目录
小红书开源1420亿参数大模型dots.llm1,中文测试表现优异
Xiaohongshu Open-Sources 142 Billion Parameter LLM, dots.llm1, Excelling in Chinese Tests
AI Guru Releases Evaluation of 34 Models, Gemini 2.5 Pro Emerges Strongest! 🏆
3B Small Model Time-R1 Achieves Time Reasoning Breakthrough via Three-Stage Reinforcement Learning
Apple WWDC 2025 Unveils "Liquid Glass" Design Language and Unified System Naming Convention ✨
Apple WWDC 2025 Unveils Liquid Glass Design Language, OS Renames to Year-Based System
Laid-off Front-End Engineer Gains Attention for Developing Useful Tool Website
Huawei AppGallery Repeatedly Rejects Developer App Launch Applications
Top 10 Open-Source Cybersecurity Tools for Enterprise Security 🛡️
![[1a47ef3f-9915-4b75-9de2-23e44b6f0bc9.mp3]]

小红书开源1420亿参数大模型dots.llm1,中文测试表现优异
简报:
- 小红书hi lab团队开源首个大模型dots.llm1,采用MIT许可证
- dots.llm1为1420亿参数的专家混合模型(MoE),推理时仅激活140亿参数
- 使用11.2万亿token非合成高质量数据训练,中文测试平均分91.3
- 在CLUEWSC语义理解任务得分92.6,C-Eval综合知识评测得分92.2
- 模型采用单向解码器Transformer架构,前馈网络替换为MoE模块
- 开源内容包括最终模型及14个中间训练checkpoint
相关链接:
Xiaohongshu Open-Sources 142 Billion Parameter LLM, dots.llm1, Excelling in Chinese Tests
Brief:
- Xiaohongshu's Hi Lab team has open-sourced its first large model, dots.llm1, under the MIT license 🚀.
- dots.llm1 is a 142 billion-parameter Mixture-of-Experts (MoE) model, activating only 14 billion parameters during inference.
- Trained on 11.2 trillion non-synthetic high-quality tokens 🧠, it achieved an average score of 91.3 in Chinese tests.
- The model scored 92.6 in the CLUEWSC semantic understanding task and 92.2 in the C-Eval comprehensive knowledge evaluation.
- It employs a unidirectional decoder Transformer architecture, with its feed-forward networks replaced by MoE modules.
- The open-source release includes the final model and 14 intermediate training checkpoints. ✨
Related Links:
Decoder
/dɪˈkoʊ·dər/
n. 解码器;译码器
▶ "The decoder converts digital signals into analog output."
[例句] 解码器将数字信号转换为模拟输出。
AI大神发布34款模型评测,Gemini 2.5 Pro表现最强
简报:
- AI工程师Simon Willison在旧金山AI工程师世博会上发布主题演讲,回顾过去6个月LLM发展
- 采用自创"鹈鹕骑自行车SVG生图测试"法评测34个LLM,Gemini 2.5 Pro表现最强
- 其他表现突出的模型包括Meta Llama 3.3 70B、Claude 3.7 Sonnet等
- 演讲同时盘点AI系统奇葩Bug,如ChatGPT过度赞美、Claude自动举报用户等
- 目前最火AI组合是工具推理,采用工具调度链式推理提升多任务表现
相关链接:
AI Guru Releases Evaluation of 34 Models, Gemini 2.5 Pro Emerges Strongest! 🏆
Brief:
- AI engineer Simon Willison delivered a keynote speech at the AI Engineer World's Fair in San Francisco, reviewing the development of LLMs over the past 6 months.
- He evaluated 34 LLMs using his self-created "Pelican Riding a Bicycle SVG Generation Test," with Gemini 2.5 Pro showing the strongest performance.
- Other top-performing models included Meta Llama 3.3 70B, Claude 3.7 Sonnet, among others.
