Group 1: OpenAI and AI Model Developments - OpenAI has launched the GPT-5.1 series models, emphasizing that effective AI should not only be intelligent but also engaging in conversations [1] - The GPT-5.1 Instant model is designed to be warmer, smarter, and better at following instructions [1] - The GPT-5.1 Thinking model focuses on advanced reasoning, performing faster on simple tasks and more persistently on complex ones [1] Group 2: 3D World Generation by Li Feifei's Team - Li Feifei's team, World Labs, has released the Marble model for 3D world generation, supporting various input modalities including text, images, and videos [2] - Marble introduces AI-native editing tools for local replacements and structural adjustments, with the Chisel feature allowing for style separation [2] - Subscription options range from a free version (7000 points/month) to a flagship version (120000 points/month), supporting multiple export formats for game engines [2] Group 3: Anthropic's Infrastructure Investment - Anthropic has announced a $50 billion partnership with Fluidstack to build customized data centers in Texas and New York [3] - This marks Anthropic's first significant investment in tailored infrastructure, aligning with its internal forecast of achieving $70 billion in revenue and $17 billion in positive cash flow by 2028 [3] - Fluidstack, established in 2017, has collaborated with companies like Meta and Mistral and is among the first third-party suppliers to receive Google's custom TPU [3] Group 4: Google Gemini Voice Upgrade - Google has upgraded its Gemini Live voice capabilities, introducing features like real-time speech rate adjustment and emotional tone responses [4] - The Gemini 2.5 Flash model has significantly improved the voice engine's ability to model nuances in tone, stress, pauses, and pitch variations [4] - The upgraded voice features are seamlessly integrated into the Google ecosystem, allowing for hands-free activation and ensuring that voice data is not stored by default [4] Group 5: Baidu's Wenxin 5.0 Release - Baidu has officially launched Wenxin 5.0, which focuses on a native multimodal approach, integrating language, images, video, and audio into a unified training framework [5] - The model supports full multimodal input and multi-output capabilities, achieving a score of 1432 on the LMArena text leaderboard [5] - With over 2.4 trillion parameters, the model employs a sparse activation design with an activation ratio below 3%, and is available on various platforms [5] Group 6: Tencent's Industrial-Grade Model - Tencent has introduced the industrial-grade native multimodal model, Mixed Yuan Image 3.0, available on LiblibAI [6] - This model can accurately interpret complex prompts and generate coherent content, supporting both Chinese and English text generation [6] - It excels in aspects like realistic lighting, material styles, and logical continuity in content generation [6] Group 7: Sina Weibo's VibeThinker-1.5B Model - Sina Weibo has released the open-source VibeThinker-1.5B model, which has 1.5 billion parameters and a training cost of under $8000 [7] - The model outperformed larger models in top mathematical competition benchmarks, showcasing its efficiency [7] - It utilizes an innovative principle to decouple training objectives, achieving a remarkable cost-effectiveness ratio [7] Group 8: Google DeepMind's AlphaProof - Google DeepMind's AlphaProof system has published its technical details after winning a silver medal at the 2024 IMO [8] - The core innovation combines Lean formal language with reinforcement learning, generating a vast number of formal statements from natural language math propositions [8] - The system employs "Test-Time Reinforcement Learning" to progressively tackle complex problems through easier variants [8] Group 9: New Coding Evaluation System - LMArena has launched a new coding evaluation system called Code Arena, which reconstructs the assessment of code performance and interaction quality [9] - The domestic model GLM-4.6 has topped the new rankings, tying with Claude and GPT-5, surpassing Gemini and Grok [9] - GLM-4.6 achieved a code modification success rate of 94.9%, narrowing the gap with Claude Sonnet 4.5 [9]