对游戏软件的深度研读

一、Unity Software（U）

• 核心定位：全球领先的实时3D内容创作与运营平台，手游引擎市占率约70%，PC端约50%，形成"Create（开发）+Grow（变现）"双轮驱动。

• 商业逻辑：

◦ Create：按订阅（Pro/Enterprise）+使用量收费，覆盖引擎、编辑器、资源商店（Asset Store，超7万插件）。

◦ Grow：广告网络+变现工具，通过Vector AI提升投放效率，与Ironsource整合强化"引擎+广告"协同。

◦ 生态飞轮：开发者越多→资源越丰富→新开发者涌入，形成强锁定效应。

• 产业链：上游（芯片/硬件厂商、AI大模型如GPT-4/PaLM2）；中游（引擎+工具链）；下游（游戏、汽车仿真、建筑可视化、影视虚拟制片）。

• 竞争格局：

◦ 直接：Epic Unreal Engine（高端3A优势）、Godot（开源轻量化）。

◦ 潜在：DeepMind Genie、自研引擎厂商。

• 优劣势：

◦ 优势：跨平台支持约30种、移动端绝对主导、AI套件（Assistant/Generators/Inference Engine）落地、非游戏领域拓展（占比约23%）。

◦ 劣势：2023年安装费争议致信任受损、与Ironsource整合阵痛、广告业务被AppLovin赶超、定价策略反复引发开发者不满。

• 股价与关键事件：

◦ 2020年纽交所上市，首日涨约30%，元宇宙热潮推高至2021年峰值。

◦ 2023年9月安装费政策引发抗议，CEO下台，股价暴跌约40%。

◦ 2026-01-30受Genie冲击跌24.22%，创2022年以来最大单日跌幅，市值约124.5亿美元。

• 前景：短期AI工具与云服务驱动增长，长期看非游戏领域与AI协同；Genie短期难撼动，但长期存在技术替代风险，需持续迭代AI能力巩固生态。

二、Take-Two Interactive（TTWO）

• 核心定位：3A游戏发行龙头，旗下Rockstar Games（GTA/Red Dead）、2K（NBA 2K/文明/无主之地）双品牌驱动，2022年收购Zynga切入手游。

• 商业逻辑：精品IP+长线运营+收购扩张；以高投入打造3A大作，通过DLC、内购、在线模式实现长期变现，收购补强IP与品类覆盖。

• 产业链：上游（自研工作室+第三方开发商）；中游（发行/营销/运营）；下游（主机/PC/移动平台、玩家）。

• 竞争格局：

◦ 直接：EA、动视暴雪、索尼/微软第一方工作室。

◦ 间接：Roblox等UGC平台、AI生成内容工具。

• 优劣势：

◦ 优势：GTA等顶级IP壁垒、开放世界设计能力、手游与主机协同、《GTA6》预期带来强业绩弹性。

◦ 劣势：3A研发成本高、周期长、业绩依赖爆款、收购Zynga后整合与盈利压力、AI降本增效落地慢于工具型厂商。

• 股价与关键事件：

◦ 2000年代上市，GTA系列驱动长期上行；2013年GTA5发布后持续创新高。

◦ 2022年收购Zynga股价波动；2026-01-30跌7.93%，市值约407亿美元。

• 前景：短期看《GTA6》释放业绩，长期需加快AI在研发/运营落地，平衡3A投入与手游增长，应对UGC与AI生成内容的竞争。

三、Roblox（RBLX）

• 核心定位：元宇宙UGC平台，以Robux虚拟货币连接创作者与用户，形成"创作-社交-变现"闭环，目标用户向青少年+年轻成人延伸。

• 商业逻辑：用户付费购Robux→创作者变现（平台分成）→广告与品牌合作增收；AI工具（Code Assist/Material Generator/Cube）降低创作门槛，提升内容供给与用户时长。

