The artificial intelligence big model has already sparked the "Hundred Model War" and even the "Thousand Model War". Under the craze, who can take the lead in landing and monetizing technology undoubtedly becomes a key competitive point on the big model field. During the 11th ISC 2023 Internet Security Conference, renowned entrepreneurs and top scholars engaged in a heated discussion on how to empower big models to enable businesses and reach households. The acceleration of the emergence of large industry models often requires individuals or enterprises to call the tax service hotline for manual assistance when conducting tax declaration, payment, and query operations, and even hire tax advisors with high salaries to take charge of related work. In the future, the tax big model will be able to share a portion of tax consulting, intelligent risk control, automatic tax calculation, and other tasks. Our tax model can achieve a score of 55% in the registered tax agent exam, exceeding 45% of the GPT, "said Chen Qiuwu, senior partner of China National Taxation Group. The industry that actively embraces big models is not just about taxation. On August 9th, eight enterprises in different fields, including China National Taxation Group, Qifu Technology, InBev Digital Science and Technology, Digital Network, and Humi Technology, signed a strategic cooperation with 360 Group to create industry models for various industries such as finance, automotive and motorcycle, industrial manufacturing, collaborative office, and digital reading through the "independent research and development+cooperative research and development" model model. In the past two months, Ctrip has released a vertical big model for the tourism industry called "Ctrip Quest", Tianyancha has launched a business search big model called "Tianyanmei", and Yunding Technology has collaborated with Huawei Cloud to develop a commercial AI big model for the energy industry called "Pangu Mine"... In the industry, vertical big models have accelerated to emerge and have already been applied in typical cases. Lowering the threshold for the implementation of large models "People generally say that partial science is not good, but it is good to be a partial student in large models." Zhou Hongyi, the founder of 360 Group, said, "For example, for the security model, does it need to understand Olympiad? Does it need to be able to write ancient poetry? Does it need automatic translation?" After the popularity of the general model for a while, everyone is reflecting on some of the problems that exist in the general model. Cost is undoubtedly the first "natural barrier" that needs to be crossed before the large-scale implementation of the universal large model. Zhou Hongyi believes that the computational power and training costs required to truly create a super strong "omniscient" universal model are very high, which will take some time for the Chinese market. For large models with a scale of over a hundred billion, training requires investment in manpower, electricity, network expenses, etc., ranging from at least $50 million to $100 million per year. "According to Fang Han, CEO of Kunlun World Wide, based on this estimation, the battle for Chinese base type large models is destined to be a game for a few players. In addition to high investment and high threshold, universal large models still need to solve many difficulties before being implemented on a large scale. Peng Hui, Vice President of 360 Group, summarized the difficulties in implementing the universal big model as seven points: lack of industry depth, lack of understanding of the enterprise, data security risks, untimely knowledge updates, "nonsense", huge investment, and inability to guarantee ownership of the core knowledge required for training the big model. Taking the demand for large models in the AI pharmaceutical industry as an example, due to the high cost of obtaining high-precision experimental data in drug research and development, and the large amount of unlabeled data in public databases, the requirements for model construction of large models will be higher. It is necessary to make good use of a large amount of unlabeled data