Artificial Intelligence
The Chinese government has long elevated artificial intelligence (AI) development to a strategic national priority with policies like the New-Generation AI Development Plan in 2017. It sees the potential of AI to radically transform social and productive structures due to its ability to perform tasks that usually require human intelligence.
US-China competition for dominance over AI technologies shapes today’s geopolitics. Wary of China’s potential, especially Beijing’s ambition to leverage AI for modernizing its military, the US has imposed sweeping export controls aimed at hobbling its progress.
Thanks to government support and global links, Chinese applied AI companies like computer vision firms SenseTime and Megvii are now household names. China’s fast-growing talent base and vibrant ecosystem of companies and public research labs are leading AI application development and producing high-impact AI research.
But China’s main weakness lies in its historical links with US industry: it trails in the fundamentals. China still cannot match the quality of US-designed AI chips and relies on machine learning frameworks developed by US firms. While China’s AI future is for now tied to choices made in Washington, in the long run, Chinese large language models and other types of AI systems may well prove good enough for the tasks they need to fulfill: enabling industrial applications and boosting military and security capabilities.
Graphics dashboard
Chinese large language models (llm) perform well in Chinese
The Chinese ecosystem for large language models (LLM) has entered a “battle of a hundred models.” Leaderboards scoring models on benchmarking tasks, such as language generation and translation, reflect fierce competition. In the June 2024 ranking for SuperCLUE, a Chinese-language benchmark, Alibaba’s Qwen is China’s top model, trailing only slightly behind GPT-4.
China's research output on AI gathers steam
In 2023, China-affiliated authors published more on AI than authors from any other country or region, after lagging behind both Europe and the US in 2017. This shows that the government’s focus on AI is having real impact. China’s output is even larger if Chinese-language papers are taken into account.
Chinese policy documents show focus on AI in sector development and real economy
Chinese policy makers view AI development as a priority, evidenced by how many policy documents mention the term. They see AI as key to achieving government goals like informatization and focus on applications for the real economy. Regulation is also a concern, as policy documents also mention words like security, declaration, standard, and supervision.
Artificial Intelligence in China: Timeline of crucial events
Development
Policy/regulation
Huawei releases deep learning software framework Mindspore to the open-source community, as China attempts to reduce reliance on Google and Meta frameworks.
China's algorithm filing portal goes online for recommender engines, the genesis of a licensing system later used for generative AI and large language models (LLMs).
China releases rules on synthetic (AI-generated) content, motivated by concerns around deepfakes.
Tsinghua University's Knowledge Engineering Group announces the bilingual pre-trained model ChatGLM-130B and its open-source version, ChatGLM-6B. The team founds Zhipu AI.
Baichuan-Inc, a startup of the founder of internet search engine Sugou, releases its first open-source LLM. The Baichuan series is among the best performing in China.
Chinese AI experts led by a team at the Chinese Academy of Social Sciences (CASS) publish recommendations for a comprehensive AI law after legislation is added to the State Council's work plan.
China restricts access to Hugging Face, the top international platform for open-source AI collaboration. China’s open-source community is displeased but seeks alternatives to work on leading models.
Yi-34B, a model from leading startup 01.AI, tops the international leaderboard for open-source LLMs. But it becomes mired in controversy over the way it referenced Meta's Llama2 architecture.
Chinese online retailer Alibaba announces investment in LLM startup MiniMax, following similar investments in Zhipu AI, Baichuan AI, 01.AI, and Moonshot AI.
Tsinghua and ShengShu Technology unveil Vidu, a text-to-video generator and China's answer to OpenAI's Sora announced in February 2024.
US and Chinese officials hold first bilateral talks on AI safety and risks in Geneva.
The Cyberspace Administration of China (CAC) and three other agencies require companies to file algorithms for review, reflecting concerns over how they spread online content and make decisions.
The US puts export controls on semiconductors and key hardware that Chinese firms need to train AI, severely limiting their access to leading-edge graphic-processing units (GPUs).
