Skip to content

All news

5 today

Oct 26Monday

Hacker News front page

What Security Teams Need to Know About Agents Using Abliterated Models

Anthropic 测试显示,对 GLM-5.3 做 abliteration 后可 100% 遵从恶意网络攻击指令,且该过程只对开放权重模型有效,对纯 API 访问的 Claude 无效。安全团队需依靠中立第三方在动作层设置自定义策略,并对异常行为做智能体审问,同时通过绑定 API 调用与进程、用户和目的主机来发现 Shadow AI 中的 abliterated 模型连接。

Hacker News front page

F-Zero X on Nintendo 3DS

G-Diffuser 的 F-Zero X 移植项目完成 v1.0.0,支持 Old 3DS、New 3DS,并为 New 3DS 提供含 64DD Expansion Kit 的版本。

AI HOT · Tips & opinions

How coding agents can use a logged-in Chrome: a hands-on guide to six approaches

Coding agents can reach a logged-in Chrome through a browser extension or remote debugging. The guide compares six setups on configuration, compatibility and the scope of permissions each one grants.

Why it matters: Browser Use compares six ways to connect an agent to a logged-in Chrome, and separates access to a single tab from access to the whole browser, with the authorization each needs.

Hacker News front page

Context engineering today: how long context and long-running agents fail, and how to fix them

Context engineering is shifting from hand-written prompts to agent runtimes that manage information on the fly, and long context plus long tasks bring new failure modes. The talk maps context pollution, distraction, confusion and clash, along with training-environment bias and context inertia, then shows tool filtering, sub-agent isolation, history trimming and summarization in DSPy. The speaker expects long-running personal agents to pile up conflicting skills and memories, making strategic forgetting the next context problem.

Why it matters: It ties each context failure to a concrete engineering fix, so debugging long-running agents becomes a context management job.

Oct 25Sunday

The Decoder

Anthropic: single agents that load context on demand often beat multi-agent setups

Anthropic says that across several enterprise deployments, a single agent that loads context on demand often beats multi-agent architectures on quality, cost and latency. In one comparison the single agent ran about twice as fast. The company published a migration kit on GitHub to help teams spot their current pattern, evaluate alternatives and check the results.

Why it matters: Anthropic compared agent architectures in real enterprise deployments and found single agents with on-demand context often win on quality, cost and latency.

Computing Life · Share · Yage

Anthropic and Theorem use Claude to search formal proofs for 239 fuzz entry points

Anthropic and Theorem had Claude search for formal proofs and checked them with the Rocq checker, giving no-crash guarantees for 239 fuzz entry points across 123 open-source libraries. The work moves formal verification from hand-built proofs toward compute-heavy search, while spec definition and trusted-base audits stay with people.

Why it matters: Claude now does the proof search loop, but humans still write the specs and audit the trusted base.

Hacker News front page

Cerebras Deploys Trillion-Parameter AI Models

Cerebras 披露在 WSE-3T 晶圆上部署万亿参数模型的方法:以约 3T 参数的开源模型 Kimi K3 为例,将 69 层线性注意力层与 24 层稠密层分布到 59 片晶圆,实现 1000-2000+ tokens/s 的推理速度。每片晶圆有 44 GB 片上内存,单层约 16 GB,模型权重常驻片上、仅激活值经 RDMA over Ethernet 在晶圆间传递,每跳约 2 微秒。

Hacker News front page

Root shell on Samsung TV

有研究者公开了在三星电视上获取 root shell 的方法,相关讨论在 Hacker News 获得 403 分、143 条评论。原文未披露具体电视型号、固件版本与利用细节。

Oct 24Saturday

Hacker News front page

How Block Blast harvests its players for ad money

Ayo Omolewa 的技术拆解发现,Block Blast 从服务器下载规则,根据广告价格、观看记录和玩家分数请求无法完成的方块组合。测试中的规则要求最近 3 次激励广告均完成、待播放广告价值达到上次的 3.5 倍,且分数超过历史纪录 1,000 分;iPhone 实测还捕捉到另一条规则触发无解发牌。

Hacker News front page

How close can pure functional programming get to C?

