Published March 30, 2026 | Version v1.0

Model Collapse as Drift Cascade: Recursive AI Training, Entropy Loss, and the Thermodynamics of Synthetic Data Feedback

Authors/Creators

  • 1. Independent Researcher, Moreright DAO

Description

Model collapse (Shumailov et al. Nature 2024) is the drift cascade operating on training data. Entropy loss 0.2-0.4/gen = drift velocity, data accumulation = constraint specification, first-order phase transition maps onto D1→D2→D3. 7/7 KC PASS.

Notes

Part of the Void Framework research project (Moreright DAO).

Files

void-framework-paper164.pdf

Files (87.4 kB)

Name Size
md5:55295645abdf6323e0e693c2676a239f
87.4 kB Preview Download