🤤Remini's Viral Clay Filter, Alibaba's New LLM, China-US AI Gap Analysis, and One Million Tokens for $0.14

Weekly China AI News from May 6, 2024 to May 12, 2024

May 13, 2024

Hello readers, in this issue I will discuss Remini’s new clay filter, which has sparked a viral sensation across China. Alibaba released its latest LLM Qwen 2.5, touted as comparable to GPT-4 Turbo in performance. A Chinese tech media founder warned that China is 10 years behind the US in AI development, offering five key reasons for this assessment. A few Chinese LLMs are being offered at the price of $0.14 per million tokens. What does this mean for AI competition within China?

Clay Craze in China: Remini’s New Filter Turns Photos into Clay Figures

What’s New: Remini, an AI-powered image-filter app, has ignited a viral trend in China with its new clay filter that transforms photos into figures reminiscent of Shaun the Sheep.

While some users deem the clay filter as “ugly,” it has not stopped them from sharing their quirky clay images online. During China’s Labor Day holiday, the app’s daily downloads skyrocketed to over 380,000 and climbed to the top free app spot on Apple’s App Store in China.

How it Works: Developed by a Beijing-based startup, which was later acquired by Italian mobile app developer Bending Spoons, Remini first gained traction for restoring old photos a few years ago.

The app’s new clay filter allows users to transform existing photos into playful, clay-style animations. This feature is exclusive to Remini Pro users charged RMB68 per week after a 7-day free trial.

Why It Matters: Remini, along with Miaoya Camera last year, highlights the ongoing interest in and the commercial potential of AI-driven image editing applications.

While this trend has exploded across Asia, it hasn’t made as significant a splash in Western markets, yet.

Alibaba Cloud Unveils Qwen 2.5, Serves 90,000 Enterprises

What’s New: Alibaba Cloud last week released its latest LLM Qwen 2.5. China’s public cloud giant claimed Qwen 2.5 matches GPT-4 Turbo in tasks like text generation, knowledge queries, life advice, and casual conversations. Both models score 50 on the OpenCompass benchmark.

Alibaba also said its Qwen series of models have been servicing over 90,000 enterprises, covering a wide range of sectors including PCs, mobiles, automotive, aerospace, astronomy, mining, education, healthcare, dining, gaming, and cultural tourism.

How it Works: Compared to the previous version, Qwen 2.5’s enhancements include a 9% increase in understanding ability, 16% in logical reasoning, 19% in following instructions, and 10% in coding skills.

Qwen 2.5 can process documents with up to 10 million words and process up to 100 documents in one go. The model also demonstrates advanced abilities in audio-video understanding and coding.

Why It Matters: While Alibaba’s consumer-facing chatbot lags behind rivals like Baidu’s ERNIE Bot and Alibaba-backed Kimi, Alibaba Cloud is leveraging its strengths in cloud computing, extensive experience in enterprise services, and a broad array of open-source models to attract enterprises and developers to use its models on a Model-as-a-Service (MaaS) basis.

Alibaba claims to already have 90,000 enterprise customers using its Qwen LLM, with over 2.2 million businesses using it via DingTalk. These numbers reflect Qwen’s growing acceptance in the enterprise domain.

China is 10 Years Behind US in AI, Media Veteran Says

What’s New: Zhao Hejuan, founder of China’s top tech media TMTPost, recently highlighted five major misconceptions regarding China’s AI capabilities and its comparison with global counterparts, particularly the U.S. This comes amid growing international focus on AI development speeds and innovation quality across different nations.

