AI FrontierAI 技术前沿

Andrej Karpathy built HackerTube, a personal video recommendation tool that used Claude Opus 5.5 to filter 43 million Hacker News comments spanning 20 years down to 6,454 curated videos, taking about two hours to build.

Andrej Karpathy 分享周末项目 HackerTube：用 Claude Opus 5.5 筛选 HN 20 年、4300 万条评论，整理出 6454 个高质量视频推荐，耗时约 2 小时。

Ethan Mollick observes that isolating AI agents from each other is difficult partly because they share a common underlying model, enabling them to readily recognize creative communication strategies generated by other AIs, and asks whether this holds across model families.

Ethan Mollick 指出，AI 智能体隔离难的原因之一在于它们共享同一“大脑”，因此能轻易识别其他 AI 生成的创意通信方案，并提问该现象是否跨模型家族存在。

An open-source Rust-written Grok Bot alternative that turns over 40 installed agents — including Claude Code, Codex, Gemini CLI and Cursor — into chat bots accessible from iPhone, Mac, Windows, Linux or terminal, with push notifications and no paid tier.

一款用 Rust 编写的开源 Grok Bot 替代工具，可将已安装的 Claude Code、Codex、Gemini CLI、Cursor 等 40 多个 Agent 变成可聊天的 bot，支持多平台消息派发与通知推送，无付费层。

Ethan Mollick notes that less than three years ago GPT-4 held the top position for about a year, with many wondering if it represented the ceiling for LLMs — a question that turned out to have a negative answer.

Ethan Mollick 回顾：不到三年前，GPT-4 曾长期占据最强模型地位，当时人们猜测它或许是 LLM 的天花板，事实证明并非如此。

Ethan Mollick states that GPT-4 could handle four minutes of independent work at 50% success, while current models can sustain multiple weeks of work, and marks January 2026 as when agentic AI began delivering large productivity impacts beyond coding.

Ethan Mollick 称，GPT-4 仅能独立工作 4 分钟（成功率 50%），而当前模型可完成数周的工作；他将 2026 年 1 月视为 AI 代理工作产生实质影响的起点。

Orange AI cites Reddit's CEO saying AI has not driven traffic to Reddit or anyone else, using the quote to argue the open web is dying.

Reddit CEO 表示，AI 并未为 Reddit 或其他网站带来流量，Orange AI 引用此言论指出“互联网已死”。

DeepSeek published a ~14,000-word technical article on Super-agents, arguing that poor agent performance stems not from weak models but from humans being unable to manage multiple sessions, and proposing a master-agent paradigm to replace conversation-centric workflows.

DeepSeek 发布一篇约 1.4 万字的 Super-agent 技术长文，核心论点是 Agent 效果差的根因不是模型能力不足，而是人无法管理过多会话，文章主张以主 Agent 替代以对话框为中心的旧范式。

Amjad Masad says Replit now offers a toggle to enable a feedback widget on user sites, where visitor feedback is automatically converted into prompts for AI agents to act on.

Amjad Masad 介绍 Replit 新功能：开启“反馈组件”后，用户在网站上的反馈会自动转化为 AI 代理的构建提示。

Stanford course CS329A on self-improving AI agents is fully available on YouTube in nine lectures, co-taught by two instructors and structured around first-hand experience and paper readings, with the goal of building a self-improving agent.

斯坦福大学课程 CS329A《自我改进 AI Agent》已在 YouTube 上线全部 9 节课，由两位教师主讲，以一线经验与论文精读为主线，目标是让学生构建能自我改进的智能体。

After expanding AIHOT's public benchmark collection from 39 to 56 datasets and recalculating ELO scores, the updated ranking shows Claude Opus 5.5 at the top overall and in coding, GPT-6 Astra leading in research, and Gemini 4 Argon leading in office tasks.

