Wes Gurnee

Sources#

• Verbalizable Representations Form a Global Workspace in Language Models

Summary#

Entity. Researcher on Anthropic's interpretability team. Co-first author (with Nicholas Sofroniew) of Verbalizable Representations Form a Global Workspace in Language Models (Transformer Circuits, July 2026), and — with Jack Lindsey — the originator of both the Jacobian Lens (J-lens) method and the conjecture that connects verbalizable representations to conscious access.

Contributions#

Per the paper's author-contributions section:

• Conceived of the Jacobian Lens (J-lens) method, and of the connection between verbalizable representations and conscious access (with Jack Lindsey)
• Developed the first implementation (with Mateusz Piotrowski)
• Led subsequent development and refinement, including methodological variants and the comparisons against the logit and tuned lenses
• Ran the early experiments demonstrating that the lens surfaces concepts used in the model's internal reasoning — the result the rest of the paper is built on

Earlier work of his is cited within the paper itself: the character-counting task used throughout the workspace experiments (the model silently tracking a running line width) is from Gurnee et al.

Connections#

• Jacobian Lens (J-lens) — co-originator, led development
• The Global Workspace in Language Models (J-space) — co-first author of the finding
• Jack Lindsey — co-originator of the method and the conscious-access framing; corresponding author on the paper
• Anthropic — interpretability team

Sources#

• Verbalizable Representations Form a Global Workspace in Language Models — Author Contributions