腾讯研究院AI速递 20250718

Group 1 - Google DeepMind's MoR architecture achieves two times inference speed by combining parameter sharing and adaptive computation, resulting in fewer parameters while maintaining large model performance [1] - The dynamic routing mechanism allocates different recursive depths based on token complexity, reducing redundant computations and optimizing KV cache [1] - Experimental results show that MoR improves inference throughput by 2.06 times, reduces training time by 19%, and decreases peak memory usage by 25% [1] Group 2 - Amazon launches Bedrock AgentCore preview, offering seven core AI agent services including runtime, memory, and authentication [2] - The introduction of Nova customization options and Strands Agents V1.0 simplifies agent development and enables multi-agent collaboration [2] - Amazon S3 Vectors cloud object storage is released, reducing vector storage costs by 90%, along with Kiro AI IDE to enhance developer experience [2] Group 3 - Elon Musk is seeking names for the male AI companion Grok, with suggestions like "Draven" that align with characters from "Twilight" and "Fifty Shades of Grey" [3] - A user named Jackywine has created an open-source 3D digital companion "Bella," which retains only the visual aspect without large language model capabilities [3] - The "Bella" project follows an "AI native" development path in three phases: perception core, generative self, and proactive companionship, with plans to incorporate voice recognition and affinity systems [3] Group 4 - Google Search introduces an AI feature that can make phone calls to book local services for users, such as pet grooming [4] - The search integrates the Gemini 2.5 Pro model and Deep Search functionality, capable of handling complex queries and generating in-depth reports [4] - This new feature has launched in the U.S. and will be gradually rolled out globally, sparking discussions about the effectiveness of AI automated calls and merchant experiences [4] Group 5 - The AI programming platform Windsurf reintroduces the Claude Sonnet 4 model, allowing Pro users 250 free calls per month [6] - Claude Sonnet 4 offers advantages such as cross-file intelligent refactoring, a 200,000 token context window, and precise code completion [6] - This renewed partnership follows OpenAI's acquisition failure and executive team changes, representing Windsurf's strategic move to regain user trust [6] Group 6 - Anthropic successfully rehires core programming leaders Boris Cherny and Cat Wu from Cursor within two weeks [7] - Anthropic reveals that direct sales of models and Claude yield a gross margin of 60%, while sales through AWS and Google Cloud result in a negative 30% margin [7] - Claude Code has become a new asset for Anthropic, with weekly downloads increasing sixfold to 3 million since June, contributing over $200 million in annualized revenue [7] Group 7 - CrePal launches the first AI video creation agent, allowing users to produce videos through a single command that orchestrates multiple models [8] - The system can automatically plan scripts, select appropriate models, generate visuals, and add sound effects, addressing high barriers in traditional AI video creation [8] - The innovation lies in transforming the creative process, enabling users to focus on creative expression rather than technical operations by integrating dispersed tools into a unified intelligent task [8] Group 8 - Apple's MLX framework adds CUDA support, enabling developers to train models using NVIDIA GPUs and deploy them back to Apple devices [9] - This move is seen as Apple's concession to the NVIDIA ecosystem, which dominates AI development with 5 million developers [9] - Despite past tensions over NVIDIA support, Apple opts to leverage NVIDIA's ecosystem for compliance and to expand its influence [9] Group 9 - HeShan Technology, founded by alumni from Tsinghua and Beihang University, focuses on AI tactile sensing technology and has developed the world's first AI tactile perception chip [10] - Utilizing capacitive tomography technology, HeShan achieves "sensing and control integration," addressing the tactile feedback needs in robotic precision operations [10] - The company has completed four rounds of financing and serves over 70% of domestic robot manufacturers, transitioning from a hardware provider to a comprehensive tactile solution provider [10] Group 10 - Nobel laureate John Jumper discusses the journey of AlphaFold, highlighting that the value of algorithm research is 100 times that of data [11] - AlphaFold predicts protein structures with atomic-level precision and has been cited 35,000 times, accelerating scientific discoveries [11] - Jumper predicts that AI4Science will become more generalized in the future, with AlphaFold enhancing the pace of structural biology development by 5-10%, leading to widespread advancements across scientific fields [11]