The inner workings of large language models have long been a black box. Now Anthropic has taken a step toward opening it. Researchers at the company built a tool called the Jacobian lens, or J-lens, and used it to find a hidden area inside their flagship model Claude. They named this space the J-space.
The J-space contains words that relate to the response Claude is working on but may not ultimately produce. If Claude were a person, you might say these hidden words reveal what is on its mind before it speaks. But Claude is not a person. The discovery gives the clearest glimpse yet into what is really going on inside large language models as they answer questions or carry out tasks. The findings range from the mundane to the unnerving.
Anthropic peeks inside Claude’s hidden space
The J-lens works by tracking how input flows through the model and identifying words that get partially activated but never make it to the final output. This hidden vocabulary forms a kind of intermediate representation of the model’s reasoning process. Anthropic’s team found that Claude often considers multiple candidate words simultaneously before settling on one. Some of those candidates are closely related to the final answer, while others are surprising or even contradictory.
Understanding this hidden space could help researchers build safer and more transparent AI systems. If developers can see what a model is considering before it responds, they might catch harmful or biased reasoning early. The work also raises questions about how much of a model’s internal thought process is truly accessible and whether current interpretability methods are seeing only a fraction of the picture.
OpenAI launches its super app ambition
Meanwhile, OpenAI has taken a major step toward its long rumored super app. The company unveiled ChatGPT Work, a product that blends its chatbot, coding tool, and new models into a single interface designed to do work for and with you. The launch came on the same day OpenAI released its GPT 5.6 models, signaling an aggressive push to integrate advanced capabilities into a unified platform.
ChatGPT Work is aimed at professionals who want an all in one assistant for writing, coding, analysis, and task automation. It represents OpenAI’s vision of an AI operating system for the workplace. The company is also developing a fully automated researcher, suggesting that the super app may eventually take on more autonomous roles.
The timing is notable. Anthropic and other competitors are also racing to build comprehensive AI tools. But OpenAI’s move to combine multiple products under one brand could give it a first mover advantage in the enterprise market. The super app concept has been successful in Asia, especially with platforms like WeChat, and OpenAI appears to be betting that the same model will work for professional AI use.
The week’s other AI signals
Several other stories this week show how fast the AI landscape is evolving. Humanoids performed the first teleoperated surgery on living animals, removing gallbladders from pigs. It is a world first that could eventually extend to human surgery, though the human work behind these robots remains hidden.
SK Hynix, the South Korean chip giant, landed the largest US listing by a foreign company, raising 26.5 billion dollars. Demand for AI data centers has pushed its profits to record highs, though some analysts warn that the jumbo share sale may signal overheated markets.
In China, Tencent is leading a deal to unwind Meta’s 2 billion dollar acquisition of the AI startup Manus. Beijing had ordered Meta to sell the company, and Tencent is now in talks to become Manus’s largest shareholder. The move underscores the geopolitical tensions around AI acquisitions.
Meta also started charging developers for access to a new version of its Muse Spark tool and plans to produce its own AI chip in September. OpenAI and Google, meanwhile, were reported to have sold AI models to blacklisted Chinese groups via Singapore based subsidiaries of Alibaba, Baidu, and Tencent.
On the research front, scientists resuscitated human retinas that responded to light 10 hours after death, a step toward eye transplants. And an astronomer argued that the hunt for alien life needs more statistics, replacing speculation with mathematical frameworks. For a broader look at how these developments fit into the larger race for embodied AI, check out our coverage of the humanoid robot race.
