《大模型通识与应用：从”会聊天”到”会执行”》核心报告显示，当前大模型正从通用对话工具向垂直领域的任务执行专家加速进化，为各行业带来效率重构的新机遇。

一、大模型能力跃迁的两大核心趋势

1. 功能进化：从”闲聊助手”到”任务执行者”

大模型的能力边界正快速突破对话场景，向精准任务执行延伸。以GPT-4为代表的通用大模型，已实现从日常闲聊到代码生成、复杂数据分析的跨越，MIT 2025年的研究数据显示，其任务完成率较前代模型提升62%。 多模态整合则进一步拓展了应用的便捷性：Lobster（龙虾AI）实现了语音指令到设计成果的直接转化，支持语音描述创意后一键生成PPT、海报设计稿，原本需要数小时的设计修改需求，现在响应时间可缩短至3分钟内。

2. 行业渗透率：垂直场景的效率革命

大模型的价值正在各行业的具体场景中落地：

– 制造业：Manus机械臂搭载自然语言编程系统后，东莞某电子工厂仅通过语音指令即可完成产线参数调整，产线切换效率较传统编程方式提升40%；

– 法律领域：合同审查AI工具可自动识别条款风险、标注合规问题，平均为律师节省87%的文书处理时间，让从业者聚焦更具价值的策略性工作。

二、三款高实用性大模型工具测评

三、场景落地的可复制方法论

1. 四步实施框架：从需求到落地的标准化路径

– 需求拆解：将业务流程拆解为”输入-处理-输出”三要素，比如客服场景可拆解为「用户语音输入→AI情绪与意图分析→标准化话术推荐」，确保AI任务目标清晰；

– 工具选型：根据任务复杂度匹配工具能力——基础文本处理、文案生成可选择ChatGPT等通用工具，工业控制、精密操作等场景则需选用Manus等具备专用API的垂直工具；

– 冷启动训练：上传企业专属数据训练垂直模型，建议至少准备500条标注数据，让AI快速适配企业业务逻辑；

– 人机协同：设置关键节点的人工复核机制，例如在合同审核、产线调整等关键决策场景，保留15%的人工干预权，平衡效率与风险。

2. 避坑指南：落地前需警惕的核心问题

– 警惕”全能型AI”陷阱：无需追求能解决所有问题的工具，单一工具对目标任务的解决率超过80%即符合落地标准，过度追求”全能”反而会增加成本与复杂度；

– 数据安全优先：涉及企业核心数据的场景，优先选择支持本地化部署的方案，如华为盘古大模型企业版，避免核心数据外溢风险。

结语

2026年大模型工具已进入”解决真问题”的务实阶段，对于企业而言，无需等待完美的技术方案，建议采用”小场景切入-快速迭代”的策略：从客服、标书制作等单一痛点切入，通过数据反馈持续优化模型，通常6个月内即可看到明显的效率提升与成本节约效果。

大模型应用实战指南：从认知到落地的全路径解析最先出现在AI Weekly。

为什么我疯狂心动？
本地运行超安心！文件死守你电脑
（再也不用担心机密表格乱飞！打工人安全感拉满）
智能到像读心术
“分析Excel，找销量TOP5+画趋势图+写报告”
它自动拆步骤：清洗→计算→生成图表→输出带结论的PPT！
音频秒变爆款素材
采访录音丢过去自动生成小红书笔记+字幕+公众号长文！
（博主/运营人直接跪谢！！省下3小时剪辑命）
还能自己加“技能”
内置超多工具，定制专属工作流～越用越懂你！

