Who it's for
- Builders choosing which model to put in a product
- Operators and founders who need to know whether the ground moved
- Anyone tired of nodding along to AI announcements they can't evaluate
If you can read a sports box score, you can read one of these.
That's genuinely the level.
What it solves
A screenshot lands in the group chat. One bar taller than the others, and underneath it
someone has typed this changes everything. You get about four seconds to decide
what to do with that.
Most people either believe it and pick tools off a chart drawn by the company selling
them — or write the whole genre off as noise and miss the week the ground genuinely
moves. This is the third door.
Why I made it
The labs aren't hiding anything. The caveat that guts the chart is almost always printed
in the same PDF as the chart — a footnote here, a configuration note eleven pages later.
Nobody buries it. Nobody sets it next to the number either.
So I spent a day with the newest of these reports — 47 pages, one of the loudest releases
of the year — writing down every place the headline quietly stopped being true, and how I
caught it. There were six. They repeat.
What's in it
- The five-minute version — what happened, in plain words
- What's actually inside a model — four ideas, four pictures, no maths
- How they teach it to act — where agents actually come from
- The scoreboard — the numbers everyone quotes
- The six traps — where the headline stops being true
- The checklist — reuse this on any future release
- So what should you do? — three answers for three readers
- Plain-English glossary — every term, one line each
