AI Daily Briefing — 2026-07-12

Anthropic has developed a new technique that offers an unprecedented glimpse into the inner workings of large language models, revealing a hidden layer of activity that can differ from what the model actually says. The AI firm created a tool called the Jacobian lens, or J-lens, to probe Claude Opus 4.6, a version of its flagship model released in February. This tool uncovered a space they named J-space, which contains words the model is considering for future responses—essentially, what it might be “thinking” before it speaks.

According to MIT Technology Review AI, the J-lens builds on previous interpretability research by adapting a logit lens to look deeper into the middle layers of the model. While a standard logit lens identifies words the model is likely to output next, the J-lens picks up words that may appear several steps ahead. This reveals concepts the model is processing but might not ultimately include in its final response, offering a clearer picture of its internal reasoning.

Anthropic claims that monitoring J-space provides a new way to understand and control its models, potentially improving safety and transparency. The company has published its findings in a paper and partnered with Neuronpedia to create an interactive demo for the public. Tom McGrath, chief scientist at Goodfire, praised the work as “very good and interesting,” noting that it exposes a deeper level of LLM computation that researchers had not previously seen.

In a significant development for U.S. energy, four nuclear reactors achieved criticality just in time for the nation’s 250th birthday, surpassing a goal set by the previous administration to have three microreactors reach this milestone. The achievement marks a positive step for nuclear power amid growing demand for clean electricity, though reaching criticality does not yet mean these reactors are ready to supply power to the grid. The companies behind them are now exploring next steps for deployment.

Meanwhile, China is reportedly preparing to allow its top artificial intelligence firms—including Alibaba, ByteDance, and DeepSeek—to purchase Nvidia H200 chips, reversing an earlier hold on approvals despite U.S. authorization. This move could accelerate China’s AI capabilities, though it also raises questions about technology export controls and national security.

In other tech news, NATO is building a sensor network using drones, satellites, and AI to detect Russian cyberattacks, while researchers are exploring solar geoengineering as a potential defense against future El Niños. Meta is also patenting an AI device that records users to analyze emotions, ostensibly for personalized workout plans, sparking fresh privacy concerns.

From chipmakers stacking transistors to keep Moore’s Law alive to paleontologists using robots to study extinct species, the week’s stories highlight how innovation continues to push boundaries—even as ethical and security debates intensify.

Anthropic has revealed a hidden layer inside its AI model Claude, offering an unprecedented look at how large language models process information before generating responses. Using a tool called the Jacobian lens, researchers discovered a space—dubbed J-space—where the model holds words and concepts it considers but may not ultimately use. This finding, reported by MIT Technology Review AI, gives a rare glimpse into the internal reasoning of AI, revealing both predictable and unsettling patterns.

Meanwhile, OpenAI has launched its long-anticipated “super app,” ChatGPT Work, which combines its chatbot, coding tools, and new models into a single platform. Designed to work alongside users or autonomously, the app aims to handle tasks from research to coding. The release coincided with the debut of OpenAI’s GPT 5.6 models, signaling a push toward more integrated and capable AI assistants.

In other AI developments, humanoid robots have performed teleoperated surgery on living animals for the first time, removing gallbladders from pigs. SK Hynix, a South Korean chipmaker, secured the largest U.S. listing by a foreign company, raising $26.5 billion amid surging demand for AI data centers. Tencent is also leading a deal to unwind Meta’s $2 billion acquisition of Chinese AI startup Manus, following Beijing’s order to divest.

From resurrected retinas responding to light after death to AI models sold to blacklisted Chinese groups via Singapore subsidiaries, the week’s news underscores AI’s expanding reach and ethical complexities. As one Harvard professor noted, despite the hype, AI often fails to deliver in practice—a reminder of the gap between promise and reality.

Four small nuclear reactors achieved a key technical milestone just ahead of America’s 250th birthday, meeting a deadline set by the previous administration. The goal was to have three new microreactors reach criticality—meaning they can sustain a nuclear chain reaction—by July 4. In a surprising turn, four companies succeeded, signaling progress for emerging nuclear technology amid rising electricity demand and climate concerns. However, reaching criticality does not mean these reactors are ready to generate power for the grid.

The Reactor Pilot Program, launched by the U.S. Department of Energy, fast-tracked prototype development by offering land and national lab support to 11 selected projects. All four successful reactors are microreactors, far smaller than traditional light-water plants. Antares Nuclear was first in June, followed by Valar Atomics, Deployable Energy, and Aalo Atomics, the latter hitting the mark in the early hours of July 4. The speed is notable, especially given the nuclear industry’s history of delays and cost overruns.

These reactors achieved only zero-power criticality—a test of chain reaction initiation without meaningful energy output. As former nuclear energy official Kathryn Huff noted, this milestone can be reached without major engineering progress on fuel or design. The companies now face significant technical hurdles, including adding cooling systems to transfer heat from the core. Aalo aims to power an on-site data center by 2027, while Deployable Energy targets commercial deployment by 2028.

Startup timelines in nuclear should be viewed cautiously, given the complexity of the technology and regulatory challenges ahead. The Nuclear Regulatory Commission has proposed a faster approval framework for microreactors, but its effectiveness remains unproven. These four companies have cleared an early hurdle, but the path to delivering clean electricity is still long.

A third-generation California dairy farmer, Anthony Agueda, is turning to an unlikely ally to tackle manure pollution: red earthworms. Using a method called vermifiltration, he and others are harnessing worms and microbes to clean wastewater from livestock operations. This natural approach could significantly reduce methane, nitrous oxide, and water contamination, offering a sustainable alternative as the industry faces mounting environmental pressure. The humble earthworm is emerging as a key player in reshaping farming practices.

Meanwhile, solar geoengineering is moving from theory toward practical application, but researchers are hitting sobering roadblocks. The concept of deliberately altering the climate to counter global warming now requires real-world infrastructure, including specialized aircraft and materials. Early deployment would demand substantial time and investment, challenging the notion of a quick fix. As scientists delve into the details, they are confronting the harsh realities of engineering such a massive intervention.

These developments highlight a broader shift in technology: from abstract ideas to tangible solutions with complex trade-offs. Whether through nature-based methods like vermifiltration or high-stakes climate engineering, innovators are grappling with the gap between promise and practice.