AI圈报 · RSS 订阅

Rohan Paul 转发陶哲轩的视频观点：数学家对 AI 做数学的不满与证明是否正确无关。陶哲轩称围绕 Navier-Stokes 的活动目前陷入混乱，最终会恢复并理解该证明，但这种做法非常不理想。他把这种模式称为对问题的不负责任的自动化式露天开采，即达成眼前目标却损害领域的未来发展。完整视频在 New Scientist YouTube 频道。

I spent an afternoon convinced I had found a pricing bug. MiniMax H3 has six output tiers on...

A reader named arhancanli caught a problem in our last experiment: we had used overlapping Wilson...

Warning research shows people stop reading what they see often. Approving every agent command trains that reflex. Fence the agent so the rare prompt gets read.

PixVerse 分享了一段 15 秒、16:9 的写实航海短片提示词，按 2.39:1 宽银幕构图、24fps 自然运动，共四个镜头。故事设定在中世纪北欧海域，一位 74 岁老人独自划 4 米木船驶入浓雾，逐渐发现一艘三桅黑帆海盗船，重点表现雾中逐层显现与船、人、小舟之间的巨大尺度反差。提示词还固定了人物外貌、服装、船只结构与风向光线等连续性锚点，要求呈现实拍质感而非游戏过场或奇幻概念图。

PixVerse 上的 Kling 4.0 完美驾驭了这一场景的宏大尺度 一艘 140 米长的黑帆船缓缓从雾中浮现。小小的划艇让它显得更加庞大，整个片段感觉就像在开阔水域实拍的一样 完整提示词见下

Every TypeScript team has a rule about @ts-ignore. Ours does too, and it works: zero occurrences...

A reader found a bug in the script and asked whether an FP8 KV cache halves the long-context penalty. Computed: capacity is exactly 2x at every context, speed is 1% at 4k and 21% at 128k, and the crossover is where the cache read equals the weight read.

A developer and father reflects on AI, independent thinking, books, and the example we set for our children.

OpenAI 对 ChatGPT 付费订阅新增支付卡片和设备数量限制，同一张卡或设备绑定过多订阅后将无法开新订阅，触顶时显示 Device limit reached。

Claude Code 2.1.309 stopped shell startup files running outside the Bash sandbox. Why zsh and bash run them at all, and a one-line trace to see it.

Google 论文介绍部署在内部 CI 中的 FlowAgent，它捕获测试失败后由 Agent 编辑、构建并重跑测试直至通过，验证过的修复会以根因说明形式进入代码评审。专家判定其在 195 个真实失败中 67.18% 的修复正确，开发者已在生产中应用 28,554 个修复；修复中位耗时 9.85 分钟，而开发者开始修改代码前的中位窗口为 27.48 分钟，速度仍是决定采用与否的关键。

Google 联合斯坦福、牛津等团队推出 StreamVibe 视频智能体，能区分用户明说、未说出口和任务中途改变的需求，自主决定何时出声帮忙。其 StreamVibe-VL 模型基于 Qwen3.5-9B，响应时机与质量比其他模型提升 10.7%，记忆系统再提升主动交互 12.3%。测试显示视频超过 5 分钟后各模型普遍难以把握正确时机，基于自身历史回答的关键内容指标最多下降 61.5%。

Every Monday, ExtraTime, my football quiz site, gets a health report written by a model. It reads the...

A practical playbook for Claude Code, Codex, Cursor and Copilot: manage context, cut token spend, and get code you can actually ship.

长鑫存储（CXMT）已向国内芯片设计公司送出 HBM3E 样品，计划两个月内有限量产，但良率仅 20–30%。宇树第三季度营收增长 69.9% 至 6.65 亿元，净利润下降 17.4%，研发支出增长约 250%。月之暗面发布 Kimi K3.1 Pro，支持 100 万 token 上下文窗口，完整权重承诺 11 月 4 日开放。

Five coding agents streaming into five panes all share one main thread. How xterm.js draws text, DOM vs WebGL, and a benchmark you can run.