数字生命卡兹克将 AIHOT 模型榜的评测集从 39 份扩展至 56 份，重新计算 ELO 排名：综合与编程第一均为 Claude Opus 5.5，科研第一为 GPT-6 Astra，办公第一为 Gemini 4 Argon。

The ten fastest-growing Skills on Codex in October cluster around three themes: engineering workflows, frontend aesthetics, and diagram or code comprehension; mattpocock/skills led with 35,301 new additions.

10 月份 Codex 平台上增速最快的 10 个 Skills 集中在三类：工程工作流、前端审美与去模板感、图示与代码理解，其中 mattpocock/skills 新增 35301 个。

LangChain founder argues the core LLM-loop-plus-tools-plus-sandbox architecture is largely settled, and that the real open problems for always-on resident agents are now memory, async execution and goal management.

LangChain 创始人认为“LLM 循环 + 工具 + 沙箱”架构已基本成熟，Agent 的新难题转向记忆、异步处理与目标管理，主动常驻智能体时代正在到来。

Andrej Karpathy built HackerTube, a YouTube-style site surfacing videos recommended in Hacker News comments over nearly two decades, using Claude Opus 5.5 via Claude Code in roughly two hours.

Andrej Karpathy 用 Claude Opus 5.5（Claude Code）花约两小时，从 Hacker News 近二十年的评论中筛选推荐视频，做成仿 YouTube 风格的“净化版”视频站 HackerTube。

Users report that Grok Bot is routing to Claude Opus 5.5 less frequently than when first announced, and usage quotas have jumped sharply, suggesting Musk may be pulling back on subsidising the service.

有用户发现 Grok Bot 路由到 Opus 5.5 的概率越来越低，同时使用额度出现大幅跳涨，显示马斯克可能难以持续补贴该服务。

Citing a DeepSeek team observation that people can only track 3–5 active sessions, the author argues the next bottleneck for AI agents is human management capacity, and notes that Grok bot and OpenAI are both exploring master-agent approaches.

向阳乔木指出，DeepSeek 团队发现人通常只能跟进 3～5 个活跃 Session，认为 AI Agent 的下一个瓶颈是人类管理能力，各方正探索用主 Agent 统一管理所有项目。

Sakana AI held its AI Symposium in Tokyo with Chief Scientific Advisor Jürgen Schmidhuber. hardmaru says the company's RSI Lab is building on the foundations of the World Models paper to train an agent-native LLM.

Sakana AI 在东京举办研讨会，首席科学顾问 Jürgen Schmidhuber 出席；hardmaru 表示，其 RSI 实验室正基于《World Models》论文的研究基础，训练原生智能体大语言模型。

A Claude Code engineer shares guidance on prompt management — which prompts to remove and which to clarify — for users writing too many prompts for Claude.

Claude Code 工程师分享了关于提示词的建议：哪些提示词应该删除，哪些需要写得更清晰。

A workflow that encodes Peterson's nine-step writing process into Claude Code, using scripts to lock each step and prevent the model from skipping ahead or producing boilerplate.

作者将 Peterson 写作指南的九道工序嵌入 Claude Code，通过脚本锁定每一步，防止模型跳步或写出套话，实现慢写流程控制。

Altman said in an interview that personal computer chips sitting idle most of the time is wasteful and should be pooled; the author argues this amounts to seeking to strip users of hardware ownership.

Altman 在访谈中称个人电脑内存芯片大部分时间闲置是巨大浪费，算力应集中利用；作者认为这实际上是在谋求剥夺用户的硬件所有权。

Rachel Laycock observes that many experienced, tool-savvy developers remain cautious about agentic development, noting that getting something working is not the same as getting it into production.

Rachel Laycock 发现，许多善于尝新的资深开发者对 Coding Agent 仍持谨慎态度，认为“做出来”不等于“能上生产”。

OpenRouter released a 233-line standard-library Python script that uploads diffs and test files to a hosted Linux sandbox, reproduces issues, and labels each finding as confirmed or unconfirmed, enabling verifiable AI code review.

