目前，这款新模型已在智谱官网上线，并在 GitHub 和 Hugging Face 平台开源，模型权重遵循 MIT License。 GitHub ： https://github.com/zai-org/GLM-5 Hugging Face ： https://huggingface.co/zai-org/GLM-5 OpenRouter ： http://openrouter.ai/z-ai/glm-5 整理 | 华卫 值得一提的是，智谱在官宣帖中特意注明"GLM-5 在 OpenRouter 上的前称是 Pony Alpha"。就在几 天前，全球模型服务平台 OpenRouter 上一款代号为"Pony Alpha"的神秘模型，因卓越性能和一系列 令人惊艳的实测表现走红。当时，该平台合作方 Kilo Code 透露，Pony Alpha 是"某个全球实验室最 受欢迎的开源模型的专项进化版"。 临近春节，智谱 AI 发布了其最新旗舰大模型 GLM-5。自 1 月初在香港进行备受关注的 IPO 之后， 这是该公司推出的首款重磅大模型。 之后，Pony Alpha 被众人猜测可能是 Anthropic 的 C ...

Core Viewpoint - The software industry is at a pivotal moment where AI programming has evolved from enhanced autocomplete to autonomous development agents, leading to a shift in hiring practices and developer roles [2]. Group 1: Junior Developer Issues - The recruitment of junior developers may decline due to AI automating entry-level tasks, but could rebound as software permeates various industries, necessitating different survival strategies [4]. - A study by Harvard found that when companies adopt generative AI, the employment rate of junior developers dropped by approximately 9-10% over six quarters, while senior developers' employment remained stable [4]. - The U.S. Bureau of Labor Statistics predicts that software jobs will still grow by about 15% from 2024 to 2034, indicating a potential demand for human developers to leverage AI opportunities [5]. Group 2: Skills Issues - As AI writes most of the code, core programming skills may degrade, or become more critical as developers need to supervise AI outputs [9]. - Currently, 84% of developers regularly use AI tools, leading to a shift in skill sets from implementing algorithms to effectively querying AI and validating its outputs [9]. - The future may see a divide among developers, with some relying heavily on AI and others advocating for foundational coding skills to handle AI-generated errors [11]. Group 3: Role Issues - Developer roles may shrink to limited auditing tasks or expand to key coordinators managing AI-driven systems, with value creation extending beyond mere coding [15]. - In a pessimistic scenario, developers may become mere auditors of AI outputs, while in a more optimistic view, they could evolve into architects or product strategists overseeing AI integration [16]. Group 4: Expert vs. Generalist Issues - Specialists in narrow fields may face risks of obsolescence due to automation, while T-shaped engineers with broad adaptability and deep expertise in one or two areas are increasingly favored [22]. - Nearly 45% of engineering roles now expect proficiency across multiple domains, highlighting the shift towards versatile skill sets [24]. Group 5: Education Issues - The traditional four-year computer science degree is being challenged by faster learning paths like coding bootcamps and employer training programs, as universities struggle to keep pace with rapid industry changes [30]. - By 2024, nearly 45% of companies plan to eliminate degree requirements for certain positions, reflecting a shift towards skills-based hiring [31].

48 小时内，xAI 两位联创出走，引发 Grok 难产猜测？ 作者 | 冬梅、高允毅 这两天，全球首富马斯克旗下人工智能公司 xAI 连丢两位联创， 一个是 Yuhuai (Tony) Wu（吴宇 怀），另一个是深度学习大神 Jimmy Ba。 他俩都曾是杰弗里·辛顿（Geoffrey Everest Hinton）的 门徒。 在推特上，他们纷纷发表了"离职"声明。 Yuhuai (Tony) Wu： "这家公司——以及我们之间如同家人般的情谊——将永远铭刻在我的记忆中。我会深深怀 念这里的人们、作战室，以及我们并肩作战过的所有战役。 我的人生新篇章即将开启。这是一个充满无限可能的时代：一支配备人工智能的小团队可以 移山填海，重新定义一切皆有可能。" 致埃隆 @elonmusk，感谢你们相信我们的使命，也感谢你们带给我们这段毕生难忘的旅 程。 Jimmy Ba： 在 xAI 的最后一天。 xAI 的使命是推动人类技术在卡尔达舍夫科技树上不断攀升。很荣幸能参与创立 xAI。还要特别 感谢 @elonmusk 感谢你们让我们齐聚一堂，共同踏上这段非凡的旅程。我为 xAI 团队的成就感到无比自豪，并将 永远以朋友 ...

