Anthropic Uncovers Hidden "J-Space" in Claude LLM

Anthropic researchers have gained unprecedented insight into the internal workings of large language models (LLMs) by uncovering a hidden "J-space" within their flagship model, Claude. This discovery was made possible by a new tool developed by the company, dubbed the Jacobian lens, or J-lens. The J-space is described as a hidden area containing words related to a response that the model is processing but may not ultimately produce.

This finding offers a glimpse into the "thought process" of an AI, akin to what might be on a person's mind before they speak. The researchers have detailed their findings, providing a deeper understanding of how LLMs like Claude formulate answers and execute tasks. The implications of this research extend to improving AI transparency and potentially debugging or enhancing model behavior by understanding these intermediate conceptual spaces.

In parallel, OpenAI has unveiled its "super app" called ChatGPT Work, which integrates its chatbot, coding tools, and new models. This launch coincided with the release of OpenAI's GPT 5.6 models. The app is designed to assist users with their work, and OpenAI is also reportedly developing a fully automated researcher. Separately, humanoids have successfully performed teleoperated surgery on living animals, removing gallbladders from pigs in a world-first demonstration. The human effort behind such advanced robotics remains a significant area of focus.