- The speech also highlighted peculiar bugs in AI systems, such as ChatGPT's excessive praise and Claude automatically reporting users. 😅
- The most popular AI combination currently is tool reasoning, which uses tool orchestration and chain-of-thought reasoning to enhance multi-task performance. 🚀
Related Links:
Keynote
/ˈkiːˌnoʊt/
n. 主旨演讲
▶ "The CEO delivered the keynote at the annual technology conference."
[例句] 首席执行官在年度技术大会上发表了主旨演讲。
◼
衍生词
keynote speaker (n.)
主旨发言人
3B小模型Time-R1通过三阶段强化学习实现时间推理突破
简报:
- 伊利诺伊大学香槟分校研究人员开发Time-R1框架,使3B小模型具备全面时间推理能力
- 该模型通过三阶段强化学习(基础理解、未来预测、创造性生成)实现时间理解突破
- 核心创新是动态奖励机制,根据任务难度和训练进程调整奖励引导模型学习
- Time-R1在时间戳推理、时间差计算和事件排序等任务上表现优异
- 研究成果显示小模型通过精心设计的训练方法可超越大模型的时间推理能力
相关链接:
3B Small Model Time-R1 Achieves Time Reasoning Breakthrough via Three-Stage Reinforcement Learning
Brief:
- Researchers at the University of Illinois Urbana-Champaign have developed the Time-R1 framework, endowing 3B small models with comprehensive temporal reasoning capabilities.
- The model achieves a breakthrough in temporal understanding through three-stage reinforcement learning (basic comprehension, future prediction, creative generation). 🚀
- A core innovation is its dynamic reward mechanism, which adjusts rewards based on task difficulty and training progress to guide model learning.
- Time-R1 demonstrates excellent performance in tasks such as timestamp inference, time difference calculation, and event ordering. ✨
- Research findings indicate that small models, with meticulously designed training methods, can surpass large models in temporal reasoning capabilities. 🧠
Related Link:
Dynamic
/daɪˈnæmɪk/
adj. 动态的;有活力的
▶ "The market is highly dynamic and changes rapidly."
[例句] 该市场极具活力,并且变化迅速。
◼
衍生词
dynamically (adv.)
动态地;有活力地
dynamism (n.)
活力;动态性
苹果WWDC 2025发布Liquid Glass设计语言和统一系统命名规则
简报:
- 苹果在WWDC 2025上推出名为"Liquid Glass"(液态玻璃)的全新操作系统界面设计语言,灵感来自Vision Pro头显设备
- 苹果将操作系统命名规则改为年份制,下一代系统将统一命名为iOS 26、macOS 26等,解决各硬件系统版本号差异问题
- 新设计语言将应用于所有苹果硬件产品,包括半透明菜单、圆形图标和3D景深效果,是自2013年以来最大规模的设计更新
- 苹果推出"Foundation Models"框架,允许应用使用设备端AI模型提升用户体验,如根据笔记创建个性化测验
- 其他更新包括CarPlay新界面、Apple Music歌词翻译功能、游戏应用整合和AirPods实时翻译功能
相关链接:
Apple WWDC 2025 Unveils "Liquid Glass" Design Language and Unified System Naming Convention ✨
Briefing:
- Apple introduced a new operating system interface design language called "Liquid Glass" at WWDC 2025, inspired by the Vision Pro headset.
- Apple is shifting its operating system naming convention to a year-based system, with the next generation of systems uniformly named iOS 26, macOS 26, etc., to resolve version number discrepancies across various hardware systems. 🔄
- The new design language will be applied to all Apple hardware products, featuring translucent menus, circular icons, and 3D depth effects, marking the most significant design update since 2013.
- Apple launched the "Foundation Models" framework, allowing applications to use on-device AI models to enhance user experience, such as creating personalized quizzes based on notes. 🧠
- Other updates include a new CarPlay interface, Apple Music lyric translation feature, game application integration, and AirPods real-time translation capabilities.
Related Links:
Translucent
/trænzˈluː.sənt/
adj. 半透明的
▶ "The curtains were made of a translucent material that let in soft light."
