Recursive training collapses LLMs by gen 9; 10% human data halts the damage
alex_verem · x · 2026-09-18
Oxford/Cambridge research on model collapse: a model trained 9 generations on its own output becomes irreversibly degraded. Keeping just 10% real human data per generation reduces damage to minor. With 74% of new web pages AI-generated (Ahrefs) and OpenAI paying News Corp $250M for content, the scarce resource is unmodelled human writing.
More from AGI Musings
- Biggest threat to AI adoption isn't the tech, it's job-loss fear and change resistance — kevinsurace · 2026-09-18
- Prediction: More Frontier Labs Will Self-Identify as Data Research Labs — ajratner · 2026-09-18
- Martin Fowler: 'I Don't Like LLMs' — Useful, but the Persona Grates — blaizedsouza · 2026-09-18
- 'AI Companies Don't Need Regulation — They Need Investigation,' Argues Viral Analogy — kevinnbass · 2026-09-18
- David Sacks: Pausing AI 'Will Just Hand the Frontier to China' — kevinnbass · 2026-09-18
- kalomaze Pushes Back on Treating Data Constraints as Inherent and Irreducible — kalomaze · 2026-09-18