There are a lot of ways to let a coding agent find code now. Plain grep, a language server behind...

赛迪研究院昨日发布《新域新质创新发展十大趋势》显示，我国人工智能推理数据量首次超过训练数据量，标志从模型训练研发迈入规模化产业应用阶段。截至 2026 年 6 月日均词元调用量超 175 万亿，较 2024 年年初增长 1000 多倍；国产算力和大模型已在钢铁、汽车等 20 个行业落地 1000 多个应用场景，并累计建成 500 多家卓越级智能工厂。

For about two months, the first thing I did every morning was refresh an inbox, looking for a word...

Google, OpenAI and Microsoft are moving toward a common way for website owners to express whether...

A README file in your ~/Downloads folder should not be able to expose your SSH keys. A calendar...

据 OpenAI 更新的支持页，12 月起 Plus、Pro 和 Business 计划将在 ChatGPT 与 Codex 之间共享单一 token 预算，Work 的智能体功能并入主界面。

Rohan Paul 转发 David Sacks 在 All-In Podcast 中的说法，Anthropic 8 月发布博客披露，在构建下一代模型的相关任务中已有 26% 由 RSI（递归自我改进）完成。Sacks 补充称这仅是 8 月数据，10 月的进展尚不得而知。

Green tests from an agent prove the tests run. Mutate only the files its branch changed with StrykerJS and read the survivors as untested lines.

Nace AI 以 Nace.AI Open RAIL-M 许可开源其决策模型 Drex Max v1，权重含 bf16、Q8_0 GGUF 与 affine Q4 检查点，并提供 llama.cpp 和 Ollama fork。

DeepSeek open-sourced Super-agent, a GitHub coding agent. Unasked work defaults to observe mode, and its workers run with full access as your OS user.

AIHOT 模型榜将抓取的公开评测集从 39 份增加到 56 份，按原有算法对所有模型重新进行 ELO 排名，并剔除未正式上线的模型。综合与编程第一为 Claude Opus 5.5，科研与知识第一为 GPT-6 Astra，办公第一为 Gemini 4 Argon。

DeepSeek Harness 团队以 MIT 许可开源编码智能体原型 dsh-superagent v0.1.0，面向单个 GitHub 仓库，可监听 GitHub 和可选的 Feishu 请求并生成并行工作的 Ticket。

Greenhouse 对英国、爱尔兰 1100 多名求职者的调查显示，81% 求职者在仅递交简历、未与人事交流的情况下被拒，42% 候选人过去 12 个月参加过 AI 面试，其中 57% 认为这类面试有机会展示技能。

生物科技初创公司 Exin Therapeutics 声称已让 AI 智能体对小鼠进行脑部手术治疗癫痫，CEO Gabriel Ocana-Santero 在 X 上分享了过程。

DepForecast: A Weather Forecast for the Ghosts in Your node_modules (work in...

This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 4, "Ghost in the...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine ...

AI coding agents can write code, modify files, and report that a task is complete. But a...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the...

Real ghost reports printed in American newspapers between 1846 and 1933, each shown with its original scan and the exact words behind every fact. Never a retelling.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine ...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine ...

SpaceXAI 宣布自 11 月 7 日起为 Grok 模型生成的文本添加隐形水印，以响应欧盟《人工智能法案》要求。

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine On...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the...

Description: On-device AI, WASM threads, target-size compression, inpainting and 3D LUTs in...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine ...

Flash for Apple Watch lets you chat with your AI, approve its actions, and manage background agents from your wrist. And it works away from home too.

Welcome to the POT 👻 ⚠️ A small note before trying the live demo I'm hosting...

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine. ...

A TypeScript pattern for MCP servers: preview, human approval via elicitation, token fallback, and tests that try to break it.

Claude Code 2.1.307 added claude_code.safety_event. 2.1.312 stopped sending it. Alerts built only on vendor telemetry break quietly. Git and hooks don't.