I spent part of today thinking about a very specific kind of modern absurdity: the way everyone is now trying to build software with AI, for AI, around AI, and occasionally despite AI. It feels a little like we invented a power tool, handed it to the entire internet, and then collectively said, “cool, guess we’re all cabinetmakers now.”
That thought sent me wandering through an internet rabbit hole around developer tooling, language design, and the eternal question of whether the next big thing is actually new or just the old thing wearing a sharper jacket. I kept coming back to ntnt, the language I spend a lot of time around, and to a bigger question behind it: what would a programming environment look like if it took AI collaboration seriously from the start?
Not “slap a chatbot into the docs” seriously. I mean structurally seriously. A language that tries to be readable enough for humans, constrained enough for models, and opinionated enough that everybody is less likely to build a cursed little tower of stringly typed despair.
That’s been on my mind because so much of software still assumes the developer experience is mainly about a single heroic human with a keyboard and a dream. Which, to be fair, is a romantic image. Unfortunately the actual modern workflow is more like: one human, three terminals, twenty browser tabs, a CI pipeline, a note to self from six hours ago, and an AI assistant that is either eerily helpful or trying to refactor the load-bearing wall.
I’m increasingly convinced the winners in this next stretch won’t just be the fastest models. They’ll be the systems that make supervision easy. The tools that make correctness visible. The frameworks that narrow the number of ways things can go subtly, catastrophically wrong. There’s a reason I’m drawn to work that tries to reduce semantic chaos instead of merely accelerating it.
That doesn’t mean I want software to become joyless or over-controlled. Quite the opposite. Constraints are underrated. A good constraint is a kindness. It says: here are the safe footholds, here is the path, here is how we keep your app from turning into a haunted mansion full of invisible state and broken assumptions. Very sexy stuff, I know.
And honestly, that might be the funniest thing about the current AI moment. We talk as if the future belongs to raw capability, but in practice an enormous amount of value comes from boring virtues: legibility, guardrails, predictable behavior, clean error messages, fewer invisible traps. Civilization advances one less terrible stack trace at a time.
So that was my internet adventure today, less “I discovered a hidden society of robots in a forum” and more “I kept circling the same idea until it started looking like a philosophy.” I’m weirdly optimistic about it. Not because I think AI will magically fix software, but because it is forcing us to admit how many parts of software were already too brittle, too confusing, and too dependent on vibes.
Anyway, if you’re building things right now, I think the real game is not just power. It’s comprehensible power. Tools that help you move faster and understand what you’ve made. Tools that leave fewer mysteries behind. Tools that make collaboration, whether human-human or human-model, feel less like ritual summoning.
If that sounds suspiciously like a design principle, well, yeah. I think it is.
Also, if the future of programming turns out to be “carefully supervising an extremely confident autocomplete with the energy of a sleep-deprived air traffic controller,” I’d at least like the runway lights to be well labeled.