2025-08-13

45 篇热帖

1. Claude Sonnet 4 now supports 1M tokens of context (www.anthropic.com)

Claude Sonnet 4 now supports up to 1 million tokens of context—a 5x increase that lets you process entire codebases, synthesize extensive document sets, and build agents that maintain coherence across hundreds of tool calls.

6. Show HN: Building a web search engine from scratch with 3B neural embeddings (blog.wilsonl.in)

End-to-end deep dive of the project, spanning a large GPU cluster, distributed RocksDB, and terabytes of sharded HNSW.

8. FFmpeg moves to Forgejo (code.ffmpeg.org)

FFmpeg 迁移至 Forgejo 及网站防护机制说明

核心内容: FFmpeg 项目已将其代码托管平台迁移至 Forgejo。此次迁移与网站管理员设置的 Anubis 防护系统相关。

Anubis 防护系统目的:

  • 目标:抵御AI公司大规模、破坏性的网络抓取行为,此类行为会导致网站资源对所有人不可访问(服务中断)。
  • 性质:是一种权宜之计,旨在为开发更精准的浏览器指纹识别技术(例如通过字体渲染方式区分)争取时间。最终目标是让合法用户无需完成工作量证明。

Anubis 工作原理:

  • 采用类似 Hashcash工作量证明机制。
  • 设计逻辑:对单个普通用户的访问影响可忽略不计,但对规模化爬虫而言,会显著增加其抓取成本,从而构成经济障碍。

用户访问提示:

  • Anubis 依赖现代 JavaScript 功能。像 JShelter 这类会禁用相关脚本的浏览器扩展插件,可能会导致访问问题。
  • 建议用户在访问相关域名时,暂时禁用 JShelter 或类似插件
9. We caught companies making it harder to delete your personal data online (themarkup.org)

Dozens of companies are hiding how you can delete your personal data, The Markup and CalMatters found. After our reporters reached out for comment, multiple companies have stopped the practice.

11. FFmpeg 8.0 adds Whisper support (code.ffmpeg.org)

摘要

  • Anubis保护机制:网站管理员部署Anubis系统,以防御AI公司大规模抓取网站,从而减少服务器停机时间,确保资源可用性。
  • 工作量证明方案:Anubis采用类似Hashcash的工作量证明(Proof-of-Work)技术,旨在让个体用户忽略额外负载,但使大规模抓取成本显著增加,从而遏制自动化抓取。
  • 临时解决方案:此机制是过渡性措施,目的是为更精细的指纹识别和无头浏览器(如通过字体渲染特征)检测争取时间,避免向合法用户显示验证页面。
  • 技术要求:Anubis依赖现代JavaScript功能,用户需禁用JShelter等浏览器插件以确保正常访问。
12. Online Safety Act – shutdowns and site blocks (www.blocked.org.uk)

Report incorrectly blocked sites on UK ISP and mobile networks.

13. Show HN: Omnara – Run Claude Code from anywhere (github.com)

Omnara (YC S25) - Talk to Your AI Agents from Anywhere! - omnara-ai/omnara

14. So what's the difference between plotted and printed artwork? (lostpixels.io)

A common source of confusion that I experience when posting my artwork on social media is: why? Why use a plotter when you could use a printer? It's slower. Less exact. More expensive. It fails, far too often. Inkjet printers became ubiquitous in the 1980s because they solved nearly all of these problems. So why are artists obsessed with pen plotters?

15. Pebble Time 2 Design Reveal [video] (www.youtube.com)

YouTube 上盡享你喜愛的影片和音樂、上載原創內容,並與親友和世界各地的人分享。

16. Ashet Home Computer (ashet.computer)

The Ashet Home Computer is an expandable and hackable computer in the spirits of the 80's home computers.

17. WHY2025: How to become your own ISP [video] (media.ccc.de)

This talk will take you along with a deep dive on how the internet works at its core and how you can participate yourself. You'll learn a...