阿里大佬原话戳中泪点：
“让AI走出聊天框，真正帮你干活！”
本打工人含泪点赞：终于不用在Excel里反复横跳了啊啊啊！！

真实使用脑补：
早上咖啡没凉
它已把周报写好+配好图+标红重点
我：？？？这真是我的电脑？？（感动到想给它颁锦旗）

划重点：
目前开放邀测！手慢无！！
（蹲到的家人们评论区吼我！！求组队冲）
指路：阿里Qoder官网（记得搜“QoderWork”！）

互动时间：
你最想让它帮你干啥？
自动回邮件？整理手机相册？写小红书文案？
评论区甩需求！点赞过百我求阿里加更教程！！

#打工人续命神器 #AI办公天花板 #阿里黑科技真香 #效率开挂日记 #职场妈妈必备

关注我，带你挖遍让生活变甜的神仙工具！

✨阿里王炸新品！打工人直接封神的桌面AI助理来了！！💻💫最先出现在AI Weekly。

它不像你手机里那些 AI 聊天 App，而是直接住进你的 Mac / 服务器里，通过 Telegram、Slack、iMessage 这些你每天都在用的聊天软件跟你对话，还能操作你的文件、终端、浏览器……甚至帮你自动填报销单！

今天就来手把手带你上车 + 拆解它为什么这么特别

它到底是什么？

Clawdbot 是由知名开发者 Peter Steinberger（@steipete）发起的开源项目，定位是 Personal AI Assistant（个人 AI 助手）The

但它最牛的地方在于：
 本地优先（Local-first）：所有数据存在你自己的电脑里
 无独立 App：通过你已有的聊天工具交互（比如 Telegram）
 能动手做事：不只是聊天，还能执行真实任务！

简单说：它不是“另一个 AI 聊天窗口”，而是你电脑里的数字员工

手把手部署指南（超详细！）

前期准备

• 一台 Mac / Linux / WSL2 机器
• Node.js ≥ v20（推荐）
• Bun（可选但超快！作者亲推）
• 一个 LLM API Key（Claude / GPT / Gemini 都行）
• 建议从 Telegram 入手！配置最简单～

三步启动

# 1. 克隆项目
git clone https://github.com/clawdbot/clawdbot.git
cd clawdbot

# 2. 安装依赖（用 Bun 更快！）
bun install

# 3. 创建 .env 文件

在 .env 里填入：

编辑 env

ANTHROPIC_API_KEY=你的 Claude Key
TELEGRAM_BOT_TOKEN=从 @BotFather 拿到的 Token
TELEGRAM_ALLOWED_USER_IDS=你的 Telegram ID（防别人白嫖！）

然后运行：

编辑 bash

bun run dev

手机远程控制？

对！你可以在手机 Telegram 里给机器人发消息：

“ping”

如果它回你 “pong” —— 恭喜！连接成功
从此你就能用手机指挥家里的 Mac 干活了！

iMessage / WhatsApp 也能接，但需要额外配置，新手先玩 Telegram 就好～

解锁“手脚”：技能系统太香了！

Clawdbot 本身是个“躯干”，真正厉害的是它的 Skills（技能包）！

比如你可以让它：

“帮我列出桌面上所有包含 ‘Confidential’ 的文件”
“打开 Chrome，登录 Notion，截图发我”
“检查 Git 状态并告诉我有没有未提交的改动”

这些能力都来自 skills/ 目录下的插件，比如：

• filesystem：读写本地文件
• browser：控制浏览器
• fetch：调用 API

而且社区还在不断贡献新技能！它会越用越聪明，越用越能干

避坑 & 安全提醒（必看！）

• 记忆存在哪？ → 默认在 ～/.clawd，删掉就“失忆”
• 网络要稳！ → 需频繁调用 Claude / OpenAI API
• 千万设白名单！ → 不设 ALLOWED_USER_IDS = 全世界都能用你的机器人！API 费用爆炸不说，还可能被黑！

它有键盘、有屏幕、有身份——一旦失控，后果严重！
请务必当成“高危实验品”谨慎使用！

为什么它值得被关注？

1⃣ 最好的 UI，是没有 UI

它不强迫你打开新 App，而是融入你现有的工作流。你在 Telegram 聊天，它就在 Telegram 回你；你在 Slack 开会，它就在 Slack 帮你查资料。

未来 AI 或许不该是个“目的地”，而是一层智能图层

2⃣ 从“聊天”到“做事”

