I’d prefer not to think about AI. I’ve lived for decades taking pleasure in (and being paid for) writing essays and code. Both those disciplines are obviously in AI’s cross-hairs. I’m old but unready to step away from caring about what’s going on in my world, then writing about it. So herewith what I hope are fresh takes, rhetoric that I haven’t seen elsewhere.
I’m not calling them “conclusions” and nor should anyone in this liminal time when nobody can claim to see (abstractly) the Big Picture, nor can anyone credibly predict (concretely) what the Hot New Stuff will cost after we escape the shadow of the rapidly-inflating galactic bubble of fear and greed.
So I hope these offerings, mostly in the software-development space, prove useful in seeing a little further into our future’s oppressive opacity.
Singleton apps · I’m going to suggest this as a label for GenAI-generated apps produced by a technically-skilled person to meet a personal need. No team, no product manager or any other kind of manager, just someone who needs their computer or phone to help them do something and generates an app to provide it.
Let’s start with Nelson Minar, an extremely proficient and now mostly-retired developer. Check out Six weeks of Claude Code (Pangram: 100% human). I have no idea what most of these things even are, partly because upon retiring I swore I’d never manage another server. But they sound useful. And Nelson, whom I know and like and have occasionally hung out with, in my experience doesn’t bullshit. Also he offers a gentle and reflective afterthought in Coding craft vs. AI.
Let’s turn our attention to Craig Mod, in particular A Swarm of Blood Robots (Pangram: 100% human). It, Dear Reader, contains fulsome AI praise which is apt to annoy some of you, but I’d recommend reading it anyhow because Craig is an awfully good writer and the stories are compelling. You might want to go deeper and read Software Bonkers (Pangram: 100% human), which is about the singleton he built to help him understand his business and pay his taxes. He’s an American writer and photographer who lives in Japan and makes his living publishing beautiful illustrated books and selling them all over the world. My gut clenches at the thought of what his tax machinations must be.
But once again, every one of the pretty-impressive list of things he’s had Claude do for him is a singleton: One author, no bosses, no team. Suppose we believe that Claude is ideal for that sort of thing. The follow-up question is: What proportion of the world’s useful software is singletons? Even if we assume it expands, it’s starting from a very small base and to me has a niche-y smell.
But you know, when Craig gets excited about something I take that as significant. Back in 2010 or so he wrote The GF1 is a near perfect travel camera. I think it’s one of the most influential camera reviews ever written by anyone. The whole notion of a compact mirrorless camera with a smaller-than-film sensor was controversial but Craig changed a whole lot of people’s minds by taking one out on a Himalayan hike and coming home with a lot of beautiful pictures wrapped in a nice piece of writing. I trust that man’s instincts.
The singleton audience · You’ll notice that my definition of “singleton” was centered on the individual unaccompanied free-agent author. What about the size of the audience? Nelson Minar’s constructs are just for himself. Craig’s also include “The Good Place”, a small, non-federated, nicely-designed social network reserved for his paying fans. He doesn’t say how many people it is, but anyhow, the number is greater than one.
And then, there’s Nate Silver’s FLIPR, an elaborate model which predicted the outcome of the 2026 US midterm election. It’s based mostly on polling data, with a bunch of other relevant inputs. Like like every other pre-midterm prognostication, it underestimated the force of the Democratic wave; for example, it had the chances of a 56-blue-seat Senate at 8.8%. Nate had ChatGPT build it for him (in Python). You had to be a subscriber to get his deep-dives into how it worked, but the basic predictions were public and were very widely read. I have no information about testing discipline or other software-engineering issues. The result is impressive, though. But, once again, not typical of what the mainstream of the software profession works on.
And you know, I’ve got a couple of ideas for singletons of my own. But I can’t get past the negative externalities so I won’t be building any of them. For now, that is. But if the post-bubble rubble is ethically habitable, I’d love to give it a try.
A question occurs to me: what is the relationship between singleton apps and what, throughout my working life, has been referred to as “software engineering”?
Enough singletons (though I like that label and hope it catches on.)
Contras · I don’t, generally speaking, dislike the LLM technology in and of itself. But I loathe the business goals and ethical obliviousness of the people who are leading the charge. So it’s not surprising that my antennae tingle whenever I hear new and interesting rhetoric that supports my position.
Cory Doctorow, in The actual epistemic crisis (Pangram: 100% human) points out that we inhabit, and always have, a world that contains lots of liars. We’ve erected defenses and countermeasures in the form of courts, democracies, and regulatory agencies. They’re not perfect, but they’re not nothing. Cory argues convincingly that AI strengthens the liars and weakens the defenses.
There’s this guy Kyle Kingsbury, goes by Aphyr, who’s a major expert on database concurrency, serializability, and integrity. He’s got this chunk of software called Jepsen that he uses to stress databases out, and you can hire him to stress your database out and report back on how it stood up.
Recently he was hired to stress out a new database named ParticleDB, which was substantially built by and for AI agents. The report is polite, professional, and savage (Pangram: 100% human). Once again, I’m moved to ask what the relationship is between GenAI and “software engineering”.
One of the big open questions in AI is how much better it will get than it is today, and how soon. At the end of that road is the notion of RSI, “Recursive Self Improvement”, where this frontier model builds the next, and then that builds the next, repeat until you’ve succeeded in constructing God. In Where’s the “intelligence explosion”? (Pangram: 100% human), Ramez Naam offers a lengthy, data-backed argument that improvement is already asymptotic and will inevitably stay that way and we won’t be building any deities any time soon.
Micah Lee offers a fresh take in Just say no to OpenAI and Anthropic (Pangram: 100% human). He starts with antifascism because, quoting, “Tech oligarchs make up an important component of the end-times-fascist coalition.” Having said that, this isn’t a polemic against GenAI, it’s an argument for open-weight models and doing all the AI work on hardware you own or control, as opposed to paying the oligarchs per-token.
His recommendations for how to go about doing this are specific and seem well-founded, and I hope he finds a way to keep updating them as the tech landscape shifts. It’s obvious that there’s this huge geek army whose hearts will swell at the thought of having a found a good reason to buy a $10K Mac Studio to sit on their desktop and host their daily work. I mean, even I find that idea attractive…
I bet it’d run Lightroom Really Fucking Fast.