What happens when a man who loves playing with fire gets his hands on AI?

💻 PEER: A New Chapter for MoE Models

🤖 Amazon Launches AI Shopping Assistant

🚫 OpenAI Whistleblower Files Complaint

📊 Japan's MIC: China Leads in GenAI

🌱 Shennong LLM 2.0 Debuts

📙 Kuaikan Tech Develops ACG Foundation Model

💼 Guofei Jiang Joins China's AI "National Team"

💰 Speak Hits $500 Million Valuation

📈 Houmo Smart Chip Closes Strategic Round Worth Hundreds of Millions of RMB

🎆 Cai Guo-Qiang, the Artist Who Plays With Fire, to Stage an AI-Powered Fireworks Show

PEER: A New Chapter for MoE Models

Xu Owen He, a research scientist at DeepMind, has proposed an innovation called the PEER architecture[1], a significant improvement on MoE models. The PEER architecture leverages product key techniques to enable efficient retrieval from over one million tiny experts, effectively solving the bottleneck that traditional MoE models face in scaling expert count.

The PEER architecture offers the AI field a new way to expand model capacity while maintaining computational efficiency. On language modeling tasks, PEER layers demonstrate superior performance-compute tradeoffs compared to traditional dense FFW and coarse-grained MoE approaches. Its introduction points toward a new direction for designing and training large AI models, particularly in scenarios involving massive parameters and data.

Amazon Launches AI Shopping Assistant

After five months of consumer testing, Amazon's AI shopping assistant Rufus[2] is now available to all US customers.

Users can ask Rufus shopping-related questions through a text chat interface in the smartphone app — everything from identifying durable outdoor speaker models to comparing running shoes to getting order updates.

OpenAI Whistleblower Files Complaint

A whistleblower at OpenAI has filed a complaint with the US Securities and Exchange Commission (SEC), alleging that the company illegally barred employees from warning regulators about potentially severe risks posed by its AI technology, and calling for an investigation into OpenAI's conduct.

The complaint states that employment, severance, and confidentiality agreements required by OpenAI are overly restrictive and could penalize employees who raise concerns about the company with federal regulators.

Japan's MIC: China Leads in GenAI

Japan's Ministry of Internal Affairs and Communications (MIC)[3] has published its 2024 Generative AI Usage White Paper on its official website. The white paper surveyed Japanese citizens on their use of more than 10 generative AI tools including ChatGPT, Copilot, Midjourney, Stable Diffusion, Gen-2, and MusicGen, while also conducting cross-country comparisons with the United States, China, the United Kingdom, and Germany.

The results show that China topped individual usage rates at 56.3%, followed by the United States at 46.3%, with the UK, Germany, and Japan trailing at 39.8%, 34.6%, and 9.1% respectively.

In enterprise adoption, China leads by a wide margin at 71.2%, reflecting its strength in productizing generative AI and driving real-world applications.

Shennong LLM 2.0 Debuts

At the 8th International Conference on Smart Agriculture Innovation and Development, China Agricultural University unveiled this agricultural AI model covering the full spectrum of applications: breeding, crop cultivation, livestock farming, agricultural remote sensing, and meteorology.

It comprises four specialized agricultural foundation models: the breeding model "Shennong · Gu Xin," the cultivation model "Shennong · Zhu Ji," the livestock model "Shennong · Qiang Mu," and the remote sensing and meteorology model "Shennong · Wen Qiong." Through multimodal interaction and intelligent reasoning, these models significantly enhance the intelligence and efficiency of agricultural production.

Shennong LLM 2.0 not only demonstrates China's technical capabilities in agricultural AI but also offers new ideas and tools for the global development of smart agriculture.

Kuaikan Tech Develops ACG Foundation Model

Kuaikan Manhua is fine-tuning an open-source foundation model, using RAG (Retrieval-Augmented Generation) technology to develop a vertical foundation model for ACG (anime, comics, and games), currently in internal testing.

