model-diffing
1 articles
A Git Diff for AI Models — Anthropic Found a Way to Compare Behavioral Differences Across Models
Anthropic Fellows adapted software diffing for AI safety, creating a cross-architecture tool that compares model behavior. It found a “Chinese Communist Party Alignment” switch in Chinese models and an “American Exceptionalism” switch in an American model.