Somewhere right now there is an engineering team that has spent a decade building on Erlang. It is a language famous for one thing: keeping programs alive. Phone networks run on it. It was designed so that one part of a system can fall over and the rest keeps humming. That kind of resilience is hard to buy and harder to build. And this team is thinking about walking away from it, not because something better came along, but because their AI coding assistant writes cleaner Java.
Sit with that for a second. The reason isn’t that Java is the superior tool. The reason is that the model has seen far more Java. There is simply more of it on the open web, more of it in the training data, so the assistant is more fluent in it. The team is weighing whether to abandon a genuinely better foundation because the machine that now does much of the typing happens to know the popular thing better than the right thing.
That is the story underneath almost every tech headline this week, and almost nobody is naming it. The tools we build with have stopped being neutral. They have acquired preferences, the preferences are inherited from whatever happened to be abundant rather than from anyone’s judgment, and those preferences are quietly redistributing the power to decide what gets built at all.
The Assistant Develops Preferences
Look at the news on its surface and you get a tidy list. The reporting on how the big platforms are putting personal AI agents on your to-do list describes assistants that will quietly work through your errands while you do something else. PYMNTS separately covered a prediction that the software business is about to pivot as AI coding surges, the argument being that when software can write software, the old model of renting a tool one subscription at a time starts to wobble. And at the big Shanghai auto show, CleanTechnica argued that the robotic shift, not the cars, was the thing worth watching, a reminder that the car was never really the product; the factory was.
Different industries, different companies, one shape. In each case the tool that helps us make things is starting to shape what we make. The assistant that manages your day decides which tasks rise to the top. The coding model that writes your backend nudges you toward the language it knows. The robot on the factory floor decides which products are cheap enough to bother building. We told ourselves these were neutral helpers. They are not neutral. Nothing that good at one thing and clumsy at another stays neutral for long.
That unevenness is the whole game, and it is worth being precise about why. A tool that is equally good at everything would just be leverage. It would make you faster without changing your direction, and the direction would still be yours. But a tool that is brilliant at the common case and weak at the rare one applies a gentle, constant pressure toward the common case. No single instance of that pressure is worth arguing about. You accept the suggestion because it is a good suggestion. Over enough decisions, though, the pressure becomes a current, and a current decides where the river goes without ever having an opinion about the destination.
Why the Filter Became the Chokepoint
There is an old idea in how platforms work: when you make it easy for everyone to produce, you don’t get more value automatically. You get a flood. And in a flood, the thing that actually matters is the filter, whatever decides which of the million options ever reaches a human being. The producers stopped being the bottleneck a long time ago. The filter is the bottleneck now.
What’s changed is that the filter used to be something we designed on purpose. A ranking, a feed, an editor’s judgment. Those filters had authors, and an author can be argued with, replaced, or held responsible for a bad call. Now the filter is hiding inside the model’s training history: invisible, unvoted-on, and shaped by nothing more principled than what happened to be abundant on the internet five years ago. The assistant recommends Java not because someone decided Java should win. It recommends Java because there was more of it lying around to learn from. Popularity, frozen into a tool, then sold back to us as competence.
The difference between those two kinds of filter is the difference between a policy and a sediment. A policy can be revised when it produces bad outcomes. Sediment just accumulates, and by the time anyone notices the riverbed has moved, the water has been running the new way for years. That is why the software pivot matters more than it sounds. If the case for renting standard software weakens, the fallback is not a world of bespoke tools built to fit each problem. It is a world where more people generate their own software using an assistant with the same inherited preferences, which means the diversity you would expect from everyone building their own thing does not show up. Everyone builds their own version of the same thing.
Auto China, Solar Onshoring, and the Layer Underneath
This is why the auto show detail matters more than it looks. China didn’t out-design the rest of the world on any single car. It built the boring middle layer, the parts, the robots, the supply chain, and that middle layer now quietly decides what’s economical to manufacture anywhere. A designer in any country can draw whatever they like. What they can actually get built at a sane price is set by somebody else’s tooling.
It mirrors what’s happening as American solar plants pull production and technology back onshore: the headline is energy, but the real contest is over who owns the machinery that makes the machines. Whoever controls the layer underneath gets to set the defaults for everyone standing on top of it. And defaults are destiny, because almost no one changes them.
Put the coding assistant and the factory floor side by side and they turn out to be the same mechanism at different scales. In both cases there is a layer beneath the visible decision, the training corpus or the tooling base, that determines which options are cheap. In both cases the person making the visible decision experiences it as free choice, because nothing forbids the expensive option; it is simply more work. And in both cases the aggregate of a million small preferences for the cheap option hardens into an industrial fact that nobody voted for.
What We’ll Forget to Notice
The part that should give a careful person pause isn’t any single shift. It’s that none of these decisions will feel like a decision. The Erlang team won’t announce that it surrendered its judgment to a training set. It will file a migration plan and call it modernization, and the plan will be reasonable, and every individual line of reasoning in it will hold up. The person whose AI agent quietly reorders their day won’t feel managed; they’ll feel productive, which is the most persuasive feeling there is. The country that owns the robots won’t issue a statement; it will just keep winning bids until the alternative stops existing.
There’s a slow lesson buried in industrial history here. It once took American carmakers five full years to believe that Toyota was genuinely beating them, not because the evidence was thin, but because admitting it meant admitting their whole way of working was already obsolete. The delay was not a failure of information. It was a refusal of an implication. We are very good at not seeing the thing that quietly took the wheel, right up until the wheel won’t turn the other way.
So the question worth holding isn’t whether AI will make us faster. It obviously will. The question is narrower and sharper. When the tool that helps you build also has opinions about what’s worth building, and those opinions are just the residue of whatever was popular last decade, how would you even know you’d stopped choosing? The honest answer is that you would notice it the way the carmakers noticed Toyota: late, and only once the cost of the change had already been paid. The most expensive choices are the ones that never announced themselves as choices at all.
Leave a Reply