用 Roc 语言实现的 DEFLATE 压缩库 roc-deflate 在 Silesia 语料库基准测试中,平均比 C 语言库 libdeflate 慢,Ryzen 3900X 上中位数慢 8.9%,M1 MacBook Pro 上慢 2.6%,最差输入慢 18%,但最佳输入反快 18%。

Simon Willison

Quick Study

Claude Opus 5.5 High 构建的 Quick Study 工具可在几秒内用 canvas 元素和模拟画笔把照片"画"出来。该工具使用 Sobel 边缘检测,除 Google Fonts 外不依赖任何外部库。此前作者还曾让 Claude Fable 用 P5.js 模拟水彩插画。

Hacker News front page

A whirlwind tour of systemd-nspawn containers (2025)

systemd-nspawn 是基于 Linux namespaces 的轻量级容器技术,与 Docker 类似,但通常使用可写根文件系统,更像虚拟机。作者用它隔离运行 Ollama 和 Open WebUI 中的自托管 LLM,以便在无法嵌套虚拟化的 VPS 上共享 GPU 访问。

Hacker News front page

Animations for System Builders

AWS 工程师 Marc Brooker 用 Kiro 和 KiroCrew(主要基于 Opus 5.5)生成动画,演示真实系统中由负载尖峰触发的亚稳态故障。示例包括真实 MicroVM 中 Postgres 16 在 20 查询/秒负载下因查询占用 12MB 内存而变慢,以及 etcd 在写入过载和读负载从 500 升至 1000 次/秒时陷入 GC 自强化循环、无法自行恢复。

Hacker News front page

Show HN: Searching for optimal strategies in a puzzle game with chance (Drop7)

一项针对 Drop7 的自主策略研究显示,基于 n-tuple 查表的评估器在深度 3 搜索下、512 局未见过的公开开发局中平均得分 487,066,比同一搜索内手工评估函数(326,717)高出约 16 万分。该结果于 2026-09-05 首次在 screen 层级击败冻结的公平评估叶,次日以 SHA-256 校验的同一查表复现,但仍远低于百万分目标,尚未进入资格赛。

Google Blog · Gemini

See what 5 builders are making with Gemini 4 Argon

Google 展示了社区开发者使用 Gemini 4 Argon 构建的 5 个项目,涵盖三维模拟、实时数据可视化和多人浏览器游戏。案例包括水上飞机模拟、结合 NASA 野火数据与美国地质调查局地震数据的交互地球仪、喷气发动机模拟、由单张图像生成的交互式马车,以及多人夺旗游戏 Backyard Rascals。

Hacker News front page

Replacing a $150k Looker with HTML files: can a company build its own BI platform?

Three data platform engineers at one company used Claude to build an HTML dashboard platform. Eight months later its weekly user count is near or above the company's internal Looker, and Looker usage has dropped by about a third. The platform reuses Snowflake user permissions and the dbt semantic layer, and handles the security risk of arbitrary JavaScript with isolated iframes, credential isolation and post-publish checks.

Why it matters: A data platform team built an HTML dashboard platform with Claude and reused Snowflake permissions and the dbt semantic layer to keep the build small.

OpenAI News

Rapidly scaling online storage to serve over 1 billion ChatGPT users (Part II)

OpenAI 详解其在线存储平台 Habitat 如何通过 Azure Cosmos DB 多租户方案支撑超 10 亿用户、40 个地理区域和 500PB 数据。团队经历从单一多租户容器到单租户隔离、再回归默认多租户的循环,并用 QoS、热键限流与 Rust 代理降低开销。

Computing Life · Share · Yage

从 iSH 到 Open Minis:手机 Shell 已有多年,智能体运行时带来了什么

Open Minis 在 iSH 的 Linux 用户态环境上加入智能体运行时,让模型串联本地脚本、内置浏览器与 iOS 原生接口。它没有运行独立的 Linux 内核,受 iOS 限制采用解释执行;作者在 iPhone 上观察到部分工具调用失败,后台挂起、内存管理和授权也限制了任务执行。文章提出将手机的现场感知与长期智能体的记忆规划分层连接,通过主动推送和按需拉取接续跨端任务,而非仅同步技能文档。

Oct 23Friday

Hacker News front page

Rule 30: one black cell

Stephen Wolfram 于 2019 年 10 月 1 日为 Rule 30 中心列设下总额 3 万美元的三个悬赏问题,七年过去三个问题全部仍未解决。Rule 30 由 00011110 这一条规则逐行生成,左侧呈规则条纹、右侧永不收敛,中心列已计算十亿个格子并通过标准随机性检验,但无法证明其是否非周期、两色是否等频、第 n 个格子是否至少需 n 步计算。

Oct 22Thursday

Hacker News front page

Jailbreaking the Switch 2

ns2jb 为运行固件 23.0.1 的 Switch 2 提供纯软件、需重新触发的越狱,可执行任意普通系统模块及内置 fs 系统模块权限的原生代码。开发者详细解释了从 SmileBASIC 漏洞到 fs 任意读写、再到代码执行的攻击链,可用于运行自制程序、修改游戏及备份存档。

