Four Chinese generative AI models pass official assessment

The first group of four large generative artificial intelligence (AI) models passed the official "Large Model Standard Compliance Assessment" in China during the weekend, paving the way for faster and better regulated development of the AI generated content (AIGC) industry in the world's second-largest economy, observers said.

The move will also establish a compliance directory for the sector, where a growing number of large language models were launched this year.

Four large models that went through government assessment are developed by 360 Group, Baidu, Tencent and Alibaba. The assessment was initiated by the China Electronics Standardization Institute under the Ministry of Industry and Information Technology. The evaluation focused on a multi-domain, multi-dimensional model evaluation framework and indicator system, covering the generality, intelligence and security of large models. It encompassed multiple modalities, including language, speech and visual content.

In a statement sent to the Global Times on Sunday, Baidu said its ERNIE Bot, which passed the official assessment, has accumulated more than 70 million users, covering 4,300 different scenarios as of November.

Throughout this year, ERNIE Bot has undergone rapid upgrading. In May, it was upgraded to version 3.5, and in October, it was upgraded to version 4.0, achieving significant improvements in four major capabilities - understanding, generation, logic and memory, the company said.

Alibaba said in a statement sent to the Global Times on Sunday that its AI model Tongyi Qianwen is the only open-source model among the first group of large AI models evaluated this time.

"Tongyi Qianwen has a broad user base among developers and corporate clients globally, and its performance and security have undergone extensive public scrutiny," Alibaba said.

The national assessment has raised the threshold for large models, and leading Chinese AI models are likely to emerge from the four companies, while other large models may focus on industry-oriented applications, Pan Helin, a professor at Zhejiang University's International Business School, told the Global Times on Sunday.

Pan noted that the early intervention of government regulation will better nurture industry growth. "Rather than disrupting and reshaping the industry in the future, it would be better to regulate it from the beginning," he said.

Chinese companies have been racing to launch AI large models to compete with OpenAI's ChatGPT and GPT-4. Earlier, Baidu, SenseTime and AI start-ups such as Baichuan Intelligent Technology and Zhipu AI released their AI chatbots to the public.

According to media reports, China has more than 130 large language models, accounting for 40 percent of the global total, only behind the 50 percent share of the US.

Zhou Hongyi, founder and chairman of 360 Security Technology, told the Global Times that as every company, both Chinese and foreign, is striving to develop AIGC models and hardware, the key to global competition at the next stage is to apply the technology in real scenarios.

Zhou indicated that this will also be the focus of 360.

"One of our goals is to make online searches more AI-powered based on the 360 Brain Model. The model, after training, will show exactly what you want to search based on the keywords you provided and subsequent interactive questions with users," Zhou said.

In terms of intelligence, which is a result of training methods, ChatGPT still leads Chinese peers by 12 to 18 months, he said.

China's top domestic AI large models have reached the level of GPT-3.5, and the technical gap with GPT-4 is narrowing, Dou Dejing, an adjunct professor at Tsinghua University's Electronic Engineering Department, said on Saturday at the 2024 Global Times Annual Conference.