Testing Anthropic's Jacobian Lens on Open Models Reveals Hallucination Routing

Core summary

One sentence to understand this update

A user tested Anthropic's new Jacobian Lens, a concept from their Global Workspace paper, on open models and found it surprisingly acted as a local-model hallucination router.

Impact & opportunity

What this could mean

This experiment highlights the potential of introspection tools like Jacobian Lens to analyze and potentially debug internal states and failure modes in open-source models, offering insights for model improvement.