Receipts, not takes
A take is a sentence. A receipt is a screenshot. The rule I use to decide whether something is worth publishing at all.
There is an enormous amount of writing about AI and very little evidence inside it.
The standard piece follows a shape: a claim about where things are heading, three examples the author read about, and a confident conclusion. Nothing in it can be checked. Nothing in it required the author to do anything other than read and rephrase.
I try to hold one rule against my own output, and it is deliberately strict.
The rule
If I can’t show a receipt, I don’t publish it.
Concretely, anything I put out has to carry at least one of four things:
- A screen recording of the thing actually running
- A number I measured myself — time, cost, error rate, output volume
- A teardown: what worked, what broke, what I would change
- My own rebuild, fork or extension of it
Commentary is allowed. Commentary without one of those four is a take, and it does not get published.
Why the bar is set there
The obvious reason is trust. It is very easy to sound informed about AI and very hard to be informed, and the gap between the two is invisible to a reader unless you show your work. A screenshot closes that gap in a way no amount of prose does.
The less obvious reason is that the rule changes what I do, not just what I write.
If I know I cannot write about a tool without running it, I run it. If I know I cannot claim a cost improvement without measuring it, I measure it. The publishing constraint reaches backwards into the work and makes the work more honest. That is the real value — the writing is downstream.
What it costs
It is slower. Considerably.
Reading about a tool takes ten minutes. Running it properly, hitting its edges, and forming a defensible view takes a day. The volume of output drops. That is a genuine cost and I want to be straightforward about it rather than pretend the disciplined path is also the easy one.
The trade is that the smaller volume holds up. A take ages badly the moment the thing it guessed about resolves. A measurement stays true — it was true when you took it, and it stays a fact about that version on that day.
The version for your own work
You do not need my rule. You need a falsifiable one.
The test is simple: could a reader prove me wrong? If the answer is no — if the claim is shaped so that nothing could contradict it — it is decoration. Decoration is fine in a conversation. It should not be the thing you build a reputation on.
Show the screenshot. Report the number, including when the number is unflattering. Say what broke.
Especially say what broke. That is the part people remember, because it is the part almost nobody publishes.