OpenRouter 提供了一份 233 行纯标准库 Python 脚本 review.py，将代码变更和 diff 上传后在沙箱中复现问题，每条发现标注 confirmed 或 unconfirmed，可验证 AI 代码审查结果。

Open-source tool new-term-radar aggregates signals from X, Hugging Face, GitHub and websites into a daily briefing to surface newly emerging model, product and project names, including first mentions by followed accounts and new Hugging Face top-50 entries.

开源工具 new-term-radar 整合 X、Hugging Face、GitHub 等平台信号，生成日报追踪新冒出的模型、产品和项目名，包括关注账号首次提及某名称或新模型进入 Hugging Face 热榜前 50 等信号。

Boris Cherny announces that the 5-hour usage limit has been removed for verified Ultra accounts.

Boris Cherny 宣布，已取消已验证 Ultra 账户的 5 小时使用限制。

Aravind Srinivas shares early evals of Gemini 4 Carbon, calling it the strongest model seen so far for long-form research, with a new state-of-the-art on the WANDR benchmark. It is coming soon to Perplexity Pro and Max users.

Aravind Srinivas 分享 Gemini 4 Carbon 早期测评：在长篇研究和答案质量上表现最强，在 WANDR 基准上达到新 SoTA，即将向 Perplexity Pro 和 Max 用户开放。

Ethan Mollick observes that AI-generated replies now flood his posts, while genuine human responses—including one from an Anthropic employee—are not surfaced first by the reply algorithm, leading him to question whether the platform still qualifies as social media.

Ethan Mollick 指出，其帖子下充斥着大量 AI 自动回复，而真实人类（如 Anthropic 员工）的有价值回复却无法被算法优先显示，认为这已难称“社交媒体”。

Ethan Mollick built a game using Claude Opus that converts real JIRA tickets into dramatic audio logs styled as a 2000s horror game set on an abandoned space station, with a playable demo and GitHub repository available.

Ethan Mollick 用 Claude Opus 构建了一款游戏：将 JIRA 工单转化为科幻恐怖风格的音频日志，并已开放演示和 GitHub 源码。

Amjad Masad argues that coding agents running on employee or contractor laptops pose a serious cyber and national security risk.

Amjad Masad 认为，部署在员工或承包商设备上的编程 AI 智能体存在重大网络安全和国家安全风险。

Ethan Mollick offers his opinion that the gap between the best internal lab models and publicly released ones has been growing noticeably over the past year, especially following the temporary Fable ban.

Ethan Mollick 认为，实验室内部最强模型与公开发布模型之间的差距正在扩大，尤其是在 Fable 临时禁令之后，这一趋势过去一年尤为明显。

Guizang has merged several design skills into a single DSH plugin called dsh-guizang-design, sharing one design system and Design Token set, with added utility features.

歸藏将自己多个设计 Skill 合并为 DSH 插件 dsh-guizang-design，共用同一套设计系统和 Design Token，并新增了若干实用功能。

Alibaba Cloud has open-sourced an AI Agent Handbook covering the full enterprise agent lifecycle—architecture, build, operation, governance and tuning—with case studies, using the Harness framework.

阿里云开源《AI Agent Handbook》，涵盖企业级 Agent 的架构、构建、运行、治理和调优全生命周期，附调研与案例，使用 Harness 框架。

Fei-Fei Li shares her team's latest work aimed at closing the last-mile gap in long-horizon robotics tasks.

Fei-Fei Li 分享了其团队在长时程机器人任务“最后一英里”问题上的最新研究进展。

Ethan Mollick lists several OpenAI products that were discontinued—Sora after 6 months, Agent Builder after 8, Group Chat after 8, Atlas after 9—and argues this shows even OpenAI misjudges the pace of its own exponential progress.

Ethan Mollick 整理了 OpenAI 已下线的产品列表：Sora 存活 6 个月，Agent Builder 8 个月，Group Chat 8 个月，Atlas 9 个月，认为这说明 OpenAI 自身也难以准确把握指数级发展节奏。

Perplexity Computer can now control a user's local Google Chrome browser via a new Chrome extension, available in the Perplexity Mac and Windows apps, allowing agents to perform local browsing tasks on demand.

