Inside Meta’s secret experiments that improve its AI models

A high-profile legal case has unearthed a trove of internal Meta communications, and one particular document has caught the eye of some AI researchers.

This reveals new insights into how models are built and could influence who gets to share in the spoils of this new technology.

Buried in these court filings is a description of how Meta researchers used a process called ablation to identify which data helped improve the company’s Llama AI models.

Ablation is a…