作者 | 褚杏娟 2 月 10 日，阿里巴巴正式发布新一代图像生成及编辑模型 Qwen-Image-2.0。据介绍，Qwen-Image-2.0 集生图和编辑于一体，在 AI Arena 文 生图评测中斩获 1029 分，超过 Seedream4.5、Flux2-Max 等模型，仅次于谷歌 Nano Banana Pro 和 GPT Image1.5。 | | BETA | | | | | Last updated: 2026-02-09 12:00 | | --- | --- | --- | --- | --- | --- | --- | | | | | Text-to-Image Model Elo Leaderboard | | | | | | | | Task: Text to Image Generation Based on Alibaba Al Arena Platform | | | | | RANK ↑ | MODEL | ELO SCORE | 95% CI | VOTES | ORGANIZATION | WIN RATE | | 1 | Gemini-3-Pro-Image-Previ ...

整理 | 华卫 ChatGPT 无广告体验的日子要结束了。 经过数周的预热，刚刚，OpenAI 宣布，将正式开始在其 AI 平台测试广告，ChatGPT 用户可能很快 就会在对话中看到广告。这些广告会以标注"赞助"的链接形式出现在 ChatGPT 回答底部，但 OpenAI 表示，广告不会影响 ChatGPT 给出的回答内容，在视觉上也会区分开来。 目前，广告仅对免费版 ChatGPT 用户以及每月 8 美元的低价订阅服务 Go 套餐用户展示，Plus、 Pro、商业版、企业版和教育版用户不会看到任何广告。也就是说，想要避开广告的用户至少需要每 月支付 20 美元订阅 Plus 套餐。OpenAI 提到，免费版用户要想退出广告，但代价是每日免费对话次 数减少；Go 套餐用户无法选择退出广告。 一位接近 OpenAI 的消息人士表示，OpenAI 预计从长远来看，广告收入占比将低于其总收入的一 半。目前，该公司还通过其聊天机器人集成的购物功能，从用户购买的商品中抽取分成。另据外媒报 道，OpenAI 首席执行官 Sam Altman 告诉员工，ChatGPT"月增长率已恢复到 10% 以上"，将于本 周部署"更 ...

开源 AI 编程工具 OpenCode 正式亮相，其具备原生终端界面（Terminal UI）、多会话支持，并广泛 兼容包括 Claude、OpenAI、Gemini 及各类本地模型在内的 75 种以上模型。除了命令行（CLI）工 具外，OpenCode 还提供桌面应用版本，并支持作为 VS Code、Cursor 等主流 IDE 的插件使用。 OpenCode 允许开发者沿用现有的付费服务订阅，如 ChatGPT Plus/Pro 和 GitHub Copilot。此外， 它还内置了 一系列免费模型，用户可以通过 LM Studio 在本地直接运行。 在功能集成方面，OpenCode 与包括 Rust、Swift、Terraform、TypeScript 和 PyRight 在内的多种语 言服务器协议（LSP）服务器实现了深度整合。通过利用 LSP 服务器输出的反馈信息，大语言模型 能够更高效地与代码库进行交互。 该智能体同时支持远程和本地的 MCP 服务器。不过，开发团队提醒道，使用 MCP 服务器会增加上 下文占用，部分服务器（特别是 GitHub MCP）往往会消耗大量的 Tokens。 OpenCo ...

Core Viewpoint - The article discusses the emergence of the role of "Vibe Coder" at Lovable, an AI-driven website and application building platform, highlighting the shift from traditional coding to a more conversational approach with AI tools [2][10]. Group 1: Company Overview - Lovable is valued at $6.6 billion (approximately 45.7 billion RMB) and has 8 million users with 517 employees as of the end of 2025, indicating a high per-employee valuation nearing 100 million RMB [2]. - The company has achieved significant growth, doubling its Annual Recurring Revenue (ARR) from $100 million to $200 million within four months [4]. Group 2: Role of Vibe Coder - The first official Vibe Coder at Lovable, Lazar Jovanovic, spends 80% of his time on planning and dialogue, with only 20% on execution, emphasizing the importance of clear communication over traditional coding skills [11][28]. - The role of Vibe Coder is a natural evolution in the AI landscape, where the focus has shifted from coding to articulating product requirements effectively [10]. Group 3: Vibe Coding Process - Jovanovic employs a unique approach by running multiple versions of ideas in parallel, using various methods such as voice brainstorming and reference images to clarify requirements before executing [12][42]. - He emphasizes the importance of creating a structured workflow, including generating Product Requirement Documents (PRDs) and maintaining clear guidelines for AI tools to follow [48][51]. Group 4: Skills and Mindset - A non-technical background can be advantageous in this new role, as it allows for a more open-minded approach to problem-solving without preconceived limitations [25]. - The article stresses the need for clarity in communication with AI tools, as the quality of output heavily relies on the specificity of input provided [32][40]. Group 5: Future of Work - The traditional job titles such as programmer, product manager, and designer are becoming less relevant as roles evolve into a combination of skills, focusing on the ability to create value through AI collaboration [68][70]. - The future workforce will likely be defined by a blend of capabilities rather than single labels, with an emphasis on judgment and aesthetic sensibility over technical skills [69][71].