• 产业链：上游（创作者/开发者、AI工具商）；中游（平台运营、支付、社区治理）；下游（C端用户、品牌广告主、教育/企业客户）。

• 竞争格局：

◦ 直接：Epic Games（Fortnite创意模式）、Meta Horizon、网易蛋仔派对。

◦ 潜在：DeepMind Genie等"凭空造世"工具降低UGC门槛，分流创作者。

• 优劣势：

◦ 优势：Z世代社交心智、UGC生态正循环、AI原生工具赋能创作、用户时长与DAU高增（2025Q3 DAU约1.515亿，同比+70%）。

◦ 劣势：盈利尚未稳定、内容审核与合规压力、成人用户拓展慢、虚拟经济监管风险。

• 股价与关键事件：

◦ 2021年直接上市，元宇宙概念冲高后回落；2023-2024年随用户与现金流改善反弹。

◦ 2026-01-30跌13.17%，市值约461.6亿美元。

• 前景：短期AI工具提升创作效率与内容多样性，长期向"AI生活空间"进化；Genie短期冲击有限，长期需强化社交与网络效应构筑壁垒。

四、AppLovin（APP）

• 核心定位：移动广告与游戏发行双龙头，以AI驱动的Axon引擎为核心，构建"广告投放+流量聚合+归因分析"闭环，游戏广告市占率约28%，iOS端约43%。

• 商业逻辑：Axon算法精准匹配广告主与流量，MAX聚合平台整合多方资源，Adjust提供归因，形成"数据→算法→效果→更多数据"飞轮；同时自研/发行游戏补充流量与变现场景。

• 产业链：上游（广告主、流量主/开发者、AI模型商）；中游（广告技术平台、发行平台）；下游（移动用户、渠道平台如iOS/Android）。

• 竞争格局：

◦ 直接：Unity Ads、Google AdMob、Meta Audience Network。

◦ 间接：ATT政策、AI广告技术新进入者。

• 优劣势：

◦ 优势：Axon 2.0算法壁垒、游戏广告市占率第一、高利润率（调整后EBITDA约81%）、轻资产高效运营。

◦ 劣势：依赖移动广告周期、隐私政策（如ATT）影响数据获取、与Unity竞争加剧、游戏发行业务波动性。

• 股价与关键事件：

◦ 2021年上市，初期破发后随AI广告业绩爆发大幅上涨。

◦ 2026-01-30跌16.89%，市值约1598.9亿美元。

• 前景：短期AI算法迭代与市场份额提升驱动增长，长期拓展品牌广告与全域覆盖；Genie对其直接冲击小，但若游戏开发模式剧变可能间接影响广告预算结构。

五、行业共性与Genie冲击总结

• 共性逻辑：均受益游戏/数字内容增长，AI是核心变量；工具型（Unity/AppLovin）看效率与生态，内容/平台型（Take-Two/Roblox）看IP与社交壁垒。

• Genie冲击：

◦ 短期：市场恐慌性抛售，股价下挫，属"先开枪再询问"的情绪反应，基本面未实质恶化。

◦ 长期：若Genie实现商业化"造世"，将重构开发流程，压缩工具与内容厂商利润空间；但现有企业已布局AI，生态与合规壁垒仍在，完全替代概率低。

• 发展空间：

◦ Unity：AI工具+云服务+非游戏领域打开增量。

◦ Take-Two：爆款IP兑现+AI降本+手游多元化。

◦ Roblox：AI UGC+社交深化+成人用户渗透。

◦ AppLovin：AI广告技术领先+全域覆盖+高盈利延续。

需要我把这四家公司的关键指标（市占率、营收结构、核心风险、催化剂）整理成一页对比清单，方便快速查阅吗？


Sent from my iPhone

全方位分析人工智能的skill 应用, 从pdf2skills开始说起

pdf2skills是文因互联（Memect）开发的文档-技能编译器，能将PDF文档自动转化为标准格式的AI技能包（skills.zip），实现"死文档"到"活技能"的数字化转生，让AI可直接调用文档中的业务逻辑而非仅做摘要。以下从五个维度展开说明。

