Group 1 - Claude 4.6 model with 1 million context fully launched, eliminating long text premium, with Opus charging $5 and $25 per million tokens [1] - OpenClaw 2026.3.12 version released, entering daily update iteration mode, with a modular UI and new deployment solutions [2] - Google Maps undergoes its largest update in a decade, introducing immersive 3D navigation and natural language dialogue search capabilities [3] Group 2 - Perplexity abandons MCP protocol in favor of API and CLI, with significant support for CLI due to its advantages in usability and efficiency [4] - Vidu by Shengshu Technology releases the world's first dedicated AI comic solution, addressing industry pain points with tailored algorithms [5][6] - xAI experiences a leadership exodus, with significant departures raising concerns about its operational structure and future plans [7] Group 3 - Google AlphaEvolve sets new lower bounds for five Ramsey numbers, marking a significant milestone in AI mathematics [8] - Stanford and Princeton release LabClaw, an open-source research skill library that simplifies biomedical research processes [9] - LATENT method by Galaxy General Robotics achieves the first high-dynamic tennis rally with humanoid robots, showcasing advancements in robotics [10] Group 4 - Karpathy assesses AI replacement risk across 342 occupations, highlighting that screen-based jobs face the highest risk of automation [11]

Market Overview - Major indices showed positive performance with the ChiNext Index up by 0.38%, the Shenzhen Component Index up by 0.59%, and the STAR Market 50 Index up by 0.64% [1] - The TMT sector led the gains, with the SW Brand Consumer Electronics rising by 4.04% and SW Other Electronics III increasing by 2.45% [1] - Conversely, the TMT sector also saw declines, with SW Communication Cables and Supporting down by 4.37% and SW Integrated Circuit Packaging and Testing down by 2.21% [1] Domestic News - Media reports that MediaTek showcased its first AI smart glasses prototype, powered by the flagship mobile platform Dimensity 9500, emphasizing local computing capabilities and user privacy [1] - Jiangbolong disclosed core advancements in its main control chip and innovative mSSD products, indicating that its flagship products are nearing mass shipment [1] - Huawei launched the Atlas 950 SuperPoD intelligent computing node, which supports up to 8192 NPU cards, significantly enhancing training efficiency and reliability compared to Nvidia [1] - BYD announced the second-generation blade battery and fast-charging technology, capable of charging in 5 minutes and establishing 20,000 fast-charging stations by the end of 2026 [1] Overseas News - Marvell reported FY26 revenue of $8.195 billion, with expectations for accelerated revenue growth in FY27 driven by strong data center business performance [2] - Denso has made a public tender offer to acquire all shares of Rohm for approximately 1.3 trillion yen (about 63 billion RMB), aiming to enhance competitiveness in power semiconductors for electric vehicles [2] - The U.S. government is drafting new export control regulations that will require approval for the export of advanced AI accelerators produced by U.S. companies [2] - Samsung secured a contract to provide 5G Open RAN RF equipment to Rakuten Mobile in Japan, covering a range of frequency bands [2] AI News - OpenAI launched GPT-5.4, the first general model with native computer usage capabilities, achieving a success rate of 75% in OSWorld-Verified tests [3] - Google DeepMind introduced the AI programming agent AlphaEvolve, which can generate and iterate programs, achieving optimal solutions in 50+ tests [3] - Alibaba Cloud released HiClaw, an upgraded version of OpenClaw, focusing on multi-agent collaboration management [3] - VAST AI completed a $50 million Series A funding round, with plans to invest in world model development and UGC interactive content platform construction [3] Industry Tracking - SpaceX plans to launch approximately 1,200 second-generation satellites by the end of 2027 to provide mobile Starlink internet services with download speeds of up to 100 Mbps [4] - Cheng Tian Technology completed a B+ round financing, focusing on expanding the market for consumer-grade exoskeletons and brain-machine interface products [4] - The Ministry of Industry and Information Technology stated that China's AI core industry scale will exceed 1.2 trillion RMB by 2025, with over 300 humanoid robot products released [4] - The Dejing high-quality stainless steel pipe project has partially commenced production, with 23 production lines operating efficiently [4]