编译｜宇琪 借助 Coding Agent 等工具，如今构建一个 AI 产品的技术门槛和启动成本已急剧降低。一夜之 间，将想法变为可交互的原型变得前所未有的容易。但一个刺眼的矛盾也随之浮现：大多数 AI 产品仍在走向失败。如果技术实现不再是瓶颈，那么问题究竟出在哪里？ Aishwarya Naresh Reganti 和 Kiriti Badam 曾在 OpenAI、Google、Amazon、Databricks 等 公司参与构建并成功推出了 50 多个企业级 AI 产品。最近，他们在播客节目中，与主持人 Lenny 细致分享了当前 AI 产品开发中的常见陷阱与成功路径。基于该播客视频，InfoQ 进行了部分删 改。 核心观点如下： AI 产品构建中的挑战 Lenny：目前 AI 产品构建的情况是怎样的？哪些进展顺利，哪些地方问题依旧明显？ 今天构建的成本已经非常低了，真正昂贵的是设计，是你是否真正想清楚了产品要解决什么痛 点。对问题本身和产品设计的执着，是被低估的，而单纯追求"快点做出来"，是被高估的。 AI 不是答案，而是解决问题的工具。 领导者需要重新回到"亲自上手"的状态，并不是要他们亲自实现系统， ...

作者 | 高允毅 OpenAI Codex 的核心研发者，竟然成了 Claude Code 的忠实用户？ Calvin French-Owen 是 Segment 联合创始人、前 OpenAI 工程师、Codex 项目的早期研发者。他 最近在一档播客中，对当前最火的代码智能体 Codex、Claude Code 和 Cursor 进行了锐评。 结论出人意料，他最常用、也最偏爱的，是 Claude Code，他表示搭配 Opus 模型更"香"。 "上下文管理" ，是 Calvin French-Owen 在整期播客中反复强调的关键词。 他认为，代码的上下文信息密度极高，只要检索方式得当，模型往往比人类更容易理解系统结构。但 与此同时，上下文窗口本身，也成为制约代码智能体发展的最大瓶颈。 提到上下文污染的问题时，主持人表示 LLM 会变笨。Calvin 趁此分享了一个非常实用的经验： 当上 下文 token 占用超过 50%，他会主动清理。 他甚至分享了一种创业者常用的 "金丝雀检测" 方法：在上下文里埋入一些无关但可验证的小信息， 一旦模型开始遗忘，说明上下文已经被污染。 在他看来，Claude Code 真正 ...

Core Insights - The article discusses the phenomenon of "AI fatigue" among engineers, highlighting that increased efficiency in task completion does not equate to reduced workload, but rather leads to greater exhaustion due to expanded task volume and constant context switching [2][6][10]. - It emphasizes that the role of engineers has shifted from creators to evaluators of AI outputs, which can lead to decision fatigue and anxiety due to the unpredictability of AI-generated results [11][13][15]. - The article warns against the "FOMO treadmill," where engineers feel pressured to keep up with rapidly evolving tools and technologies, resulting in wasted time and knowledge decay [18][20][22]. Group 1 - AI can accelerate individual tasks, but this does not reduce the overall workload; instead, it leads to an increase in the number of tasks engineers undertake [10][11]. - The shift in work dynamics means engineers spend more time reviewing and evaluating AI outputs rather than creating, which is more mentally taxing [13][14]. - The unpredictability of AI outputs disrupts the foundational assumption of certainty that engineers rely on, leading to ongoing anxiety and stress [15][16]. Group 2 - The rapid pace of technological advancement creates a "FOMO treadmill," where engineers feel compelled to constantly adopt new tools, leading to inefficiencies and superficial knowledge [18][20]. - Engineers often find themselves in a cycle of switching between tools without achieving significant improvements, resulting in wasted effort and time [21][22]. - The article suggests that focusing on foundational infrastructure rather than chasing every new tool can lead to more sustainable practices [23]. Group 3 - The "prompt spiral" trap occurs when engineers become overly focused on refining AI prompts instead of addressing the core problem, leading to wasted time [25]. - Perfectionism in engineering can exacerbate frustration with AI outputs, which are often not perfect, causing engineers to spend excessive time making minor adjustments [26][27]. - The article highlights the importance of maintaining critical thinking skills, as reliance on AI can lead to a decline in independent problem-solving abilities [28][29]. Group 4 - The article advocates for setting boundaries around AI usage, such as time limits for tasks and accepting that AI outputs do not need to be perfect [34][37]. - It emphasizes the need for engineers to protect their cognitive resources and recognize that sustainable productivity is more valuable than merely increasing output [38][39]. - The conclusion stresses that the most successful engineers in the AI era will be those who know when to stop and prioritize their mental well-being over relentless productivity [40].