跳到正文
9月3日周四
  1. 公众号:卡尔的AI沃茨67

    实测豆包当国产 Codex 用的 7 种玩法

    作者总结一周高强度使用豆包的经验,介绍如何把豆包当成国产 Codex 使用。核心差异在于豆包把本地电脑和云电脑(Windows 虚拟机)打通,切换即可,而 Codex 本地和云端是两套分开的环境;授权一次飞书后可直接用 PRD、表格、会议预约等内置工具。

  2. Rohan Paul61

    @thehypedotnews 用 Three.js 让 Fable 5.1、Fable 5 和 Opus 5 分别从单张参考图程序化重建 bag end、barad-dûr、rivendell 三个《魔戒》场景,全程经 OpenRouter 运行。

    引用thehype.@thehypedotnews

    fable 5.1 vs fable 5 vs opus 5 – three lord of the rings landmarks, built in 3d from one image the setup: one reference image per scene, one html file per build, everything procedural – no meshes, no textures, no image files, nothing past @threejs from a cdn. each model reads the picture, writes its own prompt from it, then builds to that prompt in the same turn. three named camera shots per scene on keys 1/2/3, so it can be screen-recorded. run through @openrouter tasks: 1. bag end – hobbiton from two frames, outside and in. the round green door has to open onto the room you are standing in 2. barad-dûr – the tower and orodruin from one film still. the eye has to move and track the camera, the volcano erupts on a cycle, the clouds never stop 3. rivendell – jerry vanderstelt's painting. sun shafts that shimmer, water that falls without a break, trees that sway on a gust models: @anthropicai fable 5.1, fable 5, opus 5 total cost, three builds #1 fable 5 – $14.97 #2 opus 5 – $18.53 #3 fable 5.1 – $22.38 wall clock, three builds #1 fable 5 – 38m #2 fable 5.1 – 92m #3 opus 5 – 122m output tokens #1 fable 5 – 298,592 #2 fable 5.1 – 439,435 #3 opus 5 – 724,418 lines of code shipped #1 fable 5 – 2,885 #2 fable 5.1 – 4,021 #3 opus 5 – 5,161 biggest single build, lines #1 opus 5, bag end – 2,410 #2 fable 5.1, barad-dûr – 1,375 #3 fable 5, bag end – 1,319 observations: • fable 5.1 is the only model that furnished the bag end interior – a live fire, panelling, books on the floor, leaded diamond windows, against fable 5's flat color and opus's dark tunnel. the round door outside opens onto that room, the hard part of the brief • what it costs is thinking room. the 128k output ceiling is a thinking budget in disguise: fable 5.1 burned 102,116 of it on reasoning and hit the wall mid-file. opus spent 109,241 and hit the same wall. fable 5 spent 61,240 and finished bag end in one call – the only one that did • fable 5.1's first pass is not the finished thing. its barad-dûr came back with three defects you only catch by looking at it – nothing a read of the code would have flagged • it is the best of the three at being corrected. handed a plain list of what was wrong, it returned 32 targeted patches over two rounds, every one applied first try, and it worked out one of the causes itself instead of guessing at constants conclusion: nine scenes, 12,067 lines and 1.46m output tokens for $55.88 all in – and the cheapest model was also the fastest, by 3.2x! follow @thehypedotnews for 24/7 ai news, analysis and breakdowns

  3. Meta Engineering Blog53

    Meta 构建“组织第二大脑”AI Agent,从专家反馈中沉淀机构知识

    Meta 发布技术博客,介绍其为合规领域构建的 AI Agent,用结构化知识系统加可组合 recipes 复现专家推理方式,专家反馈经诊断、编译、验证和人工审核后写入 200+ 知识文件,无需重新训练模型。六周三轮开发后,单项评估时间从数天缩短到数分钟,改进周期内零回归,每轮修复自动加入回归测试集。

  4. GitHub Blog62

    GitHub Copilot 如何在不牺牲任务质量的前提下降低 AI 编码成本

    GitHub 工程师 Erik Kristensen 分享了 Copilot 降本的四项改动:选择性压缩工具输出、移除 view 工具行号前缀(线下推理成本降约 5%,线上用户日均推理成本降约 3%)、压缩 task-tool 提示词(每轮省约 1300 token,每活跃小时归一化成本降 2.9%)、后台任务完成后直接交付结果(AI Credits 用量降约 2.3%)。