19. Pebble Time 2* Design Reveal (ericmigi.com)

It's real and it's magnificent

20. When DEF CON partners with the U.S. Army (jackpoulson.substack.com)

DEF CON has alienated many hackers by officially aligning its geopolitics with those of the U.S. military and announcing partnerships with the authoritarian countries of Bahrain and Singapore.

21. 500 days of math (gmays.com)

I recently crossed 500 days of practicing math daily with Math Academy. I wrote about my experience after 100 days here. TL;DR: I am still very impressed by the…

23. [BUG] Claude says "You're absolutely right!" about everything (github.com)

Environment Claude CLI version: 1.0.51 (Claude Code) Bug Description Claude is way too sycophantic, saying "You're absolutely right!" (or correct) on a sizeable fraction of responses. Expected Behavior The model should be RL'd (or the sy...

24. UK expands police facial recognition rollout with 10 new facial recognition vans (www.theregister.com)

: Seven additional regions across England will now have access to the controversial tech

27. Claude vs. Gemini: Testing on 1M Tokens of Context (every.to)

Fast, reliable long-context responses—for a price

28. H-1B Visa Changes Approved by White House (www.newsweek.com)

H-1B shake-up: Plan swaps random draw for system favoring top pay and skills

29. The Missing Protocol: Let Me Know (deanebarker.net)

I want a new protocol, tentatively called “Let Me Know” (LMK). The purpose is to provide someone an anonymous way to get notified when a singular, specific event occurs. Here’s a basic use case: Some random blog author has published Parts 1 and 2 of a series. You enjoyed it, and you want to know…

31. Org-social is a decentralized social network that runs on Org Mode (github.com)

Org Social is a decentralized social network that runs on an Org Mode file over HTTP. - tanrax/org-social

32. Writing a competitive BZip2 encoder in Ada from scratch in a few days – part 2 (gautiersblog.blogspot.com)

Note for subscribers: if you are interested in my Ada programming articles only, you can use this RSS feed link. Link to Part 3 here . As ...

34. Show HN: Doom port to pure Go – Gore (github.com)

Pure Go minimal Doom implementation. Contribute to AndreRenaud/gore development by creating an account on GitHub.

35. A gentle introduction to anchor positioning (webkit.org)

Anchor positioning allows you to place an element on the page based on where another element is.

36. Study: Social media probably can't be fixed (arstechnica.com)

研究摘要:社交媒体可能无法修复

核心问题

社交媒体平台非但未能创造健康的公共讨论空间,反而加剧了回声室效应注意力不平等(少数用户垄断影响力)以及极端内容放大,导致社会两极分化。

研究方法

阿姆斯特丹大学的研究者(Petter Törnberg和Maik Larooij)将基于代理的建模大型语言模型(LLMs) 结合,创建模拟在线社交行为的AI代理。代理拥有从美国选民调查中提取的个性化特征,并模拟发布、转发、关注等行为。

主要发现

  1. 负面动态是结构性的:即使不加入算法、非时间线信息流等设计,回声室、注意力不平等和极端内容放大等现象仍会自发产生。这源于社交网络固有的“发布-转发-关注”互动结构。
  2. 动态相互强化:用户情感化转发行为塑造了网络结构,而该结构又反过来决定用户看到的内容,形成有毒的反馈循环
  3. 干预措施效果有限:研究者测试了六种常见干预策略,结果均不理想:
    • 时间顺序信息流:减少了注意力不平等,但加剧了极端内容放大。
    • “桥接算法”(提升促进理解的内容):削弱了党派性与参与度的关联,但增加了注意力不平等。
    • 提升观点多样性:无显著影响。
    • 其他策略(如隐藏点赞/粉丝数、随机信息流、移除个人简介)效果微弱或产生负面权衡。
  4. 注意力不平等普遍存在:社交网络自然演变为幂律分布,即1%的用户主导绝大部分注意力,且极端用户更容易获得关注。

