ANTHROPIC OPEN-SOURCES CIRCUIT TRACING TOOLS FOR AI INTERPRETABILITY

Summary from the TLDR AI newsletter

The tools generate "attribution graphs" to trace how large language models make decisions internally, revealing the step-by-step reasoning process behind outputs. The library works with popular open-weight models and includes an interactive Neuronpedia frontend for exploring model circuits.