腾讯研究院AI速递 20260303

Group 1: Nvidia and OpenAI Developments - Nvidia will launch a dedicated inference chip based on the Groq LPU architecture at the GTC conference, with OpenAI as the first customer, providing 3GW of dedicated inference computing power [1] - The LPU uses high-density on-chip SRAM instead of GPU's HBM solution, significantly reducing latency and energy consumption, with theoretical inference speeds up to 100 times faster than GPUs [1] - Nvidia invested approximately $20 billion to acquire Groq's core technology and team, marking its first large-scale introduction of external architecture design into its core AI product line [1] Group 2: OpenAI's GPT-5.4 Leak - An OpenAI engineer accidentally leaked the "gpt-5.4" model in the Codex public GitHub repository, which was quickly modified to "gpt-5.3-codex," with rumors suggesting the new version may launch as early as next week [2] - Key upgrades focus on a 2 million Tokens context window and "stateful AI," enabling cross-session persistent memory, which retains workflow and tool invocation states, eliminating the need to repeatedly explain project backgrounds [2] - The new version includes full-resolution visual reading capabilities, allowing for pixel-level visual analysis by bypassing traditional image compression [2] Group 3: Anthropic's Claude Updates - Anthropic has introduced a "memory import" feature for Claude, allowing users to transfer their ChatGPT conversation preferences and work styles in 60 seconds through a simple copy-paste process [3] - Following a partnership announcement with the Pentagon, the QuitGPT topic surged, resulting in 700,000 users canceling their ChatGPT subscriptions and uninstalling the app, while Claude topped the App Store charts [3] - This feature significantly reduces the cost of switching for users, sparking discussions on "digital sovereignty" regarding the portability of AI memory data [3] Group 4: OpenClaw Directory Launch - The third-party OpenClaw Directory website has launched, featuring 39 ecosystem tools categorized into nine major categories, with support for sorting by popularity and ratings [4] - The top six tools include Claw for All, OpenClaw Launch, ClawTeam, and Vibeclaw, among others [4] - The site also offers a comprehensive tutorial library covering everything from introductory science to deployment selection and token optimization, allowing developers to submit their own OpenClaw tools [4] Group 5: Meituan's AI Browser Tabbit - Meituan's team has released the AI browser Tabbit, which features an "intelligent agent mode" capable of automating web tasks, extracting information, filling out forms, and exporting to Excel [5][6] - Tabbit includes "tricks" and "scripts" functionalities, allowing users to save frequent operations as shortcuts using natural language, and has integrated multiple models [6] - Meituan's AI strategy is expanding from core local life scenarios to a general internet entry point, facing the challenge of differentiation in a crowded AI browser market [6] Group 6: Tongyi's Voice Generation Models - Tongyi Lab has launched Fun-CosyVoice3.5 and Fun-AudioGen-VD models, enabling voice generation controlled by natural language commands, moving beyond traditional preset labels [7] - CosyVoice3.5 now supports four additional languages, covering a total of 13 languages, with a reduction in rare character mispronunciation rates from 15.2% to 5.3% and a 35% decrease in initial latency [7] - AudioGen-VD allows for the design of sound and scenes from textual descriptions, supporting character simulation, environmental sound layering, and spatial reverb effects, enhancing voice generation from a functional tool to a creative one [7] Group 7: Research AI Partner "Da Sheng" - A collaboration between the Institute of Advanced Intelligence, Fudan University, and Infinite Light Year has resulted in the release of the super research partner "Da Sheng," which possesses four capabilities: cognition, action, memory, and verification [8] - The platform has accumulated over 300 reusable research skills covering more than 20 categories, supported by a Git-style multi-branch collective memory architecture for long-term research [8] - It has established a closed loop of "cloud prediction → intelligent wet experiments → data feedback → model updates," improving the efficiency of some research processes by approximately three times [8] Group 8: Anthropic's AI Masterclass - Anthropic has launched a comprehensive free AI course accessible without an account, covering practical topics such as Claude Code, API development, and MCP fundamentals [9] - The courses include introductory training on Agent Skills, teaching how to build, configure, and share reusable Markdown directive skills, as well as platform integration courses with AWS Bedrock and Google Cloud Vertex AI [9] - Customized AI fluency courses for educators, students, and non-profit organizations are also available, with completion certificates for resumes, and a previously exclusive AWS employee training program is now publicly accessible [9]