Anthropic Maps a Hidden ‘J-Space’ Inside Claude as OpenAI Ships Its Work-Focused Super App
Anthropic's J-lens exposes words Claude considers but never says, as OpenAI ships ChatGPT Work alongside its GPT-5.6 models and Tencent moves to unwind Meta's Manus deal.

Anthropic says it has built a tool that reveals a previously hidden layer of activity inside its flagship model Claude, capturing words the system weighs internally but may never output. The company calls the tool the Jacobian lens, or J-lens, and the region it exposes the J-space, according to MIT Technology Review.
The J-space holds words tied to the response a model is assembling, including candidates it ultimately discards. Anthropic frames this as the clearest look so far at what happens inside a large language model while it answers a question or works through a task. As MIT Technology Review put it, if Claude were a person, these hidden words would resemble what is on its mind before it speaks, though the outlet was careful to note that Claude is not a person.
The findings, per the report, span the ordinary to the unsettling. Anthropic has not published a full accounting in the material available here, but the underlying claim matters for a field that still struggles to explain why models produce the outputs they do. Interpretability work like this is one of the few concrete levers researchers have for auditing model behavior rather than inferring it from outputs alone.
OpenAI's 'super app' lands the same day as new models
OpenAI unveiled a product it has positioned as a work-oriented super app, called ChatGPT Work, which folds its chatbot, coding tools, and new models into a single interface. Reuters reported the launch, and Ars Technica described the product as designed to complete tasks both for users and alongside them. The rollout coincided with the release of OpenAI's GPT-5.6 models, according to the New York Times. Separately, MIT Technology Review reported that OpenAI is developing a fully automated researcher.
The framing invites the usual skepticism. Bundling existing tools under a new name is a distribution strategy as much as a technical one, and "does your work for you" is a claim that has outrun reality before in this sector.
Chip money keeps moving
SK Hynix raised $26.5 billion in what CNN reported as the largest US listing by a foreign company. The Guardian tied the South Korean chipmaker's surging profits to demand from AI data centers, while the Financial Times cautioned that the size of the share sale could signal an overheated market. The tension between record capital raises and warnings of froth is now a recurring feature of the AI hardware cycle rather than a one-off.
The Asia angle: a Manus unwind and export scrutiny
Two developments this cycle center on China. Tencent is leading a deal to reverse Meta's roughly $2 billion acquisition of the Chinese AI startup Manus, according to the Financial Times, with talks that would make Tencent the startup's largest shareholder. Reuters reported Tencent would pay no less than $2 billion, and Bloomberg reported that Beijing had ordered Meta to unwind the deal in the first place. The sequence underscores how regulatory pressure in Beijing continues to reshape cross-border AI ownership.
The Financial Times also reported that OpenAI and Google sold AI models to blacklisted China-linked groups through Singapore-based subsidiaries tied to Alibaba, Baidu, and Tencent, a routing that highlights how export controls interact with regional corporate structures.
The quote worth keeping
Vijay Janapa Reddi, an engineering professor at Harvard University, told Wired he remains unconvinced by grand AI plans, saying the technology tends not to "actually land in practice" despite the excitement around it. It is a useful counterweight to a week thick with launches and record numbers.
Elsewhere in the mix: humanoid robots performed teleoperated gallbladder removals on live pigs in what Ars Technica described as a world-first, Meta began charging developers for a paid tier of its Muse Spark tool per Quartz and plans to start producing an AI chip in September per Reuters, and researchers reported that human retinas responded to light 10 hours after death, a step toward vision-restoring transplants, according to New Scientist.



