The Most General Tool Ever Built Sells One Specific Thing
An LLM is the most general technology I’ve ever touched.
The instinct everyone has on first contact is the same: this can do anything. ChatGPT made that feeling universal. One text box — write a poem, debug a function, plan a trip, talk you through quantum mechanics. A single thing that looked like it could help with whatever you threw at it.
Which is what makes the next part strange to me. If the technology really is that general, why are the products making the most money the narrow ones?
When I look at what’s actually working as a business — Claude Code, Codex — the fit didn’t come from being general. It came from taking the same engine and aiming it at one specific, painful, expensive problem: writing and maintaining software.
That’s the paradox I keep coming back to. The most general tool ever built finds product-market fit by getting specific.
Generality is not a business model
A blank box that does everything is, weirdly, hard to charge much for.
The value to any one person is real, but it’s diffuse — spread thin across a hundred occasional uses, none of which has a clear price tag on it. ChatGPT is an extraordinary consumer surface and a genuine habit; I’m not knocking it. But there’s a gap between “it helps with everything” and “I have a line item for this exact job,” and the money lives on the far side of that gap.
Product-market fit was never about what a technology can do. It’s about finding the one problem urgent enough, and frequent enough, that someone’s already reaching for their wallet. Generality is just supply. Demand is a separate question, and demand has never paid for potential — it pays for a job that got done.
Why coding won first
Through that lens it’s almost obvious why software was the first thing to really stick. Coding happens to have a rare combination of traits:
- It’s verifiable. Code compiles or it doesn’t. Tests pass or they don’t. The model gets a real signal about whether it did the right thing — and so does the person paying for it. Without that signal there’s nothing to trust, and nobody pays for what they can’t trust.
- It’s expensive. Engineers cost a fortune, so even a modest gain in their output carries an obvious dollar value. The math almost writes itself.
- It’s constant. Developers hit this problem all day, every day. That frequency is what turns a “nice tool” into something you’d fight to keep.
- It’s already budgeted. Companies already pay for IDEs, CI, and a long list of developer SaaS. The willingness to pay is just there — you’re not inventing a new budget line, you’re slotting into one that exists.
- The output is the deliverable. What the model produces is the product — the code itself — not a description of it. Nothing has to be translated from the model’s work into value.
Generality is what let the model write code at all. But it’s everything in that list — not the raw capability — that turned the ability into a business.
I pay without thinking
The clearest signal I know of is my own behavior.
I pay for Claude Code and I don’t sit there weighing whether it’s worth it. Not because it’s cheap, but because the alternative — my own hours — is so much more expensive, and I feel that every single time I use it. The return is legible. I can point at what it saved me.
That legibility is the whole game. General chat is genuinely useful too, but if you asked me to name the exact dollar a given conversation saved, I’d hesitate. The coding tool never makes me hesitate. When the customer stops doing the math because the answer is already obvious, that’s the thing.
The lesson if you’re building
If you’re building on top of these models, the gravity all pulls one way: ship “an assistant that helps with anything.” The model keeps suggesting you shouldn’t box it in — look at everything it can do.
Resist that. The model’s generality is exactly what drags your product toward vagueness, and nobody pays for vague.
Go find the problem that’s painful, costly, constant, and already has money pointed at it. That’s where the fit is. The generality is your raw material, not your product. The fit comes from the specific problem you choose to aim it at — and from how clearly the person paying can see what it did for them.
This is the shape it always takes
If this were a one-off, some quirk of LLMs, it’d be less interesting. But it’s the oldest pattern in technology. A broad new capability shows up, and it earns its keep one specific job at a time.
The personal computer was a general machine that hobbyists adored and businesses shrugged at. What changed that wasn’t a better pitch about everything it could do — it was VisiCalc, the first spreadsheet. Accountants suddenly had one job worth buying an Apple II for, and people bought a general-purpose computer to run a single program. The killer app was specific.
The laser, when it showed up, was so general that people called it “a solution looking for a problem.” Coherent light doesn’t sell. Barcode scanners, CD players, fiber-optic links, eye surgery — those sell. It found its footing one use at a time.
The web was a universal substrate, but nobody wrote a check for “hypertext.” They paid for email, then search, then commerce. Amazon didn’t open as the everything store. It sold books, then earned the rest.
Every time, it’s the same shape. The broad capability is the raw material; some narrow, valuable, repeated job is where the paying customers actually turn up. Coding is just the latest place it’s happened — and LLMs are new enough that we get to watch it play out in real time.
Where it ends vs. where it earns
None of this means the general surfaces don’t matter. ChatGPT’s open box is a colossal distribution and habit engine, and over a long enough horizon the broad surface may quietly pull the specific ones back into itself.
But that’s a story about where this ends. The story about where it earns, right now, is narrower and clearer than all that generality would ever suggest.
The most general tool we’ve ever built is making its money one specific problem at a time. That isn’t the vision falling short. It’s just what product-market fit has always looked like.