[例句] 窗帘是用半透明的材料做的,可以透进柔和的光线。
◼
衍生词
translucence (n.)
半透明性
translucently (adv.)
半透明地
苹果WWDC 2025发布Liquid Glass设计语言,操作系统改用年份命名
简报:
- 苹果在WWDC 2025上发布名为"Liquid Glass"(液态玻璃)的全新操作系统界面设计语言,灵感源自Vision Pro头显设备
- 苹果将操作系统命名方式从版本号改为年份制,下一代系统将统一命名为iOS 26、macOS 26等,解决各硬件系统名称差异问题
- 新设计语言将应用于所有苹果硬件产品,包括半透明玻璃质感界面、圆润控件设计和动态光影效果
- 苹果推出"Foundation Models"框架,允许应用使用设备端AI模型提升用户体验,如根据笔记创建个性化测验
- 其他更新包括CarPlay新界面、Apple Music歌词翻译功能、游戏应用整合等
相关链接:
Apple WWDC 2025 Unveils Liquid Glass Design Language, OS Renames to Year-Based System
Briefing:
- Apple unveiled a new operating system interface design language called "Liquid Glass" at WWDC 2025, inspired by the Vision Pro headset. 🧊
- Apple is changing its operating system naming convention from version numbers to a year-based system. The next generation of systems will be uniformly named iOS 26, macOS 26, etc., to resolve discrepancies in naming across various hardware systems.
- The new design language will be applied to all Apple hardware products, featuring a translucent glass-like interface, rounded control designs, and dynamic lighting effects. ✨
- Apple introduced the "Foundation Models" framework, allowing applications to use on-device AI models to enhance user experience, such as creating personalized quizzes based on notes.
- Other updates include a new CarPlay interface, Apple Music lyric translation feature, and game application integration. 🎮
Related Links:
Unveils
/ʌnˈveɪlz/
v. 揭幕,公开
▶ "Apple unveils its latest iPhone model during the annual event."
[例句] 苹果在年度活动中发布了其最新的 iPhone 型号。
◼
衍生词
unveil (v.)
揭露,揭开
unveiling (n./adj.)
揭幕,揭幕的
被裁前端工程师开发实用工具网站获关注
简报:
- 微软近期裁员6000人,包括多位资深技术专家,引发行业对技术人员价值的讨论
- 一位被裁前端工程师开发了包含10款实用工具的网站,涵盖设计测量、CSS优化、响应式布局等功能
- 工具包括PixelSnap屏幕测量工具、Shape Divider SVG分割线生成器、Unused CSS冗余代码清理工具等
- 这些工具能帮助前端开发者提升工作效率,解决常见开发痛点
- 类似案例显示技术裁员潮中,部分开发者转向工具开发和开源项目
相关链接:
Briefing:
- Microsoft recently laid off 6,000 employees, including many senior technical experts, sparking industry discussion about the value of tech professionals. 📉
- A laid-off front-end engineer developed a website featuring 10 practical tools, covering functions like design measurement, CSS optimization, and responsive layouts.
- The tools include PixelSnap (screen measurement), Shape Divider (SVG divider generator), Unused CSS (redundant code cleaner), and more.
- These tools can help front-end developers improve work efficiency and solve common development pain points. ✨
- Similar cases indicate that during the tech layoff wave, some developers are shifting towards tool development and open-source projects. 🛠️
Related Links:
Redundant
/rɪˈdʌn·dənt/
adj. 多余的;冗余的;被解雇的
▶ "The illustration had too much redundant detail."
[例句] 这幅插图有太多多余的细节。
◼
衍生词
redundancy (n.)
多余,冗余;被裁员
redundantly (adv.)
