Paul Sutter Says AI-Aided Galaxy-Void Code Was Very Wrong After 10 Minutes

1 articles · Updated · Futurism · Aug 29

Ten minutes into a February presentation, Paul Sutter was stopped by a collaborator who spotted that his updated galaxy-void algorithm was producing flawed results.
The code, built with extensive AI help, mishandled survey edges; Sutter said the bug was subtle but made everything downstream of it wrong.
The failed update had looked promising on paper — 10 times faster and able to handle surveys 100 times larger — which made the mistake harder to catch before the talk.
In a Nautilus essay, Sutter said the episode changed how he works: he now treats large language models as useful but fundamentally untrustworthy and says every output must be audited.