一、工具定位与核心价值

• 本质：面向AI Agent生态的知识工程自动化工具，解决"领域专家的悲剧"——业务专家不懂技能封装、开发者不懂业务逻辑的"死锁"

• 核心流程：上传PDF→自动解析→生成标准skills.zip→导入AI Agent（如Claude Code）→AI获得专业能力

• 典型应用：将《手把手教你读财报》转为财报分析技能包，可自动提取三大报表、计算28项指标、分析风险；将《哈佛经典谈判术》转为谈判模拟器

• 与传统工具区别：非简单摘要（RAG），而是结构化业务逻辑提取，输出包含技能名称、适用场景、核心逻辑、前置条件、输出结果的完整技能包

二、底层逻辑与技术实现

1. PDF解析层：处理格式多样性（文本/PDF、扫描件OCR），恢复表格与排版结构，提取原始内容

2. 语义拆解层（核心）：像编译器做词法分析，识别文档中的知识单元（方法论、流程、公式、案例），非关键词提取，而是业务模块识别

3. 逻辑建模层：建立知识点依赖关系（A是B的前置条件、C是D的异常处理分支），构建执行逻辑图，解决"知识碎片化"问题

4. 技能封装层：按标准格式（SKILL.md+资源文件）输出技能包，适配主流AI Agent技能规范，确保可被AI动态加载执行

5. 技术栈：融合大模型语义理解（识别业务逻辑）、知识图谱构建（建立关系）、文档结构化解析（处理PDF复杂格式）、标准化封装（适配AI Agent生态）

三、发展空间与市场前景

1. 市场痛点：企业80%知识存于非结构化文档，转化为AI可用能力的人力成本极高，pdf2skills可将效率提升10-100倍

2. 应用场景扩展

◦ 企业培训：将员工手册、SOP转为AI培训技能包，新人快速上手

◦ 金融分析：研报→投资分析技能，自动生成风险评估报告

◦ 医疗领域：医学指南→辅助诊断技能，提升基层医生能力

◦ 法律行业：法规/案例→合同审查技能，降低合规风险

3. 生态协同：与skills2app等工具联动，实现"书籍→技能→APP"的0代码开发，半小时完成应用构建

4. 商业模式：从免费内测到企业订阅制（按文档量/技能数收费）、行业定制化方案（金融/医疗/法律专属模板）、API服务（嵌入企业知识库系统）

5. 行业趋势：AI Agent从"通用助手"转向"专业专家"，技能市场将成万亿级赛道，pdf2skills作为"技能工厂"占据关键入口

四、竞争对手与合作伙伴

主要竞争对手

 

潜在合作伙伴

1. AI Agent平台：Anthropic（Claude Code）、OpenAI（GPT-4o）、阿里（千问Agent），提供技能生态入口

2. 文档管理系统：微软SharePoint、谷歌Workspace、国产WPS，拓展企业用户场景

3. 行业内容平台：金融（Wind）、医疗（丁香园）、法律（北大法宝），提供垂直领域文档源

4. 低代码平台：Mendix、OutSystems，打通"技能→应用"最后一公里

5. 硬件厂商：智能办公设备（如会议平板），内置技能生成能力，提升办公效率

五、上下游产业链

1. 上游

◦ PDF技术供应商：OCR引擎（合合信息、PaddleOCR）、PDF解析库（PyMuPDF、PDFMiner），保障文档处理准确性

◦ 大模型服务：GPT-4o、Claude 3、文心一言等，提供语义理解能力

◦ 知识图谱工具：Neo4j、JanusGraph，辅助逻辑关系构建

2. 中游（核心层）

◦ pdf2skills核心引擎：语义拆解、逻辑建模、技能封装三大模块

◦ 技能标准适配层：兼容主流AI Agent技能规范，确保跨平台使用

3. 下游

◦ 企业用户：金融机构、医疗机构、律所、大型企业，提升知识复用与AI应用效率

◦ 开发者生态：全栈开发者、AI应用创业者，降低专业应用开发门槛

◦ 垂直行业解决方案商：基于技能包开发行业专属AI应用，快速响应客户需求

4. 延伸生态

◦ 技能交易市场：技能包上传/下载/交易，形成知识变现新渠道

◦ AI技能评测机构：对生成的技能包质量评级，保障生态健康发展

总结

pdf2skills是AI Agent时代的知识转化基础设施，通过自动化知识工程打破业务与技术壁垒，让专业知识快速转化为AI可用能力。随着技能生态成熟，其有望从工具升级为"知识技能化操作系统"，重构知识生产与应用的全流程，为企业数字化转型提供核心动力。

