腾讯研究院AI速递 20250918

Group 1 - Li Feifei's company World Labs launched the spatial intelligence model Marble, capable of generating large-scale 3D worlds from a single image or text prompt [1] - Marble offers larger scale, more diverse styles, and cleaner geometric structures compared to previous products, supporting free navigation in browsers [1] - Users can export generated worlds as Gaussian point clouds for efficient operation on desktop, mobile devices, and VR headsets, with whitelist testing now open [1] Group 2 - Google partnered with over 60 institutions, including American Express and PayPal, to introduce the AI Payment Protocol (AP2) aimed at creating a secure standard for AI agent payments [2] - AP2 builds trust through "Mandates," using encrypted digital contracts as proof of user instructions, allowing pre-authorization for AI agents to make purchases under specific conditions [2] - The protocol supports real-time purchases and automated tasks without human involvement, with an encrypted version A2A x402 enabling stablecoin payments, and a GitHub repository is available for developers [2] Group 3 - Anthropic plans to invest $10 billion to create enterprise application clones, while OpenAI expects to spend $8 billion on data-related costs by 2030 [3] - Both companies are training AI models to operate various professional software using a "reinforcement learning environment" that simulates enterprise applications [3] - They may hire domain experts to demonstrate task execution, aiming to develop AI as "virtual colleagues" and open new revenue streams [3] Group 4 - Tencent Cloud announced the global launch of its upgraded Intelligent Agent Development Platform 3.0 (ADP3.0), which has seen nearly 600 features launched in the past three months [4] - The platform upgrade includes enhanced knowledge base management, multi-agent collaboration support, global agent visibility in workflows, and instant command capabilities [4] - Targeted industry agents for smart quality inspection and media content processing have been introduced, with Youtu-Agent framework and Youtu-GraphRAG knowledge graph framework set to be open-sourced [4] Group 5 - Disney, Warner Bros., and Universal Pictures filed a lawsuit against Chinese AI company MiniMax, accusing it of unauthorized use of IPs like Spider-Man for AI training [5] - The companies seek restitution for infringement profits and damages of up to $150,000 per infringement, along with a permanent injunction to prevent MiniMax from using related IPs [5] - MiniMax previously faced similar accusations from iQIYI regarding the drama "Canglan Jue," highlighting significant risks in IP imitation within AIGC [6] Group 6 - The AI tool ima has been updated to support audio file uploads in formats like MP3, M4A, WAV, and AAC, enabling automatic generation of transcripts, summaries, and notes [7] - The update includes a screenshot shortcut feature for desktop users, allowing direct questioning, knowledge base addition, or note-taking after capturing images [7] - Mobile note-taking now supports offline editing and creation, with automatic synchronization once reconnected to the internet [7] Group 7 - YouTube introduced a generative AI tool for Shorts creators, incorporating a customized version of Google's text-to-video model Veo 3, enabling low-latency content generation at 480p resolution [8] - The new version allows for sound addition and dynamic effects application to static images [8] - YouTube also launched a "voice-to-song" remix tool based on Google's Lyria 2 and an "AI editing" feature that automatically organizes highlights, adds music, and transitions [8] Group 8 - Figure, a humanoid robotics company, completed a Series C funding round, raising over $1 billion and achieving a post-money valuation of $39 billion, the highest in the embodied intelligence sector [9] - The funding round was led by Parkway Venture Capital, with participation from Nvidia and Intel Capital, aimed at expanding production capacity and building GPU infrastructure [9] - Figure has rapidly progressed since parting ways with OpenAI, launching the Helix end-to-end "vision-language-action" model, with robots capable of complex tasks like folding clothes and sorting packages [9] Group 9 - Huawei released two research reports, "Intelligent World 2035" and "Global Digital Intelligence Index 2025," forecasting key technological trends and their industry impacts over the next decade [10] - The reports predict ten major trends, including AGI as a transformative force, AI agents evolving from execution tools to decision-making partners, and human-machine collaborative programming becoming mainstream [10] - It is anticipated that by 2035, total computing power will increase by 100,000 times, AI storage capacity demand will grow by 500 times compared to 2025, and renewable energy generation will exceed 50% [10] Group 10 - Shopify shared insights on the evolution of its AI assistant Sidekick, recommending a simple architecture, clear tool boundaries, and a modular design approach [11] - The company suggested replacing "golden datasets" with "benchmark truth sets" that reflect real production environments, aligning large language model evaluations with human assessments [11] - Shopify warned about "reward hacking" issues and advised establishing detection mechanisms in advance, combining programmatic validation with semantic evaluation to create a multi-layer reward system [11]