In the future, users will be able to have immersive conversations with IP characters — not just purely virtual ones. Kuaikan Manhua hopes these technologies will substantially improve user experience and drive the development of IP character derivative applications, strengthening interaction between users and IP characters and delivering richer, more personalized reading experiences.

Guofei Jiang Joins China's AI "National Team"

According to reports, Guofei Jiang, former senior vice president at Ant Group and president of Ant Digital Technologies, will join China Electronics Corporation in a key role.

In his new position, Jiang will be responsible for advancing the implementation of national strategies around data elements and artificial intelligence. This marks a further deployment of his deep expertise in data science and AI, bringing fresh perspective and momentum to China Electronics Corporation and helping advance national strategic implementation in data elements and artificial intelligence.

Speak Hits $500 Million Valuation

Speak recently closed a $20 million Series B-3 round led by Buckley Ventures, with participation from the OpenAI Startup Fund, Khosla Ventures, Y Combinator co-founder Paul Graham, and LinkedIn executive chairman Jeff Weiner. Following this round, Speak's total funding reaches $84 million, with two rounds raised in 2023 alone, at a valuation of $500 million.

Speak is an AI-powered language learning app that applies AI across the entire chain, from real-person video instruction to AI-powered spoken conversation and correction. Since launching in South Korea in 2019, it has accumulated over 10 million users. Speak's global dual-platform revenue for the past 30 days was $947,000, ranking third worldwide among comparable products. With this funding, Speak will accelerate its global market expansion.

Houmo Smart Chip Closes Strategic Round Worth Hundreds of Millions of RMB

Houmo Smart Chip has received strategic investment from two digital economy industry funds under China Mobile, marking recognition of the company's innovative strength and market potential in AI chips. Houmo Smart Chip has formally become a key edge-side large model chip company supported by the China Mobile ecosystem. According to reports, Houmo's recently launched M30 delivers up to 100 TOPS of compute at 12W power consumption; its next-generation chip, built on the new "Tianxuan" architecture, will deliver another multi-fold leap in computational efficiency.

China Mobile Research Institute will leverage Houmo's technical strengths, focusing on product demand analysis and new scenario exploration while driving the formulation and promotion of software toolchain standards. Houmo will provide diverse compute support, delivering innovative solutions for deploying and applying edge-side and device-side large models. The two sides will jointly explore new product forms for edge-side large models in scenarios including enterprise-government integrated large model machines, homegrown AI PCs, and home computing boxes, while driving commercialization.

Cai Guo-Qiang, the Artist Who Plays With Fire, to Stage an AI-Powered Fireworks Show

At the crossing of technology and the humanities, a spectacular fireworks display awaits.

Soon, Cai Guo-Qiang's AI-powered fireworks show will light up the Los Angeles Memorial Coliseum. WE ARE is co-conceived and orchestrated by Cai Guo-Qiang[4] alongside cAI™, a multimodal AI model developed specifically for him. The project is commissioned and hosted by the Getty Center, realized in collaboration with the University of Southern California (USC).

The event marks the first large-scale daytime fireworks and drone fireworks performance in US history, representing a breakthrough for cAI™ in the art of fireworks. WE ARE is not merely a visual feast but a symbol of the collision between art and science, the fusion of tradition and modernity — inspiring boundless imagination about what the convergence of art and technology might become.

WE ARE is a free event; tickets will open for reservation later this summer. Below, the venue where the fireworks will take place: the Los Angeles Memorial Coliseum.

Live truthfully, and don't forget to look at the moon every day.

— Ryuichi Sakamoto

Editorial Team Editor: Yuki

Design: Ivan

For business partnerships, add on WeChat: Rwkfbcianvd

References [1] PEER architecture: https://arxiv.org/abs/2407.04153

[2] Rufus: https://www.aboutamazon.com/news/retail/how-to-use-amazon-rufus

[3] Japan's Ministry of Internal Affairs and Communications: https://www.soumu.go.jp/

[4] Cai Guo-Qiang: http://www.caiguoqiang.com/