Interpretability

Anthropic Makes Breakthrough in AI Interpretability with Sparse Autoencoders

New research from Anthropic demonstrates that sparse autoencoders can identify specific 'circuits' in large language models, opening a path to understanding how AI systems make decisions.