Baidu Baijiahao Beta Test 'Baijiahao AI Assistant' to Enhance Creator Efficiency

Chinese research finds diffusion language models learn data 3x faster than autoregressive ones under token limits. A 1B-parameter model achieved 56% (HellaSwag) and 33% (MMLU) accuracy after 480 epochs. Key insights: 1) Bidirectional modeling boosts performance; 2) Robust to data reuse. Offers new training approach for token-constrained scenarios.....
Tencent's Q2 2025 financial report shows revenue of 184.5 billion yuan (up 15% year-over-year), R&D investment of 20.25 billion yuan (up 17%), and significant achievements in the AI field. Capital expenditures of 19.11 billion yuan (up 119%) were mainly used for AI technology research and development, including data enhancement and model optimization. The Huan Yuan series of 3D models performed outstandingly, with the v2.5 version and a high-quality 3D generation model released, attracting over 230,000 downloads in the open-source version. AI applications are accelerating their implementation, with smart features such as video interpretation added to WeChat/QQ's video channels, continuously enhancing user experience.
Tencent's R&D investment in the second quarter reached 20.25 billion yuan, with significant achievements in its AI strategy: revenue of 184.5 billion yuan (+15%), operating profit of 69.25 billion yuan (+18%). The Huan Yuan AI model capabilities continue to improve, with multiple breakthroughs in the 3D series model, with over 2.3 million downloads, widely applied in fields such as games and 3D printing. AI applications are accelerating their deployment, with Yuangao upgrading its core functions and integrating into the WeChat ecosystem, promoting AI services to penetrate multiple scenarios.
Rokid launches Cantonese-supported smart glasses overseas, integrating AI models for translation and recognition, with live navigation coming in October. Powered by Qualcomm AR chip, available via CSL Mobile with a HK$910 gift pack.....
Tencent reports DeepSeek denied rumors of August release for DeepSeek-R2 model, citing CEO's dissatisfaction with performance. The team continues optimizations, delaying May plans to enhance code generation and non-English reasoning. No official release date announced.....