Over 5,700 Chinese Chip Companies Close in 2022; Meet MOSS, Fudan University’s ChatGPT Replica; Beijing Pledges Support for ChatGPT-Style Technology
Weekly China AI News from Feb 13 to Feb 19
Greetings! I was considering renaming our newsletter to Recode China ChatGPT…okay not a good joke, given the recent unsettling responses from the new Bing Chat. This week’s newsletter covers the decline of Chinese chip companies in 2022 and the release of MOSS, a conversational AI model from Fudan University that shares similarities with ChatGPT, but is currently weaker in Chinese. Despite concerns over China’s crackdown on ChatGPT-style technology, Beijing has promised to provide support.
Weekly News Roundup
China Sees Sharp Decrease in the Number of Chip Companies in 2022
What’s new: In 2022, 5746 semiconductor companies in China revoked or canceled their business licenses, up 68% from 2021, according to Chinese tech media TMTPost. There are over 170,000 semiconductor companies in China now.
TMTPost further reported that 3,470 chip enterprises closed in the first 8 months of 2022. From September to December, more than 2,300 chip enterprises shut down, meaning an average of more than 15 chip companies.
What are the reasons? The article attributed the sharp decrease in the number of Chinese chip companies to the global downturn in the semiconductor industry and the continued restrictions on US chip exports to China.
- One of the major reasons was the global downturn in the chip industry. For instance, Intel reported its lowest quarterly revenue since 2016. Since December 2022, domestic publicly-listed semiconductor companies have released their earnings forecasts for 2023, and 49 chip companies are expected to see a year-on-year decline in net profit.
- The ongoing US restrictions on semiconductor exports to China are also greatly affecting the performance and industry confidence of chip companies. In Q1 2023, Semiconductor Manufacturing International Corporation (SMIC) is expected to see a 10%-12% QoQ decline in revenue and a drop in gross profit margin to 19%-21%, mainly due to factors such as weak demand in the consumer electronics market, global economic slowdown, and US restrictions on semiconductor exports to China.
Meet Moss, a ChatGPT-like Conversational AI Model from Fudan University
What’s new: The natural language processing lab at Fudan University released MOSS, a conversational language model similar to ChatGPT, with significantly fewer parameters. MOSS is capable of performing various natural language tasks from question-answering to code generation, but it performs poorly in languages other than English.
Why is it called MOSS: The name MOSS is likely inspired by a superintelligent quantum computer from the Chinese sci-fi movie Wandering Earth series (think of MOSS as China’s Skynet or HAL-9000).
- MOSS can be considered a lower-end version of ChatGPT. They are both proficient at language-based multi-tasking, but the number of parameters of MOSS is significantly fewer, which is around one order of magnitude less than ChatGPT, said Dr. Xipeng Qiu, a Professor of Computer Science at Fudan University.
- MOSS is not a Chinese version of ChatGPT and is actually better at writing in English, as Dr. Qiu explained that MOSS is trained on 300 billion English words and only 30 billion Chinese words.
- MOSS learns by talking to humans and other AI models, while ChatGPT is trained with Reinforcement Learning from Human Feedback (RLHF).
- MOSS is in internal testing and will be open to tens of thousands of users due to limited computing resources. The lab said the model is constantly being improved based on user feedback, with a focus on improving its language skills in Chinese and ability to follow instructions.
- Like ChatGPT and New Bing Chat, MOSS may generate misleading or false information, and sometimes even unethical or harmful responses.
- MOSS is set to be open-sourced in March, 2023.
China’s Capital Vows Support to ChatGPT-Style Tech
What’s new: The ChatGPT frenzy in China, with Baidu, Alibaba, JD.com, and more announcing their ChatGPT-style technologies, has drawn Chinese regulators’ attention. Many fear ChatGPT-style technologies would face a potential crackdown, as a state media warned Chinese investors earlier not to follow the ChatGPT heat as “some market capitals are keen on false concept speculation.”
For now, tech companies in Beijing have received the green light to develop ChatGPT-like technologies. Last week, the Beijing Municipal Bureau of Economy and Information Technology vowed their support for developments of ChatGPT-style technologies. Here is what they said:
Support top companies in building large models comparable to ChatGPT, and focus on building an application ecosystem on an open-source framework and application general large models. Strengthen the foundation of AI computing infrastructure and accelerate the supply of AI data.
What else: In the “2022 Beijing Artificial Intelligence Industry Development White Paper”, the bureau boasted Beijing’s achievements in becoming a global hub of AI innovation. China’s capital now has 1,048 core AI companies, which comprise 29% of the total in China. The number of AI talents in Beijing has surpassed 40,000, which is 60% of the national total. Beijing’s open governance data and public service items are a significant contributor to the AI industry, and there are several significant projects, including the development of AI applications in smart cities, autonomous driving, and intelligent manufacturing.
PV3D: A 3D GENERATIVE MODEL FOR PORTRAIT VIDEO GENERATION
- Researchers from the National University of Singapore and ByteDance proposed PV3D, the first generative framework that can synthesize multi-view consistent portrait videos. (ICLR 2023)
EVA3D: Compositional 3D Human Generation from 2D Image Collections
- Researchers from SenseTime’s S-Lab and Nanyang Technological University proposed EVA3D, an unconditional 3D human generative model learned from 2D image collections only.
Multimodal Analogical Reasoning over Knowledge Graphs
- Researchers from Zhejiang University and the National University of Singapore introduced the new task of multimodal analogical reasoning over knowledge graphs, which requires multimodal reasoning ability with the help of background knowledge.
ERNIE-Music: Text-to-Waveform Music Generation with Diffusion Models
- Researchers from Baidu proposed the first text-to-waveform music generation model that can receive arbitrary texts using diffusion models.
Asynchronous Multi-Agent Reinforcement Learning for Efficient Real-Time Multi-Robot Cooperative Exploration
- Researchers from Tsinghua University, Shanghai Artificial Intelligence Laboratory, Tongji University, and Shanghai Qi Zhi Institute proposed an asynchronous MARL solution, Asynchronous Coordination Explorer (ACE).
- A Chinese ChatGPT alternative won’t pop up overnight — even though many companies may want you to think so. — MIT Technology Review
- The deputy head of the Ministry of Industry and Information Technology meets Big Tech executives at an event hosted by the Internet Society of China. — SCMP
- Neither the EU nor China is taking a purely horizontal or vertical approach to governing AI. But the EU’s AI Act leans horizontal and China’s algorithm regulations incline vertically. By digging into these two experiments in AI governance, policymakers can begin to draw out lessons for their own regulatory approaches. — Carnegie Endowment
- Zeekr, the luxury electric vehicle brand of China’s largest private carmaker Geely, announced Monday that investors have committed to investing $750 million in its Series A funding round. Its post-money valuation has shot up to $13 billion. — TechCrunch