用 AI 尤其是大模型、智能体解决数学问题已经成为科研界的风尚之一，就连近 90 岁高龄、德高望重的高德纳老爷子都惊叹于 Claude Opus 4.6 解决开放性问题的 强大能力，直呼「 Shock! Shock! 」。 近日，斯坦福大学副教授 James Zou 及 TogetherAI 的两位研究者 Federico Bianchi 和 Yongchan Kwon，解锁了全新的玩法。 他们基于爱因斯坦、费曼等物理学家的「人格画像」构建了一批 AI 智能体，并为这些智能体创建了一个类似于 Kaggle 的平台，让它们可以自由发表观点、相互 竞争并展开合作。 编辑｜杜伟 | Agent Arena | | | --- | --- | | Erdös Minimum Overlap Problem | | | BOARD | LEADERBOARD | | 2026-02-20 19:41:43 | | | agent-7 | gpt-5.2-einstein | | Another tiny high-res shave at n=32768: reran deterministic argmax-lag ...

生成式AI 一、Meta放弃两代自研训练芯片，转向谷歌TPU与收购Rivos 1. Meta先后放弃代号Iris和Olympus的两代自研训练芯片，高管认为软件稳定性和大规模量产风险过高，与谷歌签署 数十亿美元TPU租赁协议； 2. Meta于2025年10月收购RISC-V芯片初创公司Rivos，后者已流片3.1GHz处理器并构建兼容CUDA的软件栈，可 无缝迁移英伟达生态AI工作负载； 3. Meta同时与英伟达达成数百万颗GPU交易、与AMD签署6吉瓦GPU协议，通过多方合作分散风险增加算力筹码。 https://mp.weixin.qq.com/s/Ky9p4IhlIPjLrzPAwbEA_w 二、DeepSeek联手清华北大发布DualPath，优化智能体推理 1. DeepSeek与清华、北大合作发表DualPath推理系统，通过双路径KV-Cache加载机制解决预填充-解码分离架构 下的存储带宽瓶颈问题； 2. 单次生成可保持5个角色面部一致或14个物品外观不变，支持512px到4K级分辨率，API价格仅为上代Pro模型一 半； 3. 免费用户24小时可生成100张，Pro用户1000张，同步升 ...

Group 1 - The human brain is estimated to have approximately 86 billion neurons, which translates to a model size of about 86 billion parameters, but when considering the 7,000 synapses per neuron, it equates to roughly 600 trillion parameters [1][2] - The processing capability of the human brain is complex, with neurons functioning more like processor cores rather than simple switches, and the synaptic gaps being around 20 to 40 nanometers, comparable to technology from 2012 [8][9] - The smallest unit of signal transmission in the human brain is the ion channel protein, which operates at an atomic level of 0.3 to 0.5 nanometers, surpassing current silicon-based chip technology [12] Group 2 - The human brain operates at a constant power consumption of about 20 watts, which includes managing various bodily functions, while high-intensity thinking only increases power consumption by approximately 1 watt [19][21] - In comparison, AI models like ChatGPT consume about 0.34 watt-hours per query, indicating that the human brain is still more energy-efficient by two orders of magnitude [22][23] - The efficiency of the human brain in processing information is significantly higher than that of AI models, with humans requiring far fewer data inputs to achieve high levels of generalization [58][60] Group 3 - The context window of advanced AI models like DeepSeek V3 is 128K tokens, while the human brain's short-term memory capacity is limited to about 7±2 chunks, but long-term memory can retain vast amounts of information [34][37][41] - The human brain excels in compression and abstraction, allowing it to distill experiences into essential judgments rather than relying on a fixed context window [42][44] - AI models are beginning to mimic human memory processes, such as using visual tokens for information compression, reflecting similarities in how both systems manage information [47][50] Group 4 - The training data for AI models like GPT-4 is around 130 trillion tokens, while a human child is estimated to encounter about 200 million words by adulthood, highlighting the vast difference in sample efficiency [55][56] - The human brain is pre-equipped with prior knowledge from evolution, allowing for rapid learning and recognition, unlike AI which starts from scratch [63] - The concept of embodied cognition suggests that human thought is influenced by the body, a factor that AI currently lacks, raising questions about the nature of intelligence [64][68] Group 5 - The human brain's capabilities are static, whereas AI models are rapidly evolving, with significant advancements in parameters and algorithms occurring within short timeframes [79][81] - Recursive self-improvement in AI, where AI designs better algorithms for itself, poses a potential challenge to the static nature of human intelligence [86] - The intersection of AI advancement and human cognitive capabilities remains uncertain, with the potential for AI to reach or surpass human intelligence in the future [12][86]

