A recent AI research paper has unveiled surprising geometric structures within large language models (LLMs) like ChatGPT, revealing how these models organize information in a brain-like manner. The study, conducted by Max Techark, introduces tools called sparse autoencoders, which act like x-ray machines for AI, allowing researchers to peek inside and understand the organization of concepts. This newfound understanding is akin to looking under the hood of a well-functioning car, uncovering an unexpectedly organized engine. The researchers identified three distinct levels of organization that emerged naturally as the AI learned, rather than being programmed.
At the first level, referred to as the atomic structure, researchers discovered that the AI organizes concepts in geometric patterns. For example, when plotting relationships between words like “man,” “woman,” “king,” and “queen,” they form a parallelogram, illustrating how the AI understands gender relationships consistently across different contexts. This level of organization was initially obscured by irrelevant information, akin to light pollution, which needed to be filtered out to reveal the true patterns of concept relationships.
The second level of organization resembles the structure of the human brain, with distinct regions or lobes dedicated to different functions. The researchers identified three main lobes: the coding and math lobe, which specializes in programming and mathematical concepts; the general language lobe, which processes regular English text; and the dialog lobe, which focuses on conversational text. This organization emerged naturally, with features that activate together being physically close in the AI’s mental space, mirroring how neurons in the human brain operate.
At the third level, termed the galaxy structure, the researchers found that the AI’s knowledge is organized according to specific mathematical patterns, particularly in the middle layers of the model. These layers act as an information bottleneck, retaining only the most essential features for further processing. This efficient compression of information allows the AI to focus on key features while discarding irrelevant details, enhancing its performance across various tasks. The hierarchical structure of information follows power laws, indicating an organized and predictable pattern of feature importance.
The implications of these findings are significant for AI research, as they provide insights into how AI systems organize information, which can improve their versatility and efficiency. Understanding this internal organization can lead to targeted enhancements, reducing biases, optimizing computational efficiency, and making AI models more interpretable. While the similarities between AI structures and human cognition are intriguing, it is essential to note that these are organizational similarities rather than biological ones. The study opens the door for further research into the nature of intelligence, potentially offering insights into both AI and human cognition, as well as advancing fields like cognitive science and neuroscience.
