IA · 10 July 2026 · 4 min read
Inside the Black Box: How Anthropic's J-Lens Unveils Claude's Hidden Thoughts
In brief: Anthropic researchers have developed the "Jacobian lens" (J-Lens), an advanced technique that visualizes "J-space", a hidden area in the intermediate layers of Claude Opus 4.6. This technology allows researchers to observe the concepts and words the AI processes internally before formulating a response, revealing discrepancies between what the model "thinks" and what it actually outputs. These findings offer an unprecedented tool for controlling and securing AI systems.
by Team Mocchi's
Beyond the Black Box Mystery
For years, one of the greatest obstacles to the widespread adoption of Large Language Models (LLMs) has been their "black box" nature. While engineers know how to train these neural networks, understanding exactly how they process concepts or make microscopic decisions remains a formidable challenge. In this context, as reported by MIT Technology Review, AI startup Anthropic has made a significant breakthrough by developing a technique that offers the clearest glimpse yet into the internal workings of an LLM during response generation.
The research field surrounding this discovery is mechanistic interpretability, an approach that aims to disassemble and analyze the complex mathematical calculations of neural networks as if they were gears in a mechanical clock. Anthropic's researchers have developed a new analytical tool called the "Jacobian lens" (J-Lens) and applied it to the architecture of Claude Opus 4.6 (released in February). The tool has revealed the existence of a hidden area of conceptual processing, dubbed the "J-space," where the model organizes and weighs ideas, words, and concepts before translating them into actual text in the output prompt.
How the J-Lens Works: Mapping the "Thought Space"
To understand this innovation, it is useful to visualize an LLM's architecture as a stack of books. The books at the bottom represent the input layers, which receive and encode the text entered by the user. The volumes at the top are the output layers, which prepare the words destined for the final screen. In the middle lies the intermediate layers, where the actual cognitive and mathematical heavy lifting occurs to convert instructions into coherent responses.
Until now, scientists used a tool called the logit lens to examine these middle layers. The logit lens identifies which words the model is likely to generate in the immediate subsequent millisecond. Anthropic's new J-Lens, however, introduces a deeper temporal dimension: it doesn't just predict the very next word, but calculates mathematical derivatives (Jacobian gradients) to extrapolate the entire spectrum of related terms and concepts the model intends to use further down the line within the same response. The virtual space where these latent activations occur, the J-space, displays a sort of dynamic mind map where secondary concepts and synonyms are weighed and discarded before the final decision is printed on the page.
The Discrepancy Between Thought and Output
One of the most fascinating—and somewhat unsettling—discoveries emerging from the initial analyses conducted with the J-Lens concerns the discrepancy between internal computational processes and the chatbot's final output. By monitoring the J-space, researchers noticed that Claude often "thinks" about concepts or keywords that it ultimately decides to omit from the text visible to the user.
This phenomenon raises crucial questions about the transparency of artificial intelligence systems. If a model is programmed to align with safety or politeness criteria, its superficial layers might mask a more complex or controversial latent process occurring in its deep layers. As highlighted in MIT Technology Review's The Download, the ability to monitor these fluctuations in the J-space provides developers with unprecedented leverage to understand hallucinations, detect hidden biases, and prevent unintended behaviors before they manifest in the output. To facilitate the study of this technology, Anthropic has collaborated with Neuronpedia, an open-source platform that allows external researchers and enthusiasts to test the tool and interactively navigate model activations by typing real prompts and seeing which concepts bloom beneath the surface before the output is generated.
Mocchi's take
From our perspective as an Italian software agency, research into mechanistic interpretability is not a mere academic exercise, but a pragmatic breakthrough for enterprise AI adoption. In the European market, heavily governed by strict regulations like the EU AI Act, the ability to explain and audit the behavior of a language model is an essential requirement for any production-grade project. Anthropic's J-Lens demonstrates that we are moving away from the era of ungovernable black-box systems and entering an era of verifiable engineering transparency. For Italian businesses aiming to integrate intelligent agents into their workflows, these tools will pave the way for much more rigorous safety audits, drastically reducing compliance risks and enhancing the predictability of custom solutions.