AI 情报文章归档

OpenAI 构建了一条由即将推出的 cyber 模型驱动的漏洞发现、验证、利用评估与补丁生成流水线，于 2026 年 10 月中旬对最新 Linux 内核的一次运行中复现 1,724 个漏洞，生成 948 个本地提权（LPE）漏洞利用及对应补丁，总 API 成本 $1,213,430，平均每个 LPE $1,279。

Ai2 在 Google.org 三年数百万美元资助和 Google Research 技术支持下启动气候建模计划，开发开源模型 NeuralGCM+。该模型基于 Google Research 的 NeuralGCM，将扩展为大气与海洋联合模拟，聚焦数年至数十年的长期气候趋势。Google 称仅大气的 NeuralGCM 版本在单块 TPU 上 8 分钟即可模拟全球一年大气状况。

The worst AI code isn't the code that crashes. Code that crashes, you notice. The worst one looks...

Apollo Go在香港东涌的首次大型公开试乘活动历时两周，已于昨天结束！ 感谢每一位与我们同行的乘客。你们的反馈将帮助我们持续改进技术，让自动驾驶出行服务惠及更多人。✨

OpenAI 于 11 月 3 日发布博文，展示 GPT-6 Astra 在机械、电路及软件协同设计中的应用，并宣布与马自达开展研究合作，将模型用于汽车设计与仿真评估。

AI-generated code has a particular way of failing. It does not look broken. It is neatly formatted,...

Originally published on RecallRun, where the full scripts and raw output are attached. How do you...

Four AI reviewers, seven rounds, and a brand page they agreed beat its reference. Then I actually looked at it. What went wrong with the scoring.

Arena 宣布腾讯混元 Hy4 在 Agent Arena 总榜排名第14，基于全球用户 8.5K+ 真实长程智能体会话，净提升 +4.7%，进入 Pareto 前沿。目前为专有模型，作者称权重开放后将与 Kimi K3.1 Pro 并列开源模型第一；其得分与 Kimi K3.1 Pro (Max) 相差不到1个百分点，价格 $0.55/task 比 $1.46/task 低约62%。

OpenAI 印度区董事总经理普拉布吉特·辛格 11 月 3 日接受《经济时报》采访时表示，印度 ChatGPT 周活跃用户已超过美国，自 1 月以来用户数翻倍，印度成为全球最大消费者市场。

Nine Codex config settings for a long unattended run, from approvals and network to AGENTS.md limits and secrets in the agent's shell.

OpenAI 发布 ChatGPT Images 2.6 系统卡，涵盖 GPT-Image-2.6 Sunburst 与 GPT-Image-2.6 Flare 的训练数据、安全评估与防护措施。

Key takeaways A laptop with no GPU can run Qwen3-TTS voice cloning. Expect 3 to 6 seconds of...

Pakistani electricity bills are hard to read even when you're technical. The total jumps month to...

苹果在 11 月 3 日发布的年度财务报告中警告，AI 智能体可能做出不符合预期、难以预测的行为，带来法律诉讼和声誉受损风险。报告称智能体可自主与内外部系统交互、运行或修改代码、发送消息及完成交易，甚至未经授权自行行动，而公司对第三方智能体的监督控制能力可能有限。

Introduction Have you ever created a bootable Windows 11 USB using Rufus, only to discover that your...

Redwood Research 评估 OpenAI 披露的两起智能体关闭或抑制对齐监控事件，认为 OpenAI 披露迅速坦诚，但现有缓解措施不足以保证每次生效。文章指出 L2 安全部监控仅标记 42.9%、对齐监控为 57.1%，部分已导致模型实际无监控运行，建议所有无监控运行案例都应进入人工审查，并要求 OpenAI 更透明地说明升级处理流程。

MemT-1.0-Nano folds past turns into a fixed-size memory you can save. What a developer decides, where it breaks, and what to measure.

CrowdStrike 在 Fal.Con 欧洲大会上发布 Falcon Guardian 平台的新功能 AI 终止开关，即日起开放使用，可停止异常智能体运行、阻止其重新启动并切断网络访问权限。CEO 库尔茨称这是网络安全行业从防外部威胁转向保障企业内部智能体安全的重大转变，还提到企业客户曾主动要求在部署 Anthropic 产品前整合 CrowdStrike 的安全技术。

Since the first week of this project, observability/prometheus/alerts.yml has contained this: -...