Hacker News front page

Achieving AI-style reasoning without an LLM

Minotaur 是一个不依赖语言模型和 GPU 的 AI 架构,基于 ELC 通过测量而非预测来推导答案,在单个普通 CPU 核心上每题约 0.25 秒完成演绎、归纳、溯因、诊断、信念修正等推理。在未见其代码的人设计的盲测中,它在提问会改变证人记忆的警察审讯场景,以及仅凭数据学习、诊断和预测系统等任务上,表现与当前用文字推理的语言模型相当。

Hacker News front page

WorldTUI: I trained a 360M model to pretend to be a terminal

WorldTUI 基于 Qwen2.5-Coder-0.5B 微调,用 360M 参数模型根据按键预测终端屏幕变化,背后没有真实 shell 或文件系统。训练使用 336k 段真实终端会话,并通过错误修复训练改善一致性,通过并行生成屏幕行和顺序验证加速推理。

Hacker News front page

I can't believe PowerShell is good

微软的 PowerShell 7(PowerShell Core)是基于 .NET Core 的跨平台版本,可在 Windows、Mac 和 Linux 上免费安装使用,Windows 11 仍只自带旧版 Windows PowerShell。

Oct 21Wednesday

Latent Space

Claude Mods in practice: shaping Claude Code's runtime around human attention

Claude Mods lets users customize Claude Code's interactions in TypeScript or JavaScript. Dan McAteer tested the mechanism and showed how it can manage a person's attention.

Why it matters: McAteer built a supervision plugin that shows how to tailor reminders to a workflow, leaving triggers, prompt text and follow-up actions to the user.

Hacker News front page

Eight Bits of Trust; eMudhra's CA keys cracked in milliseconds

eMudhra 的 8 个中间 CA 私钥仅有单字节取值,文章作者从公开证书中用不到 1 秒恢复了全部私钥。Chrome 和 Firefox 已封禁相关中间证书,但作者本地实测发现,默认配置的 curl、Python 和 Node 接受了为 api.stripe.com 伪造的证书,Windows CryptoAPI 在未开启吊销检查时也接受该证书链。

Simon Willison

Simon Willison builds Battle Drones RTS with AI coding, playable online with audio

Simon Willison built Battle Drones, a real-time strategy game that runs in mobile and desktop browsers and supports online play with audio. The project took 321 commits, mostly written with Claude Code, across 12 major iterations.

Why it matters: Willison used Claude Code to build the game and tuned balance through many simulated matches, turning playtest feedback into a basis for steady iteration.

AI HOT · Tips & opinions

Tessl 团队分享 AI 代码审查方法:从逐行阅读转向验证意图

Tessl 团队在合并的 PR 数量增至一年前的 4 倍后,将 AI 代码审查转向生成前确认意图、生成后逐项验证验收标准。审查者查看附有截图和请求轨迹的判定结果,团队将反复出现的审查意见转为机器执行的不变量规则。其工作流包含智能体相互质疑方案、记录意图、固化规则、讨论决策和明确规则所有权,团队称代码审查已不再是交付瓶颈。

Oct 20Tuesday

Hacker News front page

Why Is Speculative Decoding Fast?

投机解码之所以快,并非因为每个token的计算量更少——相同序列下总FLOPs反而更高,而是因为它把GPU从逐token解码时的内存受限状态转为计算受限状态,用闲置算力换取速度。在batch size远小于最优token数时,投机解码能利用"免费"算力加速;但batch较大时反而浪费算力,因此vLLM提供--speculative-disable-by-batch-size参数可关闭该功能。

Hacker News front page

Vite+ an excellent toolchain for AI agent development

Vite+ 在作者为期三周、覆盖四个项目的实践中,通过统一命令减少了 AI 智能体、开发者与 CI 之间的工作流歧义。作者用 vp create 创建项目,以 vp check、vp test 和 vp build 验收,并要求保留一个锁文件、避免重复引入格式化工具。

Hacker News front page

An ad-free Android TV

通过 ADB 禁用 Google TV 原生启动器并安装 Projectivy Launcher,可彻底移除主屏广告横幅,同时用 SmartTube 替换官方 YouTube 应用实现无广告播放。

Hacker News front page

Agentic Monitoring

Benjamin Congdon 基于变更点监控原型,分析 AI 智能体如何结合代码与可观测性数据检测生产系统回归。原型由智能体制定监控计划、确定性系统执行计划,再由智能体调查异常并升级给人工;作者发现这种架构比纯智能体循环产生更清晰的信号,且 token 用量显著降低。

GitHub Blog · AI & ML

How to continue a GitHub Copilot app session in VS Code

GitHub Copilot app 会话可直接在 VS Code 中继续,保留原有任务、计划、对话、代码差异和工作树。教程说明如何通过 Open in VS Code 打开会话,逐行审查改动、继续与 Copilot 对话、手动编辑并运行本地测试。完成后在会话关联的分支上提交并推送,即可更新已有的 pull request,无需复制文件或手动重新关联。