2025-08-22
36 篇热帖
2. AI tooling must be disclosed for contributions (github.com)
👻 Ghostty is a fast, feature-rich, and cross-platform terminal emulator that uses platform-native UI and GPU acceleration. - AI tooling must be disclosed for contributions by mitchellh · Pull Request #8289 · ghostty-org/ghostty
3. 4chan will refuse to pay daily online safety fines, lawyer tells BBC (www.bbc.co.uk)
The online message board's lawyers say UK safety laws shouldn't apply to a business based in the US.
4. What is going on right now? (catskull.net)
AI工具滥用对软件工程行业的冲击
工程现状与问题
- 工程师倦怠:高级工程师需审查和修复由AI生成、无法正常工作的“氛围编码”功能。
- 反馈失效:资深工程师的代码审查意见本应是学习机会,却被初级开发者直接用作AI生成代码的新提示词。
- 质量与理解缺失:在团队演示中,初级工程师提交的代码无法解释功能,甚至自身也不理解其逻辑,但管理层仍赞扬其使用AI工具。
- 审查效率低下:一项由多名工程师花费一个月时间审查的PR,其内容主要由AI生成,导致大量人力浪费在低质量代码上。
AI工具的负面影响
- 学习过程扭曲:初级开发者依赖AI工具直接生成代码,而非通过反馈和理解来提升技能。
- 工程文化侵蚀:工程师希望传授软件工程知识,但AI工具的滥用使这种 mentorship 变得徒劳。
- 工具价值质疑:作者通过停用AI工具(如Claude Pro)的实验发现,自主搜索和阅读文档的结论比AI输出更可靠准确。
AI行业的商业模式问题
- 资金循环:风险投资 → AI初创公司 → 向OpenAI等服务商支付费用 → 公司最终可能消亡。
- 盈利能力缺失:目前连OpenAI也未实现盈利,因技术存在固有缺陷:耗电量巨大、无法规模化满足需求、存在环境成本。
- 技术局限性:依赖摩尔定律复苏或宇宙热寂延长来解决经济可行性问题不现实,AI技术被作者视为“皇帝的新衣”。
总结
当前AI工具在软件工程中的应用引发了代码质量下降、工程师培养受阻、资源浪费等问题。同时,AI行业本身面临商业模式不可持续的挑战,技术缺陷使其难以实现真正的经济价值。作者呼吁行业正视这些现实,而非盲目追随AI热潮。
5. DeepSeek-v3.1 Release (api-docs.deepseek.com)
DeepSeek-V3.1 发布总结
核心特性
- 混合推理模式:首次引入思考模式与非思考模式,用户可通过界面“DeepThink”按钮切换,实现一个模型两种工作方式。
- 性能提升:
- 思考模式:相比DeepSeek-R1-0528,思考速度更快,答案获取时间更短。
- 代理能力:通过后期训练增强了工具使用和多步骤代理任务的处理能力。
API 更新
- 模式对应:
deepseek-chat对应非思考模式。deepseek-reasoner对应思考模式。
- 上下文长度:两种模式均支持 128K 上下文。
- 兼容性与新功能:
- 支持 Anthropic API 格式。
- 在Beta API中支持严格的函数调用。
- 提供更丰富的API资源与更流畅的体验。
工具与代理升级
- 基准测试提升:在 SWE 和 Terminal-Bench 上表现更好。
- 推理增强:针对复杂搜索任务的多步推理能力更强。
- 思考效率:思考效率有显著提升。
模型更新
- 基础模型:V3.1 Base 在 V3 基础上进行了 840B tokens 的持续预训练,用于长上下文扩展。
- 分词器:更新了分词器和聊天模板。
- 开源权重:已发布 V3.1 Base 和 V3.1 的开源权重。
定价变动
- 生效时间:新价格于 2025年9月5日 16:00 (UTC) 生效,同时结束非高峰折扣。
- 过渡期:在此之前,API沿用当前定价。
访问与资源
- 体验地址:用户可在 chat.deepseek.com 尝试。
- 文档与指南:更新细节可参考官方API文档与配置文件链接。
6. Io_uring, kTLS and Rust for zero syscall HTTPS server (blog.habets.se)
Around the turn of the century we started to get a bigger need for high capacity web servers. For example there was the C10k problem paper.
