Paper Dives: Verbalizable Representations Form a Global Workspace in Language Models

Anthropic's 2026 Global Workspace paper claims Claude has a small internal "J-space": verbalizable representations that can be read, edited, and ablated. This is a practitioner-level deep dive into "Verbalizable Representations Form a Global Workspace in Language Models." We break down how the Jacobian Lens differs from logit-lens-style readouts, how J-space is defined, and why Anthropic argues this workspace is causal rather than just a passive monitor. You'll learn: How the Jacobian Lens maps intermediate activations to future verbal behavior What J-space is, and what it is not The five workspace-like properties Anthropic tests Why causal swaps and ablations matter How J-space could help alignment auditing Why this does not prove AI consciousness Chapters: 0:00 - The claim without consciousness hype 1:10 - Transformer setup: residual streams and logit lens 3:00 - Jacobian Lens mechanics 5:45 - What J-space is 7:30 - Five workspace properties 11:00 - Causal evidence: swaps and ablations 13:30 - Structure: middle layers and read/write access 15:05 - Alignment auditing 17:20 - Counterfactual reflection training 18:50 - Practitioner takeaways and failure modes Sources: Paper: https://transformer-circuits.pub/2026... Anthropic summary: https://www.anthropic.com/research/gl... Jacobian Lens repo: https://github.com/anthropics/jacobia... Neuronpedia: https://www.neuronpedia.org/ #AI #MechanisticInterpretability #ClaudeAI #NerdyDives #PaperDives