dangerouslyskip
← Back to writing
Essay·Apr 20, 2026·6 min

Chat Can't Be the Answer

The Star Trek computer was never the dream. It was the consolation prize.

Chat Can't Be the Answer — hero illustration

Everyone's north star for AI is the same scene. Someone on the Enterprise says "Computer," asks a question in plain English, and the ship answers. That's the dream we've been chasing for sixty years, and in 2026 we finally more or less built it. You can talk to the machine and it talks back. Roll credits.

Except — watch that scene again. The Computer answering questions is the least impressive AI on that ship.

The genuinely astonishing stuff on the Enterprise is never the chat. It's the holodeck conjuring an entire convincing world from a one-sentence brief. It's the ship flying itself through a battle while the crew argues about ethics. It's the replicator turning "tea, Earl Grey, hot" into matter. It's a thousand sensors and subsystems acting continuously, in the background, without anyone invoking them. The Computer politely answering a question when spoken to is the floor of that world's AI, not the ceiling. It's the bit so basic they barely write dialogue for it.

And that's the thing we aimed at. We took the most boring AI on the starship — the talking FAQ — and made it our entire industry. We built a beautiful chat box and called it the future.

I don't think chat is the future. I think chat is a local maximum: a peak that feels like the summit because the climb to get here was so steep. Going from "the computer cannot understand you" to "you can just talk to it" is such an enormous leap that it's easy to mistake the way station for the destination. But the chat box has structural problems that no amount of model improvement will fix, because they're problems of shape, not intelligence.

Chat is pull, not push. It only does anything when you start it. But the best assistance in the world is the thing that already happened before you thought to ask — the calendar clash quietly resolved, the failing test flagged with a fix attached, the draft sitting ready when you open the doc. A chat box structurally cannot do that, because it's sitting there waiting for you to perform the ritual of asking. The most valuable moment is the one chat is constitutionally incapable of: the unprompted one.

Chat offloads the hard part onto the user. The blank box says "type something," and now knowing what to ask, how to ask it, and when is your job. We dressed up a command line in better manners and handed people homework. (This is the same disease I ranted about in Hide the Magician — the chat box is the exposed machinery, and we made the user operate it.)

Chat is a serial bottleneck. One thread, one conversation, your attention required start to finish. That's a fine shape for one task. It's a terrible shape for the ten things AI should be doing for you simultaneously in the background. You can't have a meaningful conversation with ten agents at once, and you shouldn't have to.

Chat makes you the integration layer. The output lands in a chat window — a separate place, away from the thing you were actually doing — and now you are the one copying, pasting, reconciling and stitching it back into your real work. The interface that was supposed to save you effort quietly made you its glue.

So how did we all get stuck here? The same way we get stuck on most things: it was the cheapest thing to ship. The model emits text, so the zero-effort wrapper is a text box. Chat isn't the destination we reasoned our way to. It's the path of least resistance we fell down. It's the terminal, not the desktop.

And that analogy is the way out, actually. The command line didn't die when the GUI arrived — but it stopped being the front door. It became the thing power users drop to when the graphical layer can't express what they need. That's exactly the right fate for chat. Not gone — demoted. The universal escape hatch. The repair tool you reach for when the ambient thing gets it wrong and you need to say, in words, "no, not like that." Chat survives as the fallback. It just can't be the front door, any more than bash is how your nan checks her email.

So what is the front door? Roughly, everything chat isn't.

It's ambient — the AI acting in the background and surfacing only when it matters, instead of waiting to be summoned. It's embedded — a layer running through the tools you already use, not a separate destination you visit and then ferry things back from. Very often the correct interface for AI is no new interface at all: your existing workflow, quietly smarter, with the magic happening between the cracks.

It's generative, when it does need to show you something — the controls assembling themselves around the task at hand rather than a one-size-fits-nothing text box. (I've been bolting adaptive, agent-driven UI widgets into the things I build for exactly this reason: don't make the user describe what they want in prose when you could just render the right control.)

And yes, it's often voice — but not voice-chat. The Star Trek "Computer" was the wrong lesson there too. The dream isn't "I speak a request and wait for an answer." It's "the system was already three steps ahead, and I use my voice to nudge and confirm rather than to operate." Voice as a steering wheel for something already in motion, not an order window for something that does nothing until asked.

Underneath all of it sits the unglamorous work that makes any of this possible — the context plumbing that means the system already knows who you are and what you're doing, so it can act without a fresh interrogation every time. (That's most of why I built things like Skynet: an agent that can act usefully in the background needs to see, and chat is what you fall back to precisely when it can't.)

I want to be fair to chat, because the contrarian move would be to declare it dead and I don't believe that. Chat is a genuinely great fallback and an even better repair surface. When the ambient system gets it wrong, dropping into a conversation to correct it is exactly right. Chat is the most flexible, most forgiving, most universal interface we have — which is precisely why it's the wrong default. The most flexible tool is rarely the one you want pointed at you first. You want the specific, invisible, already-did-it thing first, and the flexible everything-box behind it for when reality gets weird.

The Star Trek computer was never the dream. It was the consolation prize — the part that talks, because talking is the thing humans find easy to imagine. The real dream was always the ship that flies itself and only speaks up when it matters.

So stop polishing the chat box. It's a beautiful local maximum and we should be proud we climbed it. But the summit's somewhere else, and you can't get there by making the FAQ more eloquent.

Computer — end blog post.

(See? Even that was a little bit annoying, wasn't it.)