Workflow
Gemma系列模型
icon
Search documents
腾讯持续推进AI布局,关注AI模型迭代
Huaan Securities· 2025-08-17 12:55
Investment Rating - Industry investment rating: Overweight [1] Core Insights - The overall financial performance of Tencent in Q2 2025 exceeded expectations, with revenue increasing by 15% to 184.5 billion yuan, surpassing market estimates of 178.94 billion yuan; net profit grew by 17% [3] - Tencent's gaming segment revenue rose by 22%, particularly strong in international markets; R&D investment increased by 17% to 20.25 billion yuan, while capital expenditure surged by 119% to 19.11 billion yuan [3] - The AI sector showed significant growth, with the AI index rising by 11.42% during the week [23] Weekly Market Review - From August 11 to August 15, 2025, the Shanghai Composite Index rose by 1.7%, the ChiNext Index increased by 8.58%, and the CSI 300 Index grew by 2.37%; the Hang Seng Tech Index rose by 1.52%, while the Nasdaq Index saw a modest increase of 0.81% [23] - The media index increased by 1.25%, and the overseas Chinese internet index rose by 3.12% [23] Company Announcements - Tencent Music reported Q2 2025 revenue of 8.44 billion yuan, a year-on-year increase of 17.9%, with adjusted net profit rising by 33% to 2.64 billion yuan [4] - NetEase's Q2 2025 revenue reached 3.827 billion yuan, with a gross profit of 1.3925 billion yuan and an adjusted net profit of 1.946 billion yuan [4] - Tencent announced the launch of the Hunyuan 3D World Model 1.0 Lite version, significantly reducing memory requirements for smoother operation on consumer-grade graphics cards [35] AI Developments - Apple announced the integration of OpenAI's latest AI model, ChatGPT-5, into the upcoming iOS 26 system, expected to be rolled out globally next month [34] - Google released the lightweight version of its Gemma series, Gemma 3 270M, designed for low-power device deployment [34] - Tencent's Hunyuan team introduced the Hunyuan-GameCraft tool for generating interactive game videos, marking a significant advancement in game video generation [36] Semiconductor Sector - TSMC announced plans to gradually exit the 6-inch wafer manufacturing business over the next two years to enhance operational efficiency [7] - Foxconn reported Q2 2025 revenue of 1.79 trillion New Taiwan dollars (approximately 427.89 billion yuan), a 16% year-on-year increase, with AI server revenue surpassing that of iPhones for the first time [38] Smart Driving - Tesla's Robotaxi service in Austin is set to open to the public in September [39] - WeRide announced a multi-million dollar investment from Grab for deploying L4-level Robotaxi in Southeast Asia [39]
精准调控大模型生成与推理!浙大&腾讯新方法尝试为其注入“行为定向剂”
量子位· 2025-06-05 10:28
Core Viewpoint - The article discusses the dilemma in controlling large AI models, emphasizing the need for a balance between intelligence and compliance, proposing the Steering Target Atoms (STA) method as a solution to create AI that is both smart and obedient [1][6]. Method & Experimental Results - The STA method allows for "atomic-level" behavior editing of large models, enhancing robustness and safety in output control [2]. - Traditional methods often couple safety defenses with general intelligence, leading to potential performance trade-offs. The STA method addresses this by intervening at the internal neuron level, identifying and adjusting specific neurons associated with harmful behaviors while preserving those linked to correct responses [4][5]. - The STA method has been tested on models like Gemma and LLaMA, showing superior detoxification performance without significant negative impact on general performance [10]. Experimental Setup - The research involved manipulating target atom directions and amplitudes to regulate model behavior, with extensive testing on various model configurations [9]. Key Experimental Results - The STA method outperformed other techniques in detoxification while maintaining general performance, as shown in the comparative results table [10]. Steering Vectors vs. Prompt Engineering - The article compares Steering Vectors with traditional prompt engineering, highlighting that Steering is more robust against jailbreak attacks and allows for finer control [12][13]. Cognitive Intervention in Large Models - The research also explored cognitive interventions in larger models like DeepSeek-R1, enhancing reasoning capabilities by amplifying weights of neurons associated with "thinking" [16][18]. - The findings indicate that while Steering techniques may lack the convenience of prompts, they offer more robust and precise intervention effects [18]. Open Source Contribution - The research team has made some intervention methods open source to encourage further exploration in the field of safe and controllable large models [19].