1. Writing Code Was Never the Bottleneck (ordep.dev)
LLMs make it easier to write code, but understanding, reviewing, and maintaining it still takes time, trust, and good judgment.
2025-07-01
45 篇热帖
2. The new skill in AI is not prompting, it's context engineering (www.philschmid.de)
Context Engineering is the new skill in AI. It is about providing the right information and tools, in the right format, at the right time.
3. Xfinity using WiFi signals in your house to detect motion (www.xfinity.com)
Find more information on using WiFi Motion in the Xfinity app.
4. Show HN: Spegel, a Terminal Browser That Uses LLMs to Rewrite Webpages (simedw.com)
Spegel is a proof-of-concept terminal web browser that feeds HTML through an LLM before serving it as markdown in the terminal.
5. OpenFLOW – Quickly make beautiful infrastructure diagrams local to your machine (github.com)
6. Claude Code now supports hooks (docs.anthropic.com)
Reference for Claude Code hook events, configuration schema, JSON input/output formats, exit codes, async hooks, HTTP hooks, prompt hooks, and MCP tool hooks.
7. I write type-safe generic data structures in C (danielchasehooper.com)
See my follow-up article: “A Fast, Growable Array With Stable Pointers in C” I write type safe generic data structures in C using a technique that I haven’t seen elsewhere1. It uses unions to associate type information with a generic data structure, but we’ll get to that. My approach works for any type of data structure: maps, arrays, binary trees… but for this article I illustrate the ideas by implementing a basic linked list.
8. Cloudflare to introduce pay-per-crawl for AI bots (blog.cloudflare.com)
Pay per crawl is a new feature to allow content creators to charge AI crawlers for access to their content.
9. Show HN: A continuation of IRS Direct File that can be self-hosted (github.com)
A free tax tool based on IRS Direct File. Contribute to openfiletax/openfile development by creating an account on GitHub.
10. Ask HN: Who is hiring? (July 2025)
11. That XOR Trick (2020) (florian.github.io)
There are a whole bunch of popular interview questions that can be solved in one of two ways: Either using common data structures and algorithms in a sensible manner, or by using some properties of...
12. Feasibility study of a mission to Sedna - Nuclear propulsion and solar sailing (arxiv.org)
Abstract page for arXiv paper 2506.17732: Feasibility study of a mission to Sedna -- Nuclear propulsion and advanced solar sailing concepts
13. Proton joins suit against Apple for predatory practices (proton.me)
Apple's anti-competitive App Store policies are gouging consumers, enabling censorship, and harming privacy-focused developers. We're joining a lawsuit to restore internet freedom.
14. Grammarly acquires Superhuman (www.reuters.com)
15. Scientists identify culprit behind biggest-ever U.S. honey bee die-off (www.science.org)
16. About AI Evals (hamel.dev)
A comprehensive guide to LLM evals, drawn from questions asked in our popular course on AI Evals. Covers everything from basic to advanced topics.
17. HN Slop: AI startup ideas generated from Hacker News (www.josh.ing)
Fresh AI generated startup ideas from the current Hacker News front page. Powered by Claude AI.
19. Ask HN: What's the 2025 stack for a self-hosted photo library with local AI?
20. Fei-Fei Li: Spatial intelligence is the next frontier in AI [video] (www.youtube.com)
21. Spending Too Much Money on a Coding Agent (allenpike.com)
On making use of large thinking models.
22. Conversations with a hit man (magazine.atavist.com)
A former FBI agent traveled to Louisiana to ask a hired killer about a murder that haunted him. Then Larry Thompson started talking about a different case altogether.
23. ASCIIMoon: The moon's phase live in ASCII art (asciimoon.com)
ASCII Moon - view & cycle through The Moon's phases - rendered in ASCII art.
