腾讯研究院AI速递 20251230

Group 1 - Nvidia acquired Groq for $20 billion through an atypical "asset acquisition + talent recruitment" model, paying nearly 3 times the premium, with about 90% of employees joining Nvidia [1] - Groq employees are expected to receive an average of $4-6 million based on the employee option pool, with vested shares paid in cash and unvested shares converted to Nvidia stock [1] - This "reverse talent acquisition" model is becoming a new norm in the Silicon Valley AI ecosystem, as seen with previous acquisitions of Inflection AI and Character.AI [1] Group 2 - Step-DeepResearch by Jieyue Xingchen uses a 32B parameter model to achieve deep research capabilities comparable to OpenAI's o3-mini and Gemini 2.0 Flash, with a single call cost of less than 0.5 yuan [2] - It employs a three-stage training pipeline (intermediate training, supervised fine-tuning, reinforcement learning) to build data around four core capabilities: planning decomposition, deep search, reflective validation, and report writing [2] - In the ResearchRubrics benchmark test, it scored 61.42, surpassing OpenAI DeepResearch and being on par with Gemini DeepResearch, at only one-tenth the cost of the latter [2] Group 3 - Tencent's Yuanbao has launched a "task" feature, allowing users to assign scheduled tasks to the AI for proactive reminders and information push [3] - Users can customize task content and execution time, marking a shift from passive response to active service by the AI [3] - This feature enhances the AI assistant's role, making it more like a personal assistant that regularly tracks and pushes information of interest to users [3] Group 4 - JD.com has quietly launched an AI-native application "JD AI Purchase," integrating food delivery ordering, product recommendations, and AI fitting, based on JD's self-developed Yansai model [4] - The primary interaction method is dialogue, where users state their needs to receive recommendations, with the homepage "Inspiration Space" covering six major life scenarios [4] - The AI fitting feature allows users to upload photos to generate fitting effect images, and the product comparison function creates tables comparing products across six dimensions, transforming "searching for products" into "stating needs" [4] Group 5 - Domestic GPU company Muxi has released the MACA 3.3.0.X version, showing that 92.94% of 4,490 CUDA projects on GitHub can run directly, achieving near seamless migration [5] - It has completed deep adaptation for PyTorch 2.8, covering all 2,650 core operators, and is compatible with mainstream frameworks like TensorFlow, PaddlePaddle, DeepSpeed, and vLLM [5] - Based on a fully self-developed instruction set and GPU core IP, it achieves "computing power autonomy + ecological compatibility," with linearity stability in thousand-card cluster training above 95% [5] Group 6 - Insta360's research team, in collaboration with several universities, has introduced DAP, the first panoramic measurement deep foundational model trained on a dataset of 2 million [7] - It constructs a three-stage pseudo-label pipeline, refining high-quality supervision signals from 1.7 million internet panoramic images, using DINOv3-Large backbone and distance-adaptive branches [7] - In multiple zero-shot tests, it has set records in Stanford2D3D and Matterport3D, providing precise depth perception for robot navigation, autonomous driving, and VR/AR applications [7] Group 7 - Kuaikan Manhua's version 2.0 has launched AI interactive comics, allowing users to "soul travel" into the comic world and interact with characters in real-time, altering the story direction with each interaction [8] - Characters come with complete backstories and personalities, anchoring dialogues within the story world, establishing long-term companionship through shared experiences and narrative context [8] - It integrates AI capabilities from Tencent Cloud's DeepSeek API, Volcano Engine's Doubao, Alibaba's Tongyi Qianwen, and others, with a nearly threefold increase in weekly paid user rates during the testing phase [8] Group 8 - Nvidia's Jim Fan reviewed the robotics sector, stating it remains chaotic, with severe hardware reliability issues hindering iteration speed, facing daily challenges like overheating and motor failures [9] - The robotics field's benchmarks are a disaster, lacking unified hardware platforms, task definitions, and scoring standards, with teams claiming SOTA based on ad-hoc benchmarks [9] - The VLM-based VLA route feels incorrect, as VLM is optimized for visual question answering rather than the physical world, suggesting that video world models may be a better pre-training target [9] Group 9 - Andrew Ng highlighted that China has surpassed the US in releasing open-source weight models, with cumulative adoption about to exceed that of US open-source models [10] - Many users are incorrectly utilizing Agentic AI, suggesting that tasks should not be completed in one go but through an iterative workflow: outlining, researching, drafting, and revising [10] - The most important future skill will be accurately communicating needs to computers, with programming knowledge significantly enhancing efficiency, contrary to the advice of "no need to learn programming" [10] Group 10 - The Information's year-end analysis of the AI industry indicates that nearly all leading AI companies are now investing in humanoid robot technology development, shifting from competing on models to competing on ecosystems [11] - Overall, Google is seen as the strongest in comprehensive strength, with Anthropic signing a $20 billion TPU chip order, Meta seeking to adopt Google's TPU, and OpenAI signing a $38 billion server agreement with Amazon [11][12] - The alliances among the nine major AI giants are tighter than ever, as companies reduce reliance on one partner while becoming entangled with another, creating a complex interdependent network [12]