China
In January, Chinese authorities announced with great fanfare that they had approved more than a dozen large-scale language models (LLMs). LLM will be used to train generative artificial intelligence software to produce human-like responses when generating text, images, and videos.
Among the 14 approved so far this year is industry-specific LLM, a deep learning AI algorithm that can summarize, predict, and generate content using very large datasets.
However, many of these indicate that China is focusing on applying generative AI, and not just building LLMs, unlike more general AI models approved in the past.
Most of the official recognition is for LLMs developed by some of China's largest private technology companies.
But recently, a state-owned company owned by the Chinese Academy of Sciences (CAS) Institute of Automation has joined a number of developers seeking approval. One of its selling points is that, as a state research agency, it can leverage government, academic and business data to develop applications, including for use in schools and universities.
The company was founded by three CAS researchers, nurtured by the CAS Institute, and is known as Zhongke Wange. It already has more than 1,000 paying customers, including state-owned enterprises, banks, and several universities, including Tsinghua University.
The company is currently seeking approval for its larger Yayi model, which was announced last June. This model can accurately answer voice questions. According to the company, it easily passed multiple tests, including AI writing, AI drawing, and conversational ability in specified scenarios.
For universities, Zhongke Wangke's video generation services as part of its content creation software can be equipped with AI presenters who can speak on any subject.
Teachers can ask the engine to research a topic and generate what the AI presenter will say out loud. It is believed that this can replace classes in some cases. The company describes this as a “complete media ecology.”
Zhongke Wenge says it has established a “panoramic international communication intelligent cockpit” that integrates information insights, news gathering and editing, AI-generated content production, communication analysis and digital asset management tools.
Last year, Wang Lei, chairman of Zhongke Banke, said that almost all large-scale models are aimed at general tasks and do not take into account deep applications in specific industries. He pointed to issues such as high computing costs, difficult localization deployments, high data security risks, and lack of domain-specific knowledge.
Yayi is designed to help you understand complex data, not just chats. You can focus on specific industries as well as general task functions. And it supports the deployment of proprietary data and private domains, he said.
Zhongke Wange is of particular interest because it offers an AI-driven multilingual service that can monitor, translate, and generate text in approximately 42 languages, according to public relations materials.
But even more controversially, the company provides services to the Chinese government's propaganda department to monitor foreign content and create content for the department aimed at overseas audiences. Thing.
One commentator on Chinese social media platform Weibo said this meant China could “flood foreign media and drown out anti-China voices.” Other Chinese Communist Party officials commented that China “doesn't have a strong enough voice” internationally and that AI could help redress the balance.
Application for exam
Among other organizations gaining attention is Chinese startup Learnable.ai, which uses an AI-assisted scoring process to revolutionize the scoring of the annual college entrance exam known as “Gao.” I'm trying to do that.
In 2023, approximately 13 million students in China received high marks. Grading and scoring many papers requires thousands of human raters, and the process takes several weeks.
Guan Wang, director of the Ningbo Artificial Intelligence Institute at Shanghai Jiao Tong University and co-founder of Learnable.ai, calls this 1 based on the company's AI technology that can quickly create scoring criteria and compile evaluations. I believe it can be dramatically shortened to less than a day. within a few hours.
The AI model has been tested as an auxiliary tool in several Chinese provinces since 2021, including Hunan, Sichuan, Guangxi, and Jiangxi, and the AI's ability to decipher handwritten characters allows it to be easily scored by humans. A higher level of accuracy has been demonstrated, Wang said.
“Gaokao is a highly confidential exam in China, so it has passed rigorous scrutiny from various government departments,” Wang said in Davos, Switzerland, last month, Bloomberg News reported.
“Expert reviews found that the accuracy of using AI models is actually higher than humans, as students' handwriting is messy and difficult for teachers to identify,” Wang added.
The company is also developing a consumer app called Dakaotong, which is based on a database of exam questions for school students. The company says the app can grade practice exams, identify mistakes, and provide personalized feedback and follow-up practice. It's currently in test mode with a few thousand users, and Wang hopes to release it to the public after getting approval from the Chinese government.
Wang said that Dagaotong City aims to reduce the current urban-rural disparity in educational opportunities in China by leveraging AI and mobile technology.
“China's best educational resources are concentrated in places like Beijing, Shanghai and Shenzhen, but there is a shortage of qualified teachers in vast rural areas,” he said. “Since most parents at least have a mobile phone, the cost of using his AI model for inclusive education will be very low.”
Generative AI technologies have become essential tools for acquiring and disseminating knowledge. It affects teaching, learning, and even school management,” Tongji University President Zheng Tsinghua said in a speech at the 2024 World Digital Education Conference in Shanghai this week.
He said that various models developed by technology companies and universities are being applied to educational activities, and that the majority of university students in China use technologies such as Erniebot, an AI chatbot developed by Baidu. He said he is doing so.
But he added, “One of the emerging goals of AI-enabled education is to innovate the theories and methods of tackling engineering and technical problems, rather than just answering common questions.” .
Numerous LLMs
According to official media reports, China has developed more than 130 LLMs, accounting for about 40% of the global total, second only to the United States with a 50% share.
Chinese companies are racing to launch LLMs to compete with OpenAI's ChatGPT and GPT-4. ChatGPT is banned in China. That's in part because it's trained on data outside China's internet firewall, which blocks out Western influence.
Dou Dejing, an adjunct professor at Tsinghua University's School of Electronic Engineering, told officials that China's top AI large-scale models have reached the level of GPT-3.5, and the technological gap with GPT-4 is narrowing. Global Times Newspaper for December.
Last March, Baidu became the first major Chinese tech company to launch its own AI chatbot, Erniebot. Erniebot also follows China's censorship rules and does not answer questions that the Chinese government deems politically sensitive or taboo.
Ernie Bot received government approval and was released to the public last August. The company announced last month that it now has a total user base of 45 million people and 54,000 developers. Baidu announced Ernie Bot 4.0 in October and claimed that OpenAI's Ernie Bot 4.0 is as powerful as GPT-4.
But experts say the country's LLM market is constrained by high costs, lack of access to advanced chips, strict regulations, censorship of sensitive topics, high development costs and a fragmented market.