传统 AI 只会回答问题，Clawdbot 却能执行任务——这才是真正的“Agent”。

它不是靠模型多强，而是靠连接真实世界的工具生态The

3⃣ 数据在你手里

所有记忆、配置都以 Markdown 存在本地，你可以随时查看、备份、删除。

私密 + 透明 + 长效 = 真正属于你的 AI 助手！

Conclusion

Clawdbot 没有炫酷的新算法，但它用产品思维回答了一个关键问题：

用户不需要新入口，只需要把 AI 能力，无缝嵌入他们 already 在用的工具里。

微信为什么难被取代？因为它够简单、够高频。
与其造一个“全能 AI App”，不如让 AI 成为你现有生活的增强插件The

而这，或许才是 AI 入口的终极形态。

实用资源（收藏！）

• Official website:https://clawd.bot
• GitHub：https://github.com/clawdbot/clawdbot
• 文档超全：https://docs.clawd.bot
• 技能市场：https://clawdhub.com
• Discord 社区（超活跃）：https://discord.gg/clawd

🦞一夜爆火！这个“住在电脑里的贾维斯”，才是真正的AI入口？最先出现在AI Weekly。

Two Models Co - exist, and the Thinking - type is More Powerful

GPT-5.2 and GPT - 5.1 are both available. The Plus version is a "thinking - type" variant - simply put, it's better at in - depth thinking! When dealing with tables, writing review codes, and processing long documents, it's incredibly fast. It can also handle complex tool calls and image analysis.

Performance Doubles, Crushing Professionals

In 44 professional task tests, GPT - 5.2Thinking was actually 70.9% superior to / on par with industry experts (previously, GPT - 5 was only 38.8%)! Whether it's creating PPTs, scheduling, or producing professional deliverables, it's more reliable than the consultants you hire, taking office automation to a new level.

A Perfect Score in Rigorous Tests, Mastering Programming and Math

• In the programming field: The SWE - Bench Pro test set a new record, far outperforming GPT - 5.1Thinking;
• In math competitions: It got a perfect 100% score in AIME2025 and 92.4 points in the GPQA Diamond logic test;
• In logic and science: There has been a significant improvement in CharXiv reasoning and ARC - AGI - 2, evolving from a basic assistant to a "digital intelligence entity".

Now it can be used on web pages / Windows / mobile devices. Experience the power of expert - level AI for free! Have you guys tried Copilot's new features? Come and share your office efficiency tools in the comments section below 👇

微软 Copilot 免费升级 GPT-5.2！专家级工作流直接起飞，比专业人士还牛？最先出现在AI Weekly。

宝子们，最近硅谷顶尖风投机构a16z发布的消费级AI市场年度重磅报告，真的太炸裂了！现在通用AI助手赛道竞争超激烈，用户往往就选一个主力产品，“赢家通吃”格局正在加速形成。

The report shows that although the usage rate of AI has increased, users' willingness to use it across platforms is extremely low. Take ChatGPT's weekly active users as an example. Less than 10% of them will use other AI services simultaneously. Among mainstream products, only about 9% of users will pay for multiple assistants.

Currently, OpenAI is still remarkable, leading with 800 - 900 million weekly active users. However, its "super - app" strategy faces challenges. Google, with its "experimental field" model, has made Gemini catch up rapidly. The number of desktop users has increased by 155% year - on - year, and the growth rate of paid subscriptions is nearly twice that of ChatGPT. 👏The

Judging from the data, ChatGPT has a leading user volume and high user stickiness. The ratio of daily active users to monthly active users is twice that of Gemini. But Gemini is growing at an astonishing rate, especially in terms of the growth of paid users, leaving ChatGPT far behind.

In terms of product strategies, OpenAI is like building a "walled garden", stuffing various functions into ChatGPT, but this makes the interface more complex. Google, on the other hand, adopts the "experimental field" model, allowing innovative products to develop independently, but its products are a bit scattered.

其他玩家也各有绝招。Anthropic的Claude专注技术用户，编程助手收入可观；Perplexity服务注重效率的非技术群体；马斯克的xAI产品Grok增长超快，功能迭代也厉害，被说是能力演进速度最快的AI产品。

The key to the future competition of AI assistants lies in who better understands users' needs and can transform them into good business models. Guys, who do you favor more? 🤔？

#AI Assistant Competition #Winner - Takes - All #OpenAI #Google #Differentiated Breakthrough

AI助手争霸！谁能成为“赢家通吃”王者？最先出现在AI Weekly。

In the past week, the aftereffects of the React2Shell vulnerability still lingered: servers were hijacked for cryptocurrency mining, cloud vendors imposed emergency bans, and even more consequences ensued. To mitigate the risks, Vercel even spent $750,000 on vulnerability bounties and emergency response costs in just one weekend. A vulnerability in a front - end framework directly penetrated the entire technology stack. React's official team has continuously issued emergency announcements, repeatedly emphasizing "please upgrade immediately." This is already the second large - scale patch update in a short period.