Moltbot（原Clawdbot）详解：AI界的龙虾助手

Moltbot是一款由奥地利工程师Peter Steinberger开发的开源、自托管的个人AI代理，主打"真正做事的AI"（AI that actually does things），而非仅提供对话功能。它在GitHub上迅速走红，几天内斩获数万星标，甚至带动Mac Mini销量飙升，成为2026年初AI领域的现象级项目。

 一、为什么改名？

核心原因：Anthropic的商标投诉

- 原名"Clawdbot"（昵称Clawd）与Anthropic的AI模型"Claude"发音和拼写相似，被Anthropic认定存在商标冲突风险

- 开发者在X（原Twitter）官方声明："Anthropic要求我们更改名称"，并表示这"不是我的决定"

- 新名称"Moltbot"（昵称Molty）延续了龙虾主题，"molt"意为"蜕皮"，象征龙虾生长时蜕壳的自然过程，寓意项目的成长与转变

- 改名仅涉及品牌标识，核心代码、功能和龙虾吉祥物完全保留

二、它到底是什么？

Moltbot是运行在本地服务器或设备上的24/7持久化AI助手，通过消息应用（如WhatsApp、iMessage、Telegram）与用户交互，而非专用APP或网页界面。它融合三大技术领域：

1. AI代理：具备任务规划、工具调用和长期记忆能力

2. 本地自动化：可执行系统命令、管理文件、运行脚本

3. 消息网关：统一管理多个通讯平台，主动推送通知

关键特性：

- 模型无关：支持Claude 3.5 Sonnet（默认推荐）、GPT-4o、本地模型等多种AI大脑，兼顾性能与隐私

- 完全本地控制：数据不离开用户设备，自托管确保隐私安全

- 社区驱动：拥有超百种社区贡献的工具，持续扩展能力边界

- 主动式交互：不像ChatGPT那样被动等待查询，可主动提醒日程、跟踪任务进展

三、核心功能

Moltbot的能力覆盖个人与工作场景，主要包括：

四、应用场景详解

1. 个人生活助手

- 智能管家：自动处理垃圾邮件、整理文件、备份数据，24/7监控系统状态

- 出行规划：查询航班/酒店、自动值机、提醒登机、安排接送机

- 健康管理：记录饮食、提醒服药、分析睡眠数据、生成健身计划

2. 专业工作助手

- 开发者工具：自动测试代码、部署应用、监控服务器、生成API文档

- 内容创作者：批量编辑视频字幕、生成社交媒体素材、跨平台发布内容

- 远程工作者：管理多项目任务、自动参加会议、生成会议纪要、跟踪项目进度

- 自由职业者：自动生成发票、跟踪付款、管理客户关系、安排工作时间

3. 企业团队应用

- 团队协作：自动同步项目文件、提醒任务截止日期、整理团队知识库

- 客户服务：自动回复常见咨询、生成支持工单、跟踪问题解决进度

- 数据分析：自动抓取行业数据、生成可视化报告、监控关键指标变化

4. 特殊场景应用

- 老年人辅助：简化数字设备操作、设置紧急联系人、提醒医疗预约

- 残障人士支持：语音控制电脑、阅读屏幕内容、自动填写表单

- 家庭自动化：连接智能家居设备、根据习惯调整环境设置、监控家庭安全

五、底层逻辑

Moltbot的技术架构分为四层：

1. 交互层（消息网关）

- 多平台适配器：统一处理不同消息应用的通讯协议

- 安全配对：通过DM验证确保只有授权用户可访问

- 消息解析：将自然语言指令转换为内部任务格式

 2. 核心引擎（AI代理）

- 任务规划器：将复杂请求分解为可执行步骤

- 工具调用器：匹配并调用合适的系统工具或社区插件

- 长期记忆模块：向量数据库存储用户偏好、历史交互和上下文信息

- 状态管理器：跟踪任务执行进度，处理异常情况

3. 执行层（本地操作）

- 沙箱系统：安全执行终端命令，限制权限防止误操作

- 文件系统接口：管理本地文件、目录操作、数据读写

- API集成器：连接外部服务（如Google Calendar、Twitter、GitHub）