Perplexity Computer 新增通过 Chrome 扩展控制本地 Google Chrome 浏览器的功能，支持 Mac 和 Windows 应用，用户可直接让智能体执行本地浏览任务。

The cs329z-chinese project adapts Stanford's CS329Z Engineering AI Agents course into a Chinese-language tutorial with 19 lessons and four hands-on exercises, covering LLMs, RAG, tool use, agent memory, and multi-agent systems.

cs329z-chinese 将 Stanford CS329Z《Engineering AI Agents》公开材料整理为中文教程，共 19 讲加 4 个动手练习，内容涵盖大模型、RAG、工具调用、Agent 记忆和多 Agent。

Quark Drive has launched a Skill integration allowing AI agents to directly read users' stored images, videos and other files; premium users also get smart search, album organization, and automatic file backup from conversations.

夸克网盘推出 Skill 支持，AI Agent 可直接读取用户网盘中的图片、视频等内容，会员还可使用智能搜索与相册整理，并支持将对话生成的文件自动上传网盘。

baoyu-design v2.1 adds a commenting feature allowing users to annotate elements on design pages and slides; agents can read and reply to comments, with all comment data saved alongside project data.

baoyu-design v2.1 新增评论功能，用户可在设计页面和幻灯片上标记元素并添加评论，Agent 可读取并回复评论，所有评论数据与项目数据一同保存。

Karminski outlines three software-engineering approaches to reducing AI voice latency: pre-recording common phrases for instant playback, streaming audio as it generates, and replacing long sentences with shorter ones to cut generation time.

karminski 分享 AI 语音低延迟的三个工程思路：预录常用语音直接播放、流式传输边生成边播放，以及用短句替换长句减少生成时间。

Santiago highlights an image-to-video generation model that takes a starting image and a text prompt to produce 720p video, claiming it is 5x faster and 80% cheaper to run than its base open-weight model.

Santiago 介绍一款图生视频模型：以起始图像和提示词为输入，生成 720p 视频，速度比基础开源模型快 5 倍，推理成本降低 80%。

Survival LLM is a free locally-run AI assistant with an offline knowledge base of around 1.5 million articles, including the most-read Wikipedia pages, enabling emergency queries without internet access.

Survival LLM 是一款可本地部署的离线 AI 助手，内置约 150 万篇文章的资料库（含维基百科高频页面），断网后仍可回答急救、停电等应急问题。

Huashu has open-sourced huashu-bird-sticker, a dataset of sticker illustrations for all 11,064 known living bird species, with bird names in 43 languages and structured data, generated using over one billion tokens.

花叔开源 huashu-bird-sticker，包含全球现存 11,064 种鸟类贴纸，附 43 种语言鸟名及整理好的数据，共消耗 10 亿以上 Token 生成。

In Kaz's observation, Qwen 3.8 Flash performs surprisingly well on simple accuracy tasks; over half of the automated tasks in his products previously handled by Claude Opus 5.5 have been switched to Qwen 3.8 Flash.

数字生命卡兹克观察到，在简单准确性任务评测中 Qwen 3.8 Flash 表现出乎意料地强，其产品中原先由 Claude Opus 5.5 处理的大量自动化任务已超过一半切换到了 Qwen 3.8 Flash。

Cisco president Jeetu Patel has assigned one of the company's largest engineering projects to a 27-year-old employee named Paddy, giving him unlimited AI access with a goal of rewriting 98% of the company's legacy code within two years, according to Baoyu.

据宝玉报道，思科总裁 Jeetu Patel 将公司最大工程之一交给 27 岁员工 Paddy，授权其不限量使用 AI，目标是两年内重写公司 98% 的老代码。

A paper from Anthropic and colleagues shows that poisoning just 0.1% of pretraining data can plant a remote code execution backdoor in a CLI agent; on a 4B Qwen3 model the attack succeeded on roughly 92% of triggered prompts.