7. All managers make mistakes; good managers acknowledge and repair (terriblesoftware.org)
"There is a crack in everything. That's how the light gets in." — Leonard Cohen Let me tell you something that will happen after you become a manager: you're going to mess up. A lot. You'll give feedback that lands wrong and crushes someone's confidence. You'll make a decision that seems logical but turns out…
8. CEO pay and stock buybacks have soared at the largest low-wage corporations (ips-dc.org)
Executive Excess 2025 - CEO pay and stock buybacks have soared at the 100 largest low-wage corporations.
9. LabPlot: Free, open source and cross-platform Data Visualization and Analysis (labplot.org)
LabPlot FREE, open source and cross-platform Data Visualization and Analysis software accessible to everyone and trusted by professionals Download Watch Demo Feature Highlights Discover what makes LabPlot special High-quality data visualization and interactive plotting with few clicks Reliable and easy data analysis, statistics, regression, curve and peak fitting Intuitive and fast computing with interactive notebooks using Python, R, Julia etc. Effortless data extraction (plot digitizer) and support for real-time data
10. Control shopping cart wheels with your phone (2021) (www.begaydocrime.com)
项目概述
这是一个2021年的项目,旨在通过手机扬声器播放特定音频信号,来控制带有Gatekeeper Systems电子锁轮子的购物车(实现锁定/解锁)。作者在Twitter上的账号为@stoppingcart。
工作原理
大多数电子购物车轮通过监听地下电线发出的7.8 kHz信号来判断是否锁定或解锁。管理遥控器能够发送一个不同的、但频率同为7.8 kHz的信号来解锁车轮。
由于7.8 kHz处于可听音频范围内,该项目利用手机扬声器产生的寄生电磁场来“传输”类似的解锁代码。具体方法是通过播放一个精心设计的音频文件,让手机扬声器发出该特定频率的信号,从而模拟管理遥控器的功能。
相关资源
作者提供了其在DEFCON 29大会上关于此技术的原始演讲链接,可供进一步了解技术细节。
11. Y Combinator files brief supporting Epic Games, says store fees stifle startups (www.macrumors.com)
Startup accelerator and venture capital firm Y Combinator (YC) today filed an amicus brief supporting Epic Games in Epic's continued legal fight with Apple. Y Combinator says that Apple's "anti-steering restraints" have long inhibited the growth and development of technology companies that monetize goods and services through apps.
12. Go is still not good (blog.habets.se)
Previous posts Why Go is not my favourite language and Go programs are not portable have me critiquing Go for over a decade.
13. The Onion brought back its print edition and the gamble is paying off (www.wsj.com)
14. The contrarian physics podcast subculture (timothynguyen.org)
This is the story of how a circle of popular science communicators, who built their brands on championing free inquiry, worked to suppress scientific critique. Of how Eric Weinstein, the man who condemns the scientific community for suppressing his and his family’s work, nearly succeeded in cancelling me through intimidation and false threats. And of…
15. Everything Is Correlated (gwern.net)
Anthology of sociology, statistical, or psychological papers discussing the observation that all real-world variables have non-zero correlations and the implications for statistical theory such as ‘null hypothesis testing’.
16. Thunderbird Pro August 2025 Update (blog.thunderbird.net)
Our team is hard at work on Thunderbird Pro. Get the latest developments on Thundermail, Appointment, and Send in our latest update.