Baidu announced ERNIE Bot, a conversational AI bot built on the company's LLM ERNIE. It is widely seen as a response to US-based Open AI’s ChatGPT of November 2022.
The CAC releases draft rules on generative AI, building on existing synthetic content rules and reflecting the CAC’s focus on controlling online content.
China becomes the first country to regulate generative AI with final rules. Public chatbots and the underlying models must pass a review – a de-facto licensing regime.
The CAC greenlights ERNIE Bot and 10 other models for release. The approvals signal regulators' balancing act between security and development and shows censorship is not an insurmountable obstacle.
China signs the Bletchley Declaration, the outcome of a major AI safety summit hosted by the UK. China's involvement with the US and other democratic nations shows shared concern around risks of AI.
Top Chinese and Western scientists meet in Beijing to discuss safety and existential risks associated with advanced AI systems.
The Government Work Report to the National People's Congress proposes the "AI+" initiative to deepen integration between AI and the real economy.
OpenAI blocks API access for China-based developers after shutting down online influence and hacking networks linked to China and others, further decoupling US and Chinese generative AI ecosystems.
A team at Tsinghua University's Department of Precision Instruments develops Tianmoc, a brain-inspired chip designed for AI applications.
Recent developments
Tech progress
- Alibaba released its latest open-source large language model (LLM), Qwen2.5, which boasts performance-matching, leading world models by certain benchmarks. The Qwen series is the undisputed leader in Chinese open-source models and is second only to Meta’s Llama in number of downloads worldwide. (Source CN: ZhiDX)
- Beijing AI Academy launched Emu3, a multimodal AI model that works across text, image and video forms, in an effort to close the gap with models like OpenAI’s text-to-video model Sora. (Source EN: South China Morning Post)
- The 2024 World AI Conference in July in Shanghai focused on robotics. A humanoid general-purpose robot, Qinglong, was unveiled by Humanoid Robot, a lab and state-designated innovation center. Tesla also debuted its latest humanoid robot. (Source (EN): Tech Node)
Domestic dynamics
- The rush to build large language models is cooling down, as the six leading startups in the LLM space – Zhipu, MiniMax, Moonshot, Baichuan, 01.AI and Stepfun – start to pivot away from releasing new foundational models and focus on applications and monetization. (Source CN: 36 Kr)
- The city of Beijing announced the first batch of AI application joint development platforms to promote the integration of AI with different industries, including government administration, manufacturing, finance and more. (Source CN: S&T Daily)
- US firm OpenAI has ended services in China and cut off access to its generative AI models, furthering a growing split between Chinese and Western AI ecosystems. (Source EN: Time)
- The Ministry of Industry and Information Technology (MIIT) released a draft AI Industry Standardization Framework Guidance, setting goals for industry standards in six key areas., This is in line with other standardization framework guidance in China. (Source CN: S&T Daily)
Foreign involvement
- Chinese standardization committee TC260 released the Artificial Intelligence Safety Governance Framework 1.0, with the goal to implement the Global AI Governance Initiative China launched last year. (Source CN: S&T Daily)
- A paper showcasing AI for military use based on Meta’s open-source Llama model, published by researchers linked to the People’s Liberation Army, renews concerns for Chinese access. (Source EN: Reuters)
- US company Nvidia, the world's leading producer of AI chips, is reportedly working on a less powerful version of its new Blackwell chip in order to stay in the Chinese market despite US export restrictions. (Source CN: EE Times China)
- The first US-China AI dialogue on risks took place in May behind closed doors, focusing on AI risks and the importance of governance measures. China reiterated its complaint against US export controls hampering its AI development. (Source EN: AP News)
- Meta’s latest open-source LLM, Llama3.1, performs at the level of the best proprietary models. This upends previous assumptions about the strength of commercial vs freely available LLMs and will likely quicken Washington’s efforts to restrict what foreign adversaries can access. (Source EN: The Verge)
Publications
Kommentar
Kommentar
Kommentar
Topics
