第 39 期
今日趋势
今天的简报内容主要围绕AI技术工具迭代、应用实践及相关行业与科研问题展开,覆盖技术更新、落地案例、争议事件等多个维度。其中一条主线是AI工具的优化与体验反馈:llm 0.33升级了依赖库与命令功能,Notion AI推出本地Agent突破额度限制,同时也暴露了Codex CLI重复回答的缺陷,不同模型在文本阅读等场景的表现差异也引发关注。另一条主线是AI相关的行业与科研争议:Uber因自动化决策违反GDPR被罚款,国产flash级模型存在推理效率低的问题,还有对一项声称“打排球缩短寿命”的糟糕科研论文的批判,反映出AI落地中的合规、性能及科研严谨性短板。整体来看,AI技术快速迭代的同时,也伴随工具缺陷、行业合规风险等问题,不同主体在AI应用中的模式选择(如AI硬件付费逻辑)也呈现出多样化的探索与分歧。
AI 技术博客
- llm 0.33 — simonwillison.net This entry covers the release of llm 0.33, a tool for working with language models. Key updates include upgrading the OpenAI Python library to version 3.x and switching the HTTP client dependency from httpx to httpx2. The release also adds support for –key parameters to llm embed and llm embed-multi commands, plus allows repeating the –template flag to combine templates for model configuration. Additionally, reasoning-capable Responses API models now support a reasoning_summary option with auto, concise, and detailed values for use with llm openai endpoint –responses.
- More than just code review — simonwillison.net This entry discusses the core skill required to use coding agents productively. The key skill is being able to confidently instruct coding agents to make desired changes, then confidently verify that those changes have been applied correctly. While reviewing every line of code written by the agent is sometimes necessary, it is not the only approach to achieving this validation goal. The text notes that line-by-line review has never been the most effective way to validate changes to software.
- Quoting Linus Torvalds — simonwillison.net This entry quotes Linus Torvalds discussing an AI-assisted debug session for a Linux kernel fix. Torvalds states the AI handled much of the grunt work, though it several times claimed the problem was impossible and suggested writing a report instead. Despite that, the AI continued adding debug code and analyzing when Torvalds pushed it to keep working. Torvalds credits the AI for writing the commit message for the fix to the drm/xe component, which prevents handing out flat CCS storage as usable VRAM.
- WorkOS: Agents Can Now Sign Up for Your App — John Gruber This entry announces WorkOS’ Agent Registration feature, promoted via a sponsorship of the Daring Fireball weekly newsletter. The feature addresses the issue where automated agents bounce off human-designed browser login flows, turning that lost traffic into actual app signups. Users can enroll via the WorkOS dashboard, and AuthKit publishes an auth.md file that agents read to register for scoped, short-lived credentials controlled by the app. The text encourages making apps agent-ready using this feature.
- Mark Carney on the U.S. Under Trump: ‘Sometimes, Its Signature Is Written in Pencil’ — John Gruber This entry reports on Canadian Prime Minister Mark Carney’s explanation for walking away from U.S. trade talks under the Trump administration. Carney stated the U.S. proposal was a bad deal, noting America has changed and its commitments are not reliable (“its signature is written in pencil”). He added Canada will not accept the U.S. offer or give what was asked, to protect its sovereignty and key industries. The text contrasts Carney’s consistent, steadfast approach with Trump’s erratic, unstable, dishonest behavior, framing cutting off such behavior as the correct response.
- Dutch Regulator Fines Uber $1 Billion for Suspending Dishonest Drivers — John Gruber This entry covers a fine imposed on Uber by the Dutch Data Protection Authority. The authority fined Uber €825 million (about $966 million) for deactivating driver accounts through automated systems without adequate warning, a violation of Europe’s GDPR. This is the second-largest GDPR fine ever issued, behind only Meta’s 2023 €1.2 billion fine. The regulator stated Uber made major, life-impacting decisions solely via algorithms without meaningful human review or appeal, while Uber has appealed the decision, disputing the number of affected drivers and its automation of permanent deactivations.
