Skip to content
Rung 05 DataSwitch rung
Rung 05 / What it learns from

Data

What the model is trained and grounded on, and what that constrains.

3
Sources
3
Concepts
4
Entities
A stack of polished steel drive platters on a copper spindle, on paper.

In scope: training-data supply and exhaustion, synthetic data, licensing and provenance, retrieval corpora, data quality effects on capability per token, contamination. Out: retrieval as a harness mechanism (harnesses), model architecture (models).

A banded drill core in a tray, the thin copper band marking the clean licensed seam.
SyntheticLicensingProvenanceExhaustion
Analysis only1Show all
Sources compiled for this topic
TypeSourcePublished
ANALYSISWhat Authors Need to Know About the Anthropic Settlement (Bartz v. Anthropic, $1.5bn class settlement)
The Authors Guild · The Authors Guild

A court-set per-unit price for pirated training data: ~$3,000 per title across ~500,000 qualifying titles from a $1.5bn fund, against 7 million copies from LibGen and PiLiMi. The fair-use ruling covered TRAINING; the $1.5bn covered ACQUISITION — the distinction most coverage collapses. Final-approval date unresolved (hearing 2026-05-14 per this source).

2025-09-25