Anthropic developed a tool called the Jacobian lens, or J-lens, to peek inside its Claude Opus 4.6 model and discovered a hidden layer they call J-space. This space contains words the AI might say next, revealing its internal thought process before it speaks. The findings are both mundane and unsettling, showing that what an LLM is actually doing can differ from what its output suggests. Anthropic claims this new technique, detailed in a paper and paired with a public demo on Neuronpedia, offers a novel way to monitor and control models. Tom McGrath of Goodfire praised the work as significant. Building on earlier mechanistic interpretability research, the J-lens digs into the middle layers of neural networks, where complex math turns prompts into responses. Why it matters: This breakthrough could be a key step toward making black-box AI systems more transparent and trustworthy, a critical need as they influence more decisions in daily life.