量子位

Search documents
游戏巨头Steam幕后:CEO隐居海上,六旬老人带79员工赚尽全球玩家的钱
量子位· 2025-07-09 01:18
Core Viewpoint - Steam is a dominant player in the $400 billion gaming market, achieving significant user engagement and profitability despite its opaque financial reporting [3][6][9]. Group 1: User Engagement and Financial Performance - As of mid-2025, Steam reached a peak of nearly 41.2 million concurrent users, with around 13.2 million players online in games, marking a historical high [2]. - Valve has maintained an operating profit margin exceeding 40% over the past decade, with employee productivity surpassing that of some of the highest market-cap companies like Amazon and Alphabet [6][9]. - In 2021, Valve had only 336 employees, with just 79 dedicated to Steam, yet it contributed to 70% of total PC game sales, indicating a highly efficient operational model [9][10]. Group 2: Business Model and Organizational Structure - Valve's business model emphasizes user-generated content (UGC) through its Creative Workshop, allowing developers to open-source their projects for community modification, enhancing user engagement [15][19]. - The company operates with a unique organizational structure where employees have significant decision-making power, promoting productivity and creativity [21][27][40]. - Valve's decision to remain private allows it to avoid external pressures from shareholders, enabling a focus on content quality and user needs [31][34][40]. Group 3: Leadership and Future Directions - Gabe Newell, co-founder of Valve, has shifted focus towards personal interests and new ventures, including Starfish Neuroscience and Inkfish, while maintaining a low public profile [44][46][48]. - The company's future direction remains uncertain, with speculation about its next steps in the gaming industry as Newell's involvement appears to be lessening [48].
4B小模型数学推理首超Claude 4,700步RL训练逼近235B性能 | 港大&字节Seed&复旦
量子位· 2025-07-09 01:18
Core Viewpoint - The Polaris model, developed by a collaboration between the University of Hong Kong's NLP team, ByteDance Seed, and Fudan University, demonstrates superior mathematical reasoning capabilities compared to leading commercial models, achieving scores of 79.4 on AIME25 and 81.2 on AIME24 [1][53]. Group 1: Model Performance and Training - Polaris utilizes Scaling Reinforcement Learning (RL) to enhance the mathematical reasoning abilities of the 4B model, surpassing various commercial models such as Seed-1.5-thinking and Claude-4-Opus [1][5]. - The lightweight nature of Polaris-4B allows deployment on consumer-grade graphics cards [2]. - The research team confirmed that Scaling RL can replicate significant performance improvements in cutting-edge open-source models like Qwen3 [5]. Group 2: Training Data and Methodology - The success of Polaris hinges on tailored training data and hyperparameter settings that align with the model being trained [7]. - The team discovered a mirrored difficulty distribution in the training data, indicating that the same dataset presents varying challenges to models of different capabilities [8][10]. - A dynamic updating strategy for training data was implemented, allowing the model to adapt as it improves, ensuring that overly easy samples are removed during training [13]. Group 3: Sampling Diversity and Temperature Control - Diversity in sampling is crucial for enhancing model performance, allowing exploration of broader reasoning paths [14]. - The team identified that common temperature settings (0.6 and 1.0) were too low, limiting the model's exploration capabilities [27]. - A three-zone temperature framework was established: Robust Generation Zone, Controlled Exploration Zone, and Performance Collapse Zone, guiding the selection of optimal sampling temperatures [28]. Group 4: Long Context Training and Performance - The model's pre-training context length was limited to 32K, but during RL training, it was extended to 52K, addressing the challenge of long-context training [37]. - The introduction of length extrapolation techniques improved the accuracy of long text generation from 26% to over 50% [41]. - A multi-stage training approach was adopted, gradually increasing context window lengths to enhance reasoning capabilities [48]. Group 5: Evaluation and Results - Polaris achieved the highest performance in most evaluations, demonstrating its effectiveness in mathematical reasoning tasks [53].
