Recode China AI

Happy new year, dear readers! This week's newsletter will cover Baidu's annual tech trends and Douyin's popular AI anime effect. We will also discuss a recent controversy where a Chinese Go player was accused of cheating with AI. Finally, Tencent's new text-to-3D model, Dream3D, allows users to create 3D objects using text prompts.

Weekly News Roundup

Believe It or Not: AI Dominates Top 10 Tech Trends for 2023

What’s new: At the beginning of each year, it is common for Chinese tech companies to release their technology trend outlooks for the upcoming year. These selected trends often reflect the companies' current areas of focus, but the list can also provide insight into the future direction of technology.

Baidu Research, the AI research arm of Baidu, recently released its predictions for the top 10 technology trends in 2023, with a strong emphasis on AI. This is the fourth year in a row that Baidu Research has released its annual outlook.

• Trend 1: Big Model Building - Big models for industries emerge, providing intelligent upgrades across a wide range of sectors.

• Trend 2: Digital-Real Convergence - The increasing demand for AI infrastructure drives deeper integration of digital technology with the real economy.

• Trend 3: Virtual-Real Symbiosis - Web 3.0 technology creates a new type of online space, leading to disruptive innovation in the metaverse industry.

• Trend 4: Autonomous Driving - Autonomous driving sees new upgrades, leading progress in intelligent transportation.

• Trend 5: Robotics - The use of industrial robotics accelerates, addressing labor shortages.

• Trend 6: Scientific Computing - AI technology has become a valuable research aid, transforming the paradigm of multidisciplinary research.

• Trend 7: Quantum Computing - Breakthroughs in core technologies continue to drive the industrialization of quantum computing.

• Trend 8: Privacy Computing - Privacy computing platforms enable data interoperability while balancing value creation with security and trust

• Trend 9: Ethics in Technology - Explainable AI technology promotes "mutual trust," making reliable and controllable technology a new competitive advantage.

• Trend 10: Sustainability - The focus on green, low-carbon, and sustainable energy grows, with key breakthroughs in edge computing and advanced computing.

My take: Several trends, including quantum computing, scientific computing, privacy computing, and autonomous driving, have been selected in the previous year. What’s noteworthy is:

• Big models, or foundation models that are more widely used in the West, are being integrated with industries. One of the biggest challenges to democratizing AI is the high threshold required to build these models. However, by using simplified pre-training AI architectures that are fed with large amounts of industry-specific data, it will become much easier to build downstream AI applications.

“Deep learning platforms, together with AI big models, have constructed a solid foundation for industrial intelligence,” said Dr. Haifeng Wang, CTO of Baidu. Dr. Hongjiang Zhang, Chairman of BAAI and Distinguished Visiting Professor of Tsinghua University AIR, also echoed “big models are a fast-developing, market-explosive opportunity technology that will bring a new wave of entrepreneurial enthusiasm.”

• I was surprised to see generative AI, or AI-generated content (AIGC), was dismissed. Instead, it was only mentioned in the “virtual-real symbiosis (虚实共生)” trend, which highlights the potential for a parallel universe of digital and physical worlds powered by Web 3.0 and other technologies.

• One new trend to note is the focus on ethics in technology, as the emergence of AI-generated art and chatbots has raised concerns about copyright infringement, job replacements, and the output of hateful or discriminatory language. Building human-centered AI that is controllable and reliable is therefore becoming a top priority.

Douyin’s AI Anime Effect Gains 27.5 Million Users

What’s new: Creative AR effects make it fun to create on TikTok and Douyin, the two short video app kings owned by ByteDance. Among all effects this year, an AI Anime effect that can turn photos into animes is the most popular one, according to Douyin.

How popular: Over 27.5 million users have used the effect - named AI Paintings (AI绘画) - to create videos. At an all-time high, the app has to process 14,000 queries per second (QPS), a daunting task to ensure a seamless user experience.

Tech deep dive: To meet this challenge, the team behind AI Paintings turned to the latest text-to-image generation technology, Stable Diffusion. The team built a dataset of over one billion images and trained a particular anime-style model. Douyin has had a lot of experience in anime effects before, and after observing user feedback on different styles, the “AI Painting” chose a refined Asian anime style.

In addition, the team employed other models to understand images better. First, the user's image is classified, such as portraits or pets. Then, if images contain people, further detection is performed on attributes such as gender, number of people, and age. In addition, AI Paintings also appear with a certain probability of special effects, such as reversing the gender of people.

In training, the team reduced the training time of Stable Diffusion on 128 A100s from 25 days to 15 days, a 40% increase in training efficiency.

Chinese Go Player Emerges Victorious By Playing Like AI; Teammate Accuses him of Cheating

What’s new: One of China’s highest-ranking Go players was accused by his teammate of cheating using AI during his last winning game in 2022. The Chinese Go association dismissed the accusation and suspended the accuser for six months due to a lack of tangible evidence.

More context: Xuanhao Li (李轩豪) is a 27-year-old 9-dan professional Go player. He was known for adopting AI-inspired strategies and became one of the fastest-improving Go players this year. In a semi-final game on December 21 at Chunlan Cup, a prestigious international Go competition, Li defeated the 9-dan Shinjin Seo from Korea. Some of his plays during the game matched AI’s predictions.

Drama: Li’s national teammate, the 9-dan Dingxin Yang (杨鼎新), suspected Li is cheating using AI. Yang said in a WeChat post,

I want to play 20 games of Go with you. No breaks, no time limit, one game per day. The playing field will block all signals and the game records will be made public for everyone to judge. If everyone thinks I've wronged you, I'll retire after the LG Cup. Do you dare accept the challenge?

Result: The Chinese Go Association conducted an investigation and reviewed the entire tournament's surveillance footage to confirm that no violations or cheating occurred. They dismissed the accusation on December 30 and suspended Yang for six months. Yang also deleted his WeChat post.

AI Coach: Go players using AI to hone their skills is nothing new. For example, Fine Art, a Go-playing computer program created by Tencent, has been utilized by the Chinese national team as a training partner since 2018.

Trending Research

• TinyMIM: Researchers from Microsoft Research Lab-Asia explored techniques to transfer the success of large masked image modeling (MIM)-based pre-trained models to smaller ones using distillation. The Our TinyMIM model of tiny size achieves 79.6% top-1 accuracy on ImageNet-1K image classification, which sets a new record for small vision models. Read the paper TinyMIM: An Empirical Study of Distilling MIM Pre-trained Models.

• Text-to-3D Dream3D: Researchers from Tencent introduced the explicit 3D shape prior to CLIP-guided 3D optimization methods in order to improve the accuracy of 3D structures generated from text prompts. Their method, Dream3D, is able to generate 3D content with higher visual quality and shape accuracy than previous methods. Read the paper Dream3D: Zero-Shot Text-to-3D Synthesis Using 3D Shape Prior and Text-to-Image Diffusion Models.