📄 中文摘要
当前大型语言模型（LLM）在通用任务上表现出色，但其高昂的训练和推理成本、资源消耗及难以部署到边缘设备的局限性日益凸显。本文提出 Interfaze 范式，倡导构建由多个任务专用小型模型（TSMs）组成的生态系统，以替代单一大型通用模型。Interfaze 旨在通过模块化、可组合的 TSMs，实现更高效、更经济、更具可扩展性的 AI 解决方案。TSMs 专注于特定任务，可显著降低模型规模和计算需求，从而降低训练和推理成本，并促进 AI 在资源受限环境中的广泛应用。Interfaze 强调 TSMs 之间的协同与集成，通过智能路由和编排机制，实现复杂任务的分解与协作处理。这种范式不仅提升了 AI 系统的效率和可解释性，还为 AI 模型的持续演进和个性化定制提供了灵活框架。Interfaze 展望了 AI 发展的新方向，即从“大而全”转向“小而精”，通过分布式、专业化的模型群，构建更可持续、更普惠的 AI 未来。
📄 English Summary
Interfaze: The Future of AI is built on Task-Specific Small Models
While large language models (LLMs) have demonstrated remarkable capabilities across various general tasks, their inherent limitations, including exorbitant training and inference costs, substantial resource consumption, and challenges in deployment to edge devices, are becoming increasingly apparent. This paper introduces the Interfaze paradigm, advocating for an ecosystem composed of numerous task-specific small models (TSMs) as a viable alternative to monolithic, general-purpose large models. Interfaze aims to foster more efficient, economical, and scalable AI solutions through modular and composable TSMs. By focusing on specialized tasks, TSMs can significantly reduce model size and computational requirements, thereby lowering both training and inference costs and facilitating broader AI adoption in resource-constrained environments. The Interfaze framework emphasizes synergy and integration among TSMs, utilizing intelligent routing and orchestration mechanisms to decompose and collaboratively process complex tasks. This paradigm not only enhances the efficiency and interpretability of AI systems but also provides a flexible framework for the continuous evolution and personalized customization of AI models. Interfaze envisions a new direction for AI development, shifting from a 'one-size-fits-all' approach to a 'small and specialized' one, building a more sustainable and inclusive AI future through distributed, expert model collectives.