- 脚本执行器：运行Python、Bash等自定义脚本

4. 模型层（AI大脑）

- 抽象接口：兼容不同大语言模型，简化切换过程

- 提示工程：优化提示词模板，提升模型推理效率

- 成本控制：监控API使用量，防止意外超支

工作流程示例：

1. 用户通过Telegram发送指令："帮我准备明天的会议材料"

2. Moltbot解析请求，规划任务：收集相关文档→整理内容→生成PPT→发送到邮箱

3. 调用文件管理工具查找最近项目文件，调用PPT生成工具创建演示文稿

4. 通过邮件API发送完成的材料，并主动通知用户任务完成

六、如何落地应用？

1. 准备工作

硬件要求：

- 推荐：Mac Mini M2/M3（性能与能耗平衡，社区最受欢迎）

- 替代：旧电脑、树莓派4（性能有限）、VPS服务器

- 最低配置：2GB内存、双核CPU、20GB存储空间

软件依赖：

- Docker（容器化部署，简化安装）

- Node.js（运行核心代码）

- 消息应用账号（如Telegram、WhatsApp）

- AI模型API密钥（Claude 3.5 Sonnet或GPT-4o）

2. 安装步骤（Docker方式，推荐）

# 1. 克隆仓库

git clone https://github.com/peterfriese/moltbot.git

cd moltbot

# 2. 配置环境变量

cp .env.example .env

# 编辑.env文件，填入API密钥、消息平台配置等

# 3. 启动容器

docker-compose up -d

# 4. 配对消息应用

按照终端提示，通过消息应用扫描二维码或发送配对码

3. 安全配置（关键步骤）

- 启用沙箱模式：限制系统命令执行权限

- 设置访问白名单：只允许信任的联系人使用

- 定期更新：关注GitHub仓库，及时获取安全补丁

- 监控日志：检查异常操作，防止未授权访问

4. 自定义扩展

- 安装社区工具：通过命令行添加新功能

- 编写自定义脚本：扩展Moltbot能力，适配个人需求

- 调整提示词：优化AI模型的响应，符合个人使用习惯

七、风险与注意事项

1. 安全风险：完全系统访问权限可能导致误操作或安全漏洞，建议启用沙箱并限制权限

2. 成本考量：使用API模型会产生费用，建议设置使用上限

3. 技术门槛：需要基础的命令行操作和Docker知识，不适合纯小白用户

4. 诈骗防范：警惕冒充Moltbot的加密货币骗局，官方项目不涉及任何代币销售

总结

Moltbot代表了AI助手的新方向：从云端对话工具转向本地执行代理，强调用户控制、隐私保护和主动服务。它不是完美的万能助手，但作为开源项目，其灵活性和可扩展性使其成为个人和开发者探索AI自动化的理想平台。随着社区壮大和功能完善，Moltbot有望成为连接用户数字生活的重要枢纽.

who is Zoominfo?

With over US$900 million in financing (approximately RMB6.3 billion), the first day of the IPO rose as high as 100%...

 It has been a long time since the US capital market has seen such an exciting sight.

 It was a SaaS company named ZoomInfo that created all of this. Its core business is to use machine learning and other technologies to organize and verify data to help sales staff find suitable targets to achieve marketing intelligence.