The two vulnerabilities disclosed this time are: the high - risk DoS (Denial - of - Service) CVE - 2025 - 55184, where a single request can cause the server to crash; and the medium - risk source code leak CVE - 2025 - 55183, which may expose the source code of React Server Components.

1 One React Vulnerability Shakes the Global Web

In the past week, a vulnerability known as React2Shell has swept through the entire Internet industry. The fundamental reason for such a significant shock is simple: React is of utmost importance. It is almost the "default foundation" of modern web.

From Meta's own Facebook and Instagram to large - scale platforms such as Netflix, Airbnb, Shopify, Walmart, and Asana, all rely on it. Not to mention the millions - strong developer ecosystem, and many frameworks depend on the vulnerable React package.

The React team numbered it as CVE - 2025 - 55182, which received a perfect severity rating of 10.0 in the Common Vulnerability Scoring System. As the creator and main maintainer of Next.js, Vercel also assigned a separate CVE number, CVE - 2025 - 66478, to this issue.

What makes it terrifying is that attackers can exploit this vulnerability with almost no pre - conditions. Cloud security vendor Wiz observed that 39% of cloud environments contain Next.js or React instances with the CVE - 2025 - 55182 vulnerability. It is estimated that when the leak occurred, over two million servers had security vulnerabilities. Even worse, in their experimental verification, they found that the exploitation of this vulnerability is "almost 100% successful" and can stably achieve complete remote code execution.

The affected component scope includes versions 19.0 to 19.2.0 of core modules such as react - server - dom - webpack, and also affects the default configurations of multiple React frameworks and packagers, such as Next.js, React Router, Vite RSC, etc. For many frameworks (especially Next.js with App Router), RSC is actually enabled by default.

When a Level 10 vulnerability is made public, it's not just about "the vulnerability being reported and fixed." There are real - world destructive impacts.

Many developers publicly shared their experiences of being affected on X. Developer Eduardo was one of them. After his server was blocked, he immediately checked the logs and found that the machine had long been "taken over" - the CPU soared to 361%, suspicious processes were frantically consuming resources, and continuous connections were being made to an IP in the Netherlands. "My server is no longer running my application. It's mining for someone else!"

Even worse, the intrusion didn't occur through SSH brute - force cracking but inside the Next.js container. After exploiting the vulnerability, attackers could execute any code they wanted on the server, then deploy more "professional" malicious programs, and even disguise the processes as web services like nginxs or apaches to reduce the risk of exposure. "It infected my entire server just through one Next.js Docker container!"

Finally, he warned, "If Docker is still running with ROOT privileges and the exploited React version hasn't been updated, you'll be hacked soon." (Because with ROOT privileges, one can install cron, systemd, and persistent scripts, ensuring their presence even after a restart.)

The non - profit security organization ShadowServer Foundation stated that since the disclosure of the vulnerability, the attack traffic from Next.js assets controlled by botnets has suddenly increased tenfold. "Like other institutions, we've also observed large - scale attempts to exploit React's CVE - 2025 - 55182, including activities related to botnets."

Why It Can Be Fixed with (Almost) "One Line of Code"

Security researcher Lachlan Davidson was the first to disclose this issue and released a detailed technical analysis. He described the vulnerability as "a serious lack of security checks intertwined with a highly creative exploitation mechanism."

The research process itself was also extremely challenging. It was revealed that he invested over 100 hours in this. And the independent researcher Maple, who was the first to publicly reproduce the attack code, successfully constructed a minimum viable attack chain within dozens of hours after the patch was made public, demonstrating the risk that the vulnerability could be quickly weaponized.

Simply put, this vulnerability doesn't lie in some "obscure edge feature" but in the core communication mechanism of React Server Components.

