Alibaba Predicts Generative AI as Top Tech Trend for 2023, China’s AI Computing Surpasses General Computing, and AI Can Recognize Lip Syncing

Weekly China AI News from Jan 9 to Jan 15

In this week’s newsletter, I will compare Alibaba’s annual tech trends with those of Tencent and Baidu. China’s doubling down on AI computing centers and we will explore more. At new year’s eve galas, you might wonder which singer pretends to sing. Now, AI can help identify lip-syncing! A new section has been added to feature stories from media outlets and research institutes. I hope you enjoy my selections.

Weekly News Roundup

Alibaba’s Top 10 Tech Trends Emphasize Generative AI, Chiplet Technology, and Cloud Security

What’s new: Alibaba’s research arm, Damo Academy, has released its annual list of top 10 technology trends, including generative AI, dual-engine decision intelligence, pre-trained multimodal foundation models, cloud-native security, chiplets, processing in memory, hardware-software integrated cloud computing architecture, predictable network powered by edge-network integration, computational imaging, and large-scale urban digital twins.

You can read the full report on Alizila or SCMP.

My take: The list, which was released last week, highlights the key technology areas that Alibaba is focusing on in order to drive its business forward. As a leading player in the technology industry, Alibaba’s focus on AI, cloud computing, and chips is not surprising, as these areas constitute the cornerstone of the company’s cloud business.

The inclusion of generative AI and pre-trained multimodal foundation models is particularly noteworthy, as Alibaba has made significant investments in large-scale AI models such as the 10-trillion-parameter model M6 and the language model PLUG. The company is already using generative AI technology to design clothes and create e-commerce marketing materials.

Chiplets, which refer to a technology that allows for the construction of larger chips from multiple smaller dies, are also gaining attention in China. Chiplets are believed to extend Moore’s Law, and have seen a huge increase in interest in recent months, with the search volume for “chiplet” on Baidu surging by 6000 times in the week of August 8, 2022, according to Chinese media TMTPost.

Alibaba vs. Tencent vs. Baidu: This release of the top 10 tech trends by Alibaba is in line with similar releases by other Chinese tech giants, such as Tencent and Baidu, which also recently released their own top 10 tech trends for 2023. A direct comparison of these trend lists reveals the areas of focus for each company and the direction they are heading in the future.

China’s AI Computing Surpasses General Computing in 2022

What’s new: More than 30 cities in China are building or proposing to build AI computing centers, according to a report jointly released by the State Information Center and Inspur. This is in line with the country’s increasing focus on AI and its potential to drive economic growth.

According to market intelligence firm IDC, China’s AI computing power continues to maintain rapid growth, with the scale of AI computing reaching 26.8 trillion FLOPS (EFLOPS) in 2022, exceeding the scale of general computing power. This growth is driven by the demand for digital services and the growing use of AI in various industries.

Inside AI computing: One of the major drivers of this growth is the use of GPU chips, which currently constitute 90% of the AI computing market share. While the training of models still mainly relies on GPU, there is a growing trend towards the use of ASIC, FPGA, and NPUs on the edge and inference.

Cloud services such as search, face recognition and recommendation engines are common applications that consume a large amount of AI computing power.

Looking ahead, cognitive computing applications such as natural language processing (NLP) are expected to become more important. This will drive the market for chip vendors such as NPUs and ASICs, which is expected to grow rapidly in the next five years as more industries adopt AI technology.

AI Can Identify Lip-Syncing Performers

What’s new: A Bilibili creator, known as “Gaotian” on the platform, has developed an AI software that can detect lip-syncing during live performances. The software uses a combination of machine learning and audio processing techniques to analyze the vocal pitch of singers during live performances.

This AI software could be a useful tool for identifying lip-syncing during live performances, particularly at high-profile events like New Year’s Eve galas

You can watch the original video on Bilibili.

How it works: The creator, who is a Microsoft software programmer, used the Spleeter algorithm to separate the human voice from the background music. He used the Python speech library Parselmouth to analyze the vocal pitch of the singer. The software then visualized the singer’s vocal pitch and compared it to the original song. If the singer’s pitch is accurate, but they are not a professional, they are likely to be lip-syncing.

Trending Research

• Text-to-SQL is a technology that converts natural language to SQL queries. This problem is extremely useful for finding information stored in an enterprise database, such as in healthcare or retail industries, by expressing commands in natural language. Researchers from the University of Science and Technology of China, the Chinese Academy of Sciences, and DAMO Academy proposed a novel SQL-guided pre-training framework STAR for context-dependent text-to-SQL parsing, which leverages contextual information to enrich natural language (NL) utterance and table schema representations for text-to-SQL conversations. STAR achieves new state-of-the-art performance on two downstream benchmarks (SParC and CoSQL). Read the paper STAR: SQL Guided Pre-Training for Context-dependent Text-to-SQL Parsing for more.

• The computing cost of running large language models (LLMs) like GPT-3 or ChatGPT is staggering. That’s why researchers train a smaller version of LLMs, which can be directly finetuned on downstream tasks and retains comparable performance. Baidu researchers proposed Multi-task Infused Task-agnostic Knowledge Distillation (MITKD), which demonstrates a much better generalization. Baidu’s method exceeds an 8x larger BERTBase on SQuAD and four GLUE tasks. In addition, by combining ERNIE 3.0, Baidu’s method achieves state-of-the-art results on 10 Chinese datasets.

Noteworthy Stories

• China is the undisputed champion in artificial intelligence research papers, a Nikkei study shows, far surpassing the U.S. in both quantity and quality. — Nikkei Asia
• As China becomes increasingly integrated into international networks on artificial intelligence research, how can democratic governments manage the geopolitical and ethical risks? — Brookings
• Beijing is among the first governments to regulate hyper-realistic, AI-generated media. — Wall Street Journal
• China’s internet users finding creative uses for ChatGPT. — South China Morning Post