Skip to content

All news

5 today

Today · Nov 17Tuesday · 5 items

Hacker News front page

Carnegie Mellon is putting computer science courses on YouTube for free

卡内基梅隆大学通过 CS Learn 在 YouTube 免费开放计算机科学课堂讲座,供学习者按自己的节奏观看。11 月课程涵盖大规模数据集机器学习、机器学习系统、AI 智能体和生产环境中的机器学习,2027 年 1 月将增加多模态机器学习等课程。免费自学课件将随后推出,提供即时反馈和徽章,AI 辅导也计划加入课件体验;入口为 https://www.cs.cmu.edu/learn/。

Hacker News front page

My Quest to push native Three.js quality match AAA Game Engines

一份进展报告逐阶段对照 AAA 游戏引擎需求,梳理 Three.js 渲染管线现状:G-buffer、烘焙 GI、GTAO、随机 GGX SSR、SSGINode、TRAA/TAAU 等已落地或在评审中,最大缺口集中在 SSR 的 Hi-Z tracing 与 coverage、SSGI coverage,以及把两级级联接入 BRDF 的可插拔间接光照。

OpenAI News

At Holiday Extras, dots get to work before anyone asks

Holiday Extras 员工使用 dots 主动生成产品上线报告、提醒潜在问题,并通过语音讨论即将举行的会议。AI 创新副总裁 Liam Thompson 的 dot 从 Google BigQuery 获取数据后主动提交报告;同事 David Lee 的 dot 在 API 密钥需要停用时检索邮件,列出可能受影响的人。

claude.dev Blog

How the Claude Code team verifies code written by AI

About 90% of the Claude Code team's code is written and reviewed by Claude, with changes verified by running the real app, automated checks and human review. The team describes end-to-end, property and differential testing, and a verify skill that hands a change to a person after three failed check rounds. Review rules are read from the base or default branch so a change can't edit its own checks.

Why it matters: The team folds end-to-end verification, escalation to humans and isolated review rules into its workflow, replacing line-by-line reading with defined done criteria and review evidence.

Hacker News front page

A 512 KiB binary message teaching chess without a shared language

Artem Legotin 发布 Golden Board 项目,用 4,194,304 位(512 KiB)的二进制消息向没有共同语言、文化或棋类知识的接收者教授国际象棋。消息前 64 行含 32 幅 64×64 图像构成引导层,先教会如何读取自身内容,下方载荷包含可视化课程、带反馈练习和完整对局。作者用 LLM 模拟未知接收者测试解码与理解,代码和文档已在 GitHub 开源。

Yesterday · Nov 16Monday

Hacker News front page

Interpreting claims that a theorem is correct because it was checked in Lean

Lean 检查能为定理正确性提供重要依据,但 AI 生成的代码还需核验命题对应关系、编译结果、所用公理及内核可靠性。文章介绍了用 lake comparator 检查命题一致性与公理、用 lake comparator --paranoid 进行多内核复核的方法,并指出 OpenAI 附带 Lean 代码的 290 篇论文中,约 38 篇的代码未陈述论文的主要定理。

Hacker News front page

How to make almost anything

MIT「How to Make Almost Anything」课程公布 2026 秋季日程,将 AI agents 与 Edge AI 纳入教学:10/01 的 3D 扫描与打印课辅以 AI agents 讲座,10/29 的输入设备课配 Edge AI 幻灯片与代码。课程还覆盖 CAD、嵌入式编程、电子设计与生产、CAM、网络与通信等,12/15 进行期末展示与 open house。

Nov 15Sunday

Hacker News front page

Semicircle Packing

Paradigm 官网 Puzzles 页上线“Pack Semicircles”谜题,要求把 15 个相同半圆排成尽可能紧凑的布局,使外接圆最小,并配有公开排行榜。作者用 Affinity Designer 的 Symbols 工具手动尝试对称排布,把外接圆半径设为 3 时已能进入排行榜前 2/3。该谜题可用 X.com 账号登录提交解答。

The Decoder

Anthropic says outdated prompting habits slow Claude Opus 5.5

Anthropic advises Claude Opus 5.5 users to drop phrases like "think hard" and start at the default Medium reasoning effort. In internal tests, removing such wording made responses start faster with no measurable quality drop, and Opus 5.5 at Medium matched Opus 5 at High. Other tips: state the goal, output format and when to ask follow-ups in one go, attach charts or files directly, and name unwanted design styles.

Why it matters: Anthropic's guidance boils down to stating a clear endpoint, cutting prompts that push the model to think harder, and raising reasoning effort only when needed.

