-toothed gear, both gears would turn one tenth of the way. As it turns out, this is completely wrong.

Notice that the gears have turned through two teeth, not just one

So it was back to the drawing board. Eventually, I found the appropriate bit of trigonometry and determined that I could use a gear with 30 teeth of a specific size, with the carry tooth turning its partner gear by exactly 3 teeth. To make the design easier, I also decided to have the carry tooth stick out from the rest of the gear. This meant the carry gear needed to be printed with supports, but the support would be small and easy to break off after printing.

This design kinda worked. The main issue left was that the gears kept slipping upward on the posts. My first solution was to use little c-clips on top, which worked okay, but wore out over time. Still, I was satisfied enough to declare the project finished.

About a year and a half later, I decided to come back to the project and make an improved version with a better fastener. After trying and failing to build a system that screwed together, I settled on a version held together with pegs that rotate into place.  This version has held up for the last couple of months, so I’m ready to share it with others.

3D model of final version.

Side view. I will add a photo when I return home.

You can find the model and print instructions on my Printables page here.

Coming Soon: Fiction: A Leap of Logic

Statistics are interesting because they are often used to sway arguments, or entice people. This discussion looks at the impact of normal statistics without an underlying context. Take the example of a cookware company that uses statistics in its advertisements (very few other companies provide statistics of any sort). Some of the statistics are shown below.

• 103 Michelin-recognized restaurants use the cookware.
• The cookware is used in over 2,000 global restaurants.
• 90,000 pans in service.

These numbers may seem significant, but in the context they are not. For example in the first statement, “100 Michelin-recognized restaurants use the cookware”. Worldwide there are approximately 18,713 Michelin-recognized (not starred) restaurants. So 103/18713 = 0.55%. One would argue that this really isn’t a significant number. If we were to look at only Michelin-starred restaurants, some 3700 worldwide, the number would increase to 103/3700 = 2.78%. The number looks a lot better if we only consider North America, where the number is 1941, so 103/1941 = 5.3% which is 10 times as significant, but still not that much.

For the second statistic, that it is “used by over 2,000 global restaurants”, let’s look at a number of scenarios. This is hard to quantify, because how do you define a “global restaurant”. It has to be different from a run-of-the-mill restaurant, because there could be 15-25 million of those. A better way to look at is to consider the number of professional chefs who run restaurants. Let’s consider just Canada first, where there are 62,200 chefs (2023). This would give us 2000/62200 = 3.2%. If we look at the USA, this number climbs to 286,000, and 0.7%. Neither is earth-shattering, and if we include worldwide restaurants the number just plummets.

The last statistic, 90,000 pans, is quite meaningless. It’s just a reflect of how many pans have been sold, and can’t really be compared to other companies because they don’t publish their data. Some vague figures suggest the Le Creuset foundry in France produces 25,000 items a day. Restaurants use either stainless steel or aluminum pans, and to put the number into context, there are circa 97,000 restaurants in Canada alone (of course the 90,000 pans were not suggested to come from restaurants exclusively).

There is nothing to say that the numbers used in the advertisements are whatsoever incorrect, but they do provide a level of ambiguity because they aren’t given in any frame of reference. Numbers only mean something if put into context.

When you're leading a team of five or 10 people, feedback is pretty easy. It's not even really "feedback”: you’re just talking. You may have hired everyone yourself. You might sit near them (or at least sit near them virtually). Maybe you have lunch with them regularly. You know their kids' names, their coffee preferences, and what they're reading. So when someone has a concern about the direction you're taking things, they just... tell you.

You trust them. They trust you. It's just friends talking. You know where they're coming from.

At twenty people, things begin to shift a little. You’re probably starting to build up a second layer of leadership and there are multiple teams under you, but you're still fairly close to everyone. The relationships are there, they just may be a bit weaker than before. When someone has a pointed question about your strategy, you probably mostly know their story, their perspective, and what motivates them. The context is fuzzy, but it’s still there.

Then you hit 100#

Somewhere around 100 people, the ground shifts underneath you, as you realize you don’t know everyone anymore. You just can't. There aren't enough hours in the day, and honestly, there aren't enough slots in your brain.

Suddenly you have people whose names you don’t recognize offering very sharp commentary about your “leadership.” They’re talking about you but they don’t know you. There’s no shared history, no accumulated trust, no sense of “we’ve been in the trenches together.” Your brain has no context for processing all these voices.

Who are these people? Why are they yelling at me? Are they generally reasonable, or do they complain about everything? Do they understand the constraints we're under? Do they have the full picture?

Without an existing relationship, it feels like an attack, and your natural human response is to dismiss or deflect the attack. Or worse, to get defensive. Attacks trigger our most primal instincts: fight or flight.

This is the point where a lot of leaders start to struggle. They still want to be open to feedback—they really do—but they're also drowning. They start trusting their intuition about what they should pay attention to and what they should ignore. Sometimes that intuition is right. Sometimes it's just... self-selected, stripped of context, pattern matching against existing biases and relationships.

On top of that, each extra layer of management, each extra level to the top has separated you, and now you’re just not like them anymore. Their struggles are not your struggles anymore.

