Somewhere right now, an engineering team is weighing a move off a language built for resilience — one famous for keeping thousands of processes alive while parts of the system fail — toward a more ordinary one they like less. Not because the new language is better. Because their AI coding tools are better at it. The frontier models were trained on mountains of the popular language and only hills of the rare one, so the assistant writes cleaner code in the thing that was never the right fit. The team is heavily invested in the old stack. They’re considering the switch anyway.
Sit with that for a second, because it inverts the usual story we tell about tools. We assume the tool serves the work. You decide what to build, then reach for whatever helps you build it. But when the tool is an AI trained on whatever the crowd already wrote, the arrow quietly reverses. The work starts bending toward what the tool already knows. Popularity becomes a gravitational field. The model is most fluent where the most people have already been, so it pulls everyone further in that direction — and the roads less traveled get a little more abandoned each quarter.
That’s the thread running under three stories that look unrelated this week. They’re the same story.
The thing that absorbs its neighbors
The first: the big platforms are turning their assistants loose on your to-do list. Not the chatbot you visit, but an agent that lives one layer back and quietly handles the errands — the booking, the reminding, the small administrative friction of being a person. The framing in the coverage is convenience. The truer read is territory. Your to-do list is a map of every decision you haven’t made yet. Whoever holds that map holds the moment before each choice, which is the only moment where a choice can still be steered.
The second: at the big auto show in China, the headline isn’t the cars. It’s the robots standing next to them. The carmakers figured out that the hard part was never the chassis — it was the perception, the motors, the decision loop that lets a machine move through a messy world without hitting things. Once you’ve built that for a car, the car is just the first body you put it in. The factory arm is the second. The warehouse mover is the third. The vehicle was never the product. It was the proof of concept for everything else that needs to sense and act.
The third: a leading AI lab’s chief executive says the traditional software business is about to pivot, because writing code is getting cheap. For thirty years the model was simple — build a tool once, rent it to a thousand companies who couldn’t build it themselves. That moat was the difficulty. When difficulty drops toward zero, the moat drains. Why rent the standard tool when your own agent can spin up the custom one over a long weekend?
Three industries, one motion. The agent doesn’t sit politely beside the work. It reaches out and absorbs whatever is adjacent — your tasks, your hardware, your software margins, your choice of language. It eats sideways.
The gap between the take and the truth
The consensus take on all of this is productivity. Faster code, easier errands, smarter machines. All true, and all beside the point. Productivity is what you measure when the categories stay fixed and you just do more inside them. What’s actually happening is that the categories are moving.
The to-do list isn’t getting faster — it’s becoming a place where decisions get made on your behalf before you notice there was a decision. The carmaker isn’t building better cars — it’s quietly becoming a robotics company that happens to sell some cars. The software vendor isn’t shipping features faster — it’s watching the reason customers needed it dissolve. And the engineering team isn’t writing better code — it’s letting the training data decide which language deserves to survive.
None of this is loss, exactly. The resilient language might genuinely deserve a smaller seat; plenty of old moats deserved to drain. The point isn’t that the change is bad. The point is that we keep describing it with the small word — helpful — when the honest word is generative. These tools don’t speed up the path you were already on. They redraw the map of which paths exist.
What it costs to be understood
Here’s the part worth holding onto. A tool that’s fluent in the popular thing makes the popular thing more popular, which makes the tool more fluent still. That loop is wonderful right up until you need the thing the crowd never wrote down. The resilient language existed because someone needed exactly what it offered and nothing else would do. If the next generation of builders reaches past it every time — because the assistant shrugs at it — we don’t lose the language. We lose the reason it was invented, which is harder to get back.
The agents aren’t coming for your to-do list. They’re coming for the part of you that decided what went on it. The convenient future is the one where you never have to choose the harder, stranger, better-fitting thing — because the tool was only ever fluent in the easy one, and fluency, repeated long enough, starts to feel like truth.
Leave a Reply