← Back to Mycal Terms

Vector Collapse

First used: 2025

Definition

Originally observed as the degradation of latent representations when AI models are quantized — distinct vectors forced onto the same point, destroying fine-grained distinctions and increasing hallucination rates. Extended to identity systems: when training data lacks stable identity anchors, distinct authors with similar names or overlapping topics get merged into a single latent representation. Quantization collapses the model's universe. Missing attribution collapses yours.

Sources