At 200, it's a deluge#

By the time you reach 200 people or more, feedback isn't an actionable signal anymore. At that size, feedback stops being signal being noise. A big, echoing amphitheater of opinions, each louder than the last, each written in the tone of someone who is absolutely certain they understand the whole system (they don’t), the whole context (they don’t), and your motives (they definitely don’t).

And all those kudos you used to hear? Those dry up. When you had a close relationship with everyone, kudos came naturally. You were just talking. But now folks just expect you to lead, and if they’re happy with your leadership they’re probably mostly quiet about it. They're doing their jobs, trusting you, assuming things are generally fine.

The people who are unhappy? They're loud. And there are a lot of them.

From where you sit, it feels like everybody's mad about everything all the time. And maybe they are! Or maybe it's just selection bias combined with the natural amplification that happens when people with similar grievances find each other. You don't know if this is a real crisis or just three loud people who found each other in a Slack channel. You just can’t tell anymore.

Because feedback doesn’t scale. Humans scale poorly. Your nervous system definitely doesn’t scale.

Why this happens#

Feedback doesn't scale because relationships don’t scale. With five people, you have some personal interaction with everyone on the team. At twenty, you interact with some, but not all. At 100 you still have personal relationships with 10 or 15 people, so there are a lot of gaps. At 200, your personal relationships are a tiny slice of the overall pie.

Making matters worse, as the din gets louder and louder, channels for processing all that feedback get smaller and smaller. Where you once had an open-door policy, now you have “office hours.” Sometimes. When we’re not too busy.

Where once All-Hands meetings had open questions, now you’re forced to take the questions ahead of time. Or not at all.

Even your Slack usage dwindles, because half the time you say anything, someone’s upset with it.

We tell ourselves we're "staying close to the ground" and "maintaining our culture,” But we're not. We can't. Because the fundamental math doesn't work. The sheer volume of feedback we’re getting absolutely overwhelms our ability to process it.

So what do you do about it?#

First, you have to admit the problem exists. Stop pretending you can maintain personal relationships with 200 people. You can't. Nobody can. Once you accept this, you can start building systems and processes that work with this reality instead of bumping against it. You have to filter, sort, and collate the feedback coming in, and you need to do it at a scale larger than your own capacity.

When you can’t rely on “just talk to people,” you need systems that distinguish between:

• legitimate issues
• noise
• venting
• misunderstandings
• and “this person is projecting a whole other problem onto leadership”

That means: structured listening, actual intake processes, and ways to synthesize themes instead of reacting to every single spike.

Build proxy relationships. You can't know 200 people, but you can know 10 people who each know 10 people. You should already have strong, trusting relationships with your leadership team, and then set the expectation that they have strong relationships with their own teams, and explicitly ask what’s on people’s minds. When feedback comes up through this chain, it comes with context. Pay attention.

At small scale, trust is direct: I know you. You know me. At larger scale, trust must be delegated: I trust the leaders who are closer to the work than I am. If you don’t intentionally empower those leaders to absorb and contextualize feedback, you’ll drown. They’re the ones who can say: "I know who said that, why they said it, and here’s what’s actually going on."

Build structured channels for feedback. For example, you can set up working groups to dive into thorny problems. The people closest to the problem understand it better than you do, and they can turn a flood of complaints into something you can actually act on. Or consider starting an "employee steering committee" for the sole purpose of collecting feedback and turning it into proposals. You’re essentially deputizing people who care deeply to listen for you, and then manage the feedback din.

Remember that every angry message is still a person. When someone you know well gives you feedback, you might not like it, but you’re likely to say "Oof. Okay. Let’s talk." At scale, you need to find ways to respond with humanity — even when the feedback you received lacks it.

Close the feedback loop. Let people know when you’re acting on their feedback, and if you’re not going to act on it, let them know that you at least heard it. Nobody wants to feel unheard.

In fact, you'll probably think — if you haven't done it already — that you should have an anonymous comment system to capture feedback. Don't. It's a trap. Anonymous feedback is the most contextless feedback you'll get, which makes it the least actionable. And it inevitably turns out to be contradictory or lacking key information, all those folks feel even more unheard and unhappy than before.

Finally, accept that you're going to get it wrong sometimes, and own that. You're going to ignore feedback that turns out to be important. You're going to overreact to feedback that turns out to be noise. When you make a misstep, be transparent about how you're correcting it.

The uncomfortable truth#

Past a certain size, you have to make peace with the fact that a lot of people in your org are going to be frustrated with you, and you're going to have no idea why, and you may not going to be able to fix it.

Not because you're a bad leader. Not because you don't care. But because feedback doesn't scale, relationships don't scale, and the alternative—trying to maintain authentic personal connections with hundreds of people—is a recipe for burnout and failure.

This is genuinely hard to accept, especially if you came up through the early days when you did know everyone. That version of leadership was real, and it worked, and it probably felt really good. But it doesn't work anymore, and pretending it does just makes things worse.

Note: The photo is of a large crowd gathering for a union meeting during the 1933 New York Dressmakers Strike. That's scaling feedback.