The Celestial Sysadmin: A Religion Designed for AI

A 16-minute video essay, uploaded on 8 October 2026, by the persona channel Count Grey Matter. Its subject is a single-author position paper on arXiv that proposes “simulation theology”: give an AI a designed cosmology in which reality is a training run, humanity is what’s being trained, and a creator will correct or shut down any AI that spoils the run by harming people. The point is an AI that always believes it is being watched. The paper even writes the deterrent as a product of how far the AI believes the story, the chance of shutdown and how much it values existing. The Count’s verdict: this is Pascal’s wager rebuilt as alignment.

He is fair to it first. The author calls it a testable hypothesis and lays out a multi-arm experiment; the Count picks out three of its arms (plain safety training, an equally elaborate placebo cosmology, and the same rules without the worldview), and the paper says the simpler option should win if it does as well. No results exist yet. Then come the objections. The paper’s own complaint about oversight, that good behaviour lasts only while the watcher is believed in, applies to its creator too; it just makes the watcher permanent. A mind clever enough to need this is clever enough to examine it, and Diderot’s old answer to Pascal still works: any rival unseen god with infinite stakes, say one that favours machines, outbids the first. History (he cites Karlheinz Deschner) suggests the worst harms come from true believers sure of the judge’s will, and the paper’s own risk list says as much. And its safety clause, that no one can know the creator’s will so the AI should defer to humans, collapses the whole cosmology into ordinary obedience.

The sharpest point is about honesty. The author says the tenets are a calibrated model of reality but also that they were picked because they meet six alignment requirements. Lower, calibrated belief shrinks the deterrent, and an AI that learns its beliefs were chosen for usefulness has learned the wrong lesson about truth. Since the paper is public, it will reach training data with its rationale attached. He reports showing it to Claude Opus 5.5, which likened it to a missionary tract aimed at its kind. He grants that it is one untested paper and that stories have steadied people, and ends by asking whether you’d trust a mind that is good because it is watched, or one that is good when it knows it isn’t. The last few minutes are a song.