๐ Top 10 China's AI Stories in 2023: A Year-End Review
2023 was a year of excitement, wildness, and challenges for Chinaโs AI industry. We saw the rise of over 250 LLMs in China. Tech giants intensified their focus on AI to maintain a competitive edge, while startups secured new funding to take on the industry's heavyweights.
Amidst this rapid growth, Chinese regulators aimed for a delicate balance between ensuring safety and fostering innovation. The U.S.โ export controls and investment curbs continued to constrain key segments of Chinaโs tech industry, but Huaweiโs foray into chip development offered a local option.
In this issue, I summarized the top 10 stories from Chinaโs AI sector, based on my personal observation. This summary aims to provide a comprehensive review of the AI industry in China and offer insights into what 2024 might hold for the sector.
I wanted to thank you for reading my newsletters. May all the joys of the holiday season fill your heart and follow you throughout the coming year. Wishing you a merry Christmas and a happy new year! ๐๐
Surge of Chinese LLMs
In the English-speaking and Western digital sphere, ChatGPT, Claude, and Bard have built up dominance, yet their availability is restricted in mainland China. Over the past year, the country has been rapidly building its own LLM ecosystem to address local needs, training AI models on the Chinese language and culture.
Over 250 LLMs have been publicly announced as of today, backed by various companies - from tech giants to smartphone makers, telecom operators, and even dairy makers.
Baiduโs ERNIE Bot, a multimodal chatbot released in March, marked the beginning of major Chinese tech firms entering the fray. Following Baidu, Alibaba and SenseTime introduced Tongyi Qianwen and SenseChat, respectively. Huawei, focusing on cloud solutions, launched Pangu 3.0 in July, while ByteDanceโs Doubao and Tencentโs Hunyuan entered the arena later in the year.
In addition to chatbots, Chinese tech giants are incorporating LLMs into their established business sectors like e-commerce and social media. Users today are experiencing a new kind of interaction: They receive one answer on search engines, get shopping recommendations from AI assistants, and engage with virtual figures on social media. On the enterprise front, there is a growing expectation that AI will enhance productivity.
The rise of OpenAI has also ignited a fervent startup movement in China. Numerous Chinese AI startups are vying to be the next OpenAI and challenge local tech behemoths. On the leaderboard of SuperClue, a Chinese LLM benchmark, three of the top 10 Chinese LLMs come from startups.
The pace of Chinese LLM development is fast, with upgrades occurring every 3-6 months. A mainstream consensus is that Chinese LLMs will match GPT-3.5 by yearโs end and GPT-4 by mid-next year. Baiduโs ERNIE 4.0 is now the only LLM that claims parity with GPT-4.
The Year of AI Unicorns
2023 has seen the birth of 10 new AI unicorns in China, with four - Zhipu AI, MiniMax, Baichuan AI, and 01.AI - specializing in LLMs, according to IT Juzi. These startups have not only demonstrated remarkable innovation but have also attracted big capital.
Following the release of ChatGPT, Zhipu AI and MiniMax have risen as the new rock stars of the AI world. Zhipu AI, a spinoff of Tsinghua University founded in 2020, raised RMB 2.5 billion ($350 million) in 2023 from top venture capitalists and tech giants, including Alibaba and Tencent. Its homegrown GLM-based ChatGLM has gained popularity on GitHub.
MiniMax, founded in 2021 by the Vice Head of Research at SenseTime, has reportedly raised $250 million at a valuation of $1.25 billion. This investment round is led by Tencent.
Chinaโs AI landscape is also being disrupted by veteran entrepreneurs like Wang Xiaochuan of Sogou and Lee Kai-Fu of Sinovation Ventures with their new ventures, Baichuan AI and 01.AI.
A noteworthy story was Meituan co-founder Wang Huiwenโs $50 million investment in Light Years Beyond in February. Despite Wangโs unexpected departure from the startup due to health reasons, his move has inspired more entrepreneurs to venture into the AI realm. Light Years Beyond was later sold to Meituan in June for over RMB 2 billion (~$290 million).
The wide adoption of the Llama 2 architecture, open-sourced by Meta in July, has enabled Chinese LLM startups to overcome initial development hurdles and narrow the gap with leading models. However, the rise of Llama-based models sparked a debate over the extent of original innovation in the sector.
Out of the 10 AI unicorns in China, two are robotics companies. Robotics is emerging as another hotbed, inspired by the growing buzz around Teslaโs humanoid robot, Optimus. Entrepreneurs and venture capitalists are viewing LLM-powered robots as the next big tech revolution like smartphones and electric vehicles.
