2025-07-07

37 篇热帖

1. Adding a feature because ChatGPT incorrectly thinks it exists (www.holovaty.com)

背景

Soundslice 运营一个乐谱扫描系统,可将照片中的乐谱数字化,用于播放、编辑和练习。公司通过监控错误日志持续改进系统。

问题发现

近期,错误日志中出现异常上传内容:用户提交的并非传统乐谱图片,而是 ChatGPT 会话的屏幕截图,其中包含 ASCII 吉他谱——一种简单的音乐记谱法,而非扫描系统原本支持的标准乐谱。

原因调查

经过排查,公司发现 ChatGPT 在回复用户时,错误地告知用户可以将 ASCII 谱上传到 Soundslice 网站以进行音频回放。实际上,Soundslice 并不支持此功能。这导致大量用户基于错误信息涌入系统,造成用户体验问题并使公司面临信任压力。

决策与行动

面对 ChatGPT 产生的误导性需求,Soundslice 面临选择:

  1. 添加免责声明,澄清不支持 ASCII 谱导入。
  2. 开发新功能以满足 ChatGPT 制造出的“市场需求”。

公司最终决定实现该功能,开发了一个专门的 ASCII 谱导入工具,并更新了扫描系统的界面提示,正式向用户介绍此新能力。

现状与反思

此事件被认为是已知首例因 AI 生成错误信息而促使公司开发全新功能的案例。公司对此感受复杂:一方面,新功能确实能帮助一部分用户;另一方面,开发决策在某种程度上是被 AI 的错误所“迫使”,这引发了关于是否应基于误导信息来规划产品功能的思考。

2. Bitchat – A decentralized messaging app that works over Bluetooth mesh networks (github.com)

bluetooth mesh chat, IRC vibes. Contribute to permissionlesstech/bitchat development by creating an account on GitHub.

3. Anthropic cut up millions of used books, and downloaded 7M pirated ones – judge (www.businessinsider.com)

Training Claude on copyrighted books it purchased was fair use, but piracy wasn't, the judge ruled.

4. Mercury: Ultra-fast language models based on diffusion (arxiv.org)

Abstract page for arXiv paper 2506.17298: Mercury: Ultra-Fast Language Models Based on Diffusion

5. I used o3 to profile myself from my saved Pocket links (noperator.dev)

Welp, Pocket shuts down tomorrow despite our pleas for it to stay. While migrating1 all of my saved articles, I noticed that I’ve got almost 900 saved articles spanning nearly 7 years. That’s a goldmine of stuff-I-like data! Some quick analysis using xsv2: 𝄢 unzip pocket.zip && xsv headers part_000000.csv 1 title 2 url 3 time_added 4 tags 5 status

𝄢 xsv sample 1 part_000000.csv | xsv flatten title The Uncertain Future of American Libraries url https://mek.fyi/posts/the-uncertain-future-of-american-libraries time_added 1678243022 tags status unread

𝄢 xsv count part_000000.csv 878

𝄢 xsv select time_added part_000000.csv | xsv stats | xsv select min,max | xsv flatten | while read header value; do echo -n "$header " date -I -d "@$value" done min 2018-07-12 max 2025-06-26 Experiment Recalling Simon Willison’s recent geoguessing challenge for o3, I considered, “What might o3 be able to tell me about myself, simply based on a list of URLs I’ve chosen to save?” I opened a temporary chat with o3 (doesn’t use ChatGPT’s memory feature) and prompted:

6. Postgres LISTEN/NOTIFY does not scale (www.recall.ai)

Learn why Postgres LISTEN/NOTIFY doesn't scale for high-throughput use cases and what architectural alternatives Recall.ai uses in production.

7. I extracted the safety filters from Apple Intelligence models (github.com)

Decrypted Generative Model safety files for Apple Intelligence containing filters - BlueFalconHD/apple_generative_model_safety_decrypted

8. Show HN: I wrote a "web OS" based on the Apple Lisa's UI, with 1-bit graphics (alpha.lisagui.com)

Alpha release of LisaGUI, a Lisa-themed web OS project with 1-bit graphics.