To make server components fast enough, React designed the Flight protocol. You can think of it as a set of "front - end - specific data channels" built into React. Instead of sending the complete page data to the browser all at once, the server sends data in batches according to the rendering tree structure. This way, the interface can first render the parts that can be rendered, and the rest will be filled in gradually.

The problem is that this ability is very powerful. The Flight protocol not only needs to transmit strings, numbers, and JSON data but also "incomplete things," such as intermediate states like Promise, and reconstruct the component tree. To achieve this, React needs to deserialize and interpret the content of requests sent from the client on the server - side, restoring them into objects that can be further executed.

This is where the vulnerability lies. Attackers can forge a special HTTP request and send content "that looks like normal Flight data" to any React Server Function endpoint. When React parses this data, it mistakes them for legitimate internal objects and continues processing according to the normal process. As a result, the data constructed by the attacker is treated as part of the code execution path, ultimately directly triggering remote code execution on the server.

The entire process doesn't require logging in, no credentials are needed, and there's no need to bypass traditional security boundaries. Just because React lacks a basic hasOwnProperty check in its internal serialization structure, the crucial runtime boundary is breached.

After Lachlan Davidson responsibly reported the vulnerability to Meta, Meta immediately collaborated with the React team and launched an emergency patch in just four days. In terms of implementation, it's almost "adding one line of code," yet it blocks an attack chain capable of destroying the server.

2.Vercel, Cloudflare, etc. Are Innocently Caught in the Crossfire

As soon as a Level 10 vulnerability is exposed, it's often not a small team that gets hit first but an entire industry chain relying on React, especially front - end hosting and Serverless platforms. Leading platforms represented by Vercel are almost inevitably at the center of the storm because they are not only the key maintainers of Next.js but also the default entry points for a vast number of applications.

During the emergency response phase, various vendors indeed rolled out WAFs (Web Application Firewalls) at the first moment. Companies like Vercel, Cloudflare, AWS, Akamai, and Fastly all deployed rules to intercept known React2Shell exploit payload patterns. This can indeed buy some time, but the problem is that WAFs can only serve as a buffer, not a solution.

The essence of WAFs is a rule - matching pattern. Attackers can completely adjust the form of the payload to bypass it. Many applications don't rely on these service providers at all. Self - hosted, privately - deployed, or publicly - running instances are even beyond the reach of WAFs. More importantly, mitigation measures on the edge side are always just one layer of a multi - layer defense, not your patching strategy. For a 10/10 - level RCE (Remote Code Execution), the only real fix is to upgrade React/Next and redeploy, completely removing the vulnerable code from the running environment.

Precisely because the statement "don't rely on WAF as the main repair method" hit a sore spot, there was quite a bit of contention in the industry. Shubham Shah, the co - founder of Assetnote, posted on LinkedIn accusing the Vercel CEO of asking him to remove the tweet about "not relying on WAF to protect against this vulnerability" in an almost bullying manner. Shubham Shah said:

"The Vercel CEO tried to deny the fact that their WAF could be bypassed. The vulnerability involves the latest Next.js/RSC remote code execution. He asked me to remove the tweet about 'not relying on WAF to protect against this vulnerability' in an almost bullying manner. My advice at that time was that users should directly patch their own systems instead of relying on WAF - because we could already bypass Cloudflare's protection at that time, and now Vercel's WAF can also be bypassed. This advice still holds true today.

WAFs do have their role, but the core solution is always to fix system vulnerabilities. Currently, many users have difficulty identifying the risk points in their own systems, and defenders need clear information to guide the patching work. WAF vendors like Vercel should not pressure researchers to cover up the fact that their WAF can be bypassed.

I just released an update for the react2shell - scanner tool, adding the --vercel - waf - bypass parameter. This feature is based on the attack payload designed by Adam Kues from the Searchlight Cybersecurity Research Team and can effectively bypass Vercel's WAF protection."

Trying to cover up problems is always in vain. As more people discovered Vercel's vulnerability, Vercel's attitude underwent a major change. The Vercel CEO has apologized for his previous attitude of questioning whether the WAF could be bypassed and expressed his respect to the Searchlight Cybersecurity Research Team.

