AI 情报文章归档

Knowing that a pedestrian is occluded is useful to a robot even when the detector cannot recover the hidden pixels, but occlusion estimators are normally trained on occlusion annotations that deployment settings rarely have. We show that o…

Grok Bot 现在可以拨打真实电话。作者以 Grok Bot 设计负责人身份介绍，可用它订餐厅、预约医生，并在等待时代替用户与保险公司周旋。

Automation is often justified as relief from dull, dirty, and dangerous work. I examine how these conditions relate to occupational AI exposure using 21 O*NET items across 754 US occupations. Higher exposure is associated with lower condit…

OpenAI 宣布收购 Code Storage 背后的 Pierre Computer Company，其开源库已被 OpenAI 编码智能体 Codex 使用，包括快速文件导航工具 Trees、可在 diff 视图中直接修复 bug 的 Diffs 和实时语法高亮 Highlights。

The best classical optimizer for a continuous black-box problem depends on its budget and landscape. Existing routers decide from the dimension and budget, past improvement or features fitted to one benchmark, and few test an optimizer's i…

LLM agents use context compaction to manage growing interaction histories, but summarizing long histories with large models can introduce substantial delays. Compaction therefore requires balancing information retention against latency. We…

Bill Gates 在 FT Live 访谈中表示，AI 一旦 properly set up，随着更好版本的到来，取代白领劳动将非常便宜，且不只限于法律或商业咨询，而是覆盖经济各领域。他认为现行税收政策把税负压在劳动力上、同时对购买机器人和 token 的资本支出很宽松，会加速企业用 AI 取代人工。完整视频在 FT Live 频道。

未来感十足 （引用推文：Grok Bot 现在可以打电话了。）

Error correcting codes are a fundamental component of communication systems, enabling reliable data transmission over noisy channels. Recently, neural decoders have demonstrated notable performance gains over classical iterative algorithms…

A strategy built from a public ledger of losses: four lanes ranked by measured EV, the gate that fences risk capital, and the arithmetic that decides whether $10/day is reachable.

In this work, we evaluate a simple inference-time scaling approach based on \textit{turn-level best-of-N sampling} (\method) for large language model (LLM) agents on complex, long-horizon tasks. We instantiate this using a \textit{generato…

NVIDIA 论文发现，AI 智能体通过频繁压缩历史保持短上下文时精度几乎不降。服务 256K 上下文窗口比 16K 的每 token 生成成本高 3.6 倍到 7 倍。

Automated identification of individual animals is moving from small curated studies toward shared benchmarks and reusable foundation models. This progress is useful, but current evaluation leaves the evidence behind a match unresolved, i.e…

Conductor 创始人 Charlie Holtz 宣布 Conductor 团队加入 Meta，将与 Nat Friedman 和 Alexandr Wang 在 MSL 合作打造下一代界面。他表示 Meta 超级智能实验室人才密度极高，并感谢早期支持者和用户反馈。

Claude 发布教程视频，演示如何在聊天中创建并共享 artifacts，让团队共同编辑文档、跟踪表和仪表盘。视频介绍了 viewer 到 editor 等访问级别的设置，以及共享跟踪表上 commenter 和 contributor 角色的用法；收到评论后可以让 Claude 直接完成内容或样式修改。

Hierarchical text embedding models offer variable-length representations, while quantization reduces numerical precision; both aim to improve computational efficiency. This work is an empirical benchmark study of how these two orthogonal, …

Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems. LLM compression offers a concrete efficiency objective: retain a pretrained model's behavior within a smaller model storage…

Grok Bot 现在可以拨打电话了。

We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual e…

Computer-use agents (CUAs) have recently achieved superhuman performance on several long-horizon benchmarks. However, their computational cost limits their use in long-horizon tasks and real-time domains such as video games. A key bottlene…

OpenAI 总裁 Greg Brockman 宣布 @pierrecomputer（Pierre Computer Company）加入 OpenAI。Pierre 的开源开发者工具目前已支撑 Codex 的部分功能，双方将共同打造用 AI 协作构建软件的未来。

Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt differ…

Faithfulness requires that algorithmic explanations accurately reflect the model's behavior. Yet existing operationalizations of faithfulness can disagree about whether the same explanation is faithful, leaving developers and auditors to c…

OpenAI 随 Images 2.6 模型为 ChatGPT 上线 GIF 生成，可单步生成动图并调整播放速度、逐帧编辑和分享，此前 ChatGPT 无法一步原生生成 GIF。

The Verge 资深评测者 Victoria Song 撰文讲述连续三周佩戴 Amazon Bee Frontier 常录 AI 穿戴设备参加心理治疗的经历。

Grain growth during heat treatment strongly affects the microstructure of polycrystalline materials and, consequently, their mechanical properties. Although numerical methods such as front-tracking or front-capturing simulations can accura…

On 28 October I posted about ClaudeArena, the free practice engine I built for Anthropic's Claude...