24. Alice's Adventures in a Differentiable Wonderland (arxiv.org)
这篇名为《Alice's Adventures in a Differentiable Wonderland》的论文是一篇关于可微编程与神经网络的入门导论。其核心内容可概括如下：
目的与背景：本文旨在为像爱丽丝这样的初学者介绍可微编程这一迷人的领域。神经网络作为大型语言模型、语音转录、分子发现、机器人等众多应用的基础，其本质是可微原语的组合。学习神经网络即是学习如何与这些模型交互和进行编程，即可微编程。
主要内容与结构：本文以直观、自成一体的方式介绍关键技术，旨在弥合理论与代码（特别是PyTorch和JAX）之间的差距。主要内容涵盖：
- 基础概念：通过自动微分优化函数的基础知识。
- 常见架构：介绍了处理不同数据类型（序列、图、文本、音频）的最常见设计。
- 关键设计模块：重点讲解了卷积块、注意力机制模块和循环块这三种最重要的设计技术。
- 进阶目标：通过掌握上述基础，使读者具备理解诸如大型语言模型和多模态架构等前沿模型的能力。
定位与目标读者：本文定位为一份入门读物，专注于提供直观解释和实用的设计技术知识，而非深入复杂的数学理论。它旨在为初学者搭建从基础知识通往理解高级模型的桥梁。
25. Next month, saved passwords will no longer be in Microsoft’s Authenticator app (www.cnet.com)
Microsoft's go-to password manager won't be the same after Aug. 1.
26. PlanetScale for Postgres (planetscale.com)
PlanetScale now supports Postgres
27. Converting a large mathematical software package written in C++ to C++20 modules (arxiv.org)
Abstract page for arXiv paper 2506.21654: Experience converting a large mathematical software package written in C++ to C++20 modules
28. The hidden JTAG in a Qualcomm/Snapdragon device’s USB port (www.linaro.org)
Learn about JTAG and EUD
29. Caching is an abstraction, not an optimization (buttondown.com)
I've always been told that caching is a tool to make software faster. That, given some careful considerations to consistency, caching makes it so that when...
30. Datadog's $65M/year customer mystery solved (blog.pragmaticengineer.com)
The internet has been speculating the past few days on which crypto company spent $65M on Datadog in 2022. I confirmed it was Coinbase, and here are the details of what happened.
31. Aging-related inflammation is not universal across human populations (www.publichealth.columbia.edu)
Inflammation may not be a universal human experience, but a byproduct of industrialized lifestyles.
32. Ubuntu 25.10 Raises RISC-V Profile Requirements (www.omgubuntu.co.uk)
Canonical confirm a technical shift for Ubuntu on RISC-V. With Ubuntu 25.10, it will only support hardware meeting the RVA23 profile spec.
33. Why email startups fail (forwardemail.net)
Comprehensive analysis of email startup failures. Why most email companies fail, burn millions in VC funding, and shut down. Learn what actually works.
34. Ask HN: Who wants to be hired? (July 2025)
35. Show HN: Jobs by Referral: Find jobs in your LinkedIn network (jobsbyreferral.com)
36. Melbourne man discovers extensive model train network underneath house (www.sbs.com.au)
Daniel Xu has loved trains since he was a child. In a surprising twist of fate, he now lives above an extensive hobby train network set.
