Spooky Halloween costume idea this year: dress as the current job market 🎃

Here’s a quick thread on how I automated a thing with Python. Every month, a junior editor on my team was taking structured data, finding book covers, and manually turning all of it into a presentation. That looked less like a workflow and more like code waiting to be written.

AI won't replace you, but your boss thinks it already did.

A slightly counterintuitive publishing example: If a reprint earns 450k when demand holds and loses 250k when it doesn’t, the break-even probability of success is only about 36%. Sometimes, a rational decision requires much less certainty than we instinctively want.

A 70% probability of success does not automatically make something a good decision. If success earns you 100k and failure costs you 1m, even an 80% chance of success has negative expected value. Probability tells you how likely an outcome is. It doesn’t tell you whether the bet is worth taking.

Several positive metrics do not necessarily mean several independent pieces of evidence. More social mentions, higher sales, and a bestseller ranking may look like three signals. But if all three were caused by the same marketing campaign, you may really have only one.

Historical data isn’t only useful for explaining the past. Comparable sales, cohort performance, launch success rates, product survival—all of these can become priors for the next decision. Past data is also a way of forming better expectations about the future.

One question I find useful when evaluating a project: “What usually happens in similar cases when this isn’t my project and I’m not the one making the decision?” It’s a simple way to put some distance between evidence and enthusiasm.

A vivid signal is not necessarily a useful one. Selling out a first print run in two weeks sounds impressive. But if the run was tiny, it may tell you very little about long-term demand. Boring repeat orders from bookstores over several weeks may be much more informative.

Suppose only 20% of comparable new books become successful. Then your new release sells 1,000+ copies in its first week—something 80% of future hits do, but only 10% of unsuccessful books do. You just got a strong reason to update. Your estimate of success jumps from 20% to about 67%.

A good manager doesn’t need to predict the future correctly on the first try. They need a reasonable starting belief, an idea of what evidence should change it, and the willingness to update when that evidence arrives. The goal is not to always be right. It’s to become progressively less wrong.

Managers rarely make decisions with complete information. So it’s not about being sure but rather how sure and what could change your mind. That, in a very practical sense, is what Bayes’ theorem is about.

Had the most hectic couple of weeks. But I’ve started “Everything Is Predictable” by Tom Chivers about Bayes’ theorem. I’ve been having as much fun reading it as I did when I first learnt about the theorem per se.

I finally got to watch the new Insidious movie. Honestly, I think they should’ve wrapped things up after Chapter 3.

Simpsons Kids React: Stop, He's Already Dead!

Alt: Simpsons Kids React: Stop, He's Already Dead!

static.klipy.com

One thing I learned from building deck-builder is that architecture starts to matter exactly when a script becomes useful. The first version only had to work. The next one had to survive new layouts, missing images, bad input, tests, and other presentation types.

I’m not a developer. I manage products and teams in publishing. But being able to write some code changes how I look at processes. You start noticing that some “workflows” are nothing but deterministic transformations waiting to become functions.

deck-builder is a small project. It reads Excel and produces PowerPoint. But small internal tools are cool precisely because the value is not proportional to the amount of code. A couple hundred lines can remove a recurring task from someone’s workday.

Most of my #PythonProgramming projects don’t start with me wanting to build software. They start with a recurring task at work. deck-builder started with a monthly PowerPoint that someone had to assemble manually from structured data. That was the signal.

Being able to write some code gives managers another option between: “change the process” and “assign the work to someone else.” Sometimes, the better solution is to remove the task entirely.

So someone regularly takes structured data, applies the same rules, and produces another structured output. At that point, the question may not be “how do we do this faster?” It may be “why is a person doing this at all?”

A surprisingly useful question at work: is this actually a process, or just a deterministic transformation we happen to perform manually?

A lower average print run doesn't tell us whether publishers reduced quantities for most books or simply published fewer high-volume editions. Before applying an industry average to my own list, I’d check which parts of the distribution drove the change.

Five thousand copies of an edition could mean one printing order or several smaller ones. Total output is the same; costs, payment schedules and cash tied up in stock may differ. To understand printing strategy, I’d look at individual orders as well as annual totals.

An edition falls from 101,000 copies to 99,000. The bracket above 100,000 loses all 101,000; the next bracket gains 99,000. Actual output fell by just 2,000. When comparing distributions, the boundaries may stay fixed while the editions inside them change.

The shelf you put a book on influences which titles publishers compare, which books become competitors, how series are built, and what gets commissioned next. A broad taxonomy eventually produces broad, and often weak, product decisions.

The problem with “business books” is not that someone is putting books on the wrong shelf. It's that the shelf has become so broad that it contains books with different subjects, audiences, competitors, life cycles, and reasons for purchase. That matters if you are trying to analyse the market.

I recently looked at a Russian top-10 ranking of “business books.” It included cognitive psychology, personal finance, philosophy, motivational self-help, and books on influence. If almost none of these books are really about running a business, what exactly does “business books” mean anymore?

“Business books” may no longer be a useful genre label. The category now contains management, personal finance, behavioral science, popular psychology, and motivational self-help. At some point, a category that broad stops explaining what the books actually are.

In publishing, I’ve found there are much more informative KPIs than total catalog revenue. A growing catalog can hide deteriorating title economics surprisingly well.

A smaller order from a bookseller doesn't necessarily mean readers are buying fewer books. The bookseller may be reducing existing stock. A larger order may be preparation for a promotion. Publisher shipments and reader purchases tell different parts of the story.