The Vercel team responded to Shubham Shah's team's report within a few minutes and deployed a fix within half an hour. Shubham Shah stated in his latest LinkedIn post:

"The Vercel CEO has apologized for his previous attitude when questioning whether the WAF could be bypassed and expressed his respect to the Searchlight Cybersecurity Research Team. He also invited us to collaborate in a shared Slack workspace.

We have submitted multiple valid WAF - bypass solutions through their special vulnerability bounty program

https://lnkd.in/gMsnZFeu
Some of these vulnerabilities allowed us to completely bypass Vercel's WAF protection layer (these vulnerabilities are very interesting!), and others benefited from our in - depth understanding of Node.js and Next.js.
So far, Adam Kues, Dylan Pindur, and I in the team have independently discovered different bypass methods. Assisting Vercel is crucial to us because many of our clients deeply rely on its infrastructure. The difficulty of bypassing the WAF is gradually increasing. The Vercel team can respond to our reports within a few minutes and deploy a fix within half an hour. Their attention to this matter is reassuring. In the end, it turned out to be a happy ending."

Under the dual pressure of the new vulnerability and the React Level 10 vulnerability, Vercel temporarily launched what could be called the most radical security patching plan in history.

On December 11th, on YouTube, a program called "The Programming Podcast" analyzed how Vercel spent $750,000 in just one weekend to stop this "perfect hacker" attack, as well as the line of code in the Dockerfile that could expose users' environments.

This podcast mentioned that after the incident was exposed, Vercel quickly launched an emergency response process, collaborating with the React team, the HackerOne community, and security researchers. They completed the investigation and repair in just one weekend and paid a total of $750,000 in vulnerability bounties. This handling speed and transparency were praised by the industry as "a highly exemplary public relations and technical response."

The reason why the incident didn't cause more extensive damage lies in the rapid response of the community and the platform. After the vulnerability was made public, Vercel collaborated with HackerOne to open up all relevant vulnerabilities and boundary conditions to the white - hat community. Within three days and nights, they received 17 to 19 repair suggestions and boundary situations, involving different levels of security risks. Eventually, Vercel paid approximately $750,000 in bounties to reward the developers and security researchers who participated in the repair at a crucial moment. Multiple teams of engineers, including those from React, Next.js, etc., were also fully engaged over the weekend to promote the rapid implementation of the patches.

Due to the extremely wide user base of React, besides Vercel being severely affected, Cloudflare was also thrown into chaos for a while.

In an attempt to remedy the impact of the React2Shell vulnerability, Cloudflare hastily introduced a change, which affected approximately 28% of HTTP traffic. A large number of websites relying on Cloudflare returned a 500 Internal Server Error, causing about a quarter of Internet traffic to be inaccessible for a time.

Dane Knecht, the Chief Technology Officer of Cloudflare, later stated that this incident was not caused by a cyber - attack but by an internal change introduced when the company was hastily dealing with the high - risk vulnerability in React Server Components.

In addition to these platforms, the National Cyber Security Center (CSOC) of NHS England also said on Thursday that there are already multiple functional proof - of - concept vulnerability exploitation programs for CVE - 2025 - 55182 and warned that "the possibility of continued successful exploitation of this vulnerability in a real - world environment is very high."

10级漏洞刚补完，React又炸了！现代Web“默认底座”因一行代码缺失引发全球地震，开发者经历最黑暗一周最先出现在AI Weekly。

GPT - 5.2 is positioned as "the most suitable model for daily professional use". After months of research and development, it aims to create more economic value for us. Compared with the previous GPT - 5.1, it has significantly improved in tasks such as creating spreadsheets and building presentations. Just like that immunology researcher who used GPT - 5.2 Pro to generate key questions about the immune system, the depth and persuasiveness were excellent 👍.The

Moreover, OpenAI has made remarkable improvements in "AI agent workflows", aiming to make ChatGPT a more powerful personalized assistant. Many companies like Notion, Shopify, etc. have already obtained the testing permission in advance. GPT - 5.2 pays more attention to practicality and structured output, and the interactive experience is also great.

