Latest posts

Nate Meyvis posted "Where is all the new, good software?", asking a question that AI skeptics have been raising: if AI-assisted coding is as effective as its advocates claim, where is all the good, new software?¹ He offers nine candidate answers, some of which he admits deserve little weight. It is a list of conjectures, not a verdict.

At least three of his candidates hold up well from inside a working environment.

The first is that much of the software development process was never about writing code. Getting a plan past senior engineers, product managers, security review, and whoever signs off on deployment runs at the speed of bureaucracy², not the speed of generation. AI can draft a diff in minutes, depending on the model and the specification, but it can't shorten the approval chain, and in sane environments the approval is nine-tenths of the battle.

The second is that software is going private: domain experts are building their own tools. Teams are writing bespoke internal utilities that would never have justified a developer’s time before³. Executives are even assembling their own dashboards⁴. That software now exists and is doing real work, but it will never show up on a launch page or in a GitHub trending list, which is where most people look when they ask whether anything new has appeared, although I suppose some developers will be sure to tell us of their private success on social media somewhere.

A third is that we are producing more software rather than better software. The industry has always had the choice between improving what exists and building the next thing, and it has usually chosen the next thing. AI makes building the next thing easier, so that’s where our focus goes, and making software better - something AI turns out to be pretty decent at - doesn’t get the clicks as often.

Those three account for most of the apparent gap. What they don't address is the question, which is malformed in a few important ways.

"Good, new software" - sorry, "new, good software" - requires visibility. Visibility is a property of the channels people watch, not of the software. The channels are tuned for products, launches, and repositories, and right now those channels seem full of one specific type of thing: reimplementations. Someone rebuilt a note-taking app "like Obsidian." Someone rebuilt a task tracker. Someone rebuilt a tax tool. Someone reverse-engineered Total Annihilation with Open Annihilation and re-implemented it in Nanolathe, or Adobe's creative suite.⁵ These are everywhere because reimplementation has a specification. The old software is the spec. You point the tooling at a known shape and you get that shape back, faster and cheaper than you might have if you were doing it by hand.

Ideation has no specification. Nobody prompts their way to a product that has not been imagined yet, because writing the prompt is the act of imagining it.⁶ AI collapsed the cost of implementation, and left the cost of ideation where it was. The predictable result is a flood of known shapes and no corresponding flood of new ones.

That flood is the noise floor, and it is higher than it used to be. Closing the gaps in existing software is good work, and Meyvis acknowledges as much. But a higher floor does not mean the signal is absent. It means the signal has more to clear before anyone notices it. Something new and good may well be out there. It is just harder to hear over the sound of everything being rebuilt.

If the test of AI coding is whether your feed looks different, you are measuring the feed.