Skip to content

External reference: model collapse and closed-loop training #1

@alyssadata

Description

@alyssadata

External Reference

Topic: Model collapse / closed-loop training / external grounding

This article is relevant to ASI Redefined because it supports the concern that closed model-output loops may become unstable without grounding outside the model’s own generated material.

Primary relevance:

  • Closed-loop training on model-generated data can lead to model collapse.
  • External grounding, even minimal, may prevent collapse in the studied statistical setting.
  • This supports caution around coherence claims that are not traceable across time, context, and pressure.
  • This is relevant to memory substrate and identity binding because ASI-level coherence claims require grounding, comparison, and attribution beyond repeated output behavior.

AI Foundations alignment:

  • Output is not provenance.
  • The model is not Source.
  • Self-reference is not sufficient grounding.

Status:

External support reference only.
Does not change ASI Redefined v2.0.
Potentially useful for v2.1 references, roadmap, or future discussion of grounding requirements.

Sources:

King’s College London article:
https://www.kcl.ac.uk/news/scientists-come-up-with-way-to-overcome-ai-data-cannibalism

Related paper / preprint:
https://arxiv.org/abs/2506.20623

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions