AI 情报文章归档

两个月前，我们上线了一个 Gemini 高级用户小组，用于测试应用中的早期功能。此后已测试了 20 多项功能，大家的反馈非常棒。谢谢你们！ 新一批用户现已提前体验 Daily Brief 与 Personal Intelligence 的后续功能，我们会继续招募更多人。 加入我们的社区：http://goo.gle/4onCRHq

Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise rests on picking the right one. Deployed harnesses route by preloading every skill's metadata into the context, which disperses the agent's attention and…

When a video model generates physically incorrect motion, did it fail to learn the correct motion, or did it learn it but fail to use it? We show the latter: the correct motion remains available inside the model and can still be made to co…

Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpen…

Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome feedback. We argue that the next frontier is a further transition: from solving and actin…

As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental superv…

Large language models (LLMs) have been widely adopted for clinical question answering (QA). Current systems can attach citations to their answers, but these often point to broad texts, leaving time-pressed clinicians unable to verify them …

A useful agent no longer just answers questions. It can qualify a lead, follow up with a customer,...

Record-level differential privacy exposes a structural misalignment in personalized federated learning when client-specific variation is low-dimensional while training repeatedly releases high-dimensional updates. In this paper, we address…

Language-model agents increasingly operate over complete software repositories, yet cybersecurity evaluations primarily measure whether they can detect, reproduce, or repair vulnerabilities rather than whether they can locate the relevant …

Scientific agents contribute to hypothesis discovery by synthesizing evidence, assessing proposals, and developing new explanations. Recent systems combine scientific agents with evolutionary search through critique, comparison, and revisi…

Recurrent GNNs iterate message passing to convergence, and their logical characterizations to date rely on multi-set aggregation, graded (counting) logics, and halting or acceptance conditions that cannot be verified from the network's par…

Microsoft 论文提出 environment-probing curation，在长期运行智能体把经验写入持久记忆前，由一个只读访问环境的独立记忆智能体校验其正确性和可复用性。

开发者发布 Kinesis，一款利用 Meta Neural Band 手势控制 Mac 的应用，支持切换桌面、打开调度中心、控制音乐，捏合转动可调音量或亮度，用户可自定义映射。应用需 macOS 14+，支持 Apple Silicon 和 Intel，需先在 Meta 应用中取消手环配对，可从 dmg 安装，也可用 Xcode 和 Swift 6 自行构建。

Cline 推出 Cline Desktop，一个面向开放权重模型的开源原生界面，可用 ClinePass 及 DeepSeek-V4.1-Flash、Musespark-1.3 等免费模型，也支持 BYOK 接入任意服务商。转发作者 Elvis Saravia 评价它简单易用，具备调度和 forking 功能，并可通过其常用服务商 OpenRouter 切换任意模型。

让两三家硅谷公司充当全世界所有政府的 AI 创造者、守门人和规则制定者--这种想法根本经不起说出来。 https://cohere.com/blog/who-gets-to-define-the-rules-for-ai

Artificial intelligence presents firms with an unusual timing problem. The technology frontier is improving rapidly, implementation is partly irreversible, and organization-specific capabilities are accumulated through action. This paper d…

中国外交部在回应有关 Amodei 文章的提问时表示，渲染恐惧、对抗和恶性竞争只会破坏全球 AI 治理进程，不符合任何一方利益。中国官方背景的《环球时报》批评该减缓 AI 计划是扩大美国对华领先优势的策略，相关报道称其为"冷战"式做法。

Claude Mods 正式落地了。已经有人做出了 Claude 内运行俄罗斯方块的 mod 🤯 详见 issue 了解最新社区进展、技术细节和更多酷炫演示 https://github.com/anthropics/claude-code/issues/91870#issuecomment-5666255143

我和 Sid 与 Robert 聊了 Claude Code 的开发：一切变化有多大，跟上模型能力有多难，以及我们怀念 AI 出现之前的软件工程的哪些方面。 https://www.youtube.com/watch?v=S-sYlFiGFv8

我从没想过自己会和 Yann LeCun 站在同一边。 【引用 @ylecun】：@PessimistsArc 没错。早在 2019 年，Dario 就声称 GPT2 开源太危险了。我当时就嘲笑他们。现在所有人都应该嘲笑他们。

Meta-reinforcement learning (meta-RL) enables agents to adapt to unseen tasks with limited experience. Despite its promise, the application of meta-RL in real-world tasks is hindered by safety requirements, which have been underexplored in…

纽约地区检察官办公室周一宣布查封12个托管AI生成未经同意亲密影像的网站，自称是史上最大规模的此类查封行动，涉及约1200人的照片被制成超逼真色情图像和视频，对象多为演员、政客、运动员等公众人物。地检官 Alvin Bragg 援引纽约州2023年将色情深伪入刑的法律，鼓励受害者联系其网络犯罪局，并表示只要网站可从曼哈顿访问即可采取行动；是否涉及未成年人图像仍在调查中。

真得非常小心才行 【引用】：METR 等评测机构将面临国家级攻击，想想都觉得不可思议？ 要做好这件事，评测机构既需要 AI 人才，也需要世界级的网络安全能力，据我的经验，这两个领域并不重合。

Slip detection is fundamental to dexterous manipulation, yet existing systems often lack precise characterization of detection latency and cross-platform generalization. We present SlipSense, a multimodal tactile slip-detection framework b…

Muse 是为妈妈们打造的！♥️ 此前引用 @OffZeroCyber：在 @Muse 之前，智能体要么看起来像团队周末临时拼凑出来的（@bot），要么得你自己动手拼（@openclaw）。两者都没让我想打电话给我妈、推荐她上手试试。

美国副总统 JD Vance 表示，众多前沿 AI 公司主动请求政府监管让他感到不安，担心这是特洛伊木马，美国需在对华 AI 长期竞赛中保持领先。引文中特朗普回应记者提问称美国正领先中国，谁赢得 AI 就赢得一切，可以设护栏，但不应因被夸大的担忧放慢。素材来自 Forbes Breaking News 视频。

DeepSeek-V4.1-Flash (Max) 进入 Agent Arena 开源模型第 3 名，净提升 +4.87%，每任务中位成本 $0.07，重塑 Pareto 前沿。其成本比第 2 名 Hy4 preview 低 68%、成绩仅差 0.09 个百分点；总榜排名第 12，Confirmed Success 信号排名 第 4（+13.75%）。

"自 ChatGPT 以来最大的消费级 AI 发布" 🎉 🎊 【引用 @SashaKaletsky】：Muse 在美国的日下载量已超过 Threads、WhatsApp 和 Facebook，距离 Instagram 只差 3，000 自 ChatGPT 以来最大的消费级 AI 发布。一个新的顶级消费级应用诞生了

Discrete diffusion and flow models are a promising alternative to autoregressive language models, but compressing many-step sampling into fewer steps typically requires distilling a pretrained teacher model. This caps the student at the te…

我已经 23 天没打开 Claude Code 了。 它曾是我最喜欢的产品之一，但现在我更喜欢 Codex。 Codex 与开源模型配合得更好。而在开源模型中，Kimi K3 在编码方面仍然无可匹敌。

作者认为开箱即用的 agent harness 过于臃肿，为不需要的功能付出高昂溢价，自建 harness 可显著降低成本--引用推文提到停用原 harness 后 token 消耗大幅下降。关键优化包括：精简系统提示词、更高效的 tool calling、更好的上下文压缩与交接、自定义路由调用不同模型、用自定义 verifier 提升可靠性。

大脑太大，难以随身携带--论机器人模型的端侧与数据中心推理、芯片与 DRAM 效率、Jetson Thor 与 B300 TCO 对比、部署情况、网络瓶颈 https://newsletter.semianalysis.com/p/a-brain-too-big-to-carry-on-device

The last decade has seen the development of powerful methods for learning complex structure from high-dimensional data. These advances have brought to the foreground fundamental connections between subdisciplines of physics, applied mathem…

Some of the sharpest challenges in sampling from the energy functions of physical systems arise at phase transitions, where the density of states changes abruptly and many sampling algorithms stall. Nested sampling is a particle method tha…

404 Media 调查发现，OpenAI 通过 Crossing Hurdles 招聘、经 Mercor 支付报酬，雇佣数百名合同工阅读真实 ChatGPT 对话，按 1 到 7 分给回复打分，任务是减少模型的过度谄媚和拟人化行为，北美审核者时薪超过 50 美元。

Deep networks trained on structural MRI for Alzheimer's disease (AD) staging often reach reasonable accuracy while attending to anatomically irrelevant regions, and multimodal models that add clinical tables frequently rely on variables th…

TL;DR GPT-6 Astra has started another familiar AI conversation. The model is more capable, Jensen...

TechCrunch 作者实测 iOS 27 公测版后表示重新使用 Siri，新版 Siri 基于 Google Gemini 模型构建，可处理多步指令、读取屏幕上下文、调用相机回答所见物体并支持草拟发送消息。

I did not build joacod Lab because the internet needed another portfolio. It already has plenty of...