Anthropic researchers say they have found a hidden representational area inside Claude where the company’s large language model appears to handle concepts and candidate words before producing an answer. For users and developers, the useful part is not the anthropomorphic gloss. Claude is not thinking. The useful part is that Anthropic says it has a sharper instrument for inspecting what a model is doing between prompt and output.
MIT Technology Review reported that Anthropic built a tool called the Jacobian lens, or J-lens, and used it to identify what the company calls J-space inside Claude. According to the report, that space contains words connected to the answer the model is preparing, including words that may never appear in the final response.
That distinction matters. A model can consider associations, routes and fragments that its final text hides. Anthropic’s claim is that J-lens gives researchers a clearer look at that intermediate machinery. MIT Technology Review described the findings as ranging from ordinary to unsettling, a fair warning that model interpretability remains more like poking around a running engine than reading a neat design document.
OpenAI folds more work into ChatGPT
OpenAI also moved to make ChatGPT less of a chatbot tab and more of a workplace shell. Reuters reported that the company unveiled ChatGPT Work, described as a long-awaited “super app” that combines its chatbot, coding tool and new models.
Ars Technica reported that the product is designed to do work both for users and with them. The New York Times reported that it arrived the same day as OpenAI’s GPT 5.6 models. MIT Technology Review has separately reported that OpenAI is building a fully automated researcher.
The company framing is familiar: put more tasks inside the assistant, then call the bundle productivity. The harder question, left for actual users and administrators, is what work gets delegated, what gets checked, and who owns the errors when the software confidently produces something plausible.
Robots, chips and AI access
Ars Technica reported that humanoid robots, controlled remotely by surgeons, removed gallbladders from living pigs in what it described as a world first. MIT Technology Review has also reported on the human labor behind humanoid robots, a useful reminder that “robot did surgery” often means “people operated a robot very carefully.”
In chips, CNN reported that SK Hynix raised $26.5 billion in the largest US listing by a foreign company. The Guardian linked the South Korean company’s surge to demand for AI data centers, while the Financial Times warned that the large share sale may signal an overheated market.
Meta’s AI business also kept shifting. Quartz reported that Meta has started charging for access to a new version of Muse Spark, with a paid tier for developers. Reuters reported that Meta plans to put an AI chip into production in September.
The Financial Times reported that OpenAI and Google sold AI models to blacklisted Chinese groups through Singapore-based subsidiaries of Alibaba, Baidu and Tencent. Separately, the Financial Times reported that Tencent is in talks to become the largest shareholder in Manus, while Reuters reported Tencent would buy the Chinese AI startup for no less than $2 billion. Bloomberg reported that Beijing had ordered Meta to unwind its $2 billion Manus acquisition.
Elsewhere, New Scientist reported that resuscitated human retinas responded to light 10 hours after death, a step toward eye transplants that restore vision. Wired quoted Harvard engineering professor Vijay Janapa Reddi’s skepticism about AI promises: “The damn thing never actually lands in practice.” It is a useful line to keep nearby this week.
This story draws on original reporting from MIT Technology Review.