    推荐理由:GitHub 用四项改动说明压缩 token 为何要看整个任务而非单次调用,并给出可复用的评估方法与踩坑教训。

  5. Claude:Blog79

    Anthropic 发布 Claude 电商智能体构建指南及参考实现

    Anthropic 发布构建电商智能体的指南,总结架构、延迟与成本优化、生产运维三部分实践,涉及购物和商家两类智能体,企业客户部署后出现购物车变大和卖家运营效率提升。核心建议包括用单个智能体加技能而非子智能体、将 UI 组件做成工具、用提示词缓存实现 90-99% 命中率、在 harness 中强制安全规则,并开源参考实现 anthropics/commerce-agents。

    推荐理由:Anthropic 根据多个企业部署经验总结电商智能体架构、延迟成本优化和生产实践,并附参考实现,可迁移到类似 Agent 项目。

  6. Google AI:DEV 作者专属66

    Google AI 团队分享如何为 LLM-as-a-Judge 评测编写可靠的评分标准

    Google AI 团队发布教程,讲解如何为 LLM-as-a-Judge 评测编写可靠的布尔式评分标准,指出模糊提示会导致评估不一致和浪费 token。文中给出四条经验:问题保持原子化且互不重叠、只让评判模型评估客观事实(可用 RFC 2119 术语如 MUST 表述)、只评 prompt 中明确要求的内容、用专家标注的 golden set 校准评判模型直至与人类评分一致。

    推荐理由:作者来自 Google 团队,把写作 LLM-as-a-judge 评分标准的经验整理成四条可迁移规则,附带校准流程,适合自己搭建评测时参考。

9月2日周三
  1. Google AI:DEV 作者专属69

    什么是 harness 工程?Google 用 ADK 2.0 与 Antigravity SDK 演示自动修复编码循环

    Google 员工 Shir Meir Lador 介绍 harness 工程,即用确定性组件包裹 LLM,包括编排层、执行沙箱、状态持久化和验证工具,让 Agent 不需逐行人工审查即可安全生成代码。

    推荐理由:原文解释了 harness 工程的构成要素,并给出可运行的沙箱与修复循环示例代码,方法可直接迁移到自己的 Agent 工作流。

  2. ByteByteGo57

    RAG 系统效果为何取决于嵌入模型:ByteByteGo 图解检索机制与选型要点

    ByteByteGo 发文解析 RAG 系统中嵌入模型的关键作用,指出语言模型再强也无法弥补错误检索,因为嵌入模型决定了哪些文本进入上下文。文章图解了索引与检索两个阶段的工作流程,列举了语义相似但答非所问的多种失败模式,包括否定句、版本差异和数字标识等,并说明更换嵌入模型需要全量重建向量、索引和权限,成本高昂,类似蓝绿部署。

  3. IT之家·人工智能48

    研究发现 AI 客服更拟人未必带来更佳体验

    发表于《欧洲营销杂志》的研究发现,在某些情况下不那么像真人的文字聊天机器人反而获得更积极的客户反馈,打字的慢节奏让消费者有时间平复负面情绪。另一项发表于《商业研究杂志》的研究显示,能准确识别消费者愤怒等情绪状态的 AI 客服会带来更高满意度。调查还显示 1031 名澳大利亚受访者中 78% 用过自动化客服,仅 39% 感到满意。

  4. Baseten 工程博客62

    Baseten 解析后训练最佳开源模型并按成本分档推荐

    Baseten 发文解析开源模型后训练的成本驱动因素,指出活跃参数量是 RL 后训练成本的最大来源,总参数和 KV cache 主要限制推理速度,并按成本档位推荐 DeepSeek-V4-Flash、GLM-5.2、Kimi K2.6、Kimi K2.7 Code、Nemotron-3-Super-120B 和 Qwen3 系列。选型建议关注库支持、基准表现和后训练循环成本三方面。

