Summary of Key Points
This news article focuses on two significant developments in the field of AI: First, the Anthropic team discovered that their AI model Claude spontaneously developed an internal thinking structure called J-space during training. This structure functions like a “draft paper” for humans, allowing it to temporarily store concepts and engage in complex reasoning, which is seen as a sign of approaching Artificial General Intelligence (AGI). Second, Turing Award winner David Wang publicly criticized the current definition of AGI, arguing that current large language models (such as ChatGPT and Claude) are far from true intelligence. He questioned the concept of “generality” (the “G” in AGI) and proposed his own direction for AI research.
1. What is J-space?
You can think of J-space as Claude’s “brain draft paper.” When Claude performs simple tasks, such as recognizing emotions, answering common-sense questions, or speaking grammatically correct sentences, it completes these tasks automatically without the need for J-space. However, when faced with complex tasks like reasoning, writing poetry, or translating, J-space comes into play to store intermediate thoughts. For example, if asked how many legs animals that can form webs have, J-space might initially suggest “spider,” leading to the answer “8.” If researchers replace “spider” with “ant” in J-space, the answer changes to “6.”
Anthropic used a tool called J-lens to translate Claude’s internal digital signals into language that humans can understand and thus discovered J-space. They also observed that J-space displays errors when Claude reads bug code and contains terms related to biological functions when reading protein sequences. These observations indicate that J-space is not just a record of results but actively participates in the reasoning process.
2. The Spontaneity of J-space: Why Does It Surprise Scientists?
The most surprising aspect for Anthropic is that J-space was not designed by them; it emerged naturally within Claude. This is similar to mitochondria in human cells, which were originally independent bacteria that became part of larger cells through symbiosis, though their origin remains unknown. The emergence of J-space suggests that AI models may be spontaneously developing internal structures similar to biological intelligence, rather than merely relying on the statistical patterns in external training data.
In the past, we judged whether a model was AGI based on its ability to solve problems or write code (external behaviors). The discovery of J-space highlights the need for an internal structure that can integrate language, knowledge, and reasoning capabilities. This structure is essential for true intelligence, as it allows an AI to think independently.
3. David Wang’s Criticism of AGI: Why “G” (Generality) Is a Myth
David Wang, one of the pioneers of deep learning, recently stated on X that the concept of “G” in AGI is nonsense. His arguments are compelling:
- Current models can only pass “tests” (such as language-based exams or math competitions), which are tailored to their strengths (statistical analysis).
- They lack real-world intelligence, such as the ability to perceive things (e.g., recognizing a pen standing on its tip) or understand causal relationships (e.g., knowing that releasing a object will cause it to fall).
- He opposes the idea that increasing computing power will lead to AGI, arguing that language data cannot capture all aspects of the real world, and pure regression-based predictions (like ChatGPT’s word-for-word generation) are not efficient methods for reasoning.
- He also dismisses doomsday scenarios involving AI destroying humanity, suggesting that current models are far from achieving autonomous intelligence.
4. David Wang’s Approach: What Is He Doing?
David Wang left Meta at the end of 2025 and founded AMI Labs, which received a $1.03 billion seed round of funding, including investments from NVIDIA and Jeff Bezos. His research focuses on enabling AI to learn causal relationships and planning from perceptual data (images, videos). He uses the JEPA architecture to make AI behave more like humans: observe the world, understand patterns, make predictions, and plan actions, rather than relying solely on language interactions. Only in this way can AI truly understand the world and stop being a mere “language repeat machine.”
5. The Intersection of J-space and David Wang’s Vision
While Anthropic sees J-space as a step towards AGI, David Wang is skeptical, arguing that it still resides within a pure language model and lacks real-world perception and interaction. In his view, this represents an internal optimization of the language model, not true intelligence capable of understanding the physical world and making autonomous plans.
This news article reflects the debate in the AI community between two approaches: one that relies on large models and increased computing power to achieve AGI spontaneously, and another that emphasizes starting with perception to enable AI to understand the world before pursuing general intelligence. Which approach do you support?