👯‍♀️Alibaba, Baidu Expand Open-Source LLMs; AI App Generates Barbie-Style Avatars; Apple Removes LLM Apps from China App Store; Human Challenges AI through Robot Proxy

Dear readers, I’ve been writing this newsletter for over a year now. I would greatly appreciate any feedback on how I can improve the content to optimize your reading experience.

In this issue, I covered how Alibaba and Baidu are expanding support for open-source large language models (LLMs). Meet an AI app that generates Barbie-style avatars from photos. Additionally, Apple removed over 100 LLM-powered apps from its China app store.

Alibaba Open Sources Downsized Tongyi Qianwen LLM; Baidu Supports 33 LLMs including Llama 2

What’s new: Alibaba and Baidu are expanding their open-source large language model (LLM) offerings to win loves from developers and compete for enterprise customers.

Meet Qwen-7B: Alibaba Cloud released Qwen-7B and Qwen-7B-Chat, two open source, free, and commercially usable versions of its massive pre-trained large language model (LLM) Tongyi Qianwen.

Qwen-7B is a multilingual LLM trained on over two trillion tokens, with a context window up to 8k tokens. Qwen-7B-Chat is a conversational model built on top of Qwen-7B. According to Alibaba Cloud, benchmark results show Qwen-7B outperforming other open source models of similar size, with strengths in math, coding, and Chinese language understanding.

Why it matters: The release of Qwen-7B marks Alibaba Cloud’s entry into the lucrative but competitive open source LLM market. The smaller versions are designed to be more accessible to enterprises and SMEs compared to the original giant model. Even with Llama 2’s release, local Chinese models remain well positioned to win the domestic market, as Llama 2 stills needs improvement in Chinese language understanding, one Zhihu blogger claimed.

Baidu adds 33 open-source LLMs: Baidu AI Cloud also announced a major upgrade of its LLM platform Qianfan (千帆), supporting 33 advanced open-source LLMs including ChatGLM2–6B, RWKV-4-World, MPT-7B-Instruct, Falcon-7B, reducing inference costs by 50%. The platform also offers 103 pre-defined prompts for dialogue, games, and coding.

Barbie-Style Photo App 45AI Gains Traction Despite Technical Glitches

What’s new: Amid the buzz around the new Barbie movie, a mini program called “45AI” has gained traction in China for generating Barbie-style portraits. For 9.9 RMB, users can submit 9–15 photos and get AI-generated glamour shots.

How it works: 45AI relies on social sharing and word-of-mouth growth like previous viral hit Miaoya Camera. Paid users get three invite codes for friends to try it free. The tech likely isn’t too complex, as many clones have emerged since Miaoya’s success.

Technical glitch: 45AI saw overwhelming demand after launching July 22, with 10+ hour queues and payment issues from traffic surges. The development team apologized, halted orders, added servers and put staff on customer support. After fixing bugs, 45AI reopened on July 27.

Apple App Store Pulls LLM Apps in China Citing Lack of Licenses

What’s new: On July 31, Apple’s China-based App Store removed more than 100 apps providing generative AI services, a majority of which are seemingly powered by ChatGPT.

How it works: In a statement, Apple said deep synthesis tech providers in China must obtain licenses, including from the Ministry of Industry and Information Technology (MIIT), which these ChatGPT-related apps lacked. The lack of algorithm filing for ChatGPT was likely the main reason for the takedown, one Chinese media reported. Developers of apps that access LLMs, apart from the model providers, should also submit filings and get approval.

Surprisingly, iFlytek’s chatbot Spark was also taken down, though the app restored on the second day.

Why it matters: According to Article 20 of the Interim Measures for the Administration of Generative Artificial Intelligence Services, which will take effect on August 15, for generative AI services provided from outside the People’s Republic of China to within China that do not comply with laws, administrative regulations, and the provisions of these measures, the national cybersecurity authority shall notify the relevant agencies to take technical measures and other necessary actions to address the issue.

Robot Boxer Copies Human Moves in Real Time During Man vs AI Fight

What’s new: Three developers in China created a teleoperated robot that can mimic human boxing moves in real time, making the movie Real Steel a reality. With a latency of just 12ms — a huge improvement over similar tele robot with 100ms of latency — developers claimed QIBOT may be the quickest teleoperated robot built so far.

How it works: QIBBOT was optimized for speed through mechanical design and a specialized controller system. It combines reactive feedback controls with proactive feedforward controls that shortens latency. This allows the robot to mimic a human operator’s punches and blocks as they control it remotely using VR.

Human vs AI: The researchers tested QIBBOT in an entertaining boxing match against an AI opponent. The 2-meter-tall, 140 kg robot threw punches as quickly as its human operator. It’s punches could be even faster if controlled directly by AI. While created for entertainment, the low-latency control system could have applications in other areas needing quick teleoperated machines.

The three-man development team, which includes a father and son, created QIBBOT purely out of interest and have invested hundreds of thousands of dollars into the project since 2019.

Weekly News Roundup

🚕 Volkswagen Group announced an investment of around $700 million in XPeng Motors to enhance VW’s R&D capabilities for smart connected vehicles in China. The two companies plan two new NEV models to make debut in 2026. Within the same week, the company announced that its Vice President of Autonomous Driving, Xinzhou Wu, will leave the company for personal and family reasons. Wu is rumored to join NVIDIA.

🚗 Baidu Apollo announced Great Wall Motors and ECARX as first batch of partners to explore applications of its ERNIE LLM in smart cockpits. Some functions may first launch in Great Wall, Lynk & Co, and Smart production models.

🏭 Chipmaker Loongson announced it has taped out its new 3A6000 Quad-Core CPU, based on the LoongArch microarchitecture. Read more on Tom’s Hardware.

💊 Meet a LLM specialized for traditional Chinese Medicine: Qihuang Wendao (岐黄问道·大模型), developed by a Chinese medicine diagnosis and treatment platform named Dajing TCM. The LLM consists of a clinical diagnosis and treatment sub-model based on confirmed diseases, a clinical diagnosis and treatment sub-model based solely on symptoms and signs, and a traditional Chinese medicine health preservation and treatment sub-model.

💰 Hangzhou plans to provide up to 50 million yuan in subsidies for qualified foundation model R&D, and up to 5 million yuan annually for up to 10 excellent specialized models successfully implemented in Hangzhou.

🤖 Geely plans to release the automotive industry’s first full-scenario AI model, covering capabilities like drawing, music and language.

🌤 Huawei Cloud’s climate model, Pangu-Weather, has officially launched on the European Centre for Medium-Range Weather Forecasts website, offering free 10-day global weather forecasts.

🚕 Pony.ai, Toyota Motor China, and GAC Toyota signed an agreement to set up a joint venture to support mass production and deployment of robotaxis. The JV will launch within the year with over 1 billion RMB investment.

Trending Research

DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models

• Existing large text-to-image models can generate high quality images but lack precision when editing generated or real images. The proposed DragonDiffusion method enables drag-style editing on diffusion models. It constructs classifier guidance to transform editing signals into gradients that modify the diffusion model’s intermediate representation. This allows various precise editing modes like moving, resizing, replacing objects and content dragging without extra model tuning or components. (Affiliations: School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University; ARC Lab, Tencent PCG)