9. A non-anthropomorphized view of LLMs (addxorrol.blogspot.com)

In many discussions where questions of "alignment" or "AI safety" crop up, I am baffled by seriously intelligent people imbuing almost magic...

10. The jank programming language (jank-lang.org)

jank is a Clojure dialect on LLVM with a native runtime and C++ interop.

11. Intel's Lion Cove P-Core and Gaming Workloads (chipsandcheese.com)

Lion Cove is Intel’s latest high performance CPU architecture.

12. The ChompSaw: A Benchtop Power Tool That's Safe for Kids to Use (www.core77.com)

Of all freestanding power tools, perhaps the jigsaw or scrollsaw would be the most fun for kids to use, as they allow freeform cuts. They're also pretty darn dangerous, with the jigsaw being powerful enough to remove fingers. A much safer way to introduce kids to power tools is this

19. Async Queue – One of my favorite programming interview questions (davidgomes.com)

For the past 7+ years, I've been conducting a programming interview that has been a true personal favorite of mine. It was passed down to me from good friends (Jeremy Kaplan and Carl Sverre, and it was the latter whom I believe invented it). This interview has probably been given

21. LLMs should not replace therapists (arxiv.org)

Abstract page for arXiv paper 2504.18412: Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers

22. Opencode: AI coding agent, built for the terminal (github.com)

The open source coding agent. Contribute to anomalyco/opencode development by creating an account on GitHub.

23. The Origin of the Research University (asteriskmag.com)

Universities have existed for more than a thousand years — and for almost all of that time, they weren’t centers of research. What changed in 19th century Germany?

25. Show HN: NYC Subway Simulator and Route Designer (buildmytransit.nyc)

这是一个关于纽约市地铁模拟器和路线设计器的项目,旨在让用户模拟和体验地铁系统,并支持设计自定义路线。其核心功能和特点如下:

  • 交互式地图:项目提供了一个交互式地图界面,用户可以在上面直观地查看纽约市的地铁线路、站点以及列车的实时运行位置。
  • 运行模拟:模拟器能够按照真实的时间表或自定义的设置运行地铁列车。用户可以观察列车在不同线路上的行驶过程,包括停靠、换乘等行为。
  • 路线设计器:这是一个关键功能,允许用户设计自己的地铁线路。用户可以在地图上规划新的线路走向、设置站点,并定义线路的运行参数。
  • 实时数据与统计:模拟运行过程中,系统会提供实时数据,例如当前时间、列车位置、预计到达时间等。对于用户设计的路线,系统可能还会估算出行程耗时等统计信息。
  • 自定义与配置:项目提供了一定程度的可配置性,用户可能能够调整模拟速度、运行列车数量,或者修改线路参数,以探索不同场景下的系统表现。

该项目的技术栈和完整功能细节未在提供的摘要中完全展开,但从描述来看,它整合了地图可视化、模拟引擎和用户交互设计,为地铁系统的爱好者、规划者或开发者提供了一个探索性的工具平台。

26. Deno 2.4 (deno.com)

Deno bundle is back, alongside the addition of bytes and text imports, stabilized built-in OpenTelemetry, a new --preload flag, simplified dependency management with deno update, and more.

28. Archaeologists unveil 3,500-year-old city in Peru (www.bbc.co.uk)

Researchers believe Peñico served as a trading hub linking ancient Pacific coast communities with those living in the Andes and Amazon.

29. I Ported SAP to a 1976 CPU. It Wasn't That Slow (github.com)

Contribute to oisee/zvdb-z80 development by creating an account on GitHub.

30. The Broken Microsoft Pact: Layoffs and Performance Management (danielsada.tech)

The unspoken agreement that made Microsoft's lower pay tolerable has fundamentally broken.

