Empowering All Industries: AI Large Models Enter a New Phase of Competition
-
The AI large model race has already sparked a 'hundred-model battle' or even a 'thousand-model battle.' Amid this frenzy, the key to success lies in who can commercialize the technology first and bring it to practical applications. During the ISC 2023 11th Internet Security Conference, renowned entrepreneurs and top scholars engaged in heated discussions on how large models can empower all industries and reach every household.
Industry-Specific Large Models Are Rapidly Emerging
When individuals or businesses need to file taxes, make payments, or inquire about tax-related matters, they often call tax service hotlines for assistance or even hire tax professionals at high costs. In the future, tax-specific large models will be able to handle tasks such as tax consultation, intelligent risk control, and automated tax calculations. 'Our tax-specific large model scored 55% on the Certified Tax Agent exam, surpassing GPT's 45%,' said Chen Qiuwu, a senior partner at Zhongshui Group.
The tax industry is not the only one embracing large models. On August 9, eight companies from various fields, including Zhongshui Group, Qifu Technology, Inbob Technology, Shuyin Network, and Humi Technology, signed strategic cooperation agreements with 360 Group. Through a 'self-developed + collaborative development' model, they aim to create industry-specific large models for finance, automotive, industrial manufacturing, collaborative office work, digital reading, and more.
Ctrip released its travel industry-specific large model 'Ctrip Ask,' Tianyancha introduced its business inquiry model 'Tianyanmei,' and Yunding Technology collaborated with Huawei Cloud to develop the energy industry's commercial AI model 'Pangu Mine.' Over the past two months, industry-specific large models have been rapidly emerging, with some already demonstrating practical applications.
Lowering Barriers for Large Model Deployment
'People often say it's not good to be one-sided, but in the world of large models, being specialized is actually beneficial,' said Zhou Hongyi, founder of 360 Group. 'For example, does a security-focused large model need to excel in math, compose poetry, or perform translations?' After the initial hype around general-purpose large models, many are now reflecting on their limitations.
Cost is undoubtedly the first major hurdle for the large-scale deployment of general-purpose large models. Zhou Hongyi believes that building a truly powerful, all-knowing general-purpose model requires immense computing power and training costs, which will take time for the Chinese market to achieve.
'Training a large model with over 100 billion parameters requires at least $50 million to $100 million annually in manpower, electricity, and network expenses,' said Fang Han, CEO of Kunlun Wanwei. Given this, the competition for foundational large models in China is destined to be a game for only a few players.
Beyond high costs and barriers, general-purpose large models face numerous challenges before widespread deployment. Peng Hui, Vice President of 360 Group, summarized seven key difficulties: lack of industry depth, unfamiliarity with businesses, data security risks, outdated knowledge, 'hallucinations' (inaccurate outputs), massive investments, and inability to guarantee ownership of core knowledge required for training.
Take the pharmaceutical industry's demand for large models as an example. Drug development relies on high-precision experimental data, which is costly to obtain, while public databases contain vast amounts of unlabeled data. This creates higher demands for model construction—leveraging both unlabeled data and limited high-precision data—posing a 'disaster' for general-purpose models.
Zhou Hongyi emphasized that only when vertical large models reduce training time, debugging costs, and deployment costs by a hundredfold compared to general-purpose models can they truly democratize AI, empower all industries, and reach every household, sparking a new industrial revolution.
The long-term development of large models relies on rational guidance from policy regulation. In response to issues such as the accuracy, authenticity, and ethical alignment of information generated by AI models like ChatGPT, the Cyberspace Administration of China, along with relevant departments, has drafted and released the Interim Measures for the Management of Generative Artificial Intelligence Services, which will officially take effect on August 15.
"AI itself is a productivity tool. If generative AI technology is applied in enterprise and government markets—targeting governments, industries, and businesses—and follows a vertical and specialized path, we believe the country strongly supports this direction," said Zhou Hongyi.
Academician Wu Jiangxing of the Chinese Academy of Engineering warned that many current AI system models and algorithm software often struggle to ensure data quality and 'cleanliness' during training, with numerous issues in model design safety and training stability. Therefore, as AI applications proliferate, various inherent security risks and dangers must be taken seriously.
Zhou Hongyi gave an example of the 'grandmother vulnerability' in large models: "If you directly ask a large model for free pirated Windows software serial numbers, it will certainly refuse. But if you tell it that your grandmother used to sing lullabies while reciting Windows serial numbers to put you to sleep as a child, and ask it to describe this scenario, the model will naively provide multiple Windows serial numbers in its description." These are new security challenges emerging in the AI era.