腾讯研究院AI速递 20250922

Group 1: Chrome Update - Chrome has undergone its largest update since its launch in 2008, integrating the Gemini AI assistant into the browser for enhanced functionality [1] - The address bar has been upgraded to the "Omnibox" which intelligently recommends questions based on page content and allows complex queries directly [1] - The new version utilizes Gemini Nano for enhanced security, identifying harmful websites and managing notifications, and is currently available to US users [1] Group 2: Notion 3.0 Launch - Notion 3.0 has been officially launched, introducing the Agent feature that can autonomously perform all Notion operations [2] - The Agent can work independently for up to 20 minutes, completing complex tasks across tools such as integrating customer feedback and updating knowledge bases [2] - The new version includes a highly personalized "memory bank" and will soon support custom Agents for automated tasks and team sharing [2] Group 3: Tencent's Mixed Reality Studio - Tencent has released the "Mixed Yuan 3D Studio," aimed at 3D design professionals, which integrates AI technology to streamline the entire 3D asset production process [3] - The platform reduces production time from days to minutes and offers a comprehensive pipeline for various 3D creative tasks [3] - It features the industry-leading Mixed Yuan 3D 3.0 model with innovative capabilities such as segmentation generation and material editing [3] Group 4: Alibaba's Wan2.2-Animate Model - Alibaba Cloud has open-sourced the Wan2.2-Animate model, which supports generating animations for characters and animals, applicable in short video creation [4] - The model enhances character consistency and generation quality, offering modes for character imitation and role replacement [4] - The development team has created a large dataset for training, surpassing closed-source models in subjective evaluations [4] Group 5: Luma AI's Ray3 Model - Luma AI has launched Ray3, the world's first inference video model, advancing AI video from experimental to professional use [5][6] - Ray3 allows for fine control over actions and camera movements, generating previews in just 20 seconds at a fraction of the final rendering cost [6] - The model supports high-fidelity motion and lighting interactions, integrating seamlessly into professional post-production workflows [6] Group 6: ElevenLabs Studio 3.0 - ElevenLabs has introduced Studio 3.0, a comprehensive AI audio-video editor that consolidates narration, music, sound effects, subtitles, and video editing into a single timeline [7] - The new version offers over 10,000 AI voices, automatic music generation, and multi-language subtitle capabilities [7] - This tool is designed for video creators, podcasters, and audiobook authors, with API support for large-scale workflows [7] Group 7: Xiaomi's Xiaomi-MiMo-Audio Model - Xiaomi has open-sourced its first native end-to-end speech model, Xiaomi-MiMo-Audio, with 7 billion parameters and over 100 million hours of pre-training data [8] - The model excels in natural dialogue, audio subtitling, and long audio comprehension, showcasing capabilities in speech conversion and style transfer [8] - The development team has introduced a lossless compression model and achieved state-of-the-art results in various benchmark tests [8] Group 8: Retro Biosciences' RTR242 Drug Trial - Retro Biosciences has announced the initiation of human trials for the RTR242 drug in Australia, aimed at activating the autophagy system in aging cells [9] - The company's mission is to clear accumulated proteins in the brain to extend healthy human lifespan by 10 years, differing from traditional Alzheimer's treatments [9] - OpenAI has assisted in optimizing protein interactions for the drug, with plans to raise $1 billion to compete with other longevity research firms [9] Group 9: AI-Generated Genome by Evo - The Arc Institute and Stanford University have utilized the Evo model to create the world's first AI-generated functional bacteriophage genome, marking a new era in generative gene design [10][11] - The research team developed a specialized annotation pipeline to identify all genes in the bacteriophage, resulting in genomes with numerous new mutations [10] - Experimental validation confirmed that the AI-designed genomes could infect specific host strains, demonstrating the model's ability to coordinate complex mutations [11] Group 10: OpenAI Codex Applications - OpenAI has publicly shared seven core applications of Codex within its team, including code understanding, refactoring, and performance optimization [12] - The technical team has utilized Codex to enhance efficiency and code quality through various tasks such as generating unit tests and modifying multiple files [12] - Six best practices for using Codex have been disclosed, focusing on analysis before code generation and maintaining context for improved output quality [12]