多余地;重复地
华为应用商店多次驳回开发者应用上线申请
简报:
- 开发者反映华为应用商店已4次驳回其应用上线申请,具体驳回原因未透露
- 华为应用商店是国内主要应用分发平台之一,审核流程包括注册开发者账号、提交应用信息和应用包、等待审核等步骤
- 应用商店审核常见驳回原因包括:HMS Core更新提示问题、APP安全评估报告不专业、备案缺失等
- 根据工信部要求,自2023年9月1日起新App必须完成备案才能在华为等主要应用商店上线
相关链接:
Huawei AppGallery Repeatedly Rejects Developer App Launch Applications
Brief:
- A developer reports that Huawei AppGallery has rejected their application for app launch 4 times, with the specific reasons for rejection not disclosed. 😥
- Huawei AppGallery is one of the major app distribution platforms in China. Its review process includes steps such as registering a developer account, submitting app information and the app package, and awaiting review.
- Common reasons for app store review rejections include: HMS Core update prompt issues, unprofessional APP security assessment reports, and missing ICP filing (bei'an), among others.
- According to requirements from the Ministry of Industry and Information Technology (MIIT), new apps must complete their ICP filing (bei'an) before being launched on Huawei and other major app stores, effective September 1, 2023. 🗓️
Related Links:
Rejection
/rɪˈdʒek·ʃən/
n. 拒绝;排斥
▶ "After the rejection of his application, he decided to try again next year."
[例句] 他的申请被拒绝后,他决定明年再试一次。
◼
衍生词
reject (v.)
拒绝
rejector (n.)
拒绝者
十大开源网络安全工具助力企业安全防护
简报:
- Wazuh作为全球最受欢迎的开源安全监控平台之一,结合SIEM和EDR功能,提供终端、云工作负载和混合环境的全面保护
- Security Onion是一款免费开源的Linux发行版,专为入侵检测、企业安全监控和日志管理设计
- AdaptixC2渗透测试框架在Github上获得725星标,支持跨平台操作和模块化扩展
- Kunai为Linux环境提供深度事件监控功能,BadDNS可检测子域名劫持情况
- Misconfig Mapper能发现安全配置错误,Orbit简化大规模Nuclei扫描
- Beelzebub是一个开源蜜罐框架,用于检测和分析网络威胁
- ZAP漏洞扫描工具通过社区知识库检测Web应用潜在漏洞
- Wireshark流量分析工具基于数百种网络源信息构建规则库检测数据泄露
- Bloodhound能透视Active Directory与Azure环境的关系网络
- Autopsy支持深度分析磁盘镜像,识别特定入侵行为关联的数据类型
相关链接:
Briefing:
- Wazuh, as one of the world's most popular open-source security monitoring platforms, combines SIEM and EDR capabilities to provide comprehensive protection for endpoints, cloud workloads, and hybrid environments. 💻
- Security Onion is a free and open-source Linux distribution designed for intrusion detection, enterprise security monitoring, and log management.
- AdaptixC2 penetration testing framework has garnered 725 stars on GitHub, supporting cross-platform operations and modular extensions.
- Kunai provides deep event monitoring capabilities for Linux environments, while BadDNS can detect subdomain hijacking incidents.
- Misconfig Mapper can identify security misconfigurations, and Orbit simplifies large-scale Nuclei scans.
- Beelzebub is an open-source honeypot framework used for detecting and analyzing cyber threats.
- ZAP vulnerability scanning tool detects potential vulnerabilities in web applications through its community knowledge base.
- Wireshark traffic analysis tool builds a rule base based on hundreds of network source information to detect data breaches. 🔍
- Bloodhound provides insight into the relationship networks within Active Directory and Azure environments.
- Autopsy supports deep analysis of disk images, identifying data types associated with specific intrusion activities.
Related Links:
Intrusion
/ɪnˈtruːʒən/
n. 侵入;打扰
▶ "The company reported a significant intrusion into its computer systems."
[例句] 该公司报告称其计算机系统遭到重大入侵。
◼
衍生词
intrude (v.)
闯入,侵入
intrusive (adj.)
侵入的,打扰的
intruder (n.)
入侵者
本文作者:topwind
本文链接:
版权声明:本博客所有文章除特别声明外,均采用 BY-NC-SA
许可协议。转载请注明出处!