    推荐理由:原文把后训练成本拆解为活跃参数、总参数和 KV cache 三个因素,并按成本档位比较了多款开源模型,方法可直接迁移到选型。

  5. 公众号:千问APP41

    天体物理学博士刘博洋用千问复现Druckmüller算法,日全食后3小时产出高清日冕图

    8月12日西班牙出现欧洲大陆27年来首次日全食,全食持续101秒,天体物理学博士刘博洋在日食结束后不到3小时产出了高清日冕精细结构图。他此前把Druckmüller的五篇论文交给千问「工作助理」分析核心思路,悟出大师方法的本质是模拟人眼的局部差分视觉,并用过去日冕素材迭代100多个版本复现了整套算法。观测现场他还用千问根据低仰角与高仰角曝光数据推算补偿档数、写脚本测试新快门线的参数组合效率。

  6. Rohan Paul45

    AI/ML API 用同样的提示词对 Anthropic 的 Fable 5.1 和 OpenAI 的 GPT-5.6 Sol 做了一次成型的 Three.js 测试,5 个独立场景组成私岛豪宅项目。

    引用AI/ML API@aimlapi

    Fable 5.1 charges 6× more than GPT-5.6 Sol anthropic's Fable 5.1 went live. we threw it straight at openai's GPT-5.6 Sol. same prompts, one shot. the test: five self-contained three.js scenes, one continuous project — a futuristic mansion on a private island. exterior orbit, grand hall with a diamond chandelier, pool with a water park, garage with 10 lambos, helicopter landing on the roof cost: Claude Fable 5.1 — $5.69 GPT-5.6 Sol — $0.88 what we saw: • fable goes for realism — waves, depth, a sand ring around the island. sol goes low-poly and clean, and its style never breaks across all 5 scenes • fable's chandelier is thousands of individual crystals. sol's hall is honestly just a staircase and a wall • fable's helicopter touches down on the lit helipad much smoother than sol's • fable built a triple-slide water-park tower — and shipped the pool with no water. sol's park is tiny — but the pool is full 6× the price buys you 6× the detail — and the same one-shot bugs. what's your choice? try Fable 5.1 on aimlapi[.]com

  7. Rohan Paul58

    Satya Nadella 表示 Microsoft 最新威斯康星 AI 数据中心 Fairwater 每年用水量约等于一家本地餐厅,冷却回路一次性注水后可近乎零耗水运行。

    引用Rohan Paul@rohanpaul_ai

    WSJ just published a piece, on how Data centers can be a huge win for local communities. In Loudoun County, removing them could mean about $5,800 more in annual property taxes for homeowners just to maintain the same services. In Louisiana, Meta’s data-center tax revenue helped fund teacher bonuses of up to $50,000. "Data centers also generate scant pollution. Their CO2 emissions mainly stem from the natural gas plants that power them. But modern gas generators produce upward of 90% less fine particulate matter, sulfur dioxide and carbon monoxide than a steel mill."

  8. OpenAI:官网动态46

    OpenAI 解析 Basis、Clay、Exa 如何把工作流变成 AI 原生运营能力

    OpenAI 发布文章,介绍 Basis、Clay 和 Exa Labs 用 Agent 承接员工入职、账户管理和开发者集成的工作流。Basis 把入职从 2 小时缩短到 30 分钟,Clay 的子智能体每天夜间更新各账户上下文,为销售省下约 1 小时邮件分拣,Exa 用 Codex 监控集成机会并创建 PR、跑测试。文章还给出企业从试点到规模化的六步方法。

  9. Tessl:产品与工程博客66

    Tessl 设计师总结如何把设计判断教给 AI 智能体

    Tessl 设计师复盘在 agent 软件工厂 Kikimora 中把品牌与设计规则编入 AI 工作流的方法,指出设计系统只提供组件、教不会构图。

    推荐理由:作者给出从 Figma diff 反推设计规则、再以技能和全量扫描落地的可复用路径,可迁移到任何用 agent 生产界面的团队。

9月1日周二
  1. The Decoder:AI News67

    Google AI 搜索按国籍给出不同安全建议,修复后仍对 Facebook 相关提问发出警告

    Futurism 测试发现,Google AI Overviews 对用户称与非洲、印度或巴基斯坦裔独处的提问会建议报警或前往安全地点,对英国人则建议喝茶聊天气。Reddit 用户首先发现差异,Google 承认结果不符合其标准并称触发词是 alone,修复后国籍类提问恢复正常,但询问与来自 Facebook 的人独处时仍会建议离开并拨打 110。