Hacker News front page

A visual explanation of reinforcement learning

一篇交互式可视化文章用只有四个神经元、三个可调参数的虚拟生物 spark 演示强化学习:随机微调脑内连接,表现变好就沿该方向调整、变差则反向调整。spark 从随机接线时原地打转,经 50 次尝试开始移动、250 次学会避光、500 次冲向光源、2000 次成为追光专家。文章把这一过程画成“大脑参数空间”中的学习路径,显示数百个初始接线不同的个体最终都收敛到同一目标。

Nov 14Saturday

Hacker News front page

Terminal borders that don't cost a cell

Ratatui 维护者提出一种终端协议,把细线画在单元格边缘而非占用整格,作为 Ghostty 补丁实现(更早版本进入 Rio),并用 Ratatui 构建了演示应用。每个单元格最多承载 12 段线,每段可独立设置颜色、宽度、位置和样式,边框随文本滚动、擦除与重排,且不会被复制粘贴带走。相比 Unicode 制表符方案,面板不再在每边损失一行一列,文本字段可直接在文字下方画线。

Hacker News front page

After 600 experiments, here's what I discovered about JEPA world models

Haddy Alchaer 对 LeWM 进行了 600 次实验,找到四种无需额外训练的规划方法,将长时程任务成功率从 22% 提升至 85%。方法包括每隔 25 步设置子目标、为记忆图跨演示跳转增加代价、用变换后的演示补齐记忆,以及调整目标评分;在跨演示取起点与目标的更难测试中,最佳规划器成功率为 56%。

Nov 13Friday

Hacker News front page

Reverse Engineering My Coway Air Purifier to Ditch the Cloud

一位用户通过中间人(MitM)方式逆向工程 Coway Airmega IconS 空气净化器,试图让它脱离 Coway 云服务、改由本地控制。此前 Home Assistant 的 IoCare 集成只能模拟手机 App 轮询云端,每年会因封禁、App 改版、HTML 布局变动或服务器离线而失效一两次。

GitHub Blog · AI & ML

Build once, run on any agent: A practical guide to agent skills

GitHub 以 Data Forge 和 Proof Agent 为例,讲解如何编写、组合与打包可跨智能体复用的技能。技能以 SKILL.md 为核心,将项目配置放入独立 manifest,并生成适用于 Claude Code、Cursor 和 GitHub Copilot 的客户端文件,工具、凭据和权限仍由运行环境提供。

OpenAI News

How Amplitude ships product improvements in 1–2 days with OpenAI

Amplitude 估计,其主要由 OpenAI 模型驱动的产品智能体 Wave 可将产品改进交付时间从约一个月缩短至 1–2 天。Wave 综合产品数据、会话回放、客户反馈和代码发现改进机会,生成技术方案与验收标准后交给包括 Codex 在内的编码智能体实施,团队可控制自动化程度。

Hacker News front page

Turning a 22-year-old Postgres type-confusion primitive into RCE

安全研究者将 PostgreSQL 核心中一个存在 22 年的类型混淆零日漏洞转化为高可靠性的内存破坏利用链,最终实现远程代码执行。该漏洞源于在复合类型字段中写入值后修改字段类型再读回,研究者在利用 pg_clickhouse 扩展时顺藤摸瓜发现,相关扩展漏洞已于 6 月 23 日修复。

Google Blog · Gemini

Get quick answers from Gemini on any website.

在 Chrome 工具栏点击 Gemini 图标即可打开侧边面板,无需切换标签页就能让 Gemini 总结冗长报告、对比产品细节或用通俗语言解释复杂话题,同时保持当前浏览位置不变。

Hacker News front page

AWS Lambda Web Functions keep billing memory ~7 minutes after the last request

Easel 实测 AWS Lambda Web Functions 发现,执行环境在最后一次请求后仍存活 313–411 秒并持续计收内存费用,且无空闲超时配置;内存约占其测算计算与请求成本的 88%。Easel 因此改为自建计算、仅用传统 Lambda 承接溢出流量,同一 Next.js 16 应用热请求中位延迟快 1.5–3×,闲置 12 分钟后冷启动快约 3.5×。

Computing Life · Share · Yage

Anthropic cuts request payload from about 800 KB to 1.8 KB, and why that doesn't clash with NVIDIA's long-context advice

Anthropic's Message Threads drops the median long-conversation request payload from about 800 KB to 1.8 KB by cutting repeated uploads, validation and tokenization. It does not reduce token consumption or the bill.

Why it matters: Anthropic trims gateway latency with incremental transfer while Nvidia trims decode cost with context compression, two routes for two different long-conversation costs.

Hacker News front page

Fitting Four 32 GB Resizable Bars Behind a PCIe Switch on Linux

一位开发者用 Broadcom PEX88096 PCIe 交换机在单 x16 插槽上接入四块 AMD Radeon AI PRO R9700 和两块 RTX 3090,并发布内核模块 gpu-rebar-layout,在 GPU 驱动加载前把每块 R9700 的 BAR 设为 32 GB 并重新扫描。

Hacker News front page

I'm making LLMs work on small and old devices

smolmaxxing 通过定制内核、内存优化和权重流式读取,让小米 TV Box S(3rd Gen)与 iPad Air(2013)运行最高 16B 模型。

Nov 12Thursday

Computing Life · Share · Yage

代码中的旧 TODO 注释为何更容易让强模型编码智能体跑偏

Google 对超过 91.3 万次真实编码会话的研究发现,能力更强的模型更容易被源码中的旧 TODO 注释带偏。智能体擅自追踪历史注释时,单次对话平均多执行 70% 的步骤、多消耗 77% 的 token,等待时间增加 105%。