百亿机器人独角兽冲刺IPO,细分赛道收入第一
量子位· 2025-07-08 09:11
Core Viewpoint - Megatech is a leading autonomous intelligent agent supplier in China, recently filing for an IPO with a valuation exceeding 10.5 billion RMB, backed by major investors like Bosch and WuXi AppTec [2][61]. Company Overview - Founded in 2016, Megatech specializes in autonomous intelligent agents in the robotics field [3]. - The company's mission is to create robots that can "liberate scientists" from tedious foundational tasks in laboratories [4]. Autonomous Intelligent Agents - Autonomous intelligent agents are systems that can perceive their environment, make decisions, and take actions without direct human intervention [5]. Key Application Areas - Megatech focuses on two main sectors: life sciences and smart manufacturing [6][12]. - In life sciences, Megatech has developed intelligent laboratory agents to automate basic tasks, allowing scientists to focus on experimental design and data analysis [8]. - In smart manufacturing, the company offers solutions for high-precision detection and cutting technologies [13]. Product Development and Performance - Megatech's products have shown significant efficiency, exemplified by a rapid development cycle for a COVID-19 sample processing system, which took only 25 days compared to the industry average of 4 months [10]. - The company has served over 880 clients, with a customer retention rate of 74% and a revenue retention rate of 115% in 2022 [20]. Financial Performance - Megatech's revenue grew from 455 million RMB in 2022 to 930 million RMB in 2024, with a compound annual growth rate of 43% [22]. - The revenue from smart manufacturing accounted for 68.3% of total revenue in 2024, indicating a strong market position [24]. - The gross profit margin has shown stability, with a slight increase in the smart manufacturing segment's margin from 23.1% in 2022 to 32.3% in 2024 [30]. Losses and Cash Flow - Despite net losses of 759 million RMB in 2022, the net loss rate improved from -167% to -83.9% by 2024 [32][33]. - The company's cash and cash equivalents decreased significantly from 1.125 billion RMB in 2022 to 458 million RMB in 2024, raising concerns about liquidity [39]. Funding and Growth - Megatech has raised over 2.7 billion RMB through 8 funding rounds, attracting investments from major players like Goldman Sachs and Innovation Works [61]. - The company plans to use IPO proceeds for operational expenses, R&D, capacity expansion, and strategic partnerships [41]. Market Position - Megatech ranks first among autonomous intelligent agent suppliers in China by revenue and is among the top ten globally [65]. - The company competes with other Chinese firms like BGI and Jingtai Technology, both of which are already publicly listed [69].
腾讯3D生成模型上新!线稿可变“艺术级”3D模型,鹅厂内部设计师也在用
量子位· 2025-07-08 09:11
Core Viewpoint - Tencent's Hunyuan3D-PolyGen introduces an advanced 3D generation model that significantly enhances the efficiency of 3D modeling, achieving over 70% improvement in productivity for artists in game development [2][19]. Group 1: Model Features and Performance - Hunyuan3D-PolyGen supports the generation of complex geometric models with thousands of polygons, transforming 3D models into assets [1][2]. - The model's topology function is now available on the Hunyuan3D platform, allowing for 20 free uses per day [3]. - The model distinguishes itself from standard 3D modeling by focusing on aspects such as polygon count, wireframe quality, and component structure, which are crucial for game rendering [4][19]. Group 2: Technical Implementation - Hunyuan3D-PolyGen utilizes a self-regressive mesh generation framework that processes vertices and faces for spatial reasoning [24]. - The model converts mesh structures into token sequences, which are then processed by a self-regressive model before being reconstructed into mesh format [27][30]. - The technology employs a high compression rate for mesh representation, reducing the number of tokens needed to represent a face from 9 to an average of 2.3, allowing for more complex models with over 20,000 polygons [36][37]. Group 3: Stability and Quality Improvements - The model incorporates a reinforcement learning framework to enhance generation stability, ensuring consistent quality across multiple outputs [40][43]. - The training framework uses artistic criteria such as wireframe neatness and geometric consistency as reward metrics to guide the model towards better results [41][43].
基于能量的Transformer横空出世!全面超越主流模型35%
量子位· 2025-07-08 07:30
Core Viewpoint - The article discusses the introduction of the Energy-Based Transformers (EBT) architecture by a team from the University of Virginia, which surpasses the Transformer++ model across multiple dimensions, including data, parameters, computation, and model depth, through a novel energy mechanism [1][3][28]. Summary by Sections EBT Architecture and Performance - EBT achieves approximately 35% improvement over Transformer++ in various dimensions such as data volume, batch size, parameter count, computation, and model depth [3]. - During inference, EBT shows a 29% enhancement in performance compared to Transformer++ [7]. - EBT is designed to simulate human-like thinking by minimizing energy through a gradient descent process, allowing the model to determine the number of "thinking steps" dynamically [13][14]. Energy-Based Models (EBM) - EBT is developed based on the principles of Energy-Based Models (EBM), which assign a scalar value to each input configuration through an energy function [15][16]. - Lower energy indicates higher compatibility or probability among input variables, while higher energy suggests lower compatibility [17][18]. - The challenge of large-scale training in EBM remains unresolved, with two primary training methods identified: contrastive learning and regularization methods [19][20]. Training and Scalability - The research team transformed EBM learning into an optimization problem, effectively avoiding the curse of dimensionality and enabling scalable learning [22]. - EBT includes two variants: bidirectional EBT, which is simpler to implement, and autoregressive EBT, which is more complex due to information leakage issues [26]. Comparative Analysis - EBT consistently outperforms Transformer++ across six different dimensions, becoming the first model to achieve multi-dimensional superiority without changing the tokenizer [27][28]. - As training time increases, EBT's thinking capability improves, with performance gains rising from 4%-8% to 10%-14% [28]. - EBT outperforms diffusion models in image denoising tasks while reducing the required forward computation by 99% [32]. Implications and Future Directions - EBT introduces a new approach to implementing System 2 thinking through an energy-based optimization mechanism, demonstrating strong scalability and generalization capabilities [34].