37. Price of rice in Japan falls below ¥4k per 5kg (www.japantimes.co.jp)
38. In a milestone for Manhattan, a pair of coyotes has made Central Park their home (www.smithsonianmag.com)
39. Context Engineering for Agents (rlancemartin.github.io)
上下文工程概述
上下文工程是为AI智能体（Agent）在每个决策步骤中精确管理上下文窗口信息的艺术与科学。类似操作系统管理CPU的RAM，上下文工程负责为大语言模型（LLM）筛选并组织其有限上下文窗口中的信息。无效的上下文管理会导致性能下降、成本增加等问题。
当前针对智能体的上下文工程主要可分为四大策略：写入、选择、压缩和隔离。
1. 写入上下文 指将信息保存在上下文窗口之外，供智能体在任务中使用。
- 便签/草稿纸：智能体在任务执行中将重要信息（如计划、中间结果）保存到外部存储（如文件、状态对象）。例如，Anthropic的多智能体研究员会将计划保存到“记忆”中，以防上下文窗口溢出。
- 记忆：用于跨会话保留信息。这包括基于反思生成的记忆（如Reflexion）或定期合成的记忆。ChatGPT、Cursor等产品具备自动生成长期记忆的能力。
2. 选择上下文 指将相关信息拉入上下文窗口，以协助任务完成。
- 从便签/记忆中选择：根据任务需求，智能体从已保存的便签或长期记忆库中检索相关信息（如示例、指令、事实）。例如，代码智能体常使用固定的规则文件（如
CLAUDE.md）作为始终加载的上下文。 - 工具选择：当工具过多时，可通过检索增强生成（RAG）技术，基于语义相似度检索最相关的工具描述，以提高工具选择的准确性。
- 知识检索：RAG是关键，尤其对于代码智能体等大规模生产系统。挑战在于随着代码库扩大，需要结合代码解析、知识图谱、重排序等多种技术来提升检索效果。
3. 压缩上下文 旨在保留任务所需的最少token。
- 上下文摘要：智能体交互可能长达数百轮。摘要是一种常见管理方式，如Claude Code在上下文窗口使用率达95%后自动压缩对话历史。摘要可在智能体轨迹中递归或分层进行，也可用于处理token密集的工具调用结果或在多智能体间传递知识时。
- 上下文修剪：通过启发式规则（如移除旧消息）或训练的修剪模型（如Provence）来过滤上下文，与摘要（保留精华）相比更侧重于删除无关内容。
4. 隔离上下文 将上下文分割以优化智能体性能。
- 多智能体系统：通过分解任务到多个子智能体，每个子智能体拥有独立的上下文窗口、工具和指令，实现关注点分离。这能提升性能，但可能导致token消耗激增且需要精心的提示工程和协调。
- 环境隔离：例如，HuggingFace的深度研究智能体使用代码智能体（CodeAgent），该智能体生成包含工具调用的代码，并在沙箱中运行，仅将选定的返回值（上下文）传回LLM。这有效隔离了图像、音频等大容量对象。
- 状态对象：智能体的运行时状态对象（如带有特定模式的Pydantic模型）也可用于隔离上下文。仅部分字段（如
messages）需暴露给LLM，其他字段可存储选择性使用的信息。
总结 上下文工程是构建高效智能体的核心。其四大策略——写入、选择、压缩、隔离——提供了系统性的方法来管理有限的上下文窗口，确保智能体在复杂、长程任务中保持性能并控制成本。这些模式仍在发展中，但已在当前主流智能体系统中得到广泛应用。
40. Small language models are the future of agentic AI (arxiv.org)
Abstract page for arXiv paper 2506.02153: Small Language Models are the Future of Agentic AI
41. Evidence of a 12,800-year-old shallow airburst depression in Louisiana (www.scienceopen.com)
42. Show HN: I built the tool I wished existed for moving Stripe between countries (www.stripemove.com)
该项目是一个专为跨国迁移Stripe账户设计的自动化工具，旨在帮助企业无缝地将Stripe业务迁移到另一个国家，确保运营不中断。该工具可自动转移关键数据，包括客户、订阅、产品、优惠券和支付方式。
43. Sony DTC-700 audio DAT player/recorder (kevinboone.me)
What offers the convenience of a cassette tape, and the sound quality of a CD? Digital Audio Tape (DAT) does. So why wasn’t it more successful?
44. The original LZEXE (A.K.A. Kosinski) compressor source code has been released (clownacy.wordpress.com)
Last year, I discovered that the Kosinski compression format is actually LZEXE, which was used for compressing DOS executables back in the 90s and the late 80s. Its developer catalogues three versions on his website: v0.90, v0.91, and v0.91e. While only binaries of v0.91 and v0.91e can be found on the website, v0.90 can be…