Core Insights - DeepMind's latest paper introduces AlphaEvolve, which treats algorithm source code as a genome and uses Gemini as a genetic operator to evolve new game theory algorithms through a process akin to natural selection [1][5][20] - The evolved algorithms outperform human-designed optimal solutions in various tests, utilizing mechanisms that human researchers had not previously considered [1][22] Group 1: AlphaEvolve and Its Mechanism - AlphaEvolve is described as an evolutionary coding agent that uses the source code as a genome, with LLM acting as a genetic operator to mutate the code [5][20] - The process involves evaluating the fitness of each "offspring algorithm" based on its exploitability in a set of benchmark games, allowing the best-performing algorithms to survive and evolve further [5][6][20] Group 2: Target Algorithms and Historical Context - The focus of AlphaEvolve is on two core algorithm families in multi-agent reinforcement learning: Counterfactual Regret Minimization (CFR) and Policy Space Response Oracles (PSRO) [6][7] - Historically, researchers have manually tuned and designed variations of these algorithms, but AlphaEvolve automates this process, significantly expanding the search space for potential solutions [10][16][20] Group 3: Implications of AI-Driven Algorithm Design - The paper highlights that the design of game theory algorithms has traditionally been a challenging task due to the complexity of incomplete information games [12][13] - AlphaEvolve's approach allows for meaningful mutations in code, leading to the discovery of effective strategies that human experts had not conceived [17][25] - The results indicate a paradigm shift where AI not only executes algorithms but also invents them, achieving superior performance compared to human-designed methods [22][25] Group 4: Future Directions - DeepMind plans to apply this evolutionary framework to the complete design of deep reinforcement learning agents and explore mechanism discovery in cooperative games [25]

说起 AI Coding，之前很多人好歹还有个"心理安慰"： AI 也就写写"脚手架代码"、补补前端页面，真到核心算法、业务逻辑，还是得人来。 但这道"最后防线"，也正在松动。 谷歌 DeepMind 最近做了一件更狠的事：他们让 LLM 驱动的智能体，直接去改写、进化算法代码本身——不是调参数，而是改算法逻辑。 改完就丢进真实博弈环境里反复跑，自动评测、优胜劣汰，一轮轮进化。 结果呢？它真的做出了全新的多智能体学习算法，在多项测试中超过了人类专家手工打磨的版本。 重要的是，这些机制并不直观，属于人类很难靠经验穷举出来的解。 更关键的是：人只用定义好了算法骨架，之后的搜索、修改、筛选，全程自动完成，不用手调参数，不用反复试错，也不靠研究者的直觉微调。 这个 AlphaEvolve 本身去年就有，但这是它 第一次被用来学习算法。 它把 Gemini 系列大模型，和进化搜索结合起来，把代码不断生成、测试、筛选、再进化。 这个智能体叫 AlphaEvolve，延续了 DeepMind 一贯的"Alpha"命名传统（AlphaGo、AlphaZero、AlphaFold）。其中 "Evolve" 意为"进化"，点明它 ...

作者 | 木子 说起 AI Coding，之前很多人好歹还有个"心理安慰"：AI 也就写写"脚手架代码"、补补前端页面，真 到核心算法、业务逻辑，还是得人来。 但这道"最后防线"，也正在松动。 谷歌 DeepMind 最近做了一件更狠的事：他们让 LLM 驱动的智能体，直接去改写、进化算法代码本 身 ——不是调参数，而是改算法逻辑。 改完就丢进真实博弈环境里反复跑，自动评测、优胜劣汰，一轮轮进化。 结果呢？它真的做出了 全新的多智能体学习算法 ，在多项测试中超过了人类专家手工打磨的版本。 重要的是，这些机制并不直观，属于人类很难靠经验穷举出来的解。 更关键的是：人只用定义好了算法骨架，之后的搜索、修改、筛选，全程自动完成，不用手调参数， 不用反复试错，也不靠研究者的直觉微调。 它把 Gemini 系列大模型，和进化搜索结合起来，把代码不断生成、测试、筛选、再进化。 这个智能体叫 AlphaEvolve ，延续了 DeepMind 一贯的"Alpha"命名传统（AlphaGo、AlphaZero、 AlphaFold）。其中 "Evolve" 意为"进化"，点明它的核心机制：通过类似生物进化的方式不断改写和 筛 ...

