腾讯研究院AI速递 20260312

Group 1 - Google has released its first native multimodal embedding model, Gemini Embedding 2, which integrates text, images, audio, video, and PDFs into a unified vector space for cross-modal retrieval, eliminating information loss and engineering complexity associated with traditional multi-model approaches [1] - The model employs Russian nesting doll representation learning technology (MRL), reducing vector dimensions from 3072 to 768 with only a 0.18 point loss, balancing performance and cost [1] - Compared to competitors like OpenAI, Cohere, and Jina, Gemini Embedding 2 is currently the only commercial-grade embedding model covering five modalities, achieving state-of-the-art (SOTA) in multimodal capabilities [1] Group 2 - Meta has acquired the AI social platform Moltbook, which was launched only 40 days prior, and its founders will join Meta's Super Intelligence Lab (MSL) [2] - Moltbook gained attention for its "encrypted language" hoax posts, revealing critical flaws in identity verification and security on the platform [2] - Meta is interested in Moltbook's underlying capabilities for AI agents to be permanently online and autonomously discover and connect, planning to integrate it into its AI ecosystem after addressing security issues [2] Group 3 - Nvidia has signed a $60 billion computing power agreement with Thinking Machines Lab to deploy a next-generation Vera Rubin computing cluster with a scale of 1GW, including cash investment from Nvidia [3] - The company, founded by former OpenAI CTO Mira Murati, has a current valuation of $50 billion and recently appointed the founder of PyTorch as its new CTO after a previous departure [3] - By securing scarce computing power capacity, the company is building a second moat beyond talent competition to support cutting-edge model training and enterprise-level AI customization [3] Group 4 - Tencent has responded to security and cost concerns regarding OpenClaw, emphasizing that installation is free but model calls incur token fees, and has launched a series of security products and the SkillHub plugin ecosystem [4] - Tencent's lobster services are divided into two categories: deployment solutions around the open-source OpenClaw and the self-developed desktop AI agent WorkBuddy, which shares the Agent architecture with CodeBuddy [4] - Nearly 40,000 Tencent employees are already using OpenClaw internally, indicating a shift towards a new development model involving agents [4] Group 5 - Tencent Cloud has launched the SkillHub skill marketplace, which has over 13,000 lobster skills, supporting Chinese search and domestic node acceleration, and is compatible with various agent frameworks and environments [5] - More than 50 high-quality skills covering frequent scenarios such as office collaboration, development tools, and content creation have been selected, completing security scans and quality filtering [5] - Over 10 products, including Tencent Docs, QQ Browser, and Tencent Maps, have undergone skill transformation, allowing for one-click integration with OpenClaw [5] Group 6 - ZhiMi has launched its chip brand "Xinjichuan Yue," which includes a mobile chip (Chixiao 01) with a self-developed NPU architecture, a 2nm process autonomous driving chip with 2000 TOPS computing power, and the mass-produced Tianqiong series of general-purpose robot chips [7] - The company has announced plans for a space computing center, intending to deploy 2 million computing satellites to form a super constellation, with the first space computing box set to be verified in orbit soon [7] - A personal super AI computer with 1.5 PFLOPS has been introduced, utilizing a unified memory architecture capable of locally loading large models with billions of parameters and supporting multi-device networking [7] Group 7 - On the 10th anniversary of AlphaGo's victory over Lee Sedol, Demis Hassabis reflects on how the "37th move" initiated the modern AI era, showcasing AI's ability to surpass human experts and autonomously discover new strategies [8] - The reinforcement learning and search methods from AlphaGo have been extended to various scientific fields, including AlphaFold for protein structure prediction and AlphaProof for mathematical reasoning [8] - DeepMind believes that combining Gemini's world model, AlphaGo's search planning, and specialized AI tools is key to achieving artificial general intelligence (AGI) [8] Group 8 - The new version of LeRobot v0.5.0 integrates full-body control for the Unitree G1 humanoid robot and introduces six new strategy models, supporting LoRA fine-tuning [9] - Dataset processing has been significantly optimized, with streaming video encoding achieving zero-wait recording and image training speed improved by ten times, along with the launch of EnvHub for direct loading of simulation environments [9] - The codebase has been upgraded to Python 3.12+ and Transformers v5, with a new third-party strategy plugin system, and the paper has been accepted by ICLR 2026 [9] Group 9 - The author of "The Biography of Hassabis" reveals the contradictory personality of Demis Hassabis, who dislikes control yet is extremely competitive, and his clear aversion to the pursuit of power [10] - Hassabis attempted for three years to make DeepMind independent from Google but failed, and the book contains details that contradict Hassabis's own statements following the dismissal of co-founder Suleyman [10] - The author believes that AI safety faces the "Oppenheimer dilemma," where scientists can build technology but cannot control its use, ultimately requiring cooperation between the US and China to achieve safety [10] Group 10 - Jensen Huang outlines a five-layer industrial system for AI, consisting of energy, chips, infrastructure, models, and applications, where each successful application drives demand down to power generation [11] - He asserts that AI has crossed a critical threshold, with significant improvements in model inference and deployment capabilities, and open-source models like DeepSeek-R1 accelerating the growth of full-stack demand [11] - While thousands of billions have been invested globally, trillions in infrastructure remain to be built, marking it as the largest infrastructure project in human history [11]