Anthropic found a hidden space where Claude puzzles over concepts (www.technologyreview.com)

Anthropic has unveiled a breakthrough in understanding large language models (LLMs) with its new tool, the Jacobian lens (J-lens), which exposes the hidden "J-space" within Claude Opus 4.6. This J-space consists of related words that the model considers in anticipation of future responses, shedding light on its internal decision-making process. The findings suggest that LLMs like Claude often have complex thought processes that diverge from their final outputs—offering insights into how they generate responses and the reasoning behind them. This development is significant for the AI/ML community as it enhances mechanistic interpretability, allowing researchers to probe deeper into the inner workings of LLMs. By identifying how certain words and concepts are prioritized during computation, the J-lens provides a new method to understand, assess, and potentially control AI behavior, marking a step forward in tackling the challenges associated with opaque AI systems. While the J-lens is not a comprehensive understanding of LLMs, it serves as a valuable tool in monitoring model performance and behavior, encouraging further exploration into the cognitive-like processes of these advanced systems.