Synthetic Data & the Data Wall
Concept
Training & Inference
When the internet runs out, models train on model output.
You're looking at a concept explainer: what this term means in plain English, and the records it ties together.
Frontier labs have essentially consumed the high-quality public web — the 'data wall.' The response is synthetic data: stronger models generating curricula for the next generation (Phi's textbooks, reasoning traces for RL). Done carelessly it degrades models; done well it is now standard — and it shifts value from scraping toward licensed and generated corpora.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Related
Sources
Record reviewed as of 2026-01-15 — see Methodology.