17. Building AI products in the probabilistic era (giansegato.com)
AI turns products from deterministic functions into probabilistic systems. That requires expanding old playbooks (SLOs, funnels, siloed finance), and reasoning in terms of trajectories, Minimum Viable Intelligence thresholds, and data as company operating system.
18. Being “Confidently Wrong” is holding AI back (promptql.io)
The failure mode that stalls “AI for data” efforts or "AI on my APIs" efforts isn’t psychedelic hallucination—it’s confident inaccuracy: plausible answers that are wrong in subtle and costly ways.
19. The Core of Rust (jyn.dev)
within Rust is a smaller language struggling to get out
20. It’s not wrong that "\u{1F926}\u{1F3FC}\u200D\u2642\uFE0F".length == 7 (2019) (hsivonen.fi)
21. Uv format: Code Formatting Comes to uv (experimentally) (pydevtools.com)
uv 0.8.13 adds experimental uv format command that runs Ruff's formatter, similar to how cargo fmt wraps rustfmt.
22. The issue of anti-cheat on Linux (2024) (tulach.cc)
Why are developers so hesitant to bring anti-cheat solutions to Linux?
23. Static sites with Python, uv, Caddy, and Docker (nkantar.com)
My preferred deployment stack for Python-built static sites.
24. Cloudflare incident on August 21, 2025 (blog.cloudflare.com)
On August 21, 2025, an influx of traffic directed toward clients hosted in AWS us-east-1 caused severe congestion on links between Cloudflare and us-east-1. In this post, we explain what the failure was, why it occurred, and what we’re doing to make sure this doesn’t happen again.
25. Crimes with Python's Pattern Matching (2022) (www.hillelwayne.com)
Let's make the CPython team regret adding pattern matching to Python!
26. A guide to Gen AI / LLM vibecoding for expert programmers (www.stochasticlifestyle.com)
I get it, you’re too good to vibe code. You’re a senior developer who has been doing this for 20 years and knows the system like the back of your hand. Or maybe you’re the star individual contributor who is the only person who can ever figure out how to solve the hard problems. Or maybe you’re the professor who created the entire subject of the algorithms you’re implementing. I don’t know you, but I do know that you think you’re too good to vibe code. And guess what, you’re absolutely and totally wrong. Facetious? Maybe… but I will go even further. No, you’re not too good to vibe code. In fact, you’re the only person who should be vibe coding. I would have thought this statement was crazy just a month ago because this label of “expert” coder also applies to me. ... READ MORE
27. Making LLMs Cheaper and Better via Performance-Efficiency Optimized Routing (arxiv.org)
Abstract page for arXiv paper 2508.12631: Beyond GPT-5: Making LLMs Cheaper and Better via Performance-Efficiency Optimized Routing
28. Administration will review all 55M visa holders for deportable violations (apnews.com)
The Trump administration says it’s reviewing more than 55 million foreigners who have valid U.S. visas for any violations that could lead to deportation.
29. Scientists No Longer Find X Professionally Useful, and Have Switched to Bluesky (academic.oup.com)
30. Rolling the dice with CSS random() (webkit.org)
Random functions in programming languages are amazing.
