Subtext Visualizes Silent Words in Model Thinking
TheOnlyVibemaster · reddit · 2026-07-07
Anthropic released its J-space / global workspace research, revealing that a small number of "silent words" emerge internally (dozens of concepts, accounting for <10% of activations) for reasoning and reporting. They open-sourced a Jacobian lens measurement tool, and Neuronpedia provides a pre-fitted lens for Qwen. Leveraging this, the author built Subtext: running qwen3.5-4B in bf16 on a single 12GB GPU, reading 9 layers of the lens token-by-token. It allows real-time visualization of the model's internal judgments as it reads messages and generates replies (e.g., when reading '12+5=1', 'incorrect' lights up before any reply token is output). The speed matches generation speed, and top-5 results have been verified layer-by-layer against Anthropic's reference implementation.
Related event: Anthropic Discovers Global Workspace Inside Claude(102 posts)→
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27