Observing Open-Source Model Internal States via J-space

Murky-Sign37 · reddit · 2026-07-12

This post covers two main layers: Anthropic's discovery of J-space, an internal working space where Claude "thinks in silence," and the author's application of this analytical approach to open-source model Qwen3-8B via local experiments.

Key takeaways include:

Meaning of J-space: It is not readable chain-of-thought, but a hidden conceptual space within activations. The model may pass through a series of intermediate internal states before outputting the final answer.
Experimental method: The author used the open-source Jacobian lens (J-lens) to observe the internal states of Qwen3-8B, locally capturing "prose drift" prior to tool calls—such as a tendency to generate natural language prefixes instead of JSON.
Engineering utility: This signal was integrated into an agent guard pipeline to halt, cancel, or preserve useful internal states, and the recovered samples were distilled into LoRA data.
Supplementary materials: The post includes an 8-minute blackboard lecture video and a terminal demo to help understand the probe setup and guard loop implementation.

Original post →

Observing Open-Source Model Internal States via J-space

More from coding & agent