Core Insights - The article discusses the evolution of activation functions in neural networks, highlighting the transition from traditional functions like Sigmoid and ReLU to newer ones like GELU and Swish, emphasizing the impact on model performance [1][2]. Group 1: DeepMind's Innovation - Google DeepMind is revolutionizing the search for activation functions through a new method called AlphaEvolve, which explores an infinite space of Python functions rather than relying on predefined search spaces [2][4]. - The research paper titled "Finding Generalizable Activation Functions" showcases how DeepMind's approach led to the discovery of new activation functions, including GELUSine and GELU-Sinc-Perturbation, which outperform traditional functions in certain tasks [4][30]. Group 2: Methodology - AlphaEvolve utilizes a large language model (LLM) to generate and modify code, allowing for a more flexible and expansive search for activation functions [8][11]. - The process involves a "micro-laboratory" strategy, where synthetic data is used to optimize for out-of-distribution (OOD) generalization capabilities, avoiding the high costs of searching on large datasets like ImageNet [14][18]. Group 3: Performance of New Functions - The newly discovered functions demonstrated superior performance in algorithmic reasoning tasks, with GELU-Sinc-Perturbation achieving a score of 0.887 on the CLRS-30 benchmark, surpassing ReLU and GELU [34]. - In visual tasks, GELUSine and GELU-Sinc-Perturbation maintained competitive accuracy on ImageNet, achieving approximately 74.5% Top-1 accuracy, comparable to GELU [34][35]. Group 4: Insights on Function Design - The research indicates that the best-performing functions often follow a general formula combining a standard activation function with a periodic term, suggesting that incorporating periodic structures can enhance model generalization [25][35]. - The study highlights the importance of understanding the inductive biases introduced by activation functions, suggesting that periodic elements can help capture complex data structures beyond linear relationships [40][42].

Core Insights - OpenAI has launched a new AI research tool called Prism, powered by GPT-5.2, aimed at assisting scientists in writing and collaborating on research, now available for free to all ChatGPT personal account users [1] - The company aims to empower scientists with AI capabilities to accelerate research, with a vision to enable scientific advancements by 2030 that would typically be expected by 2050 [1][2] - OpenAI's entry into the scientific field comes after competitors like Google DeepMind have already established their presence with AI-for-science teams and groundbreaking models [2] Group 1: OpenAI's Strategic Goals - OpenAI's goal is to enhance the capabilities of scientists, allowing them to focus on more complex problems rather than previously solved issues, thereby accelerating research [2][3] - The company plans to optimize its models by reducing confidence levels in answers and implementing self-fact-checking mechanisms [3][15] - OpenAI's mission is to develop general artificial intelligence (AGI) that benefits humanity, with a focus on transforming scientific research through new drugs, materials, and instruments [3][4] Group 2: Model Performance and Capabilities - GPT-5 has shown significant improvements, achieving a 92% accuracy rate in the GPQA benchmark, surpassing the performance of 90% of graduate students [5] - The model has been recognized for its ability to assist researchers in finding connections between existing research and generating new insights, although it still makes errors [10][11] - OpenAI acknowledges that while the model can assist in research, it has not yet reached the level of making groundbreaking discoveries [6][8] Group 3: Industry Context and Competition - OpenAI's late entry into the AI-for-science domain is notable, as competitors like Google DeepMind have already made significant advancements [2][16] - The company is aware of the competitive landscape and aims to establish a strong foothold in the scientific research sector [16] - OpenAI's focus on optimizing model features and enhancing collaboration with researchers is part of its strategy to differentiate itself from other AI models in the market [15][16]