AI Models Exhibit Growing Incoherent Failures

Recent research highlighted by Claudius Papirus defines "incoherence" as errors from random variance rather than systematic bias and shows these dominate failures in complex, multi-step AI tasks. The study finds task complexity and longer reasoning amplify incoherence, and that scaling models can increase such chaotic failures; it recommends redundancy, majority voting, real-time error correction, and rollback mechanisms to improve reliability.