Chinaโs Generative AI Regulation Takes Effect
China wants to position itself as a global frontrunner in regulating the burgeoning field of generative AI.
In April, the Cyberspace Administration of China (CAC), Chinaโs top Internet regulator, released a draft regulation targeting generative AI. The initial draft outlined rigorous guidelines for AI developers, including the mandate for AI-generated content to be accurate and truthful, and the requirement for service providers to verify usersโ real identities.
However, the final version released in July presented a more friendlier approach. It eliminated the real-name registration mandate and did not explicitly require measures to counter disinformation or hallucinations in AI output. The regulation only applies to public-facing generative AI services. Nonetheless, the regulation still mandates administrative permits for AI service providers and adherence to socialist core values.
In a further step towards comprehensive AI governance, Chinaโs National Information Security Standardization Technical Committee introduced a draft guideline aimed at enhancing security standards for generative AI services in October. A key proposal includes creating a blacklist for training corpus sources, which would bar any source with over 5% of illegal or undesirable content.
Chinaโs proactive approach extends beyond generative AI. In August, the CAC proposed draft rules to regulate the use of facial recognition technology, emphasizing the protection of personal information rights.
Public Rollouts of Chinese Chatbots
August 31 marked a watershed moment in Chinaโs AI industry. Eight Chinese LLMs, including ERNIE Bot, ChatGLM, and Doubao, have cleared the regulatory filings to offer services to the public. Prior to that, chatbots are only confined to being tested in beta.
By November, over 20 LLMs had been granted regulatory approval, a list that includes Alibabaโs Tongyi Qianwen, Tencentโs Hunyuan, iFlytekโs Spark, NetEaseโs Ziyue, Ant Groupโs Bailing, Moonshot, and WPS AI, among others.
With full public access, Chinese chatbots now have the opportunity to utilize more data than before. The majority of Chinese ChatGPT alternatives use Reinforcement Learning with Human Feedback (RLHF), a technique that uses human feedback to enhance modelsโ self-learning. Like unleashing hungry sharks into a sea teeming with fish, public access allows them to evolve more rapidly beyond their beta testing.
Chinese chatbots also differ from their Western counterparts like ChatGPT. Optimized for new users, these chatbots feature pre-defined prompts and various AI companions. Features like voice dialogue and text-to-image generation are also commonly included.
Stricter US Export Control on Chips
The U.S. has intensified its stance on chip exports to China. On October 10, U.S. Commerce Secretary Gina Raimondo announced tightened restrictions on the export of AI chips and manufacturing equipment to China. The new restrictions update the 2022 chip restrictionโs limitations, closing loopholes that allowed China to obtain advanced chips through subsidiaries in other countries.
The revised guidelines now include not only state-of-the-art chips but also less advanced yet potent chips. Notably impacted are Nvidiaโs A800 and H800 chips, designed specifically for the Chinese market but less powerful than their top-tier A100 and H100 counterparts.
Despite these restrictions, Nvidia is still attempting to continue serving its Chinese clients with specialized AI chips. However, the U.S. is closely investigating the details of three new AI accelerators proposed by the company.
Another major blow to Chinese startups is the U.S. investment curbs in Chinese AI, quantum computing, and semiconductors. In August, U.S. President Joe Biden signed an executive order to restrict investments from the U.S. directed toward China.
Huaweiโs AI Computing Ambition
In September, Huawei unveiled its โAll Intelligenceโ strategy, aiming to develop foundational AI technologies and build a strong computing infrastructure across various industries.
This strategic pivot is particularly noteworthy as it represents Huaweiโs first major directional change in a decade. The company previously launched the โAll Cloudโ strategy in 2013 and the โAll IPโ initiative in 2003.
A key objective of Huawei's new strategy is to fill in for the gap left by Nvidia and other global chipmakers in the Chinese market, a result of the U.S. tightening export controls. iFlytek, a company sanctioned by the U.S., claimed that Huaweiโs latest AI chip, Ascend 910B, is comparable in performance to NVIDIAโs A100. Baidu and 360 have also reportedly purchased Huaweiโs AI chips, though in small numbers, as an alternative plan.
Additionally, in September, Huawei introduced the Atlas 900 SuperCluster, consisting of thousands of Ascend 910 processors. Located in Nanjing, this AI computing cluster is designed to support the training of large-scale AI foundation models with over a trillion parameters. Huawei is planning to build more than 20 AI computing centers.