17岁少女推翻40年前数学猜想,师从北大校友张瑞祥,即将攻读博士学位
量子位· 2025-07-08 07:30
明敏 时令 发自 凹非寺 量子位 | 公众号 QbitAI Mizohata-Takeuchi猜想 ,诞生于上世纪80年代,是连接调和分析、偏微分方程和几何分析的核心桥梁。 它提出只要每条直线方向的权重积累都不太大,傅里叶传播也不会非常集中。一直以来人们都认为这一猜想是正确的,它也被视为通向解决傅 里叶限制猜想的希望之一。 如果它被推翻,几十年来关于傅里叶限制、PDE良性等核心问题的思考,也要重新更改思路。比如Stein猜想也将不成立。 才17岁的汉娜·凯罗,在今年扔下了这枚"重磅炸弹"。 原本她只是完成导师安排的家庭作业,论证 Mizohata-Takeuchi猜想的更简单形式 。没想到,她却找到了Mizohata-Takeuchi猜想的反例。 17岁少女只是做了一份家庭作业,40年前数学猜想便被推翻。 这结果有多不可思议呢?她也是花了很长一段时间,才成功说服导师 张瑞祥 ,自己提出的反例是对的。 张瑞祥,本科毕业于北大数院,博士毕业于普林斯顿。现在是UC伯克利数学系助理教授,是现代调和分析、偏微分方程、组合几何领域的重 要学者。2023年获得有"菲尔兹风向标"之称的SASTRA拉马努金奖。 值得一提的是, ...
多模态模型学会“按需搜索”,少搜30%还更准!字节&NTU新研究优化多模态模型搜索策略
量子位· 2025-07-08 07:30
MMSearch-R1团队 投稿 量子位 | 公众号 QbitAI 多模态模型学会"按需搜索"! 字节&NTU最新研究, 优化 多模态模型搜索策 略 —— 通过搭建网络搜索工具、构建多模态搜索数据集以及涉及简单有效的奖励机制,首次尝试 基于端到端强化学习的多模态模型自主搜索训练 。 经过训练的模型能够自主判断搜索时机、搜索内容并处理搜索结果,在真实互联网环境中执行多轮按需搜索。 实验结果表明,在知识密集型视觉问答任务 (Visual Question Answering, VQA) 中,MMSearch-R1系统展现出显著优势: 其性能不仅超越同规模模型在传统检索增强生成 (RAG) 工作流下的性能,更 在减少约30%搜索次数的前提 下 , 达 到了更大规模规模模 型做传统RAG的性能水平。 下文将详细解析该研究的研究方法以及实验发现。 具体怎么做到的? 近年来,随着视觉-语言训练数据集在规模和质量上的双重提升,多模态大模型 (Large Multimodal Models, LMMs) 在跨模态理解任务中 展现出卓越的性能,其文本与视觉知识的对齐能力显著增强。 然而,现实世界的信息具有高度动态性和复杂性,单 ...