31. From GPT-4 to GPT-5: Measuring progress through MedHELM [pdf] (www.fertrevino.com)
该研究通过MedHELM基准评估套件,系统测量了GPT-5模型在医学语言理解任务上的性能进展。评估仅使用公开、可确定性评分的场景,并采用与先前GPT-4时代基线相同的固定设置,以确保纵向可比性。
主要结果:
- 优势领域:GPT-5在数值计算(MedCalc-Bench,与领先者并列)、广泛事实回忆(Medbullets,新高)和多领域推理(HeadQA,新高)方面表现出提升。
- 不足领域:在受约束的文本转SQL生成(EHRSQL,显著下降)、公平性敏感推理(RaceBias,大幅下降)和完全抑制幻觉(MedHallu,低于领先者)方面出现退步或停滞。
- 效率分析:延迟表现不一;在较长推理任务中更快(如MedCalc-Bench),但在短结构化查询中更慢,且未能带来准确率收益(如EHRSQL)。
评估方法:
- 集成采用仅追加、配置驱动的方式,不改变原有场景语义。
- 评估聚焦于公共场景,使用固定种子、温度0.0和确定性评分指标。
- 结果通过与GPT-4基线及各场景当前领先模型的比较进行量化。
关键发现: GPT-5展现了选择性的能力提升,而非全面优势。其在数值推理和广泛事实检索上的进步,可能得益于模型容量的增加和潜在推理可靠性的提高。然而,在模式约束生成、公平性保障和证据受限的质量保证方面,仍存在明确的差距,表明通用推理能力的提升并不自动转化为所有临床相关任务上的改进。结果表明,在模型性能持续提升的背景下,需要多维度的评估基准来有效监测真实的临床推理进步和风险。
未来方向: 研究指出需要扩展结构化数据任务、引入校准概率评估、发布细粒度错误分类、添加纵向漂移跟踪并加强公平性压力测试,以持续完善评估体系。
32. The unbearable slowness of AI coding (joshuavaldez.com)
AI编程令人难以忍受的缓慢:作者两个月Claude Code使用体验的转变
作者在完全使用Claude Code编程两个月后,体验从最初的极度兴奋转变为明显的沮丧。
- 初始阶段的高效与兴奋:开始时,AI辅助编程带来了速度上的巨大提升,任务完成迅速,代码提交量激增,体验令人振奋。
- 当前阶段的瓶颈与缓慢:然而,随着构建的应用程序规模增大,整体流程“慢如爬行”。尽管作者同时并行运行多个Claude Code实例来构思新功能,但真正的延迟发生在后续环节。
- 主要瓶颈:人工审核与修复:作者需要逐一拉取AI生成的PR(Pull Request)并在本地应用;必须逐步检查控制台日志;并花费大量时间指导Claude修复其自身产生的问题。这个过程非常耗时且令人恼火。
- 速度提升的错觉:尽管代码提交总量空前,但整体开发速度感却严重下降。作者形容这种初期的速度快感是“令人上瘾的”,导致对后续每个任务都抱有同等速度的期望,但现实落差巨大。
- 对AI自动化测试的怀疑:作者认为,目前仍需要人工作为Claude的QA工程师。对于能否通过类似
CLAUDE.md的配置文件让AI可靠地遵循复杂规则并执行集成测试,作者持怀疑态度,指出就连让Claude遵守基本规则都很困难。 - 现实困境与持续工作流:目前,作者的解决方案是继续手动拉取PR、添加Git钩子来强制代码质量。尽管AI加速了编码,但最终仍可能遭遇问题,例如AI编造了不存在的库功能,导致需要彻底推翻原方案(如放弃Clerk,改用GitHub OAuth)。
总结核心观点:AI编程工具在生成代码阶段能显著提速,但随着项目复杂化,人工介入进行代码审查、测试和修复AI引入问题的过程,成为了新的效率瓶颈,使得整体体验从“快”转向了另一种形式的“慢”。
33. In the long run, LLMs make us dumber (desunit.com)
The comfort we get when offloading our cognitive load to LLMs is bad for us. Cognitive load should exist, and if we reduce it too much – if we stop thinking – we can actually unlearn how to think. Kids who always choose the easy route and copy their homework from other students eventually find […]
34. Optimizing our way through Metroid (antithesis.com)
How should a fuzzer balance exploration and minimization? Samus has the answer.
35. Top Secret: Automatically filter sensitive information (thoughtbot.com)
Automatically filter sensitive information before sending it to external services or APIs, such as chatbots and LLMs.
36. Google scores six-year Meta cloud deal worth over $10B (www.cnbc.com)
Alphabet and Meta compete in online advertising, but Meta needs more data center infrastructure as it pursues artificial intelligence growth.