Inside the Black Box: Anthropic’s Clearest View Yet of Claude’s Internal Processing
Anthropic has developed a tool that reveals what its flagship large language model, Claude, appears to be working through before producing a response. The tool, called the Jacobian lens – or J-lens – exposes a previously unmapped region inside the model that researchers have named the J-space. What lives there is a set of words and concepts related to Claude’s eventual output, but which may never actually appear in that output.
The discovery lands somewhere between scientifically mundane and genuinely unsettling. If Claude were a person – and Anthropic is careful to say it is not – these intermediate words would function like unspoken thoughts: the mental churn that precedes speech. The company’s researchers say this is the clearest window yet into what large language models are actually doing as they generate answers or complete tasks.
OpenAI Counters With ChatGPT Work and New Model Releases
On the same day Anthropic’s research drew attention, OpenAI rolled out what it had been calling a “super app” – ChatGPT Work – combining its chatbot interface, coding assistant, and a set of new models into a single product. The platform is designed to operate both alongside users and independently on their behalf, blurring the line between assistant and autonomous agent.
ChatGPT Work launched alongside OpenAI’s GPT-5.6 models, compressing what might have been two separate news cycles into one. OpenAI is also separately developing a fully automated research tool, adding to a product roadmap that has accelerated significantly in 2025 and into 2026. The pace signals that OpenAI is treating workplace productivity as a primary battleground, not a secondary feature.
Meanwhile, Harvard engineering professor Vijay Janapa Reddi offered a pointed counterweight to the week’s product enthusiasm. “When we’re talking about AI, we love the hype, we get excited about it. The damn thing never actually lands in practice,” he told Wired. That skepticism is worth noting as both OpenAI and Anthropic push narratives about AI’s expanding capabilities.
Meta, for its part, moved to monetize AI access this week by introducing a paid developer tier for a new version of Muse Spark. The company also announced plans to begin producing its own AI chip in September – a move that would reduce its dependence on Nvidia hardware and give it tighter control over inference costs as it scales its AI products.
Chip Money, Surgical Robots, and a Geopolitical Tangle
South Korean memory chipmaker SK Hynix completed the largest US listing by a foreign company, raising $26.5 billion. Demand for AI data center infrastructure has driven the company’s profits sharply higher – though some analysts have characterized the scale of its share sale as a potential indicator of market overheating rather than just corporate confidence. SK Hynix’s bet is that AI infrastructure spending has enough runway to justify the raise; the reception to that bet remains mixed.
In a more direct sign of geopolitical friction inside the AI industry, Tencent is leading a deal to unwind Meta’s $2 billion acquisition of Chinese AI startup Manus. Beijing ordered the divestiture, and Tencent is now in talks to become Manus’s largest shareholder – reportedly at a floor price of $2 billion. Separately, the Financial Times reported that OpenAI and Google sold AI models to entities on US blacklists through Singapore-based subsidiaries of Alibaba, Baidu, and Tencent, raising compliance questions about how model access is distributed through corporate structures. Those findings add pressure to an already complicated relationship between US AI companies and Chinese technology groups.
On the robotics front, humanoid robots performed teleoperated surgery on living animals for the first time, removing gallbladders from pigs. The procedure was conducted under human control, not autonomously – but it marks a practical step toward robotic surgical systems that could operate in environments too hazardous or physically constrained for human surgeons.
From Dead Retinas to Pigeon Missiles: The Week’s Stranger Stories
Researchers successfully prompted resuscitated human retinas to respond to light ten hours after the donors’ deaths – a development described as a meaningful step toward eye transplants that restore functional vision. A separate device capable of reviving degraded eyeballs has also been reported, suggesting that vision restoration research is moving on multiple parallel tracks.
A daughter’s account of using an AI “death bot” – a model trained on her deceased father’s voice, writing, and conversational patterns – ran in The New Yorker this week. The experience, she reported, produced both comfort and a persistent unease that she could not fully resolve. The technology surfaces a question the AI industry has not yet answered cleanly: at what point does grief support become something more ethically complex.
And then there is the older thread that connects to all of it. In 1943, psychologist B.F. Skinner ran a classified government project aimed at making bombs more precise by training pigeons to guide missiles – rewarding birds with food when they pecked the correct targets on a screen inside the warhead. The project was ultimately shelved. But the reinforcement learning logic Skinner used with those pigeons – reward the right behavior, repeat, refine – is structurally identical to the training methods powering the same Claude model that Anthropic is now peering inside with its J-lens. The company’s relationship with its own creation keeps generating new questions; the J-space research just made some of those questions considerably harder to avoid.
What Anthropic found in that hidden space ranged, by its own description, from mundane to unnerving – and the company has not yet said which parts kept its researchers up at night.
