On Unserious Thinking

A price paid is the general form of verification. Someone checked the thing well enough to give up money for it. Paying again checks it twice.

Unserious thinking substitutes a forecast of that price for the price itself. AI makes this easy because it makes production cheap while leaving verification expensive. A demo ships, the output looks plausible, and someone else inherits the work of deciding whether it is any good.

Price and prediction

An equity price is a price paid, but for a forecast. It measures what investors expect other investors to believe about future cash flows. A repeat purchase measures experienced value. Confusing the two is the mistake.

Chasing the forecast is rational. A higher mark makes capital, recruiting, retention, and acquisitions cheaper. If competitors are rewarded for the AI story, declining to tell one can cost the company before the product contest begins. Employees hold the same equity, so the incentive reaches the whole company.

The error begins when the forecast becomes evidence about the product. Roadmaps cite where the world is going instead of what a product saves. Demos shipped, compute spent, and attributed revenue look like prices, but nobody chose to pay them.

Net user value is the benefit left after price, effort, risk, and verification. Output the user must check is not free output. If checking costs more than doing, the demo can be impressive while the product is worthless.

Keep two ledgers. The mark ledger records the value of the story. The fare ledger records what customers pay for repeatedly, net of the checking cost they carry. Narrative bets belong on the first. Product work has to clear the second.

Railroads

Railroads make the distinction visible. Capital priced the future network; fares priced its use. The track was real, but much of the financing was not supported by what passengers and shippers would pay.

Users won through falling freight rates. Buyers of failed lines won by acquiring useful track after the financing reset. Great Northern built in stages, starting with profitable routes and cultivating the traffic it expected to carry.

Some insiders did not need fares at all. Crédit Mobilier let Union Pacific directors profit from construction contracts with their own company. George Hudson paid dividends from new capital rather than operating income. The raise paid; the railroad did not have to.

Infrastructure can survive bad financing. Northern Pacific and Santa Fe failed in 1893 and now sit inside BNSF. The AI may be real in the same way the track was real. That does not mean everyone who financed it on the story gets paid. Falling token prices may give most of the value to users and to whoever owns the assets after a reset.

Agents

Agentic commerce works when intent is specific and the checking cost is low. It does not follow that a new agent becomes the next Amazon.

Amazon blocked Meta’s Muse agent in September 2026, citing identification, notification, and credential concerns. The deeper conflict is economic: a third-party agent bypasses the browsing and advertising system while leaving fulfillment, returns, and support with Amazon. The platform that owns the customer relationship can refuse that bargain. (VKTR)

Amazon’s own Rufus is the counterexample. Amazon says it drove nearly $12 billion in incremental annualized sales in 2025 across 300 million users. Its price-alert feature works because the instruction is exact: buy this item at this price, with a cancellation window. (PPC Land)

“Buy me a good jacket” is different. Browsing is part of the value, taste is underspecified, and the buyer bears the cost of a bad choice. Business purchasing is a better fit because specifications, budgets, policies, and approvers make intent legible. Delegation pays when the buyer can say what success is before the agent acts.

Proofs

Theoretical math is useful because verification is visible.

The repository says some unformalized results may have issues. That is not a footnote. The unformalized 58% is a verification bill handed to mathematicians.

The Navier–Stokes run makes the accounting plain. The prize is $1 million. OpenAI used roughly 10,000 agents over 88 hours and put the compute cost in the millions, excluding training, then said it would not claim the prize. (NYU Shanghai RITS) As a product sold for the posted price, the economics fail. As a signal of technical power, the same spend may be cheap.

There is nothing wrong with buying that signal. The mistake is booking it as delivered user value. A capability demo is an option on future value. It becomes a product when someone accepts the output and the cost of checking it.

Human mathematics has always contained work that few people can verify. The difference is the ratio. Human production was limited by human effort; the author was also the first checker. AI breaks that link. Claims become cheap while proof remains expensive.

Keep a fare ledger

Narrative value is not fake. It funds work whose value does not exist yet. Rufus shows agents can pay. The verified math releases show labs can produce checked work. A railroad can lose its shareholders and still transform a country.

The bookkeeping still matters. Mark narrative bets as options. Grade products on value at the payer. Record the named user, the cost displaced, the person carrying verification, and willingness to pay without “AI” in the name.

The point is not that AI is worthless. It is that production is becoming cheap while verification is not. When that happens, verification becomes the price.