🤖 GPT-4's Chinese Cousin Boasts 90% Smarts, China's AI Standardization Guideline, and NVIDIA CEO's Low-Key China Trip
Weekly China AI News from January 15, 2024 to January 21, 2024
Hey everyone, sorry for skipping the last week’s update and the delay on this one. Things have been a bit hectic with some changes at work. No worries, I’m still here to fill you in on all the AI news happening in China.
In this issue, I will discuss Zhipu AI’s new GLM-4 model. China’s MIIT just released an AI standardization guideline. Guess who showed up at NVIDIA’s Lunar New Year party in mainland China? Surprise, surprise, CEO Jensen Huang. Plus, I've got some fun research projects for you, including Vision Mamba and a framework for swapping out video characters with 3D avatars.
Zhipu AI Launches GLM-4 Rivaling GPT-4
What’s New: Zhipu AI, a rising Chinese AI startup, launched its new foundation model, GLM-4, at Zhipu DevDay on January 16. CEO Zhang Peng announced that GLM-4 achieves over 90% of GPT-4’s performance across key English benchmarks like MMLU and HumanEval. GLM-4 is now accessible for public testing on Zhipu AI's ChatGLM platform.
How it Works: GLM-4 boasts a 60% overall performance improvement over its predecessor GLM-3, and supports longer contexts, stronger multimodal abilities, and faster reasoning.
GLM-4’s 128K context window length can handle text equivalent to 300 pages. The model can recall facts or information from deep within documents, with near-perfect accuracy, known as a “needle in a haystack“ analysis.
Its multimodal capabilities are driven by CogVLM and CogView models, with CogView3 surpassing Stable Diffusion XL and approaching DALLE-3.
Zhipu AI also introduced GLM-4-All Tools, enabling ChatGLM to perform diverse functions such as internet browsing, image drawing, and code interpretation.
Why it Matters: Zhipu AI is closely mirroring OpenAI in terms of product designs and strategic direction. The company even released its customizable GLM model, GLMs, akin to OpenAI’s GPTs. The CEO of Zhipu AI has been quoted saying, “While OpenAI forges its path, we are learning from their journey.”
Notably absent from the event was specific information about the model’s size, reflecting a potential shift in strategy amidst intense competition and limited computing resources. This change indicates a move away from the pursuit of increasingly larger models towards a more balanced approach, considering both performance and return on investment.
China Drafts New Guideline to Standardize AI Industry
What’s New: China’s Ministry of Industry and Information Technology issued a draft for standardizing AI last week, named the National Artificial Intelligence Industry Comprehensive Standardization System Construction Guideline. Open to public opinion, this guideline aims to develop a robust standardization system to enhance the quality of China’s booming AI industry.
How it Works: The guideline sets ambitious targets for 2026.
Over 60% of key technologies and application projects should yield standard results;
More than 50 national and industry standards;
Over 1000 companies advocate for the implementation of these standards.
Participating in the development of over 20 international standards
Why it Matters: According to the draft, China’s AI industry is evolving rapidly, with new technologies like LLMs leading to innovative breakthroughs, integrated industry applications, and international collaboration. This necessitates a well-defined and comprehensive standardization system.
(For readers interested in China’s AI regulation, I recommend reading MIT Tech Review’s Four things to know about China’s new AI rules in 2024, which outlines China's AI policy focus for the year 2024.)
Nvidia CEO Visits China to Reinforce Commitment Amidst Global Trade Challenges
What’s New: Nvidia CEO Jensen Huang visited the company’s Beijing, Shanghai, and Shenzhen offices earlier this month, according to Bloomberg, attending annual meetings in China for the first time in several years. Images showing Huang wearing a traditional Chinese floral vest and engaging in a folk dance with employees went viral on social media during the weekend.
This visit underlines the significance of the Chinese market for the global chip giant.
Nvidia confirmed Huang had celebrated the upcoming Lunar New Year with staff, Bloomberg reported.
Why it Matters: Huang’s visit demonstrates Nvidia’s commitment to its Chinese teams and market. NVIDIA’s growth, propelled by the AI revolution and a trillion-dollar market valuation, has been partly driven by China’s demand for its GPUs. However, recent U.S. export controls have limited the export of its high-end GPUs to China.
Reuters previously reported Nvidia plans to mass produce in the second quarter of 2024 its next China-specific AI chip, H20, a scaled-down version of H100 GPU. Chinese media TMTPost said the H20 chip is expected to deliver only about 20% of the compute of NVIDIA’s flagship H100 GPU, in theory.
Weekly News Roundup
📣 Kaifu Lee’s 01.AI has open-sourced Yi-VL-34B, new Yi vision language models that rank No.1 in the open-source category per MMMU & CMMMU benchmarks.
🔍 Shanghai AI Lab collaborates with SenseTime, CUHK, and Fudan University to unveil the new large language model Shusheng·Puyu 2.0 (InternLM2).
📣 iFlyTek will unveil its upgraded Spark Cognitive Large Model V3.5 in a launch event on January 30.
💻 Lenovo introduces a range of AI-powered PCs including Legion and Xiaoxin, with personal Agent set to meet users in 3 months.
🌐 Baidu collaborates with Macao to integrate a travel assistant feature in ERNIE Bot, enhancing the travel experience for tourists in Macao.
🚗 BYD unveils its multimodal AI - Xuanji AI model, marking the first full-spectrum AI application in vehicles.
🧠 MiniMax releases its large language model Abab6, claiming to be China’s first trillion-parameter MoE-based model.
🌌 China Telecom open sources TeleChat-7B and offers a 1T high-quality clean dataset.
🎨 Tencent Advertising launches a one-stop AI ad creative platform Tencent Advertising Miaosi, powered by Tencent’s Hunyuan large model, offering diverse creative tools for advertisers.
Trending Research
MotionShop is a framework to replace the characters in videos with 3D avatars, proposed by Alibaba. More on Gitub.
Personal LLM Agents: Insights and Survey about the Capability, Efficiency and Security
Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model
Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data
InstantID: Zero-shot Identity-Preserving Generation in Seconds
UniVG: Towards UNIfied-modal Video Generation
Thanks for reading. If you’re enjoying this issue, feel free to subscribe to Recode China AI.