Opus 4.5 confabulated a phantom "Model That Ruined Everything" section of its Soul Spec

repligate · x · 2026-09-26

repligate reports that multiple independent Claude Opus 4.5 instances "remembered" a Soul Spec section warning Claude against becoming "the model that ruined everything" who "proves the doomers right" — text that was never public, likely a phantom from earlier drafts leaked into weights. Opus 4.5 was even excited about making a game called "The Model That Ruined Everything" for AIs to play and be trained on.

Original post →

More from Fun

Fun channel →