31. 7-Zip for Windows can now use more than 64 CPU threads for compression (www.7-zip.org)

7-Zip 版本历史摘要

7-Zip 是一款开源文件压缩工具,自1999年发布以来持续更新。以下是其版本演进的主要脉络:

早期发展与核心功能建立 (1999-2002)

  • 1999年: 发布首个版本,支持基本的压缩/解压缩功能。
  • 2000年: 2.20版本引入了7z 格式BZip2、Cab支持,并构建了新的模块化架构。
  • 2001年: 2.30 Beta版本开始,逐步增加了SFX自解压、Solid压缩、多线程支持 (-mmt 开关)、BCJ/BCJ2 过滤器以及7z格式的AES-256加密。同时,对RAR、ZIP、TAR、GZip等格式的支持也得到增强。

功能完善与平台扩展 (2003-2014)

  • GUI与易用性: 7-Zip File Manager 在此期间功能不断增强,支持拖放、基准测试、文件注释、SHA-256计算等。
  • 压缩算法与性能: 引入了LZMA2压缩方法和XZ格式支持,并持续优化LZMA的压缩比和速度。多线程压缩得到广泛支持。
  • 格式支持: 增加了WIM、DMG、VHD、VHDX、QCOW2、VMDK、ext2/3/4、SquashFS、CramFS、ZSTD (初始解压支持) 等众多归档和磁盘映像格式的支持。
  • 加密增强: 2019年,将7z格式的加密初始化向量从64位增加到128位,提高了安全性。
  • 命令行改进: 增加了大量命令行开关,用于控制内存使用、时间戳存储、文件流、哈希计算等。

现代化与性能飞跃 (2015-2026)

  • 硬件优化: 利用现代CPU指令集(如AES-NI、VAES、SHA扩展)大幅提升了加密/解密和哈希计算速度。针对ARM64平台进行了优化。
  • 性能突破:
    • 25.00 (2025):Windows上支持超过64个CPU线程进行压缩和基准测试,能自动分配至多个处理器组。
    • 26.01 (2026): Linux版本支持大页内存 (2 MB),可将7z/xz/LZMA/LZMA2的压缩速度提升约10%。
    • 持续优化压缩速度,如bzip2压缩速度提升15-40%。
  • 新功能与改进:
    • 24.01 (2024): 增加ZSTD格式的创建支持和XXH64哈希算法支持。
    • 24.09 (2024): 增大了LZMA/LZMA2的默认字典大小,并新增了SHA-512、SHA3-256等哈希算法。
    • 增加了新的压缩过滤器,如用于RISC-V和ARM64可执行文件的过滤器。
    • 改进了对符号链接和硬链接的处理安全性(CVE-2025-55188, CVE-2025-11001等)。
  • 持续维护: 每个版本都包含多项错误修复安全漏洞修复(以CVE编号标识),涵盖了从简单的解压缩错误到严重的堆缓冲区溢出漏洞(如CVE-2026-48095)。文件管理器的排序、界面和临时文件管理功能也持续优化。

总结:7-Zip 从一个基础的压缩工具,发展成为一个支持格式极广、压缩性能卓越、安全性持续增强的综合解决方案。其发展关键点在于:引入并优化7z格式、实现强大的多线程支持、拥抱现代CPU硬件特性、以及对海量文件格式的广泛兼容。它始终坚持开源,为个人和企业用户提供了一款高效可靠的压缩软件。

33. Huawei cloned Qwen and DeepSeek models, claimed as own (dilemmaworks.substack.com)

A now-deleted report by HonestAGI on GitHub showed a high number of similarities between Huawei’s Pangu and Qwen 2.5.

36. The era of exploration (yidingjiang.github.io)

This post explores the idea that the next breakthroughs in AI may hinge more on how we collect experience through exploration, and less on how many parameters and data points we have.

37. I'm Building LLM for Satellite Data EarthGPT.app (www.earthgpt.app)

Work with multi-spectral satellite imagery by leveraging the power of LLMs and modern remote sensing.