 Its listing prospectus disclosed that it has served more than 15,000 corporate customers, and Zoom, another conference software supplier, is also one of its customers.

 In 2019, ZoomInfo's revenue reached 293 million US dollars, equivalent to about 2.1 billion yuan.

 The value given by the capital market is $13 billion, equivalent to RMB 92.5 billion, which is more than 44 times its 2019 revenue.

 So, what is the origin of this company?  How does it work?  Why can it be favored by the market?  And, is it worth learning from other players?

 Today, let's dig up ZoomInfo, a company that has been criticized by many media as "boosting US stock IPO" and "making US stock IPO rejuvenate".

 Being sold twice in less than 3 years, ZoomInfo's market value doubled 54 times

 ZoomInfo was founded in 2000, just when the SaaS boom just started in Silicon Valley.

 Salesforce, a SaaS company with a valuation of over 100 billion US dollars, is only a year old, and its founder Marc Benioff also held a "No Software" protest in Silicon Valley.

 Unlike most SaaS companies that want to replace old software, ZoomInfo is positioned as a company that provides sales and market intelligence. The initial business model was to sell access to the information database to business people in need, such as HR, headhunting, and sales.  and many more.

 Although ZoomInfo was established very early, compared with many SaaS software peers, it didn't usher in a real fast development lane until 2017.

 In the next 3 years, it has undergone two changes of reborn.

 In August 2017, ZoomInfo, which was founded 17 years ago, also ushered in a moment of self-sale-was acquired by private equity firm Great Hill Partners for $240 million in cash.

 Afterwards, the development speed of ZoomInfo has obviously accelerated a lot.

 In September 2018, it acquired Datanyze, a provider of technical graphics data and platform, and Y Labs in Israel to improve the construction of data centers.

 In February 2019, Great Hill Partners sold ZoomInfo to another B2B company, DiscoverOrg (founded in 2007), for more than US$500 million.

 Immediately afterwards, DiscoverOrg was renamed to ZoomInfo, which is ZoomInfo, a company with a market value of up to $13 billion.

 From the first sale in 2017, the value doubled 54 times.

 So, what kind of ZoomInfo is the current ZoomInfo?  Why is it worth so much money?

 From crawling data to selling intelligence, ZoomInfo's way to make money

 For ZoomInfo, its core asset is the information database.  In the early days, it mainly used a special crawler software to obtain data from the network to complete the database.

 According to the prospectus, it has information on 14 million companies and 120 million people.

 There are two main sources of information in the database, one is through crawlers, they monitor 45 million Internet domain names.  On the other hand, they will also record the information and data of customers using the platform, and improve the database through feedback.

 The database constructed from this constitutes one of ZoomInfo's cash cows, which can provide company and contact information to the outside world, and on this basis, provide integration, identity resolution, email verification, and alarm functions.

 In addition, they also conducted in-depth analysis of the data and provided market intelligence clues to external sales, such as the procurement needs and financing needs of some companies, to collect corresponding fees.

 Therefore, they also recruited a large number of employees related to data analysis to support the operation of the business.

 The prospectus disclosed that the data analysis team of ZoomInfo is composed of 300 research analysts and 40 data scientists, which account for more than 30% of its total employees.

 The business model is the same as most SaaS software. ZoomInfo adopts a free value-added model. A total of 4 versions have been released: from elite version, advanced version, professional version to community version, the functions change from more to less.

 In 2019, 99% of the $293 million in revenue obtained by ZoomInfo was obtained through subscriptions.  And this is also a business with a very high gross profit margin-its gross profit margin in 2019 is 76.6%.

 In ZoomInfo's view, this is an industry full of development prospects.

 ZoomInfo, a case of marketing intelligence

 According to data from Capital IQ, the total expenditures of the world's 2000 largest listed companies on sales and marketing activities alone exceeded US$ 2 trillion in 2018.

 Focusing on the ZoomInfo track, ZoomInfo believes that the target market size is $24 billion.

