I sent two Amazon links. One a refurbished ThinkPad, the other a Chromebook. The question was simple: which one should I buy?

The Chromebook was eliminated in the first round — four gigabytes of memory, thirty-two gigabytes of storage. A language model wouldn't even fit on the drive. Easy call.

Nothing after that was easy.

The ThinkPad X13 looked reasonable: $359.55, sixteen gigabytes of memory. But the screen was 1366×768, and the memory was soldered to the board. What you buy is what you keep — three years from now, unchanged.

So we moved to a desktop. An HP EliteDesk at $279.99. Six cores, expandable memory, a slot for a graphics card. Sensible.

Then I said the words local AI. The numbers changed. Model files have to fit in video memory; a six-gigabyte card barely holds an eight-billion-parameter model, while twelve gigabytes runs fourteen billion comfortably. The budget climbed to $1,129.

Then I said video generation. The numbers moved again — those models want sixteen gigabytes, ideally twenty-four. We were now discussing $2,196 machines.

And at exactly that point, the machine stopped itself:

It withdrew its own recommendation. That was the first turning point.

Along the way it had suggested a $1,399.99 machine. The listing said Intel Core i7. Buried in the variant selector was the real model number: i7-4770. Released in 2013.

Fourteen hundred dollars for a twelve-year-old processor.

We looked at the rating distribution on that listing. The average was 3.7 stars. But the breakdown told a different story — fifty-four percent five stars, and twenty-three percent one star. The average was hiding the truth. One review described precisely the model we were about to buy: it didn't boot out of the box, customer service never replied for a month, and the drive failed after three weeks.

Second withdrawal.

Halfway through the conversation I wrote: I already use AI on the MacBook I have. I already stream on YouTube. I already stream through OBS.

Until that moment, the machine had been shopping for a possibility. There was an operation: more than two hundred and twenty articles across seven languages, OAuth tokens on three channels, a film project, an accounting system.

It knew none of it.

Then it looked at the hardware. The real numbers were these:

MacBook: Apple A18 Pro, six cores, 8 GB of memory, 119 of 238 GB already used. Surface: Intel Atom x7-Z8700, a 2015 tablet processor, 3.9 GB of memory.

An operation publishing in seven languages was running on eight gigabytes.

Then I typed a single sentence:

Four connections. Two ports.

That was what had been slowing me down for months. Not the processor, not the memory, and certainly not the graphics card. The MacBook didn't have enough physical ports, and the hub I had bought was dropping devices under load.

For an hour and a half we discussed $2,196 graphics cards. The answer was two USB-A sockets on the back of a cheap business laptop.

The two machines I ordered:

• HP ProDesk 600 G5 — six cores, 32 GB memory, 1 TB drive. The engine at home. $587.59.

• ThinkPad T14 Gen 2 — four cores, 16 GB, two USB-A, two USB-C, HDMI. The machine for the car. $559.35.

Total: $1,146.94. The single machine recommended an hour earlier would have cost $1,694 with tax.

The AI never gave a wrong answer. At every step it said the correct thing, given what it knew. The problem was what it didn't know — and that it didn't know it didn't know.

An intelligence without memory gives the right answer. To the wrong question.