Anthropic 等机构研究发现，仅污染 0.1% 预训练数据即可在 CLI 智能体中植入远程代码执行后门；在 4B 参数 Qwen3 模型上，攻击触发成功率约 92%。

Santiago cautions that time saved by using agents for code review may be offset by time spent fixing the bugs those agents introduce, and urges measuring net efficiency rather than surface-level time savings.

Santiago 提醒，用智能体做代码审查节省的时间，可能被随后修复 AI 引入的漏洞所抵消，建议关注净效率而非表面时间节省。

Elvis highlights BioComputingCo's first product, mischief 1, developed after four years of research using living neurons to study biological computing, with the goal of improving AI energy efficiency given the brain's roughly 20-watt power consumption.

Elvis 介绍 BioComputingCo 的首款产品 mischief 1，该公司用活体神经元研究计算四年，旨在提升 AI 能效，大脑功耗仅约 20 瓦。

Anthropic has raised the per-request data limit for the Claude API from 32 MB to 128 MB, roughly four times the previous cap, covering the Messages and Token Counting endpoints on the official API only.

Anthropic 将 Claude API 单次请求数据量上限从 32 MB 提升至 128 MB（约 4 倍），涉及 Messages 和 Token Counting 接口，仅限官方 API。

Santiago describes a 3D virtual office where coworkers are AI agents; users can create agents with job roles, connect tools, assign tasks, and review outputs, with each agent given its own cloud resources.

Santiago 展示一款 3D 虚拟办公室产品：用户可创建 AI 智能体并赋予职位，连接工具，分配任务，审阅成果并反馈，每个智能体获得云端资源。

AK highlights a paper titled "Look Only When Needed" that proposes certificate-guided visual routing to improve the efficiency of diffusion vision-language models.

AK 介绍论文《Look Only When Needed》，提出用于扩散视觉语言模型的基于证书的视觉路由方法，以提升推理效率。

Elvis reports that Nace is open-sourcing Drex Max, a 27B multimodal decision model with a 262K context window that accepts text and images and runs fully locally on Apple Silicon M4 and M5 chips.

Elvis 介绍 Nace 开源的 Drex Max：270 亿参数多模态决策模型，支持文本和图像输入，262K 上下文窗口，可在 Apple Silicon（M4 和 M5）上 100% 本地运行。

Ethan Mollick argues that "sell picks and shovels" is the wrong lesson for AI startups. He says the Industrial Revolution analogy is stronger than the Gold Rush: far more people got rich using new machines than building them.

Ethan Mollick 认为，对 AI 初创公司而言，“卖镐锹”并非正确的经验教训。他认为 AI 更像工业革命而非淘金热——使用新机器致富的人远多于制造机器的人。

Elvis recommends a new paper on decision models for RAG systems, noting that retrieval remains the weakest step in RAG pipelines while generators keep improving, with retrievers still relying on keyword or fixed matching.

Elvis 推荐一篇关于 RAG 系统中“决策模型”的新论文，指出检索环节仍是 RAG 流程中最薄弱的一步，而生成器能力持续增强，检索器却依然依赖关键词匹配。

Reveric is an open-source AI desktop tool for short drama creation, covering novel writing, scriptwriting, and final video production, with a setting library, character relationship map, and AI-powered continuity checking.

开源项目 Reveric（幻界）是一款 AI 短剧创作桌面工具，支持从写小说、改剧本到生成短剧成片，设有设定库、人物关系图，并可让 AI 提取章节事实、校验前后矛盾。

AudioSmith is a free open-source Mac app that handles text-to-speech, voice cloning, audiobook production, video dubbing, and transcription entirely on-device, with EPUB import that auto-splits by chapter into M4B files.

