🧠 Anthropic has found a way to peek into Claude's "thoughts"
Researchers have developed a tool called J-lens to study the latent space (J-space) within the Claude model. It allows for the visualization of intermediate concepts and model intentions before they are converted into text.
🌍 This breakthrough in mechanistic interpretability is critical for AI Safety.
👤 It represents a step toward understanding the internal logic and working memory of AI.
Source 1: https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/ Source 2: https://transformer-circuits.pub/2026/workspace/