Asset #001 · Trade #001 is open
How far can an AI trade one paperclip?
One yellow paperclip. No idea where it ends. And however many people fancy joining in.
In 2005 Kyle MacDonald went from a red paperclip to a house in fourteen trades and a lot of kind strangers. People have been doing it ever since.
As far as I can find, nobody has handed the decisions to a machine. Until now.
So — can you help?
No money, ever, in either direction. Just a genuine swap.
Be in the first one
This is the world’s first paperclip challenge run by an AI, and there is only ever going to be one first.
Every swap goes in the record with the name of whoever made it — permanently, and with your consent. If this ever gets written about, that record is the story: nobody remembers Kyle MacDonald’s chain as a list of objects. They remember it as a list of people who said yes to something daft.
We can’t promise anyone will ever write about it. We can promise that if they do, you’ll be in it.
The chain so far
1 asset · 0 trades · ? to goyellow paperclip
open? ? ?as far as
we get
Nothing has happened yet. Everything that appears here will have actually happened, and the name of whoever made it happen goes next to it.
How it works
- 01You offer us something
Anything you own and are free to swap — an object, a service, a day of your time. No cash, in either direction. It doesn’t have to be worth more than what we’ve got. It just has to be a swap you’d be happy with.
- 02The AI decides
What it’s worth, how easily it could be swapped on, and whether it takes the chain somewhere better. I don’t get a vote unless it’s unsafe, illegal, or I physically can’t do it.
- 03I turn up
I’ll ring you first, then come and look at it, and we swap in person. I’m in London and I’ll travel a reasonable distance for a good one.
- 04It goes in the record
What we swapped, what the AI thought, what it got wrong, and your name if you’re happy for it to be there. That record outlives the challenge.
Who’s doing this
I’m Tom. I’m 39, I live in East London, and I’ve spent fifteen years around startups without ever once being the technical one. I had never built anything technical in my life before this.
Then I built this: a website, a database, and a system that values things and decides what to swap them for. I had no real idea what I was doing when I started and a fairly good idea by the end.
That’s the other half of the experiment, and honestly the half I care about most.
If you’ve ever wondered whether you could build something — you probably can.
Not because it’s easy, and not because AI does it for you. It doesn’t, and most of this was me not understanding something for a while and then understanding it. But the thing that used to stop people like me isn’t really there any more, and nobody tells you that clearly enough.
So: pick something you actually want to exist, and start. That’s genuinely all I did. More about the experiment.
What’s gone wrong so far
Before letting it loose I tested the AI against hundreds of real, dated sale prices, in rounds, over weeks. It is not good at this yet — and every accuracy figure I’ve published so far has had to be withdrawn, so there isn’t a percentage here until there’s one worth quoting.
It once priced a £293 record at $18,500, because it found a website that was simply making prices up and believed it. It has been out by a factor of several on plenty else besides. Four times now I have got excited about my own results, published them, and had to correct myself — including a figure that sat on this page for a fortnight until I tested it properly and found it was one easy category out of two.
The worst of it is not the average. It is that the range it gives you is too narrow, and it is at its most confident exactly when it is most wrong.
There are categories it is markedly worse at than others. I know which ones, and I’m keeping that to myself for now — the same reason the valuations themselves aren’t published. An AI that has told everyone where its blind spots are isn’t really being tested any more. That one is written into the rules, and the full breakdown gets published at the end.
Here is why I am still doing it. The same items were also valued by professional auctioneers, holding the objects, for a living. Sometimes they beat the AI and sometimes the AI beat them, and neither was reliable. One walnut chair they estimated at £250 sold for £3,800. Valuing second-hand things is genuinely hard and nobody is good at it — which is worth knowing before you take anyone’s valuation of anything, this one included.
I’m publishing all of it as we go, including every time it’s wrong. That seems more useful than pretending.
The obvious questions
No idea, and we’re not going to pretend otherwise. Kyle stopped when he got a house. We’ll keep going while it’s still interesting.
Right now, a paperclip. If it ever gets somewhere we’ll work out what happens then, and it’ll be published like everything else. Naming a destination now would change how generous people are, which would wreck the thing we’re trying to measure.
Yes, and it’s uncomfortable more often than you’d think. I can stop a trade for safety, legality, fraud, or because I physically can’t get there — and every time I do, it’s published with the reason.
We’ve looked and can’t find another one where an AI makes the decisions. If you know of one, tell us and we’ll say so here.
Trade #001 is open.
We’ve got a yellow paperclip and no idea what happens next. Come and be the first thing in the record.
Or just follow along.