 Currently, they use the ZoomInfo platform and have locked in more than 740,000 potential customers worldwide.  This means that ZoomInfo now has 15,000 customers with a penetration rate of only 2%.

 This is an excellent development opportunity for ZoomInfo, but it is also true for other players in the market.

 So, how to get a slice of it?

 Although it is difficult to imitate the 20-year-old ZoomInfo, from its development process and current development strategy, it can also summarize the experience of "coming people" that can be used for reference.

 For ZoomInfo, the period in which the valuation has changed the most is also a period in which it is constantly enriching its capabilities based on databases and technologies. Of course, it is also a period in which it is constantly intelligent.

 This is also the focus of ZoomInfo in the prospectus-the data engine driven by machine learning is constantly digging out new insights and intelligence that customers can adopt from the data.

 How to do it intelligently, the story of ZoomInfo is not outstanding:

 It has a large amount of data, continuous investment in technology, combined with professionals, to work together to create its own data engine.

 What is more critical is how to sort out various unstructured data into machine-able "food" and generate value.

 So, what is the core of intelligence?

 A group of good technical experts?  Suggested some good models?  Found a suitable landing scene?  Have you figured out Know-How in the scene?

 These are important, but in the case of ZoomInfo, the more important thing is the data, and the hard work around the data.

 In this process, ZoomInfo itself did 20 years, acquired ZoomInfo, and became the new ZoomInfo DiscoverOrg for 13 years.

 Who is ZoomInfo in China?

 Now that the capital market is so valued and given such a high valuation, the data of ZoomInfo and the intelligent value behind it are clearly recognized.

 Overall, there does not seem to be a company that directly targets ZoomInfo in China.

 But from data to data analysis capabilities, Chinese companies are not lacking, but they are not concentrated in the hands of one company.

 For example, the data terminal, Maimai and various recruitment companies, such as Lagou and Boss direct recruitment, etc., have a lot of information about the company and employees.  However, the data they possess has not been more directly transformed into intelligence and services to the B side like ZoomInfo.

 And in the future planning, ZoomInfo also put recruitment on the agenda.

 The technical side is not difficult. Companies that force intelligence to acquire customers, such as Bailian Intelligence, are doing similar things. They want to use text analysis to help sales find leads, but the richness of data and technical capabilities are still certain.  gap.

 In May 2019, UiPath completed a $568 million Series D round of financing with a valuation of $7 billion, which directly set off a wave of RPA+AI in China. Who will be the next?

Joined force of Slake and Amazon

 It was reported On June 5, Thursday local time, the office communication application Slack and Amazon announced a new cooperation aimed at attracting more corporate customers.

 The deal comes at a time when Slack is facing increasing competition from Microsoft Teams (team collaboration tools).  Slack's expanded cooperation with Amazon will allow the two companies to unite against their common rival, Microsoft. ​​

 It is reported that Microsoft's Azure cloud platform is competing with AWS, and Microsoft Teams is also fighting Slack directly in collaboration technology. ​​

 Slack is already a loyal user of AWS. According to the agreement reached before, it promises to invest $50 million annually on Amazon's cloud platform. ​​

 For a long time, Slack has been using Amazon's AWS to provide support for some of its chat applications.  Now, the company is committed to using Amazon's cloud services as its preferred partner in storage, computing, database, security, analytics, machine learning, and future collaboration capabilities.

To build a data service platform for the electronics manufacturing industry, "Han Han Xin City" won a new round of financing of tens of millions of yuan

 The electronic industry service platform "Yunhan Xincheng" has recently completed a new round of tens of millions of RMB financing. The investor is the listed company Torch Electronics (603678.SH). This round of financing will mainly be used for platform operations and big data construction.  Torch Electronics, as a capacitor manufacturer, has synergy with the electronic industry platform of "Yunhan Xincheng". Next, Yunhan will continue to promote cooperation with original manufacturers such as components.

