Abstract Current artificial intelligence, represented by large language models, follows a "general-purpose universal" path – using massive parameters to cover all scenarios. This path is functionally successful but energy-unsustainable. Nature offers another answer: model organisms. The fruit fly accomplishes complex behaviors with 139,000 neurons; bats use sonar rather than vision; mice use smell rather than hearing – they are not universal, but they are optimal within their ecological niches. This paper proposes a 3+2 structural framework to systematically analyze why model organisms are efficient: Three Layers of Static Structure: Model Organism Structure (niche specialization), Fractal Structure (self-similar organization), Pathway Structure (fixed conduction pathways) Two Types of Dynamic Processes: Dynamic Modulation (real-time scheduling), Structural Change (long-term evolution) Based on this framework, we synthesize four insights from model organisms for AI: ① Shift from general-purpose universality to scenario specialization; ② Structural health matters more than parameter count; ③ Synergy between fixed hardware and flexible modulation; ④ Evolutionary perspective in AI design. SC‑Net as the engineering implementation of this concept, demonstrates how to achieve superior functionality with minimal energy consumption in edge-side scenarios. This paper, starting from brain science, neuroscience, and computational neuroscience, integrates multiple lines of evidence from the quantum to the clinical, providing a new intellectual resource for low-energy artificial intelligence. 摘要 当前人工智能以大语言模型为代表,走的是“通用万能”道路——用海量参数覆盖所有场景。这条路在功能上成功,但能源不可持续。自然界给出了另一种答案:模式生物。果蝇用13.9万神经元完成复杂行为,蝙蝠用声呐而非视觉,老鼠用嗅觉而非听觉——它们不是万能的,但在自己的生态位里最优。 本文提出一个3+2结构框架,系统解析模式生物为何高效: 三层静态结构:模式生物结构(生态位特化)、分形结构(自相似组织)、通路结构(固定传导路径) 两种动态过程:动态调制(实时调度)、结构变化(长期演化) 基于这一框架,我们归纳模式生物给人工智能的四重启示:①从通用万能转向场景特化;②结构健康重于参数数量;③固定硬件与灵活调制的协同;④演化视角下的AI设计。SC‑Net20作为这一思想的工程实现,展示了如何在端侧场景中以最小能耗实现更优功能。 本文以脑科学、神经科学和计算神经科学为起点,整合从量子到临床的多重证据,为低能耗人工智能提供一个新的思想资源。 关键词:模式生物;结构解码;分形几何;动态调制;快慢分离;高阶关联;场景特化;低能耗AI;Transformer;人工智能;神经连接结构;脑科学;神经科学;计算神经科学;神经连接组;神经可塑性;脑机接口;阿尔茨海默病
zhengda li (Fri,) studied this question.