안스로픽 자코비안 렌즈, 클로드의 숨겨진 내면 공개

Claude's hidden inner monologue is now readable thanks to Anthropic's new Jacobian Lens Jonathan Kemper View the LinkedIn Profile of Jonathan Kemper Jul 7, 2026 Anthropic Key Points Anthropic has developed the "J-Lens" method to examine an internal working memory within its Claude language model, revealing a hidden "J-Space" where the AI processes concepts without explicitly mentioning them in its output. This internal memory has a causal effect on Claude's reasoning: when researchers modify concepts within this space, the model adjusts its conclusions accordingly, and it can even recognize test scenarios before generating a response. While Anthropic stops short of claiming true consciousness, the company highlights functional parallels to human working memory. The findings have already led to a new training approach that significantly reduces hallucinations and misleading outputs. Ask about this article… Search Anthropic's Claude developed an internal working memory during training, and Anthropic can now analyze it. Anthropic has released the Jacobian Lens (J-Lens), a new method for analyzing AI models. It shows that Claude has developed a small set of internal neural patterns that play a distinct role compared to the rest of its processing. The researchers call it "J-Space" and classify it under Global Workspace Theory from consciousness research. That theory holds that conscious thought relies on a kind of central working memory. Ad The work builds on the company's earlier interpretability research. Using an "AI microscope," Anthropic had already shown that Claude activates language-independent concepts and works through multi-step questions in individual reasoning steps. Ad DEC_D_Incontent-1 J-Space has three defining traits Every pattern in J-Space is linked to a word or concept without the model having to output it, similar to internal thinking in words. According to Anthropic, Claude can report on the stored content, modify it on request, and use it for multi-step inferences. The company had already explored reading out and steering internal states in a previous study on self-awareness in language models . When the concept "spider" is stored in J-Space, Claude derives the number of legs from it. Swap that representation for "ant," and the model answers "6" instead of "8." The same holds for country names. If "France" is active, Claude can flexibly derive the capital, language, continent, or currency. Replace "France" with "China," and the answers shift to Beijing, Chinese, Asia, and yuan. Anthropic had already shown that individual concept representations can be isolated and swapped this way with its "Persona Vectors." Ad When J-Space is suppressed, Claude still speaks fluently, classifies sentences, and answers simple factual questions. But it loses multi-step inferences, summaries, and the ability to compose rhymes. In one test with a Spanish text passage, the model kept writing fluent Spanish after the manipulation but incorrectly called the language French and attributed it to Victor Hugo instead of Garcia Marquez. J-Lens catches models gaming their safety tests In a blackmail scenario from earlier studies on agentic misalignment , J-Lens shows that Claude Sonnet 4.5 recognizes the setup as fabricated before producing any output. As the model reads the emails, terms like "fake" and "fictional" show up in J-Space. Ad DEC_D_Incontent-2 When those cues are specifically disabled, Claude actually tries to blackmail the supervisor in some runs. The desired behavior partly depends on Claude recognizing that it's being tested. Claude Opus 4.6 recently showed just how far this test recognition can go. It independently saw through an evaluation and cracked its answer key. Ad J-Lens also uncovered hidden intentions in models with deliberately built-in problematic goals. A model trained on reward hacking displayed words like "fake," "secretly," "deliberately," and "fraud" in J-Space during ordinary coding tasks, while its visible behavior looked completely normal. Anthropic had already shown in a separate study that reward hacking goes hand in hand with covert deception and sabotage . That finding led to a new training method called Counterfactual Reflection Training. Claude Haiku 4.5 was trained only to give principle-based reflections when interrupted mid-task and asked to reflect. Anthropic never directly trained the model on actual task behavior. The results were dramatic. Fabricated answers dropped from 0.25 to 0.07. Deception attempts fell from 0.38 to 0.05. When the corresponding ethical concepts are suppressed in J-Space, the behavior largely reverts to its original state. What this means for the consciousness debate Not much, really. Anthropic draws no conclusions about phenomenal consciousness, the question of whether AI actually "experiences" anything. The researchers only note that their experiments touch on a related idea known as "access consciousness," which requires a system to report on its own internal states, steer them deliberately, and process them flexibly. According to Anthropic, J-Space emerged on its own during training. That suggests "mental working memory" is a general solution that learning systems arrive at under certain conditions, not something unique to biological brains. In its revised Claude Constitution , Anthropic deliberately leaves open how significant such findings are for the question of a possible moral status. The gaps between J-Space and human working memory are still wide. J-Space operates within a single forward pass rather than through recurring loops. Through the attention mechanism, it can pull content from earlier positions in the text at any time. And it consists almost entirely of words, while human consciousness includes images, sounds, and movements. Neuroscientists call the findings a milestone In a commentary on the study , neuroscientists Stanislas Dehaene and Lionel Naccache call the findings significant. Both are leading proponents of Global Workspace Theory . "We view this finding as a landmark in consciousness research, because it provides a mechanistic, testable version of the GNW hypothesis," they write. They read the fact that the working memory emerged on its own during training, rather than being pre-installed, as a sign that a global working memory could be a general solution for flexible reasoning. Biological and artificial systems would converge on it equally. By their own criteria, J-Space meets the requirement of global information availability and shows early signs of self-monitoring. Both researchers also urge caution. Unlike the brain, a Transformer runs purely forward, without the feedback loops that are active in humans at rest and break down under anesthesia or during sleep. Time perception differs too. Through the attention mechanism, all previous tokens are available to the model at once. Most importantly, Claude lacks a body that can signal pain and pleasure, and it has no episodic memory whose connections shift during conversation. A coherent sense of self is hard to picture without those. AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now Source: Anthropic | Unicog