突破全模态AI理解边界:引入上下文强化学习,赋能全模态模型“意图”推理新高度
量子位· 2025-07-08 07:30
Core Viewpoint - The article emphasizes the increasing need for deep understanding and analysis of human intent in the context of multimodal large language models (MLLMs) and highlights the challenges faced in applying reinforcement learning (RL) effectively to complex multimodal data and formats [1][4]. Group 1: Challenges in Multimodal Reasoning - Insufficient global context understanding leads to incorrect answers when models fail to accurately identify or misinterpret multimodal evidence and contextual information [3]. - The shortcut problem arises when models overlook key clues and provide answers without fully considering multimodal information, resulting in suboptimal or partial outcomes [4]. Group 2: Innovations and Advantages - HumanOmniV2 introduces a mandatory context summarization before reasoning, ensuring models do not skip critical multimodal input and providing comprehensive global background support [12]. - A multidimensional reward mechanism is implemented, including context reward, format reward, and accuracy reward, to guide models in accurately understanding multimodal context [13][14]. - The model encourages complex logical reasoning by evaluating whether the reasoning process successfully integrates multimodal information and employs advanced logical analysis techniques [15]. Group 3: Model Design and Training Strategies - The model is based on Qwen2.5-Omni-Thinker, with improvements to the Group Relative Policy Optimization (GRPO) method to enhance training efficiency, fairness, and robustness [19][20]. - Token-level loss is introduced to address the imbalance in long sequence training, ensuring balanced optimization for each token [19]. - The removal of question-level normalization terms promotes consistency in the optimization process across different problem difficulties [19]. - Dynamic KL divergence is utilized to enhance exploration capabilities and training stability throughout the training cycle [20]. Group 4: High-Quality Datasets and Benchmarks - A comprehensive multimodal reasoning training dataset has been created, incorporating image, video, and audio understanding tasks with rich contextual information [23]. - IntentBench, a new multimodal benchmark, evaluates models' abilities to understand human behavior and intent in complex scenarios, featuring 633 videos and 2,689 related questions [23]. Group 5: Experimental Results - HumanOmniV2 achieved breakthrough results across multiple benchmark datasets, attaining 58.47% on Daily-Omni, 47.1% on WorldSense, and 69.33% on the newly introduced IntentBench, outperforming existing open-source multimodal models [24].
AI版三个臭皮匠!ChatGPT/Gemini/DeepSeek合体拿下AGI测试最高分
量子位· 2025-07-08 07:30
不圆 发自 凹非寺 量子位 | 公众号 QbitAI ChatGPT的对话流畅性、Gemini的多模态能力、DeepSeek的长上下文分析…… 能不能让它们强强联合,共同解决问题呢? 那个由Transformer作者之一Llion Jones创立的明星AI公司 Sakana AI ,提出了新方法 AB-MCTS ,核心思想是: 最伟大的成就往往源于不同思想的协作,我们相信这一原则同样适用于人工智能。 AB-MCTS,全称为自适应分支蒙特卡洛树搜索(Adaptive Branching Monte Carlo Tree Search),是一种 使多个人工智能模型同时处理 问题 的算法。模型之间交换并完善建议,协同工作,就像人类团队一样。 在具有挑战性的ARC-AGI-2基准测试中,多LLM AB-MCTS解决的问题比单独工作的任何单个模型(Single-LLM AB-MCTS)都多。 有几种情况下,只有不同模型的组合才能得出正确答案。 Sakana AI已将该算法以 TreeQuest 的名称开源,链接可见文末。 两种搜索策略 AB-MCTS结合了两种不同的搜索策略:它可以完善现有解决方案(深度搜索),也可以尝 ...
苹果庞若鸣也被小扎挖走!Meta AI天团开会直接用中文吧
量子位· 2025-07-08 03:31
Core Insights - Meta has successfully recruited several prominent AI researchers, including Pang Ruoming from Apple, indicating a strategic move to enhance its AI capabilities [2][3][23] - The majority of the known members in Meta's AI team are Chinese, highlighting a trend of attracting talent from this demographic [4][31][33] Group 1: Recruitment and Team Composition - Pang Ruoming, previously leading Apple's foundational model team, has joined Meta, which raises questions about Apple's AI prospects [7][23] - Meta's AI team now includes 14 known members, with 9 being of Chinese descent, showcasing a significant representation [4][31] - Other notable recruits include Yuanzhi Li from OpenAI and Anton Bakhtin from Anthropic, further strengthening Meta's AI research capabilities [3][28] Group 2: Pang Ruoming's Background - Pang Ruoming has an extensive background, having worked at Google for 15 years, where he led significant projects in AI and machine learning [9][10] - His expertise in machine learning and infrastructure was likely a key factor in Apple's initial recruitment of him [11][13] - Under his leadership, Apple's foundational model team has grown to approximately 100 members, focusing on core AI functionalities for Apple devices [13][14][17] Group 3: Implications for Apple - Pang's departure from Apple is seen as a setback for the company's AI ambitions, especially given the critical role he played in their AI strategy [22][23] - The competitive landscape in Silicon Valley is intensifying, with companies like Meta actively poaching top talent from rivals [23][24] - The situation draws parallels to historical instances of talent poaching in the tech industry, raising concerns about Apple's ability to retain key personnel [24][27]