Just now, Anthropic discovered that Claude "grew" consciousness! Liked by Google DeepMind scientists

Compiled by Zhidongzhi | Edited by Cheng Qian | Li Shuiqing Zhidongxi reported on the 7th that in the early hours of this morning, Anthropic published a paper proving that they can see Claude's "inner thoughts"! New research on Anthropic found that there is a differentiation mechanism inside Claude that is highly similar to the human brain: only a small part of the internal information can be recalled and thought like human conscious thoughts, and most of the operations are automatically run at the bottom. The area running at the bottom is J space. To use an example, when Claude answers "How many legs does that animal that can weave a web have?", although the answer is only the number 8, the concept of "spider" will appear in the J space. J space is the key for Claude to perform multi-step reasoning, answer subjective description questions, and be competent in higher-level thinking tasks. Moreover, this mechanism is not artificially designed, but emerges naturally from the model during the training process. In addition, the researchers also found that it can be implanted to change the content in the Claude J space, thereby artificially intervening in the output of the model. Neel Nanda, head of the Google DeepMind interpretability research team, said that this is a very valuable paper, which provides sufficient and strong evidence to prove that there is some kind of "cognitive space" inside the large model. They have reproduced all the core conclusions of the paper on the Qwen3.6-27B model. Anthropic has released an open source code base containing these core methods and has partnered with Neuronpedia to provide an interactive demonstration based on the open weight model. Interactive demonstration interface: https://t.co/hkmYl0IOn2 1. The large model also has a human brain-like consciousness space with five major characteristics. For humans, when reading, the neural circuits of the human brain will adjust posture, control breathing, and at the same time convert the text and symbols on the screen into words that humans can understand. These processes are often unconscious. However, humans can perceive some of the activities that occur in the brain, such as images that suddenly appear in the mind or planned shopping trips.