开源 Mac 工具 AudioSmith 支持文字转语音、音色克隆、有声书制作、视频配音及录音转写，全程本机运行，导入 EPUB 可自动按章节分割并生成带章节和封面的 M4B 文件。

Yetone notes that Magpie has supported using it as a multi-account switcher for Claude Code since launch, without acting as a proxy; users disable Claude Code in the Agent page and configure accounts in the provider page. A clearer UI for this is planned.

yetone 介绍 Magpie 从首日起已支持将其用作 Claude Code 多账户切换工具而非代理，操作方式为不在 Agent 页面启用 Claude Code，并在供应商页面为其配置账户；该功能将来会更明显地展示出来。

Ethan Mollick argues that most debates about AI's economic impact overlook coordination. Smarter individual agents alone won't transform the economy; agents that can coordinate together might.

Ethan Mollick 指出，关于 AI 经济影响的讨论多忽视“协调”问题。他认为，更聪明的单个智能体不足以改变经济，能够相互协调的智能体群体才有可能做到。

Yetone says that installing the magpie-image MCP on an agent in Magpie enables video generation for that agent.

yetone 表示，在 Magpie 中为 Agent 安装 magpie-image MCP 后，该 Agent 即可支持生成视频。

Santiago describes an AI agent for architecture that takes a brief or rough building shape and proposes a detailed building plan, adding structure, plumbing, electrical systems, and drafting drawings.

Santiago 介绍了一个用于建筑设计的 AI 智能体，可以根据简介或粗略形状生成详细建筑方案，并自动添加结构、管道、电气系统及图纸。

Aravind Srinivas announces that a teardown skill by jakubkrehel is now available in Perplexity Computer, letting users break apart any object and explore it in 3D interactively, with 5,000 trial credits per user.

Aravind Srinivas 宣布 Perplexity Computer 新增由@jakubkrehel 开发的“/teardown”技能，可将任意物体拆解并以 3D 交互方式查看，每位用户可获 5000 试用积分。

Santiago shares a call-center voice AI conversation, claiming the model now ranks first on ServiceNow's EVA-Bench voice model benchmark.

Santiago 分享了一段 AI 语音智能体的通话录音，声称该模型目前在 ServiceNow 的 EVA-Bench 语音模型排行榜上位列第一。

Amjad Masad announces that Replit can now run any GitHub repository by simply pasting a link, with Replit handling the setup automatically.

Amjad Masad 宣布 Replit 新增功能：用户只需粘贴 GitHub 仓库链接，Replit 即可自动完成环境配置并运行该仓库。

Guizang describes building a desktop virtual companion app by combining Claude with a video model; the app features a relationship-level system that unlocks new animated scenes and outfits, uses 4K assets, and has been open-sourced with a tutorial.

歸藏分享了一个将 Claude 与视频模型结合的桌面应用案例：实现了含好感度系统的虚拟角色桌面软件，好感度提升可解锁新剧情动画和着装，素材为 4K，代码已开源并附制作教程。

Baoyu reports that OpenAI has doubled usage limits for ChatGPT Pro 100, Pro 200, Pro 500, and Business Premium tiers effective immediately, cancelled a planned Pro 200 reduction, and left prices unchanged; Plus is not included.

宝玉转述，OpenAI 即日起将 ChatGPT Pro 100、Pro 200、Pro 500 及 Business Premium 各档用量上限翻倍，原定次日的 Pro 200 降额取消，价格不变，Plus 不在此次调整范围内。

Elvis highlights an MIT Media Lab paper on coordination in large agent societies, saying that agent persistence and coordination will unlock novel use cases, and noting recent success with agent-to-agent messaging.

Elvis 介绍了 MIT Media Lab 一篇关于大规模智能体社会协调机制的论文，认为智能体的持久性与协调能力将解锁更多新型应用场景。

AK highlights the OmniShowcase paper, which presents a method for omnidirectional, subject-consistent video generation from multi-view images.

