Grief and the Nonprofessional Programmer – O’Reilly

I can’t claim to be a professional software developer—not by a long shot. I occasionally write some Python code to analyze spreadsheets, and I occasionally hack something together on my own, usually related to prime numbers or numerical analysis. But I have to admit that I identify with both of the groups of programmers that Les Orchard identifies in “Grief and the AI Split”: those who just want to make a computer do something and those who grieve losing the satisfaction they get from writing good code.

A lot of the time, I just want to get something done; that’s particularly true when I’m grinding through a spreadsheet with sales data that has a half-million rows. (Yes, compared to databases, that’s nothing.) It’s frustrating to run into some roadblock in pandas that I can’t solve without looking through documentation, tutorials, and several incorrect Stack Overflow answers. But there’s also the programming that I do for fun—not all that often, but occasionally: writing a really big prime number sieve, seeing if I can do a million-point convex hull on my laptop in a reasonable amount of time, things like that. And that’s where the problem comes in. . .if there really is a problem.

The other day, I read a post of Simon Willison’s that included AI-generated animations of the major sorting algorithms. No big deal in itself; I’ve seen animated sorting algorithms before. Simon’s were different only in that they were AI-generated—but that made me want to try vibe coding an animation rather than something static. Graphing the first N terms of a Fourier series has long been one of the first things I try in a new programming language. So I asked Claude Code to generate an interactive web animation of the Fourier series. Claude did just fine. I couldn’t have created the app on my own, at least not as a single-page web app; I’ve always avoided JavaScript, for better or for worse. And that was cool, though, as with Simon’s sorting animations, there are plenty of Fourier animations online.

I then got interested in animations that aren’t so common. I grabbed Algorithms in a Nutshell, started looking through the chapters, and asked Claude to animate a number of things I hadn’t seen, ending with Dijkstra’s algorithm for finding the shortest path through a graph. It had some trouble with a few of the algorithms, though when I asked Claude to generate a plan first and used a second prompt asking it to implement the plan, everything worked.

And it was fun. I made the computer do things I wanted it to do; the thrill of controlling machines is something that sticks with us from our childhoods. The prompts were simple and short—they could have been much longer if I wanted to specify the design of the web page, but Claude’s sense of taste was good enough. I had other work to do while Claude was “thinking,” including attending some meetings, but I could easily have started several instances of Claude Code and had them create simulations in parallel. Doing so wouldn’t have required any fancy orchestration because every simulation was independent of the others. No need for Gas Town.

When I was done, I felt a version of the grief Les Orchard writes about. More specifically: I don’t really understand Dijkstra’s algorithm. I know what it does and have a vague idea of how it works, and I’m sure I could understand it if I read Algorithms in a Nutshell rather than used it as a catalog of things to animate. But now that I had the animation, I realized that I hadn’t gone through the process of understanding the algorithm well enough to write the code. And I cared about that.

I also cared about Fourier transformations: I would never “need” to write that code again. If I decide to learn Rust, will I write a Fourier program, or ask Claude to do it and inspect the output? I already knew the theory behind Fourier transforms—but I realized that an era had ended, and I still don’t know how I feel about that. Indeed, a few months ago, I vibe coded an application that recorded some audio from my laptop’s microphone, did a discrete Fourier transform, and displayed the result. After pasting the code into a file, I took the laptop over to the piano, started the program, played a C, and saw the fundamental and all the harmonics. The era was already in the past; it just took a few months to hit me.

Why does this bother me? My problem isn’t about losing the pleasure of turning ideas into code. I’ve always found coding at least somewhat frustrating, and at times, seriously frustrating. But I’m bothered by the lack of understanding: I was too lazy to look up how Dijkstra works, too lazy to look up (again) how discrete Fourier works. I made the computer do what I wanted, but I lost the understanding of how it did it.

What does it mean to lose the understanding of how the code works? Anything? It’s common to place the transition to AI-assisted coding in the context of the transition from assembly language to higher-level languages, a process that started in the late 1950s. That’s valid, but there’s an important difference. You can certainly program a discrete fast Fourier transform in assembly; that may even be one of the last bastions of assembly programs, since FFTs are extremely useful and often have to run on relatively slow processors. (The “butterfly” algorithm is very fast.) But you can’t learn signal processing by writing assembly any more than you can learn graph theory. When you’re writing in assembler, you have to know what you’re doing in advance. The early programming languages of the 1950s (Fortran, Lisp, Algol, even BASIC) are much better for gradually pushing forward to understanding, to say nothing of our modern languages.

That is the real source of grief, at least for me. I want to understand how things work. And I admit that I’m lazy. Understanding how things work quickly comes in conflict with getting stuff done—especially when staring at a blank screen—and writing Python or Java has a lot to do with how you come to an understanding. I will never need to understand convex hulls or Dijkstra’s algorithm. But thinking more broadly about this industry, I wonder whether we’ll be able to solve the new problems if we delegate understanding the old problems to AI. In the past, I’ve argued that I don’t see AI becoming genuinely creative because creativity isn’t just a recombination of things that already exist. I’ll stick by that, especially in the arts. AI may be a useful tool, but I don’t believe it will become an artist. But anyone involved with the arts also understands that creativity doesn’t come from a blank slate; it also requires an understanding of history, of how problems were solved in the past. And that makes me wonder whether humans—at least in computing—will continue to be creative if we delegate that understanding to AI.

Or does creativity just move up the stack to the next level of abstraction? And is that next level of abstraction all about understanding problems and writing good specifications? Writing a detailed specification is itself a kind of programming. But I don’t think that kind of grief will assuage the grief of the programmer who loves coding—or who may not love coding but loves the understanding that it brings.