- The difference orbit inclination makes — John This entry discusses the impact of Mars’ orbital inclination on the distance between Earth and Mars. It simplifies the planets’ orbits to circles (Earth at radius 1, Mars at 1.5) and compares distance calculations with and without accounting for Mars’ 1.85° orbital tilt relative to Earth’s orbit. The analysis finds the inclination effect is small—about three orders of magnitude smaller than the main distance variation—but erratic. The post includes code for the calculations and plots illustrating the difference.
- Coming soon — John This entry explores modeling the time until a business labeled “coming soon” actually opens. It uses a local pizza shop that has been “coming soon” for over a year as an example, noting the expected time to opening does not decrease over time in such cases. The text explains two key probability distributions: the probability the business will eventually open, and the time to opening if it does. It concludes that if there is any small chance the business never opens, the expected time to opening becomes infinite.
因果推断与营销科学
- This is one of the worst scientific papers I’ve ever seen. — Andrew 本文是一位评论者对题为《运动与长寿:国际运动员的观察研究》的论文的批评,称其为糟糕的定量社会科学研究,核心结论为打排球会使寿命缩短5年,该结果具有统计学显著性(p值2.4e-11)。作者还吐槽该研究使用纳税人资金,且期刊未公开评审意见,这类无依据的科学主张传播会误导公众,值得关注其对不良科研风气的批判。
- Building a Proper Backend for My LangGraph AI Agent — Soner Yıldırım 本文是LangGraph AI代理系列的内容,作者将原本用内存存储的演示版预订代理,转向构建带实际业务数据的产品级后端,计划替换为Postgres数据库,设计了统一的BookingRepository协议,支持内存与Postgres两种存储模式,解决了内存存储导致的数据丢失、多会话不同步等问题,项目源码可在GitHub获取,对构建产品级LangGraph代理有实践参考价值。
- Multi-Document RAG: A Folder of Unrelated PDFs Is One Long Document with a Nested Outline — angela shi 本文介绍多文档RAG的一种思路,将包含无关PDF的文件夹视为带嵌套大纲的长文档,每个文件对应一条摘要及自身目录,检索时分为两个层级进行,该方法解决了无共享字段导致的索引构建难题,为处理分散的多文档集合提供了可行方案。
- Why We Fine-Tuned SigLip (And Why That’s Not Always the Right Call) — Miikka Silfverberg 本文来自Alma Media团队,介绍为解决房地产图片标注的标签不足问题,采用LoRA微调SigLIP模型,同时指出是否选择微调需考虑三个问题,还对比了用外部API与自建模型的成本差异,为企业场景下构建图像分类模型提供了决策依据。
即刻简报
