data drift

2 items

RESEARCHarXiv CS.CL·4/13/2026

Drift and selection in LLM text ecosystems

This paper introduces a mathematical framework to analyze the recursive process where AI-generated text re-enters and shapes the public record from which LLMs learn. It distinguishes between "drift," which removes rare forms through unfiltered reuse, and "selection," which filters content based on criteria like quality, showing normative selection preserves deeper linguistic structures.

Text Ecosystems data drift model collapse large language models

ARTICLEDEV.to AI·4/16/2026

"Why Most AI Systems Fail in Production (And No One Talks About It)"

This article explains why most AI systems fail in production, contrasting perfect demos with the reality of noisy data and unpredictable environments. Failures occur due to changes in data and patterns (drift), causing the model to no longer solve the problem it was trained for.

MLOps data drift AI deployment