Anthropic Reveals Hidden Thought Space in AI Language Models

As MIT Technology Review reports (https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/), the AI company Anthropic has developed an innovative analysis technique that enables a detailed view of the internal workings of large language models for the first time. This method, called the “Jacobian Lens,” allows researchers to visualize the complex computational processes a model undergoes when answering questions or performing tasks.

Insight into the Hidden Thought Structure

Large language models like Claude, developed by Anthropic, are based on billions of parameters whose interplay has so far been largely opaque. The Jacobian Lens now shows that there is a “hidden space” within the model where concepts and meanings are processed in a way that cannot simply be traced back to individual neurons or layers. Instead, dynamic patterns emerge there that the model uses to understand and generate complex relationships.

From Everyday to Troubling Insights

The analysis reveals both expected and surprising aspects. Researchers were able to trace how the model retrieves simple facts or draws logical conclusions. At the same time, it became apparent that Claude also struggles with uncertainties and ambiguities in this hidden space – similar to a human pondering difficult concepts. This finding raises questions about the reliability and interpretability of AI systems, especially when used in sensitive areas.

Significance for AI Research and Application

The ability to better understand the internal processes of language models is an important step toward developing transparent and controllable AI. Through the Jacobian Lens, developers can more precisely identify and address weaknesses and biases in the model. Moreover, the method opens new ways to design AI systems that better comprehend and support human thought processes.

Outlook

While the Jacobian Lens is currently primarily a research tool, it could contribute in the long term to making AI applications safer and more trustworthy. Especially in fields such as medicine, law, or education, where decisions made by AI systems have far-reaching consequences, a deeper understanding of the models is essential.

With this development, Anthropic demonstrates the importance of not only building powerful AI but also making its inner workings transparent. This is a crucial step on the path to responsible AI technology.