美国中期选举中，13个州40多场涉及数据中心的公投提案里有38项被否决，6个城市直接禁止数据中心，12个对用电兆瓦设限。支持数据中心的候选人普遍落败，包括推动得州数据中心发展的州长 Gregg Abbott，他虽获数据中心相关捐款超过 900 万美元仍以 70 多万票差距败给 Gina Hinojosa；Elon Musk 在选举夜发帖称太空数据中心“看起来很有吸引力”。

Philipp Schmid 分享一个提示词模板，用于在没有物理麦克风的情况下端到端测试项目中的语音或音频功能。

Philipp Schmid 分享用 Gemini API CLI 配合编码 Agent 测试语音、音频等多模态功能的方法。步骤为：用 Gemini API CLI 生成语音，通过 Chrome 的 --use-file-for-fake-audio-capture=voice.wav%noloop 参数注入音频，用 Playwright 或 browser-use 点击麦克风按钮让 getUserMedia 从 0.0s 读取 WAV，最后在 DOM 中校验转录和 UI …

Axiom Math 发布对整个 Stacks Project 的形式化，共 16M 行 Lean 代码，由其 agent swarm 以类似工程组织的方式分工完成，Lean 逐条检查每个证明。62,975 个 Lean 文件、663,090 个声明建立在 Mathlib 之上，每个带 tag 的结果都链接到原文和依赖。这是免费开源库，后续研究者或 AI 可直接复用已验证的定义和定理来证明新结果。

By Spacexcx · November 2026 I built a small tool that reads a git diff and flags possible logic...

Dharma-AI 在 Hugging Face 发布 Dharma Field Extractor，一个 20 亿参数模型，基于 DharmaOCR-LITE v1.1 训练 LoRA 适配器（43,384 条样本），支持开放和按 schema 两种请求方式，将文档图片转为带类型的结构化 JSON，缺失字段返回 null。

Agent Arena 统计近 14 天近 30000 个任务运行超过一小时，平均模型每小时发起 332 次工具调用。长程任务上模型选择最多可省 66% 成本：Gemini 4 Carbon 每小时任务中位成本 $10.51，Claude Opus 5.5 为 $21.13，GPT-6 Astra 为 $31.08。

Perplexity 的智能体为商家投放广告，效果比广告公司还好。

Arena 基于 Agent Arena 数据称，Google DeepMind 的模型在成本上扩展性更好，Gemini 4 Carbon 是典型代表：典型任务上 GPT-6 Astra 成本为其 3.4 倍、Claude Opus 5.5 为 3.9 倍，GPT-6.1 Sol 约为其一半（0.54 倍）。

Apple 警告假日季毛利率将从 9 月季度的 48% 降至 45%，原因是内存芯片价格大涨。CEO John Ternus 在首次财报电话会上称定价“前所未有”，并预计成本在 12 月后继续上升；内存一项使 9 月季度毛利率约降 1 个百分点，营收同比增长 12% 至 $114.5B。

Arena 发布 Agent Arena 评测观察，称任务越复杂时 GoogleDeepMind 和 Xiaomi MiMo 是最大赢家。便宜任务上 OpenAI 和 Anthropic 占据 11 个前沿席位中的 10 个，但在最贵 5% 会话中只占 7 个，空出的席位由 Gemini 4 Argon 和 MiMo-V2.6 Flash/Pro 补上。

Computer 为 Get Maine Lobster 管理了 3 万美元的 Meta 广告支出，表现优于其代理机构。 与其常规广告相比，Computer 带来了 3.5 倍的收入，ROAS 高出 63%。它自动化了创意分析、广告制作和媒介购买。 阅读更多：https://pplx.ai/gml

OpenClaw 宣布支持腾讯混元 Hy4，模型具备视觉能力、1M-token 上下文窗口以及改进的编码和长程智能体能力。用户可通过 Tencent TokenHub、TokenPlan 和 OpenRouter 使用，详情见 https://hy.tencent.ai/research/hy4。