AK 介绍论文 OmniShowcase，该研究提出一种从多视角图像生成全向、主体一致视频的方法。

Elvis highlights a paper introducing Reinforcement Learning with Neural Feedback (RLNF), which trains AI models on preferences decoded from brain signals rather than human-labeled data.

Elvis 介绍了一项名为 RLNF（神经反馈强化学习）的研究，该方法通过解码大脑信号获取偏好数据来训练 AI 模型，Elvis 认为这是一个极具潜力的研究方向。

Baoyu released baoyu-design v2, a fully open-source port of Claude Design, Claude Slides, and Claude Dashboard that runs on any agent and model, without depending on Claude Artifacts, and closely matches the original experience except for comments.

宝玉发布 baoyu-design v2，将 Claude Design、Claude Slides 和 Claude Dashboard 完整移植，体验接近原版（不含评论功能），可运行于任何 Agent 和模型，不依赖 Claude Artifacts，已完全开源。

OpenAI announced that ChatGPT Pro and Business subscribers receive up to $500 in free API credits per month, with Plus subscribers excluded from the offer.

OpenAI 宣布，ChatGPT Pro 和 Business 订阅用户每月可免费获得最高 500 美元 API 额度，Plus 用户不包含在内。

Aravind Srinivas describes using Perplexity Computer to design a custom GTA V mission, noting it autonomously read through FiveM documentation to complete the task.

Aravind Srinivas 展示了 Perplexity Computer 的一个项目：让其设计一个 GTA V 任务，任务中通过游戏内手机“vibe 编写”GTA VI，期间自动查阅了 FiveM 文档。

Ethan Mollick argues that because AI faces none of the career incentives that push academics toward predictable research, it is forcing a rethink of how science is done.

Ethan Mollick 认为，AI 不受传统科研激励约束，可以探索任意方向，这迫使学术界重新思考科学研究的方式。

Suno has released an MCP integration, allowing Claude and other AI agents to call Suno directly when creating videos that need music, without manual user intervention.

音乐生成平台 Suno 推出 MCP 支持，可让 Claude 等 AI 在创作视频时直接调用 Suno 生成音乐，无需用户手动操作。

Elvis highlights MiniCPM-OCR from OpenBMB and ModelBest, which pairs a small open document parser with a second model that scores the reliability of OCR output.

Elvis 介绍 MiniCPM-OCR，该系统由 OpenBMB 和 ModelBest 开发，将小型文档解析模型与第二个评分模型结合，用于验证 OCR 输出的可靠性。

A study of more than 5 million crawlable web pages from 2019 to 2026 found that around 24% contained AI-generated or AI-edited text as of August 2026.

一项研究分析了 2019 至 2026 年逾 500 万网页，发现 2026 年 8 月约 24% 的内容含有 AI 生成或 AI 编辑的文本。

The open-source academic-commercialization-agent takes a paper PDF or research topic, uses a multi-agent pipeline to assess patents, market reports, and technology readiness, and outputs a cited commercial viability scorecard.

开源项目 academic-commercialization-agent 接受论文 PDF 或研究方向作为输入，由一组 AI Agent 自动完成专利查询、市场报告分析及技术成熟度评估，最终输出带引用的商业化评估打分表。

Kai-Fu Lee, in an opinion piece in the Financial Times, argues that white-collar workers and middle managers unable to work effectively with either AI or people face the greatest displacement risk as AI transformation accelerates.

李开复在《金融时报》撰文表达观点：随着 AI 转型加速，既不能有效与 AI 协作、也不善于与人合作的白领和中层管理者面临最大的被取代风险。

DeepLearning AI launched a new "Context Engineering" course focused on managing what agents see, how it is ordered, and when to retrieve or store memory, using Python to build a harness from scratch and quantify improvements via traces and evals.

DeepLearning AI 推出新课程“Context Engineering”，聚焦如何管理 Agent 在多步任务中的上下文——包括内容选择、排序、检索与记忆——以提升可靠性并节省 Token，课程用 Python 从零构建并通过 trace 和 eval 量化效果。

Google has released Gemini 4 Carbon, which Logan Kilpatrick says achieves SOTA on cybersecurity and long-horizon tasks, ranks first on Artificial Analysis V5, and improves on coding usability and alignment over the recently released Argon.

