Last month Robbie Tilton released Compositor, a Photoshop alternative for Mac. One person. Native Swift, MIT licensed, thousands of GitHub stars within days of going public. The whole app is 12MB. Photoshop on his machine was 6,455MB.
The number that gets shared is the size difference. The one I keep thinking about is the headcount.
The Moat Was Never the Ideas
Building something in Photoshop's category used to require a company, and not because the ideas were hard. Layers, masks, curves, content-aware fill — none of that is a secret. It required a company because somebody had to write and maintain that much code, and one person couldn't. That was the moat. Not the design, not the algorithms. The sheer volume of unglamorous work standing between an idea and a thing people can actually install.
That constraint is the one that moved, and I want to be precise about how, because there's a lazy version of this argument going around. Tilton didn't prompt an app into existence. He planned the feature mapping with a model and then implemented and hand-tuned the details himself — the muscle memory especially, the ⌘-drag free distort, the export shortcut, all the things that make it feel like Photoshop rather than merely list the same features. That part is craft, and craft is still the whole product.
The Same Pattern, From the Other Side
I've been living in this from the other end. I built a lightweight self-hosted DevOps platform, Gisia, alone — Git hosting, CI/CD, merge requests, the usual — because the established ones wanted 12GB of RAM and a search cluster to do what I needed on a 2GB box.
The single hardest piece was the CI config parser: stages, jobs, rules, all the small behaviors people expect to just work. AI did not make that easy. I still had to hold the whole state machine in my head, still had to chase the edge cases myself, still had to decide what correct even meant. What it did was collapse the surrounding 80% — the scaffolding, the migrations, the sixth CRUD screen, the test setup, the "I know exactly how to do this and it'll cost me two days" work.
What Actually Changed
That's the real shift, and it's smaller and stranger than the headlines. The hard parts stayed exactly as hard. The tedious parts got cheap. And it turns out the tedious parts were most of what made these projects not worth starting — you'd sketch out an idea, estimate eighteen months of grinding, and quietly drop it. Now the estimate comes back at three months, and you start. Half of what looks like acceleration is just projects that previously died at the estimate stage.
So What Happens to the Incumbents?
Probably nothing dramatic, which is the boring honest answer. The big design suites, the big DevOps platforms, the big everything-in-one-tab SaaS — none of them are going anywhere. What they lose isn't customers, it's the status of being the default.
These products all grew the same way, by absorbing every request from every enterprise account until they could answer yes to any question on a procurement checklist, and the price of that is a tool where most users touch maybe 5% of it. Every year a few more of those users notice their 5% fits in something small, maintained by one person who answers GitHub issues personally.
The floor dropped. The ceiling didn't move. A solo tool is still narrower on purpose, still has a bus factor of one, still can't sell to anyone with a procurement process.
Where the Line Falls
My rule ended up simple: if I already know how to do it, I hand it off. If I'd have to think hard about it, I still think hard about it. That line has held for a year now and I keep expecting it to move.
If you want to see what came out the other side of it, Gisia is at gisia.dev and the code is on GitHub.