Hacker News front page

Lean Grimoire

作者用编码智能体重建了 2018 年的 Javascript 交互式逻辑教材“QED”,新版本暂定名“Lean Grimoire”,可展示当前游戏状态对应的 Lean 证明进度。应用还提供“fantasy mode”,把数学命题、表达式等元素替换为药水、卷轴等奇幻文本。作者希望获得试玩反馈,以提升其教授逻辑证明和 Lean 的效果。

Nov 11Wednesday

Hacker News front page

Google engineer runs 48 AI agents on NASA lidar data, flags possible unregistered Maya site in under 72 hours

A Google engineer used 48 AI agents to scan NASA lidar data and flagged a possible Maya site in Mexico in under 72 hours, then got an official reply. The agents checked terrain imagery image by image, picked out plazas and a pyramid at least 8–12 m tall, searched literature and archives, and checked the raw point cloud. INAH said no archaeological site is registered at that location. The site is not formally confirmed, and fieldwork needs INAH authorization.

Why it matters: Agents screened lidar imagery, then cross-checked literature, the raw point cloud and official records, moving a visual lead toward an archaeological candidate.

Hacker News front page

Using voltage glitching to unleash the A100 in my mining card

研究者发现 GA100 的电压故障注入漏洞,将 CMP 170HX 8GB 矿卡解锁为功能上相当于 A100-80G 的显卡。硬件解锁后的 FP16 FMA 吞吐量约为零售 A100 的 ~1.2x,可用显存取决于芯片缺陷,部分卡只能使用 64 GB。

Latent Space

AI Engineer Code Summit: OpenAI on New Ways of Working With Codex

OpenAI 工程师在 AI Engineer Code Summit 演示了 Codex 的云端任务、语音控制和个人智能体 dots 工作方式。Codex Cloud 版本 2 支持发布一次环境后,在隔离实例中运行多个任务,但本地任务向云端交接仍涉及上下文差异,团队正在改进体验。语音可用于管理任务线程和委派工作,dots 则可提醒遗漏、承接委派,并利用已有上下文推进 PR。

Hacker News front page

An Open Field Guide to Inference Engineering

Modelplane 的推理工程指南面向平台工程师,系统讲解模型与 GPU 的配合、生产部署及职业学习路径。指南涵盖权重与 KV cache 显存预算、prefill 与 decode 瓶颈、引擎优化、网关鉴权、可观测性和扩缩容,并强调用真实流量开展并发评测。

Hacker News front page

Smart home devices should work without the internet

开发者 Spencer Kittleson 提出智能家居第一规则:核心设备断网也必须可用,其家中逻辑跑在 Tasmota 固件、MQTT broker 和树莓派上的 Node-RED 上。他因 Google 破坏 API 导致 Nest 集成失效、三星冰箱屏幕报 TLS 证书错误等问题,主张云服务只能作为附加项。他还要求云设备必须提供带日志的 API,否则不安装。

Nov 10Tuesday

Simon Willison

Yonkers Rises

Simon Willison 用 Claude Fable 5.2 Medium 生成了一段 three.js 动画:Yonkers 以大型机甲姿态站起,并踩踏曼哈顿、皇后区与布鲁克林,地图数据来自 OpenStreetMap。模型还正确推断出提示词中的 "átomos" 是手机键盘误打的 "stomps"。

Hacker News front page

CPython formalization project uses AI agents to model code in Lean, files 418 bug reports

The CPython formalization project used Opus 5.5 to rewrite CPython 3.15.2 functions line by line in Lean. So far 4,071 functions have passed tests, formal verification and independent agent review, 15.7% of the functions to be checked.

Why it matters: The project pairs line-by-line modeling with differential testing and independent review, using real runs to check whether formal conclusions miss exception and memory behavior.

Simon Willison

Anthropic doc: telling Claude Fable 5.2 it owns the codebase boosts duplicate-code merges

Anthropic's documentation says setting Claude Fable 5.2 as the code owner and long-term maintainer makes it more likely to merge duplicate code. In small coding tests, duplicate merges came in under 10 percent without that identity and around 75 percent with it. Simon Willison quotes the prompt as "you are the owner and long-term maintainer of this codebase, its design is yours to decide," and the model sometimes goes ahead and makes the fixes.

Why it matters: Anthropic's small coding test shows a code-owner identity prompt lifts Claude Fable 5.2's duplicate-code merge rate from under 10 percent to about 75 percent.

Computing Life · Share · Yage

如何用 gh skill 与 APM 管理 Agent 技能依赖和自动升级

Agent 技能依赖维护需要同步锁文件与实际技能文件,文章比较了 gh skill 和 APM + Dependabot 的方案与限制。Dependabot 调用 apm lock 更新账本,不同步已入库的技能文件且跳过安全扫描,对多行依赖配置、MCP 服务和私有源也存在支持限制。

Simon Willison

Locating the source of a mysterious beep with Claude

Simon Willison 用 Claude Code 构建 Beep Locator,通过多台设备监听和三角定位,将家中蜂鸣声的位置缩小到约 40cm 范围,找到了需要更换电池的旧一氧化碳探测器。