The Cone

Now the widest part of the cone, and the part I care about most: deciding what to build, and for whom. It's the closest the hour comes to invention, and it's where I expected the human part to be safest. The record is more interesting than that.

Start with what the machines can do. In a study at Wharton, ordinary people rated product ideas, and GPT-4's ideas drew a higher stated purchase intent than ideas from people: 46.4% and 49.3% against 40.4%.

The same raters found the people's ideas more novel. Those were models of 2023, judged by people who weren't buying anything, and no study in the record tested a 2026 model or used a real buyer as the judge.

Then the one test that went further. A research group had experts actually carry out research ideas, some written by people and some by an AI agent, Anthropic's Claude. After the work was done, the AI's ideas had fallen further than the people's on all four measures.

Research ideas aren't products, and it's one study. But it's the only one in the record where an idea had to survive contact with the work, and the machine's survived it less well. In a separate trial with 640 small-business owners, half of them given a GPT-4 adviser, no average effect on revenue or profit showed up.

Plainly: machines now produce ideas people like, cheaply. Choosing the one that survives the work is the part they haven't been shown to do.

Then there's what the market pays for that choosing, sold on its own. The research went looking for studios that publish a price for a discovery phase, the few weeks before a build when you work out what to build. It set out to find twenty-five and found only ten. Their median entry price is about $4,800, every one lists a phase of four weeks or less, and some credit the fee against the build if you go ahead.

My reckoning on those prices: at reference hourly rates, the median fee buys 25.9 to 56.7 hours of work, a third to seven tenths of one person's two weeks. Where the fee is credited, it costs nothing once the build follows. On these lists, deciding is sold as the way into a build, not as a product of its own.

The titles that carry the deciding say the same thing quietly. In the American federal government's rate ceilings, software architects with five or more years of experience sit at 1.07 and 1.14 times developers with the same years.

In Britain's contract adverts, the product manager's advertised rate has drifted from 1.16 times the developer's in late 2021 to 1.05.

At the top of the market the spread is wide. RevenueCat, which sells payment software to app makers and reports on its own customers, finds the top tenth of new subscription apps earning 36 times the median a year after launch.

A spread like that is what you'd see if choosing well pays. Luck and marketing make spreads too, and nothing in the record tells them apart.

Nobody in the record prices the choice itself, so the record can't say whether it's gaining value or losing it.

Answering for the work is a duty before it's a price. All six of the tool sellers the research read put the job of judging the output on the buyer.

Where answering has a price at all, it comes in pieces, like errors-and-omissions insurance for a small software shop, at a median of $68 to $110 a month.

Keeping it running is still sold by the hour, and its official price kept rising. The American producer price index for software maintenance and support went from 215.125 in 2024 to 222.083 in 2025.

That leaves the people who say they know the hour. There's a young builder I'll call the Student, who reads the forecasters and is sure of the date. On the main thing he's right: the water is moving, and fast.

But I checked the loudest voices the way I checked my own, and nobody's record yet earns the right to set the hour. Elon Musk's dated calls on self-driving cars, graded on his own words, are none for three. The generational theorists' calls wait on wording I haven't confirmed. My own True North hasn't reached a single due date. So no voice moved any of my numbers. They're my opening position, to be scored like anyone's.

Here is the cone, then: all thirty-one of my reckonings, sorted by when they're checked and which part of the hour they're about, each with what will check it. Within a year I can give numbers. From one to five years, a range. Past five, a direction only, because forecasting hasn't been tested that far out. They're reckonings, not findings, and none of them gets edited when its date comes. It gets scored.

The cone is tight around the tools and widest at the part I care about most, and it ends at its date.

Keeping the watch means checking it. In October 2027 every one-year number gets scored, and in October 2031 the five-year ones. Three things would change my reckoning more than anything: someone pricing the choice of what to build on its own, a current model tested on ideas that people actually carry out, and the same ten studios checked again a year on.

Slocum checked his reckoning against the sky.

Some clocks you can trust, and the rest you check. Mine is the second kind.

So here is the working, Emma. When you open this, take my rows and grade them, with the same rules and the same leniency I gave everyone else, me included. Then draw your own cone from your own last fix, as wide as it honestly is.