How It Works:

Misconception of US-China AI Gaps: Zhao said the perceived 1-2 year gap in AI development between China and the U.S. is misleading. This timeframe mistakenly equates the rapid adoption of existing models with genuine innovation capability.
If GPT-5 is not released now, we might fall behind for 10 years. However, with its release, we could catch up in 2-3 years. It's crucial to understand that the advancements of the GPT-5 model represent their milestones, not our capabilities, highlighting the fundamental gap.
Overstated AI Market Dominance: Despite leading in AI patent filings and producing a vast number of AI-majored graduates, China trails in producing top-tier, impactful research (the top 1% most highly cited) and retaining high-level AI talent compared to the US and other countries.
Overly Focus on Computational Limitations: The focus on computational power as a bottleneck overlooks broader challenges in model innovation and data capability. Many of China’s large-scale data are either unusable or unorganized, hindering true computational advancements.
Open vs. Closed Source AI Models: The debate continues over which model type is superior. While each has its merits, the focus should rather be on the appropriateness and innovation within these models rather than their source availability.
The most important thing is whether we have our own innovation capabilities and originality, rather than engaging in low-level imitation.
If China’s battles of LLMs, whether they involve hundreds or thousands of models, lack any innovative elements and are merely low-level imitations and duplications, leading to internal redundancy, then indeed, not a single one is needed.
Hyped AI Industry Explosions: Expectations of rapid breakthroughs in industry-specific AI applications are premature. Industry transformation needs significantly capable foundation models before such advancements can truly take place.

Strategic Recommendations:

Enhance Long-term Innovation Capacity: There is a critical need to strengthen foundational innovation and build long-term capabilities in AI.
Patience with AI Integration Cycles: Every industry transformed by AI will start a new cycle from basic technological changes. These transformations won’t be instantaneous or occur overnight.
Open Attitude Toward Global AI Competition: Embrace global AI competition and challenges with a more open attitude. Avoid creating self-imposed barriers.

What's New: DeepSeek, an AI subsidiary of Chinese hedge fund manager High-Flyer Quant, has launched its latest open AI model, DeepSeek-V2. What’s staggering is, the model offers input/output processing at RMB1/2 per million tokens. This pricing strategy not only makes it more accessible but also poses a competitive challenge to other market players.

This 2nd-gen Mixture of Experts (MoE) model boasts 236 billion parameters, of which 21 billion are activated for each token. DeepSeek-V2 was trained on 8.1 trillion tokens. It achieves comparable performance with Llama 3 on MMLU and HumanEval.

Chinese LLM startup Zhipu AI has also reduced the price of its GLM-3-Turbo model from RMB5.00 to RMB1.00 per million tokens. Baidu last month also released a variant of ERNIE, named ERNIE Tiny, that only charges RMB1.00 per million tokens.

Why It Matters: DeepSeek-V2 is setting new lows for cost without compromising on performance. This strategic pricing move could trigger a price war in the AI sector, compelling other companies to innovate further or adjust their pricing strategies to remain competitive.

Weekly News Roundup

The Biden administration is reportedly preparing to introduce restrictions on the export of proprietary AI models, such as those used in systems like ChatGPT, to protect U.S. AI tech from China and Russia, according to sources familiar with the Commerce Department’s plans. (Reuters)
ByteDance’s AI-powered visual generation platform Dreamina has officially rebranded to Ji Meng. Alongside its new identity, the company announced that its AI-powered image and video creation features are now fully operational and available to all users.
Kaifu Lee’s 01.AI has unveiled Wan Zhi, a one-stop AI work platform that offers a variety of functionalities, including meeting summaries, weekly reports, writing assistance, financial report analysis, academic paper interpretation, and PowerPoint creation.
In China, a unique blend of technology and tradition is giving rise to a new market: AI-driven digital avatars that allow people to interact with deceased relatives. (MIT Technology Review)

Personal note

Finally, some readers may be familiar with my professional background. This past week has been messy, full of shock, pressure, and frustration. It’s uncommon to find oneself so closely connected to someone at the center of a negative news cycle in China. Rumors have been spread everywhere—some true, others wildly inaccurate. Humor abounds too; our team, department, and company have become the butt of jokes. This is understandable. We have no grounds for complaint. We simply have to swallow it and move on.

I cannot speak for anyone, but I believe my colleagues and I may share the same thoughts. We are committed to maintaining our professionalism and respect for our jobs and industry. We will rebuild trust and continue to communicate in a transparent, honest manner.

Recode China AI

Discussion about this post