物理学模型揭示人工智能学习新机制

Artificial intelligence systems built on neural networks—like OpenAI’s ChatGPT, Anthropic’s Claude, DeepSeek models, or Google’s Gemini— have reached a level of performance that rivals or exceeds humans in many tasks. Yet despite their capabilities, the internal processes that drive their behavior remain difficult to interpret. Researchers often describe these systems as “black boxes,” where inputs and outputs are visible but the reasoning in between is not fully understood. A new physics-inspired model is beginning to change that. By applying concepts from statistical mechanics and energy systems, researchers are developing simplified frameworks that explain how neural networks learn and organize information. …