Now it will be gradually launched on ChatGPT, first available to paid users. GPT - 5.1 will be taken offline in about three months. OpenAI will also deploy it "progressively" to ensure our experience. Guys, with such an amazing new model, start looking forward to it right away 💗!！

#GPT - 5.2 #OpenAI #AI Assistant #Office Efficiency Improvement #New Model Release

GPT-5.2来袭！办公效率原地起飞💥最先出现在AI Weekly。

会上发布了《北京人工智能产业白皮书（2025）》。这份白皮书可太有料了！它指出2025年全球人工智能从单点技术突破向生态协同创新演进。北京的人工智能产业发展成果那叫一个显著The

The industrial scale has seen a rise in both quantity and quality. In the first half of 2025, the core industrial scale reached 215.22 billion yuan, with a year - on - year growth of 25.3%. It is expected to exceed 450 billion yuan for the whole year! There are over 2,500 AI enterprises, and 183 large - scale models have been filed. Moreover, universities, research institutes, new - type research and development institutions, and core enterprises have produced numerous cutting - edge results. The policy system is rich, the industrial ecosystem is full of vitality, investment and financing are active, and the enthusiasm for international cooperation has also increased.

未来，人工智能会推动生产力变革、拓展认知边界、促进技术普惠。北京在人工智能领域的地位那是杠杠的！咱一起关注这个充满潜力的行业吧，说不定能抓住新的机遇呢！

#Beijing Artificial Intelligence #Industrial White Paper #Core Output Value #AI Development #Cutting - edge Technology

2025北京AI白皮书发布，产值将超4500亿！最先出现在AI Weekly。

以前用传统TTS模型，起音时间动不动就1 – 3秒，那种卡顿感真的太影响体验了，这就是我们在语音模型使用中的痛点。而VibeVoice – Realtime – 0.5B完美解决了这个问题，它从输入文本到发声平均仅需300毫秒，几乎零延迟，就像和真人对话一样，你打字对方就开始回应，超丝滑The

Its capabilities don't stop there! It can generate an ultra - long audio of up to 90 minutes at one time, and the whole process is smooth and natural, just like a professional broadcaster reading. Moreover, it natively supports up to 4 characters to have a conversation simultaneously, with smooth emotion transitions. The built - in emotion perception module can automatically recognize emotions without manual annotation, and it's ready to use as soon as you get it 👍.The

I tried it myself. I used it on HuggingFace to read the first chapter of "The Three - Body Problem". There was no voice break, and the effect was excellent. Its English performance is close to the commercial level, and it's also very good in Chinese. Although there is still room for improvement in the handling of some polyphonic characters and neutral tones, the official will release a fine - tuned version. With its lightweight design, it can run at real - time speed on an ordinary laptop and can already be integrated into many tools.
Currently, this model has been completely open - sourced and supports commercial use. There are also many interesting demos in the community. Guys, don't miss it. Hurry up and give it a try 👇!！

300ms开声！微软实时语音模型绝了✨最先出现在AI Weekly。

这次大会在山东青岛市崂山区召开，发布的《2025全球独角兽企业500强报告》含金量超高。2025年全球独角兽企业总估值达39.14万亿元，比去年增长30.71%，这数字太惊人了，甚至超过德国GDP。不过市场环境有点严峻，只有12家独角兽成功上市，但并购活动明显增加。

In terms of enterprise distribution, China and the United States are absolute "kings", accounting for 74.8% of the number of enterprises and 86.8% of the total valuation. The United States has strong innovation capabilities in fields such as artificial intelligence, while China has an excellent performance in advanced manufacturing and automotive technology. The number of enterprises in the advanced manufacturing field in China is more than six times that of the United States, and the total valuation is nearly 2 trillion yuan higher 👏.The

In terms of urban distribution, the head - clustering effect is obvious. The top ten cities have gathered more than half of the unicorn enterprises. This also shows the important position of these cities in the global innovation ecosystem.

I have to say that these unicorn enterprises are really amazing! Let's all feel the power of this innovation together 🤩.The

#The Global Unicorn 500 in 2025 #Chinese and American Enterprises #Unicorn Enterprises #Advanced Manufacturing #Innovation Ecosystem

2025全球独角兽500强，中美闪耀！最先出现在AI Weekly。