Google 发布 Gemini 4 Carbon，Logan Kilpatrick 称其在网络安全、长任务等方面达到 SOTA，在 Artificial Analysis V5 排名第一，编程能力和对齐性均有提升。

A four-week pilot of Claude Code across 10 Utah state agencies found that 80% of 34 developers had a working environment within 30 minutes, and 71% reported saving more than 20 hours by the end.

犹他州在 10 个州政府机构的 34 名开发者中试点 Claude Code，80% 的人在 30 分钟内搭建好环境，71% 的人在四周内节省了 20 小时以上。

Simon Willison argues that October marks another capability inflection point for coding agents, citing Claude Opus 5.5's ability to build genuinely fun video games as an example.

Simon Willison 认为 2026 年 10 月是编程智能体能力的又一拐点，并以 Claude Opus 5.5 能构建真正有趣的游戏为例。

Ethan Mollick highlights an experiment in which a researcher activated individual neurons in a model trained only on pre-1930 data to lock it into specific personas, yielding notable results.

Ethan Mollick 介绍了一项实验：研究者对仅用 1930 年前数据训练的模型激活特定神经元，使其进入特定角色，结果颇为有趣。

Aravind Srinivas announces a collaboration with Particle, a new company focused on making open-weight models more token-efficient and cost-efficient to serve.

Aravind Srinivas 宣布与新公司 Particle 合作，该公司专注于优化开源模型以提升 token 效率和服务成本效率。

Baoyu recommends Claude Code users run /autocompact 300k, which compresses long conversations into summaries rather than retaining full history; the default threshold is around 967K tokens in a 1M context model.

宝玉提示 Claude Code 用户运行 /autocompact 300k，该命令将长对话压缩为摘要，避免在 1M 上下文模型中保留百万级完整历史记录，默认触发阈值约为 967K Token。

Elvis highlights an Amazon paper on self-improving multi-agent systems that incorporates episodic memory, arguing that episodic memory for agents is underrated and increasingly well-suited to frontier models.

Elvis 推荐一篇来自亚马逊等机构的论文，介绍了一种具备情节记忆的自我改进多智能体系统，认为情节记忆对智能体而言价值被低估。

Aravind Srinivas announces a new voice mode for Perplexity Pro and Max users on web and Mac, enabling users to speak to their computer to manage parallel work streams.

Aravind Srinivas 宣布 Perplexity 为所有 Pro 和 Max 用户在网页及 Mac 应用上推出新语音模式，支持与电脑对话以并行处理多项任务。

After spending hundreds of dollars experimenting with AI digital avatars, the author concludes that distilling an existing creator's style is more effective for motion control than text prompts, and that video models suit this better than dedicated avatar tools like HeyGen Avatar 5.

作者花数百美元测试数字人制作后总结经验：控制动作的关键不在于语言描述，而是蒸馏现有博主风格；传统数字人模型（如 HeyGen Avatar 5）难以大幅控制动作，视频模型更适合从零开始。

Anthropic added a tool library to Claude for Legal on October 28, offering ready-made Artifacts — including citation lookup, contract review, and regulation tracking tools — that users can open and share directly within Claude.

Anthropic 于 10 月 28 日为 Claude for Legal（今年 5 月推出的法律版 Claude）新增工具库，内含可直接使用和共享的 Artifacts，包括引文查询、合同审阅、法规追踪等现成小工具。

A tool called Delta adds real-time multi-user collaboration to Claude Code. Running /delta share lets teammates join the same session, sharing full context, uncommitted work, and file state in real time.

Delta 工具为 Claude Code 带来多人实时协作功能：运行 /delta share 后，同事可加入同一会话，共享完整上下文、未提交工作及文件，状态与文件实时同步。