The gap

There's a shape I keep sketching on whiteboards when this comes up, and it's three curves. AI adoption, going up and to the right. Hands-on coding ability in people who started after 2023, going down. And a third curve almost nobody puts on the chart: what the AI actually costs to run, versus what you pay for it. Those two lines are not the same line, and the space between them is borrowed money.

Play the three curves forward and you get a period — I don't know if it's five years or fifteen — where the people who could build software from their own heads are retiring, the people who should have replaced them never learned, and the tool that was supposed to make the difference irrelevant has been repriced out of casual reach. I call that period the gap. This post is me trying to argue myself out of it, and mostly failing.

The first curve: prompting is not authoring

Let me be precise about what juniors are and aren't learning, because "kids these days can't code" is a complaint as old as COBOL and usually wrong. Today's juniors are not lazy and not stupid — many of them ship more in their first year than I did in my first three. What they're not doing is authoring: staring at a blank function, holding the problem in their head with no autocomplete to lean on, being wrong alone, and debugging their way out. They review, they steer, they prompt again. The output is real. The formation isn't happening.

I've watched it up close. A junior on a recent project produced a clean, working feature in an afternoon — and couldn't answer why one of the two API calls in it needed a retry and the other one must never have one. Not because they hadn't read the code. Because that distinction lives in a layer of understanding you only build by having been the author when it went wrong. Skill isn't downloaded from output. It's deposited by struggle, in small amounts, over years. Remove the struggle and the deposits stop, even while the output keeps flowing. That's the trap: the metrics all look fine.

The second curve: you are not paying the real price

Here's the part of my argument people push back on hardest, so let me make it carefully. The twenty dollars a month you pay for an AI assistant is not a price. It's a bet — someone else's bet, placed with venture and cloud-provider capital, that subsidizing your dependency today buys them your budget tomorrow. The major labs lose money serving you. The GPUs are scarce, the energy is real, the datacenters are being financed like railroads. We have seen this movie: rides across Paris for six euros until the day Uber needed to become a business, then the same ride for fourteen. WeWork desks, meal delivery, scooters — every VC-funded convenience of the 2010s followed the same arc. Cheap while they're buying the habit, repriced once the habit is bought.

The honest counterargument is that inference cost per token has been falling fast, and that's true — models get distilled, hardware improves, yesterday's frontier becomes today's commodity. But three things eat that gain. Usage explodes faster than unit cost falls: agents that run for hours, codebases fed whole into context windows, one prompt fanning out into a thousand calls. The capability you actually depend on for real work stays at the frontier, and the frontier stays premium — nobody bets their production system on the free tier. And once an industry has burned hundreds of billions, the money wants to come back; it comes back through the price, the usage caps, the enterprise tier, the "your plan has changed" email. Cheap tokens for toy tasks, expensive tokens for the work that matters. The subsidy doesn't end with a bang. It ends with a pricing page that gets one row longer every quarter.

Where the curves cross

Neither curve is a catastrophe alone. A generation that codes less but has cheap abundant AI forever? Fine — that's just a new abstraction layer, and we've absorbed those before; nobody mourns hand-rolled assembly. Expensive AI in a world full of people who can code without it? Also fine — annoying, but the skill is there to fall back on. The gap is what happens when both arrive at once: the crutch gets expensive at the exact moment an entire cohort needs it to walk.

Picture a mid-size company in that world. Its senior engineers — the ones formed before the shift — are in their fifties, expensive, and leaving. Its ten-year veterans came up prompting; they're excellent orchestrators of a tool the CFO now rations, because the AI line item has quietly become the third biggest cost after salaries and cloud. There's a production incident in a system that was 80% generated three years ago, the agent budget for the month is spent, and the person on call has never once debugged something this deep by hand. That's not science fiction. That's just the three curves, extended, meeting in one on-call rotation.

The uncomfortable historical rhyme: we've lost operational knowledge this way before. Ask anyone who maintains mainframe COBOL what it costs to hire for a skill the industry stopped teaching forty years ago — the code outlived the people, and the people who remained could name their price. The gap is that dynamic, but not for one legacy niche. For the general ability to write software unassisted.

What the gap is not

I want to be fair to the future, because doom-scented essays are cheap. The gap, if it comes, is a transition, not an end state. Markets route around scarcity: if frontier AI gets rationed by price, open-weight models running on owned hardware become the diesel generator of software teams — worse, but yours. Salaries for humans who can author unassisted would spike, and spiking salaries are the best recruitment poster ever printed; the skill would get relearned, the way it always does when it pays again. And it's possible the cost curve genuinely wins — that inference becomes too cheap to meter and this whole essay ages like predictions of peak oil. I'd take that outcome happily.

But notice that every one of those escape routes takes years to travel, and none of them helps the person who is on call the night the curves cross. The gap isn't "software ends." It's a decade where the muscle is gone, the prosthetic is expensive, and everything built during the cheap era still needs an adult. Transitions are where the damage happens. Nobody drowned in the sea level; they drowned in the flood.

Positioning for it

If you buy even half of this, the hedges are cheap and mostly unfashionable. If you're junior: use the tools — refusing them is career malpractice — but ring-fence real authoring time the way athletes lift weights they'll never lift in a match. Build something with the AI turned off, regularly, and let it hurt. You're not doing it for the output; you're doing it because in a market where nobody can do it by hand, being the person who can is the trade of the decade. Scarcity is a career strategy, and hand-authoring is on sale right now precisely because everyone thinks it's obsolete.

If you run a team: treat AI spend like debt, not income — it compounds, and the rate can change. Keep an exit ramp: know today which open-weight model plus which hardware covers 70% of your usage if the invoice doubles, because negotiating that migration during a price shock is how you end up paying anything. And keep giving humans work the machine could do — on purpose, as a training budget, not a productivity failure. The companies that cross the gap intact will be the ones that kept the skill alive in humans while it was economically irrational to do so.

I might be wrong about the prices. I hope I am. But between "the subsidy lasts forever" and "keep the muscle," only one of those is a bet I have to win.