Key developments in AI and technology over the past 24 hours (October 24-25, 2025)

1. Model Releases and Updates
• China Genos Model Release: BGI-Research and Zhejiang Lab jointly launched Genos, the world's first deployable genome foundation model with 10 billion parameters, supporting genome data analysis and biomedical applications. The model emphasizes efficient computing and data privacy, with partial code open-sourced. 
• GLM 4.6 Model Release: A Chinese team introduced GLM 4.6, supporting 200k token long-context memory, capable of autonomously writing/fixing code, generating slides, dashboards, and documents. It is 4x faster and cheaper (input token cost ~$1/million) than ChatGPT. Fully open-sourced, it is considered one of the biggest AI breakthroughs of 2025. 
• Mem-α Model Release: Stanford AI Lab and Google DeepMind collaborated to apply reinforcement learning to memory management systems for the first time, solving the memory bottleneck issue in large language models, improving efficiency by 25-100x. The model is suitable for long-context reasoning and is open-sourced.
• SpikingBrain Model Release: A Chinese team launched the "brain-like" AI model SpikingBrain, designed for long-text processing (e.g., legal archives, medical records), with speeds 25-100x faster than standard systems, lower power consumption, and using domestic chips. The model "wakes up" only when needed, with partial frameworks open-sourced.
• DeepSeek-R1 and Other Chinese Open Models: DeepSeek released R1, expanding open-source AI access; Baidu invested ~$1.6B (110B RMB) in cloud and AI infrastructure; Alibaba launched the New Quark AI super-app, integrating multi-functionality based on the Tongyi model.

2. New Papers
• AI Native Daily Paper Digest: A collection of new papers, including:
◦ "Every Attention Matters": Efficient hybrid architecture for long-context reasoning, improving model attention mechanisms.
◦ "BAPO": Balanced policy optimization for off-policy reinforcement learning in LLMs, enhancing stability.
◦ "Language Models are Injective and Hence Invertible": Exploring the invertibility of language models.
◦ "LoongRL": Reinforcement learning for advanced long-context reasoning.
◦ Others like "Attention Sinks in Diffusion Language Models," "GigaBrain-0" (world-model-driven vision-language-action model), etc. These papers focus on long-context, reinforcement learning, and multimodal AI, open-sourced on Hugging Face.

3. Open-Source Projects
• China Analog AI Chip Project: A new analog AI chip prototype was released, potentially 1000x faster than Nvidia GPUs, focusing on efficient computing. The design framework is open-sourced, supporting domestic production.
• Tencent AI Code Generation Project: Tencent announced that 50% of new code is AI-assisted, open-sourcing related tools to improve development efficiency. 87 
• Jiu’an AI Model Deployment: China accelerated the deployment of the Jiu’an model for workplace safety and disaster response, open-sourcing partial APIs as the national emergency command "smart brain." 

4. Other Updates and Announcements
• China AI Policy Announcement: The Minister of Science and Technology stated plans to accelerate AI development, seek breakthroughs, strengthen chip resource design, AI governance, and promote AI as an international public good. Meanwhile, China's global AI policy trajectory emphasizes risk-based regulation (e.g., mandatory ethics committees) and AI exports.
• Ecosystem Trends: China's open-source ecosystem (e.g., DeepSeek, Qwen, Kimi) outperforms ChatGPT/Claude/Gemini in performance, with Qwen growing 68% and DeepSeek 20%. Alibaba Cloud launched the "Aegaeon" GPU pooling system, saving 82% resources.
• Global Progress: Anthropic released Claude "skills"; Microsoft integrated AI into Excel/Word; Google Video 3.1 improved text-to-video control; Reflection AI raised $2B for open-source frontier AI labs.