🧐Shocking Deepfakes Resurrect Celebrities, AI's Red Lines, and China's Race for Long-Text LLMs
Weekly China AI News from March 11, 2024 to March 24, 2024
Hello, dear readers! In this issue, I discussed the legal and ethical concerns raised in China due to the creation of deepfake videos featuring deceased Chinese celebrities. A group of leading AI academicians, business leaders, and politicians gathered in Beijing to discuss the “red lines” of AI. Moonshot, a Chinese LLM startup, introduced a 2 million context window for its LLM.
Deepfake Videos of Deceased Chinese Celebrities Spark Ethical Concerns
What’s New: A deepfake video of Coco Lee, a renowned Chinese-American singer known for voicing Mulan who died from suicide last year, has gone viral on Chinese social media and ignited a firestorm of legal and ethical concerns.
The video was created by a Chinese blogger who claimed to fulfill Lee’s fan requests, featuring Lee warmly responding to fan concerns. The creator charged 588 yuan (~$82) for producing a deepfake video.
Mixed Feeling: Other Chinese celebrities who have passed away, such as Hong Kong singer Leslie Cheung and Chinese actor Kimi Qiao, have also been “brought back to life” as digital clones. Following the spread of a video that “revived” his son, Qiao’s father strongly objected and urgently demanded its removal as the video “reopened old wounds”.
Still, some embrace deepfake for personal reasons. For example, renowned musician Bao Xiaobo “revived” his daughter who died at the age of 22 using AI. Bao even sang a birthday song with his digital daughter on his wife’s birthday,
Why it Matters: This viral deepfake video of Coco Lee and others has not only opened Pandora’s box of tech capabilities but also highlighted the urgent need for legal and ethical guidelines in the age of AI.
Creating deepfakes of the deceased can affect their families and loved ones, who may find such representations disturbing.
The images and likenesses of deceased celebrities used to endorse products or services without permission are also concerning.
China’s Civil Code protects individuals from unauthorized use of their likenesses. However, the legal framework becomes murky as the deceased cannot consent to their likeness being used under these contexts.
Chinese and Western Experts Establish AI Development Red Lines
What’s New: 25 AI experts and business leaders from China and the West signed an open letter that calls for international cooperation to prevent AI from autonomous evolution or seeking power.
The letter was drafted at the International Dialogue on AI Safety in Beijing, a unique gathering on March 10-11 that brought together global experts including Turing Award winners Yoshua Bengio, Geoffrey Hinton, and Andrew Yao.
How it Works: The letter, named Consensus Statement on Red Lines in Artificial Intelligence, draws a parallel with the international cooperation during the Cold War that helped avoid nuclear disaster, suggesting that humanity faces a similar need for collaboration in the face of unprecedented technological challenges.
It proposes specific red lines that should not be crossed, including prohibitions on AI systems replicating or improving themselves without human approval, seeking power, assisting cyberattacks, or engaging in deception
The consensus outlines a roadmap for ensuring these red lines are respected. It stresses the need for comprehensive governance to prevent the development or deployment of systems that violate these red lines and calls for the immediate implementation of national registration requirements for AI models exceeding certain computational or capability thresholds.
Why it Matters: The statement emphasized the critical need for international collaboration in AI governance, which aligns with China’s growing interest in assuming a critical role in this field. The participation of veteran Chinese diplomat Fu Ying further underscored the Chinese government’s support for the dialogue and its resolutions.
Chinese LLMs Compete in Long Document Processing
What’s New: Last week, Chinese LLM startup Moonshot AI announced that its Kimi LLM can now accept inputs of over 2 million characters, leaping from 200K five months ago. Currently in beta testing, this new feature can for example transform Kimi LLM into a Poker expert that offers strategic card-play recommendations after learning hundreds of Poker tutorials.
Game on: Other Chinese companies have been rolling out similar functionalities. On March 14, Alibaba’s chatbot Tongyi Qianwen introduced a long document processing feature capable of analyzing documents over ten thousand pages long or about 10 million Chinese characters, possibly through Retrieval-Augmented Generation (RAG).
Baidu’s ERNIE Bot is also reportedly set to upgrade next month to support input lengths between 2 million and 5 million characters.
Why it Matters: LLMs are constantly improving to process and understand long texts. This leap forward is partly thanks to techniques like RAG, which enables LLMs to retrieve documents from extensive databases for generating responses.
Another more significant technique is the “long context window” — the length of text an LLM can analyze for its outputs. GPT-4 Turbo has increased from 32K to 128K, and so has Baidu’s ERNIE Speed model (128K). Gemini 1.5 and Claude 3 now support up to 1 million tokens, with Kimi stretching to 2 million Chinese characters. I found this blog particularly insightful in analyzing the long context windows. Shortly speaking:
Pros: Long context windows allow LLMs to reference more information, understand narrative flow, and maintain coherence over longer texts.
Cons: Expanding the context window size requires more computational power and memory. For example, 72 hours after the new feature was released, users found the KIMI mini-program, app, and web version down.
Weekly News Roundup
This week, a group of Chinese AI policy scholars unveiled their proposed draft for China’s AI Law. The document serves as a glimpse into the future direction of AI governance in China. Matt Sheehan from the Carnegie Endowment for International Peace posted a great summary of the draft law.
Last Thursday in Beijing, AMD’s Chair and CEO Lisa Su unveiled a series of partnerships aimed at enhancing the local ecosystem for AI PCs. (Shanghai Daily)
On Thursday, the United Nations unanimously adopted a resolution led by the U.S. and co-sponsored by China and over 110 countries to promote the safe and trustworthy development of AI. (Bloomberg)
Chinese Premier Li Qiang Li vowed more support for AI on March 13 during his visits to Baidu’s self-driving park and the Beijing Academy of Artificial Intelligence (BAAI). (SCMP)
Kevin Xu, founder of Interconnect, spotted a slide from BAAI that outlines “3 fundamental challenges to China’s domestic large AI model development.”
A fascinating story about the competition between Beijing and Shanghai to become the top LLM city in China. (TMTPost)
Apple reportedly engaged in preliminary discussions with Baidu about incorporating the Chinese company’s generative AI technology into its devices in China. (Wall Street Journal)
Trending Research
AnimateDiff-Lightning: Cross-Model Diffusion Distillation
Adding NVMe SSDs to Enable and Accelerate 100B Model Fine-tuning on a Single GPU
DreamReward: Text-to-3D Generation with Human Preference
LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models