Anthropic Reveals Claude's Internal 'J-space' Workspace

Anthropic published research on July 6, 2026, revealing that its Claude AI model contains a small set of internal neural patterns called J-space, identified using a Jacobian-based technique named J-lens. Editing these patterns altered Claude's reasoning and outputs, such as changing an arithmetic answer from 8 to 6 by swapping an internal 'spider' pattern to 'ant'. The findings, covered by VentureBeat and Indian Express, include an open-source code release and an interactive demo with Neuronpedia.

For AI practitioners, a tool that reliably surfaces and edits a model's latent "scratchpad" would change interpretability, red-teaming, and targeted interventions across production LLM stacks. Anthropic published a research paper on July 6, 2026, presenting evidence that Claude contains a small set of internal neural patterns it calls J-space , found using a Jacobian-based technique the team names the J-lens Anthropic research post . The paper and coverage by VentureBeat and others report experiments where editing or deleting J-space patterns changed multi-step reasoning and specific answers for example, swapping an internal "spider" pattern to "ant" changed an arithmetic reply from 8 to 6 and that the patterns "emerged on their own," per the authors. Indian Express and Anthropic note an open-source code release and an interactive demo on open-weight models in partnership with Neuronpedia.