 "Yunhan Xincheng" was established in 2002, and the e-commerce transaction of electronic components began in 2011.  The company has gradually formed an industrial Internet platform that connects chip manufacturers, agents and downstream electronic product manufacturers in the electronics industry chain, and makes profits through big data services such as platform collection and precision marketing.

Google uses AI to train the "headphone cable" to realize most functions of touch screen

Google has never stopped developing wearable devices, such as the smart jacket Commuter Trucker launched in collaboration with Levi 's.

 A sensor is added to the cuff on the clothes, and the user can interact with it through a Bluetooth link.

 You can double click, slide and other operations to cut songs and other operations.

 To make persistent efforts, Google hopes to make the device smaller and more functional.

 Google then stared at the headphone cable.

 Google AI engineers have developed an electronic interactive knitting (E-Textile), which allows people to realize most of the functions of previous touch screens by pinching, rubbing, holding, and shooting gestures.

 Operations such as volume control and changing songs are not to mention. Google 's new features point to the next step of perceptual interaction, and the ultimate goal is to liberate our hands.

 Gesture dataset training process

 The device developed by Google is a combination of machine learning algorithms and sensor hardware, and the headphone cable is just the load.

 In fact, the cable is not an ordinary headphone cable, it is a flexible electronic material, and the sensor is woven into it, so human-computer interaction is possible.

 If you like, hoodies can also be transformed.

 First, Google recruited 12 participants for data collection, made 8 gestures each and repeated 9 times, a total of 864 experimental samples.

 In order to solve the drawback of too small sample size, the researchers used linear interpolation to resample each gesture time series.

 Each sample extracts 16 features, and finally obtains 80 observation results.

 Each user's trained gesture recognition can enable 8 new discrete gestures.

 Not only are there quantitative figures, but also the personal experience of the participants, the researchers hope to provide a human-centered interactive experience.

 Participants also provided qualitative feedback through rankings and comments. Participants also proposed a variety of interaction methods, including sliding, flicking, pressing, pinching, pulling, and squeezing.

 Quantitative analysis results show that the perceived speed of the interactive knitwear is faster than the existing headset button controls, and the speed is comparable to that of the touch screen.

 Qualitative feedback also shows that electronic textile interaction is more popular than headphone wire control.

 Considering different usage scenarios, researchers have developed different devices for different usage scenarios:

 Electronic textile USB-C earphones are used to control media playback on mobile phones; hoodies draw cord to add music control to clothes invisibly.

 Algorithm for precise recognition of gestures

 Google 's ability to make an electronic braid is not a machine learning algorithm, but a gesture capture and interaction on the headset line.

 Due to volume considerations, braids such as earphone cords cannot be equipped with large and numerous sensors, and their sensing and resolution capabilities are very limited.

 The second is the ambiguity and ambiguity of the hand gestures, such as how to distinguish between pinch and grab, and how to distinguish between slap and pull?

 Google engineers use 8 electrodes to form a sensor matrix, and divide the data set into 8 times as training data and 1 time as test data, and get 9 gesture transformations.

 They found that there is an inherent relationship in the sensor matrix, which is very suitable for machine learning classification algorithms, which allows the classification algorithm to be trained with a limited data set. It takes only about 30 seconds to realize a gesture recognition.

 The final accuracy rate is 93.8%. Considering the size of the data set and the training time they use, this accuracy is enough for daily use.

 The next step in headset control

 Google's training of the headset line involves gesture gesture recognition and micro-interaction.

 On touch screen devices, the space below the screen can accommodate many sensors, such as Apple's 3D Touch recognition module.

 But in external devices such as earphone cables, it may not be so easy, because the number and volume of sensors must be limited.

 During the experiment, the engineers found that multiple trainings for multiple gestures were required, and different individual gestures required multiple captures of motion.

 This study shows the possibility of achieving accurate small-scale movements in a compact form factor, and we can look forward to the development of intelligent, interactive braids.

 one day.  The micro-interaction of the wearable interface and the smart fabric can be used arbitrarily, and finally the external device can follow the shadow, interact at any time, and finally liberate our hands.

 Are you looking forward to this day?