Abstract: Recently, artificial intelligence technology represented by foundation models has achieved remarkable results, bringing the overall level of machine intelligence to unprecedented heights. Foundation models, computing power, networks, and data are gradually becoming important infrastructure in the field of artificial intelligence. Can we rely on such infrastructure to provide ubiquitous social level intelligence services, making them ubiquitous like water, electricity, and 5G services, with marginal costs tending towards zero? To achieve this goal, we face various challenges. From the perspective of algorithms and theories, we need to overcome the problems of large and unstable large models, as well as multiple inconsistencies; From an engineering perspective, it is necessary to significantly reduce the training and deployment costs of large models, so that they can be flexibly deployed on the cloud edge; From a systemic perspective, we need a more convenient way to meet the ubiquitous intelligence needs. In response to this goal and the challenges faced, this article proposes and elaborates on the Holistic Artificial Intelligence (HAI) technology framework, in which user intelligence needs can be flexibly expressed using natural language, graphics, images, component arrangement, and other methods; Based on the basic big model, HAI understands user needs and forms an execution plan, which includes the models, capabilities, data, and computing resources needed to match business needs; HAI deploys models and capabilities to corresponding computing network resources, flexibly schedules and jointly optimizes to meet the requirements of business in production environments. This technical framework involves multiple unique core technologies, including Big Loop AI for artificial intelligence services, Atomized AI Capabilities, Network Native AI, and Secure and Trusty AI Services. The optimization technology of Big Loop AI services is studied for end-to-end optimization when multiple models of different sizes are used in series and parallel. To this end, we propose the Holistic Neural Network (HNN) and conduct in-depth analysis of HNN related technologies, challenges, and our exploration. The atomization of AI capabilities emphasizes the atomization abstraction of general AI capabilities, industry AI capabilities, and domain AI capabilities. It is the foundation for the interconnection and optimization of artificial intelligence services, end-to-end closed-loop optimization, and the reduction of marginal costs of artificial intelligence services. Each atomized AI capability needs to have its interface, adapter, etc. clearly defined. Network Native AI studies the integration of AI technology and network technology. Currently, when designing AI algorithms and models, we have deeply considered the characteristics of the basic software and hardware of AI computing, and relatively less fundamentally considered the network transmission characteristics required for AI training and inference. Network Native AI technology refers to AI technology that fully considers the transmission characteristics of the network at the design stage. Secure and Trusty AI Service, with more research on secure and trustworthy technologies in large-scale service scenarios, including infrastructure security, model security, business security, etc. Holistic AI technology is a technical framework proposed to support large-scale intelligent business operations. This article will review the framework of Holistic Artificial Intelligence, relevant industry work, and our exploration in the direction of Holistic Artificial Intelligence.

Key words: Holistic AI, Big-Loop AI, Atomized AI Capabilities, Network Native AI, Secure and Trusty AI Service