LLMs are not the black box you were promised

Researchers can now peer inside AI models and trace their actual reasoning — and it's more structured than expected. Anthropic's "circuit tracing" technique identifies discrete, human-recognizable concepts like "Texas" or "Austin" firing in sequence, revealing genuine multi-step reasoning. Surprisingly, models often use algorithms that differ from how humans think, meaning this research could help detect dangerous behavior and build smarter AI systems.