China Restricts Overseas Travel for AI Researchers — Exit Permits Now Required for Alibaba and DeepSeek Staff
China has reportedly introduced restrictions requiring top AI researchers at private companies such as Alibaba and DeepSeek to obtain official permission in advance before traveling abroad.
According to The Decoder, Beijing is tightening its grip on the domestic AI industry out of concerns over data leaks, technology theft, and talent poaching. The move appears to reflect a recognition that securing AI talent is directly tied to the national interest. On Reddit's LocalLLaMA community, users have raised concerns that the restrictions could negatively affect open-source model research coming out of China.
As the AI race intensifies, moves to restrict the free movement of talent could affect international collaboration within the research community.
UMG and TikTok Renew Deal with AI Music Safeguards
Universal Music Group (UMG) and TikTok have signed a new deal that includes measures against unauthorized AI-generated music.
According to TechCrunch, UMG has long pushed platforms, streaming services, and AI companies to adopt stricter content moderation policies. The renewed deal is notable in that it includes concrete countermeasures against the impact of AI music generation on the industry.
The music industry and technology platforms are steadily building rules around the handling of AI-generated content.
Qwen 3.6 27B Shows Stunning Coding Performance in the 27B Class
Despite its relatively compact size of 27B parameters, Qwen 3.6 27B is demonstrating Opus-level coding capability.
According to a report on Reddit's LocalLLaMA community, one user provided three API documentation files for an HTML5 game console and instructed it to "make a breakout game" — and the very first generation produced a playable, finished product. Controller input, graphics, sound, and integration with the console API all worked, and it was completed with just one round of additional customization and one bug fix.
The poster rated it as "a level no model other than Opus had reached," indicating that the practical capability of locally runnable mid-sized models has improved dramatically.
Keye-VL-2.0: First Multimodal Introduction of the DSA Attention Mechanism
Kwai (快手中控) has released Keye-VL-2.0-30B-A3B. The model pushes the frontier of long-form video understanding while also featuring the first agentic capabilities in the Keye series.
Its biggest highlight is that it is the first multimodal model to adopt the DSA (Dynamic Sparse Attention) mechanism. Compared to conventional standard attention mechanisms, it is designed to improve long-context processing efficiency while boosting video understanding accuracy. It is publicly available on HuggingFace and positioned as a flagship model in the 30B class.
The exploration of efficient attention mechanisms in multimodal models marks an important step toward making long-form content processing practical.
WAVE: A Cross-Vendor Portable GPU ISA Arrives
A researcher who read through 16 GPU microarchitectures and more than 5,000 pages of specification documents has developed WAVE, a portable GPU ISA that supports four vendors: NVIDIA, AMD, Intel, and Apple.
As reported on Reddit's MachineLearning community, the project lets you write a kernel once, and thin backends then convert it to Metal, PTX, HIP, and SYCL. The same binary has been verified to run on three platforms: Apple M4 Pro, NVIDIA T4, and AMD MI300X. PyTorch integration is also provided, and it has been confirmed that identical training results are obtained across all backends.
Efforts to reduce vendor dependence in GPU programming could enhance the flexibility of inference infrastructure.