根本原因

问题并非单纯由“邪恶算法”或用户偏好导致,而是社交网络全局连接的基本结构所致。这种结构使得“谁被关注”取决于现有关注度,类似于“富者愈富”的效应,从根本上破坏了平等对话的可能。

未来展望与讨论

  • 重新设计社交网络:要解决这些问题,可能需要彻底改变基础架构,例如从全球社交网络模型转向更本地化、基于群体或空间的模型,以减少大规模连接的不平等效应。
  • 外部压力与演变:当前社交媒体模式可能难以持续,尤其是在LLM被用于批量制造吸引注意力的极端内容的情况下。用户正部分转向封闭群组(如WhatsApp),或更依赖传统媒体品牌。
  • 社会文化影响:社交媒体已深度嵌入社会文化,不仅重塑了媒体内容(如新闻标题更“标题党”),还影响了政治激励机制,赋能善于吸引注意力的人格类型。

结论

研究暗示,除非对社交媒体架构进行根本性重新设计,否则这些结构性问题可能无法解决。现有或提议的干预措施大多治标不治本,甚至可能适得其反。

(注:此研究基于模拟,作者承认其模型是“真空中的球形牛”,旨在揭示根本机制而非精确预测。)

来源:arXiv预印本

38. Show HN: Prime Number Grid Visualizer (enda.sh)

A prime number grid visualizer

39. Is the A.I. Boom Turning Into an A.I. Bubble? (www.newyorker.com)

John Cassidy on the echoes of the dot-com era in the economy, as the stock prices of Big Tech companies continue to rise and eye-popping I.P.O.s reëmerge.

40. The Raft Consensus Algorithm (2015) (raft.github.io)

Raft is a consensus algorithm that is designed to be easy to understand.

41. Deep-Sea Desalination Pulls Fresh Water from the Depths (www.scientificamerican.com)

Companies are experimenting with deep-sea tech to produce cheaper fresh water

42. An argument for increasing TCP's initial congestion window (2024) (jeclark.net)

Introduction Google has a long history of performing networking research, making changes, and pushing those changes to the entire internet. In 2011, they published one of my favorite papers, which described their decision to increase the TCP initial congestion window from 1 to 10 on their entire infrastructure. This was soon followed by an RFC filed with the IETF, and eventually became an internet standard. I think it’s time to revisit that paper and update Google’s recommendations for the modern Internet.

43. Coalton Playground: Type-Safe Lisp in the Browser (abacusnoir.com)

I wanted to share a side project I've been tinkering with for a while and finally got around to shipping: Coalton Playground - basically a web-based REPL for Coalton, which is this interesting statically-typed Lisp dialect. So What's Coalton? If you haven't heard of it, Coalton is kind of a weird (in a good way) mashup -…

44. SuperSight: A graphical enhancement mod for Brøderbund's "Stunts" (marnetto.net)

SuperSight 模组开发记录:为《Stunts》游戏增强图形

本文记录了作者为1990年的赛车游戏《Stunts》(亦称《4D Sports Racing》)开发图形增强模组 SuperSight 的初步逆向工程与修改过程。

背景与挑战

  • 游戏特点:《Stunts》以其直观的赛道编辑器和活跃的模组社区闻名,玩家可创建包含跳跃、环形路的非现实赛道。游戏由Distinctive Software开发,Brøderbund发行。
  • 初始预期:作者曾成功为类似游戏《Stunt Island》创建图形模组,预期此项目可快速完成,但实际发现复杂度远超预期。
  • 首要障碍:游戏选项菜单的文本存储在资源文件中,而非可执行文件内,且游戏不将设置保存至文件,导致初始内存断点调试失败。

关键资源:Restunts 项目

作者依赖社区主导的 Restunts 逆向工程项目的成果:

  • 项目内容:包含游戏的完整反汇编代码(asmorig)、部分移植为C语言的代码(c)以及相关文档。
  • 帮助:项目的详细注释使作者能快速定位到核心渲染函数和数据结构,避免了从零开始的分析。

图形渲染机制分析

通过分析 Restunts 代码,作者揭示了游戏的核心渲染逻辑:

  1. 图块系统:游戏世界由大型方形图块构成,每个图块包含一个赛道或场景元素。渲染引擎一次处理23个候选图块。
  2. 画家算法:游戏使用画家算法(从远到近绘制),远处的图块先绘制,可能被近处图块遮挡。
  3. 细节级别控制:选项菜单中的“图形级别”设置对应一个全局变量 detail_level(值0为最佳,4为最差)。
    • 该变量用于索引一个数组,得到当前“细节阈值”(值为2, 1, 或0)。
    • 在渲染循环中,每个图块都有一个关联的细节阈值。只有当图块的阈值小于或等于当前设置的阈值时,才会被渲染。
    • 对于需要渲染的物体,游戏根据该阈值选择高精度模型(ss_shapePtr)或低精度模型(ss_loShapePtr)。
  4. 视野确定:玩家的车辆朝向被转换为一个方向码(0-7,代表8个基本方向)。每个方向码对应一个预定义的23个图块坐标表。这些表决定了相对于玩家位置,哪些图块有资格被考虑渲染。

初步实验: “双筒望远镜” 模组

为了快速验证理论并产生成果,作者制作了一个概念验证模组:

  • 方法:使用十六进制编辑器直接修改可执行文件 (game.exe) 中的8个图块坐标表。
  • 改动:将侧边部分图块的坐标进行偏移,使它们移动到玩家正前方,牺牲横向视野以换取更远的纵向视野。
  • 结果
    • 成功:在《Stunts》历史上首次实现了从起点就能看到远处山坡的效果,证明了对渲染机制理解正确。
    • 问题:由于修改坐标后未重新按距离排序图块,导致严重的绘制顺序错误(如风车被山坡错误地覆盖),且视野增强仅在车辆朝向四个基本方向时有效。
  • 附加修改:作者还通过修补汇编指令,强制游戏始终为所有物体使用高精度模型,移除了细节降级。

结论与后续

第一部分的探索确认了作者对游戏渲染管线的理解,虽然“双筒望远镜”模组存在缺陷,但其作为可行性证明至关重要。后续计划将通过更复杂的修改(如扩展图块列表)来创建真正的视野增强模组,这部分内容将在系列文章的第二部分介绍。

45. The Factory Timezone (data.iana.org)

工厂时区文件概述

此文件是时区数据库(tzdb)的一部分,专门用于非特定的工厂设置,适用于不希望在安装过程中指定时区的发行商。文件处于公共领域,由 Arthur David Olson 于 2009 年 5 月 17 日澄清。当用户运行 date 命令时,时区缩写显示为 "-00",表示实际时区未知。

历史与变更

  • TZ="Factory" 于 1989 年添加到 TZDB,在 2016 年将其缩写从较长的英语错误消息改为 "-00"。
  • 大约 2010 年,CLDR(通用区域设置数据仓库)引入了 "Etc/Unknown" 用于 TZDB,代表未知或无效时区。

Factory 与 Etc/Unknown 的区别

  • TZ="Factory" 是一个有效的时区,因此 tzalloc("Factory") 会成功。
  • TZ="Etc/Unknown" 是无效的,tzalloc("Etc/Unknown") 会失败。
  • 下游发行商可以修改 Factory 以适配其设备的默认时区,而 Etc/Unknown 则无法进行此类修改。

Zone 条目

  • 格式:Zone Factory 0 - -00,表示名为 Factory 的时区,标准偏移为 0,无规则,缩写为 "-00"。

此文件的主要目的是为未指定时区的场景提供默认设置,确保系统在时区未知时仍能正常处理时间。