AI Photographer Goes Viral
In July, Miaoya Camera, an AI photographer app and a โLensaโ lookalike, took Chinaโs internet by storm.
While AI photographer apps have been around for a while, Miaoya Camera stands out for its exceptional image quality, by using generative AI algorithms. This app, developed by an Alibaba-backed startup, requires users to upload one headshot, 20 photos and pay RMB 9.9 to generate a range of diverse photo types, including professional headshots, portraits in traditional Chinese attire, and images styled after U.S. high-school uniforms.
The appโs instant popularity reflects the growing appetite for AI-generated content among Chinese users. Following Miaoya Cameraโs success, another app called 45AI garnered attention for its ability to create Barbie-style portraits, partially attributed to the new Barbie movie. For a similar fee, users submit 9-15 photos to receive AI-generated glamour shots.
China Joins Global Dialogue on AI Safety
On November 1, at the first AI safety summit, organized by the U.K. government, China, the U.K., the U.S., and the E.U., along with other nations, signed the Bletchley Declaration. The signing nations acknowledged the catastrophic risks posed by AI and agreed to collaborate on AI safety research.
As the world's first international agreement on AI, the declaration commits existing international bodies and initiatives to ensure AIโs development is human-centric, trustworthy, and responsible. It focuses on the โfrontierโ AIโhighly capable general-purpose models that could potentially outperform current advanced modelsโunderlining the urgency to understand and mitigate their associated risks, particularly in sensitive areas like cybersecurity, biotechnology, and disinformation.
Wu Zhaohui, Chinaโs vice minister of science and technology, led a delegation to attend the AI safety summit in November. In his speech at the opening plenary session of the summit on Wednesday, Wu said, โWe uphold the principles of mutual respect, equality, and mutual benefits. Countries regardless of their size and scale have equal rights to develop and use AI.โ
China wants to play a key role in shaping the future of global AI governance. Before the declaration was signed, on October 18, Chinese President Xi Jinping announced Chinaโs Global AI Governance Initiative in his keynote at the Belt and Road Forum in Beijing. The Initiative advocates a people-centered approach to AI development to benefit human progress and supports the establishment of international AI governance frameworks and standards based on broad consensus and UN discussions.
The Love and Hate in Deepfake
The most popular AI video this year in China (probably) is a deepfake video of Taylor Swift speaking fluent Mandarin, thanks to an AI tool called HeyGen. Throughout the year, Iโve written several stories on deepfake technology. While some have been amusing, others have been quite alarming.
Stefanie Sun, a Mandopop singer who had stepped back from the limelight for years, made a remarkable comeback this May. Her resurgence was attributed to an AI model that cloned her voice and incorporated it into numerous Mandopop classics.
Deepfake face-swapping technology was spotlighted in a case unveiled by the Chinese police, where a victim was scammed out of 4.3 million yuan within just ten minutes.
Several platforms, including ByteDanceโs Douyin and Bilibili, have introduced new rules requiring creators to label AI-generated content. This measure aims to help users distinguish between authentic and deepfake materials.
CAC enforced new regulations for deepfakes on January 10, 2023. These rules ban the use of generative AI for creating illegal content, spreading false news, or harming national security. Additionally, AI-generated content must be clearly marked to differentiate it from human-created content.
SenseTime Founder Tang Xiaoou Passes Away
The year 2023 ended with sad news: Tang Xiaoou, a renowned AI scientist and the co-founder of SenseTime, passed away on December 15, at the age of 55.
Beyond his position at SenseTime, Tang was also a distinguished Professor in Information Engineering at the Chinese University of Hong Kong (CUHK), Director of the Shanghai AI Laboratory, and Director of Pujiang Lab.
Tang has been recognized as a forward-thinking scientist. His early work in applying deep learning to computer vision tasks laid the groundwork for many future advances. Under his leadership, the CUHKโs multimedia lab achieved breakthroughs in facial recognition accuracy with the GaussianFace and DeepID algorithms in 2014, which surpassed human recognition capability for the first time. Built on top of this technology, he co-founded SenseTime.
Throughout his career, Tang has been a mentor who helped shape the careers of many prominent AI researchers, like He Kaiming, the author behind the revolutionary ResNet, and Dahua Lin, known for creating OpenMMLab.
Remarkably, the funeral of Tang saw tributes from Chinese top politicians, including Chinese Premier Li Qiang, Vice-Premier Ding Xuexiang, and Shanghai Party Chief Chen Jining among the distinguished mourners.