Before publishing anything here, I checked the story this publication is named after.
It did not survive intact. That seems like the right way to start.
The story as usually told
In 2005, an online chess tournament let anyone enter in any configuration — humans alone, engines alone, or humans working with engines.
It was won by two amateurs from New Hampshire running ordinary desktop computers. They beat grandmasters with better hardware and better chess.
The lesson drawn from this, endlessly, in business books and conference keynotes: the pairing beats either half. Man plus machine defeats man or machine alone.
I had that sentence in my draft About page. Then I went looking for where it came from.
What the primary source says
ChessBase reported the result the week it happened, in June 2005. The tournament was the PAL/CSS Freestyle event, played on the Playchess server.
The winning team was Steven Cramton, rated 1685 USCF, and Zackary Stephen, rated 1398 USCF. They ran three machines — an AMD 3200+ and two Pentiums — and used four engines: Fritz, Shredder, Junior, and Chess Tiger. No grandmasters or international masters were on their team. They won the final 2.5–1.5.
So the core of the story holds. Better than holds, actually.
“Amateurs” undersells it. A 1398 rating is a novice. These were not strong players with good tools; they were weak players with good process. The advantage came from knowing which engine to trust in which position, and when to overrule all of them.
That is a more interesting finding than the version in the keynotes.
What does not hold
Three things.
The timeline. Deep Blue beat Kasparov in 1997. Kasparov’s “advanced chess” — one human, one engine — came a year later, against Topalov at León. The freestyle format, where any combination could enter, arrived with the PAL/CSS tournaments running from 2005 to 2008. Popular retellings compress all of this into a single moment. My draft did too.
A name. One widely-used AI-generated reference source gives the winner as “Zachary Bekkedahl.” It also claims a 7/8 score against the Hydra supercomputer and a performance rating above 3100. The contemporaneous report supports none of this. The surname does not appear anywhere in the primary sources. It was invented — confidently, plausibly, in the exact register of a real fact.
The conclusion. This is the serious one. Kasparov was still asserting centaur superiority in 2017, telling Tyler Cowen there was no doubt a human with engines beat the best engine alone. He was probably right when he said it.
He is not right now. Engines improved to the point where the human contribution went negative — overriding a modern engine is almost always an error. The last meaningful centaur rankings date to around 2014. Where the old human job survives at all, selecting among engines that disagree, the selecting is now done by another program.
The punchline of the story I named this publication after has an expiration date, and it has passed.
Why keep the name
Because research is not chess.
Chess has an evaluation function. Every position resolves to a number, every move is checkable, and once the machine calculates better than you do, your judgment is overhead.
There is no such function for whether Michigan’s road funding formula is the binding constraint on its infrastructure, or whether decontrolling New York’s rental market before supply arrives is a policy or a catastrophe. Nothing scores those. Nobody can run the tablebase.
What the machine gives me is enormous: drafting speed, source retrieval, structural proposals, consistency across a hundred pages, counterarguments I would not have thought of. What it cannot give me is the decision about what is true.
And the 2005 lesson survives its own obsolescence. Cramton and Stephen won on process — on a disciplined method for when to trust the machine and when to override it. That question is dead in chess and entirely alive everywhere else.
The standard
I found the invented surname the only way anyone finds them. I opened the primary source.
Everything published here is drafted by a large language model working under my direction. That is disclosed on the About page in detail, including the failure modes. This post is one of them, caught in the wild, on the first thing I tried to write.
The rule I work under: every factual claim traces to a document I can open myself. Not a summary of the document. Not a confident restatement. The document.
That rule is slow. It killed a sentence I liked about man and machine beating either alone. It will kill others.
Before any of that, this seemed worth establishing: when the evidence does not support the story, the story goes.
Including when it is the story on the masthead.
Corrections welcome and will be made in public.