ARTIFICIAL INTELLIGENCE Anthropic Found a Hidden Space Where Claude Puzzles Over Concepts Will Douglas Heaven
- J-space is a hidden representational space inside Anthropic’s Claude where individual word-like vectors (hidden “words”) correspond to what the model is poised to produce next. These vectors reveal the model’s short-term predictive inclinations — i.e., what it’s “thinking” about before it generates text [MIT Technology Review].
- Anthropic researchers found that inspecting J-space can show clusters of related words and concepts that the model “puzzles over” prior to output, making it possible to see internal, pre-output structure that anticipates the response content [MIT Technology Review].
Follow-up Questions:
1. How did Anthropic discover and visualize J-space?
2. Can J-space be used to control or steer Claude’s outputs?
3. What are the safety or interpretability implications of exposing J-space?
4. How does J-space relate to other interpretability techniques (e.g., attention, activation atlases)?
5. Has Anthropic published a technical paper or demo showing J-space analyses?
Related questions
- How did Anthropic discover and visualize J-space?
- Can J-space be used to control or steer Claude’s outputs?
- What are the safety or interpretability implications of exposing J-space?
- How does J-space relate to other interpretability techniques (e.g., attention, activation atlases)?
- Has Anthropic published a technical paper or demo showing J-space analyses?