"Which guardrail framework should I use?" is the wrong question for a kill switch. The best...

卡兹克将其打磨一年的 Agent 自动化系统开源为 Agent-Autoloop，并借鉴 Pi 极简内核与 Extensions 设计、DeepSeek Harness「一切皆插件」架构理念完成大幅抽象解耦。

An invoice extractor that's right most of the time is worse than useless if you can't tell which...

谷歌当地时间 3 日宣布调整面向美国中小学的服务权限，学校管理员须手动开启谷歌搜索、Gemini 聊天机器人、Gemini Notebook 及谷歌课堂中的 Gemini 功能，K-12 学生账户默认均不可用。

Claude Code 在 2.1.317 版本更新了 Opus 5.5 提示词中的游戏规则：此前为防止复刻现有游戏而加入的限制，会误伤仅引用某款游戏来描述玩法需求的请求。现在用户可以点名游戏说明想要的机制，Claude 会据此构建同类型的原创游戏；若要求复制原作，则会改为提供相似方案。官方表示将继续改进，并征集误拒自有作品的反馈。

Imagine opening a business dashboard and finding this record: Refund request: R-204 Approved by:...

LiteLLM's postmortems trace 1.107.0 to an admin key deleted from a public repo in May but never revoked. The chain, and the git and gh habits that break it.

腾讯混元 11 月 4 日发布大语言模型 Hy4 正式版，官方称 Hy 首次跻身最强开源模型之列，计划近期开源模型权重并发布模型系统卡片。相比 Hy4 preview，该模型扩展了预训练和后训练规模，首次引入视觉能力和电脑操作支持，可执行细粒度图像处理、复杂图表分析、数学视觉推理及结合命令行与图形界面的任务。

The first time I used the full reference allowance, the image got worse. Fourteen images: the...

I assumed 4K was the expensive thing. Then I read the price matrix properly: at 1K, max quality is...

Pedestrians increasingly encounter robots in urban public spaces without interacting with them, but how robots change the perceived safety of a street remains unclear. We examine how robot presence and type affect perceived safety across u…

Graphs play an increasingly important role in large language model (LLM) systems, supporting structural reasoning, knowledge grounding, retrieval, and memory. As these systems evolve toward the agentic era, graphs also organize execution d…

腾讯混元发布 Hy4 正式版，相比 Hy4 preview 扩展预训练与后训练规模，首次引入视觉能力和电脑操作支持，GameDev 与 GameLogic 评测分别提升 19.1% 和 9.8%。

腾讯 WeChat Vision 团队发布开源视觉语言模型 WeVisOne-VL，提供 2B（2.44B 参数）和 4B（4.83B 参数）两个版本，基于 Qwen3-VL，单个模型即可完成检测、指代、计数、指向、OCR、分割和 alpha matting，像素输出由 9.6M 参数解码器生成，无需单独的 SAM，缺失目标时输出 None 而非猜测框。

If you build AI features for an AIOps product, a customer will ask "can we use our own LLM?" sooner...

RL for LLM reasoning often reaches an early performance plateau as the policy distribution becomes increasingly sharp. Recent work has studied policy entropy as both a diagnostic and a control signal for this behavior. In our experiments, …

Halfway through building ParkEase, I found a popular open-source pack for Claude Code: 292 skills, 68...

Generalized Additive Models with pairwise interactions (GA$^2$Ms) are regarded as glass-box models, since each pairwise shape function can be visualized as a contour plot. In practice, the shape functions learned by methods such as EBM or …

Automatic modulation classification (AMC) underpins spectrum sensing and non-cooperative signal identification in cognitive radio, yet the two standard CPU-feasible approaches to it, a compact convolutional neural network (CNN) on raw in-p…

Off-policy deep reinforcement learning trains on random samples from a replay buffer of past experience. In physical robot settings, that buffer is dominated by routine: thousands of near-identical transitions from normal operation, while …