The Chinese AI Iceberg

Summary

The video explores the shifting dynamics in AI research, with a focus on Chinese companies like Huawei, SenseTime, and others, making significant strides in open-source model development. It highlights the internal struggles within Huawei’s AI lab, the rise of Shanghai AI Laboratory and Ant Group, and the increasing dominance of Chinese firms in publishing cutting-edge research. The video also touches on U.S. companies pulling back from open-source contributions, with Chinese institutions filling the void. It reflects on how these shifts may alter the global AI landscape and research power balance in the future.

Takeaways

• 😀 Huawei's AI research division, Noah's Arc Lab, faced internal struggles, including technical setbacks and credit theft, after attempting to pass off open-source models as proprietary ones.
• 😀 SenseTime, an AI company previously known for facial recognition, has developed impressive multimodal models, like Senova V6.5, but their models remain inaccessible, making it hard to verify their claims.
• 😀 Shanghai AI Laboratory, established by the Shanghai Municipal Government, has become a powerhouse in AI research, publishing around 20 papers per month and releasing top-tier models like Intern S1.
• 😀 Ant Group, affiliated with Alibaba, has made significant strides in AI, especially with large language diffusion models, and recently released the trillion-parameter model Ring 1T.
• 😀 Cherry Studio is an open-source app that seamlessly connects AI models from China and abroad, providing easy access to different LLMs with a clean interface, which is gaining popularity.
• 😀 Chinese companies, including Xiaomi and MTOIN (a food delivery company), have entered the AI field with competitive open-source models, challenging established players in the space.
• 😀 The open-source AI landscape is shifting in favor of Chinese institutions, with several of them leading in publications and advancements, while U.S. companies are slowing down their open-source contributions.
• 😀 U.S. companies like Meta are stepping back from open-source AI development, potentially causing a shift in the global power balance, with Chinese institutions taking the lead in cutting-edge AI research.
• 😀 Ant Group's open-source initiative, Inclusion AI, is helping promote transparency and wider access to advanced AI models, pushing the open-source movement forward in China.
• 😀 The global AI development landscape is becoming more decentralized, with Chinese institutions emerging as the primary source of new open-source models and research, challenging the dominance of Western institutions.

Q & A

What technical setbacks did Huawei face with their Pangu LM project?

Huawei's Noah's Arc Lab encountered severe technical setbacks with the Ascend chip-based Pangu LM project. Whistleblowers revealed that the team was secretly wrapping rival open-source models, such as Quinn and Deepseek, releasing them as their own. They would then fine-tune these models to match Huawei's advertised parameter sizes, without creating truly original models from scratch.

What happened to the internal team at Huawei working on Pangu LM models?

The internal team at Huawei faced months of struggles trying to train original models from scratch with parameters ranging from 38 billion to 718 billion on Ascend hardware. However, a politically favored small lab bypassed Huawei's traceability rules, took their data, and claimed credit for the models, which led to fake performance reports and unearned bonuses for the smaller lab.

Why is SenseTime's Sense Nova V6.5 notable?

SenseTime's Sense Nova V6.5 is a multimodal reasoning model that performed better than competitors like Gemini 2.5 Pro and Claude on Sonnet. The company, which evolved from facial recognition research, is gaining attention for its impressive benchmarks. However, it remains unclear if the model's claims are true, as there are no ways to use it publicly yet.

What is the Shanghai AI Laboratory known for?

The Shanghai AI Laboratory is a leading R&D institute started by the Shanghai Municipal Government in 2020. It connects top talent from Chinese universities and is known for publishing cutting-edge AI research. Their 'Intern' LM series, especially the Intern S1 released in 2025, demonstrates significant advances, including vision capabilities, a feature not found in most other models.

How has Ant Group contributed to AI research?

Ant Group, an Alibaba-affiliated fintech company, has made significant strides in AI research. They pioneered concepts like large language diffusion models and introduced models like Mingai Tuni. In 2025, they released the Ring 1T, a one trillion parameter model that competes with some of the best open-source models available. They also launched the 'Inclusion AI' initiative to promote open-source releases.

What role does Cherry Studio play in the AI landscape?

Cherry Studio, an open-source app created by Ching Huay Technology, serves as a platform that connects and integrates language models from various providers, both inside and outside China. It makes it easy for users to switch between models and access tools in a clean, user-friendly interface. With over 32,000 stars on GitHub, it has gained popularity for bridging services across global AI ecosystems.

What is the significance of Shia Hong Su's Retino High Lab and its DOTS LM1 model?

Shia Hong Su, through the Retino High Lab, launched the DOTS LM1 model in June 2025, with 142 billion parameters and 14 active parameters. The model has gained attention for its promising cost-performance ratio and is open-source, making it a valuable addition to the AI community. The model's success indicates that platforms with no prior history in AI development can still make significant contributions.

What is Xiaomi's involvement in AI and its Mimo model series?

Xiaomi, traditionally known for smartphones and tech products, has entered the AI landscape with its Mimo model series. Launched in May 2025, Mimo includes a 7B vision reasoning model. Xiaomi’s involvement in AI is a natural extension of their hardware capabilities, especially for on-device language models, and their open-source approach has been well-received, with significant download counts.

How did MTOIN, a Chinese food delivery company, impact the AI scene?

MTOIN, a Chinese food delivery service, surprisingly entered the AI race by launching its MTON Longat lab, releasing state-of-the-art open-source models. Their Long Flash model, with 560 billion parameters, has been highly downloaded and comes with a groundbreaking research paper. MTOIN's entry into AI reflects how companies outside the traditional tech space are innovating in AI.

What are the implications of Chinese companies' increasing role in AI development?

Chinese companies are playing a growing role in AI development, particularly in the area of open-source models. As US companies like Meta slow down on open-source releases, Chinese institutions have become leaders in publishing papers and releasing advanced models. This shift in the AI ecosystem could alter the power balance in the industry, with Chinese firms gaining an edge in AI research and development.