- 发布了: Notion AI 现在出了一个把 Notion AI 技能下载到本地 Agent 的功能。 意思是你可以用 Notion 编辑器,高效的写一个技能(读哪个数据库的信息源、按哪个… — 即刻·评论尸 本文介绍Notion AI推出的本地Agent功能,用户可通过Notion编辑器创建技能,涵盖数据库信息源读取、页面模板处理等设置,将技能一键复制到Codex,用Codex额度执行,不受Notion AI额度限制,该功能也适用于Claude Code,因Notion的页面与数据库交互更便于上下文管理。该内容能帮助AI工具使用者规避Notion AI额度限制,优化技能执行流程,具有实用参考价值。
- 发布了: Codex CLI在compact之后,有时会把一个我之前问过、也已经回答过的问题,重新翻出来再答一遍。甚至下一次compact的时候,同一个问题还会再冒出来。 我一… — 即刻·李自然 本文指出Codex CLI在compact操作后,会重复回答用户此前已提问并得到过答案的问题,甚至在后续compact时仍会重复出现该情况。作者虽认可OpenAI模型的性能,但认为其配套工具的水平有待提升。该内容能让使用Codex CLI的用户了解其存在的重复回答问题的缺陷,便于提前规避相关问题,具有实用参考性。
- 发布了: Ox Alpha 多模态和代码能力真的牛皮 给了他一张图片,让他用 WebGL 还原这个网页。 他真的做出来了这种 3D 玻璃材质的质感和透视,非常清晰。 最牛逼的… — 即刻·歸藏 本文分享了Ox Alpha的多模态与代码能力表现,用户给该模型一张图片,要求其用WebGL还原对应网页,模型成功做出了具备3D玻璃质感与透视效果的还原内容,文字位置、排版等细节与原图高度一致,仅字体存在差异。该内容能让AI工具使用者直观了解Ox Alpha的代码与多模态结合能力,具有参考价值。
- 发布了: AI 硬件的付费点应该是让用户付费拿到自己用硬件获得的上下文,而不是反过来。 最近这个 AI 硬件反正出的越来越多了,但是大家的思路好像都没有转变过来… — 即刻·歸藏 本文提出AI硬件的付费点应是让用户付费获取自身使用硬件产生的上下文,而非让用户将数据留在硬件方,当前多数AI硬件的思路仍未转变。作者常用飞书录音豆,因它可将录音转写文字拉到本地上下文,在Agent集中化、模型价格变动阶段,提供丰富Agent或Skill助力用户获取自身上下文更具价值。该内容能为AI硬件开发者提供付费模式的新思路,具有启发意义。
- 发布了: 开发一个隐藏Twitter顶部蓝色通知条的功能 用 luna terra deepseek pro 开发了好几遍,试了三种方法,各种搞不定 那一刻觉得 AI 太笨了,这么个小功能啊… — 即刻·橘AI 本文记录了开发隐藏Twitter顶部蓝色通知条功能的过程,用户用luna terra deepseek pro多次开发该功能,尝试三种方法均未成功,一度觉得AI能力不足。后改用sol,先删除所有旧代码再重新开发,一次就完成了功能,用户认为sol物有所值,其高定价有合理之处。该内容能为开发者在AI工具选择与使用上提供实际案例参考,具有实用价值。
- 发布了: 同样的skill /ljg-book, Opus 4.6 解读的文本阅读体感明显好上很多。 — 即刻·李继刚 本文对比了同一skill /ljg-book在不同模型下的文本阅读表现,结果显示Opus 4.6的文本阅读体感明显优于其他模型。该内容能帮助用户在选择文本阅读类AI模型时提供参考依据,具有一定的实用价值。
- 发布了: 给所有想做出海 Web 即友的建议,可以先从我#公众号:哥飞 的文章看起,定个目标,两周内看完100篇,如果越看越想要看更多,那说明你真的对这事有兴趣。… — 即刻·哥飞 本文给出想做出海Web即友的建议,建议从公众号“哥飞”的文章入手,设定两周内看完100篇的目标,若越看越想看则说明对该领域有兴趣,若看不下去则说明志不在此,并非能力不足。该内容能为有意向进入出海Web领域的人提供明确的入门判断标准,具有指导意义。
- 发布了: 昨天和归藏师傅交流的时候聊到 国产的 flash 级别的模型 ox 和 ds flash 都有一个严重的问题 就是经常疯狂思考,最严重的有时候思考十几分钟才吐一个字 … — 即刻·橘AI 本文分享了与归藏师傅交流的内容,提及国产flash级模型ox和ds flash存在严重问题,常出现疯狂思考的情况,最严重时十几分钟才输出一个字,越复杂任务体验越差,作者怀疑是过度后训练超出模型参数承受极限。该内容能让关注国产AI模型的人了解其存在的性能缺陷,具有参考价值。
本期有 9 个源不可用:lcamtuf.substack.com、garymarcus.substack.com、rachelbythebay.com、joanwestenberg.com、dfarq.homeip.net、gwern.net、worksonmymachine.substack.com、chiark.greenend.org.uk/~sgtatham、tedunangst.com。