AI 情报文章归档

东京初创公司 Sakana AI 聘请 Jürgen Schmidhuber 出任首席科学顾问，称其为现代 AI 之父。Sakana 认为其 1990 年世界模型和 1991 年深度学习方面的工作以及 1987 年关于元学习与递归自我改进的论文，直接影响了 Darwin Gödel Machine 和 The AI Scientist 等项目。

快就是好。 Grok 将在 2 到 3 个月内达到 Fable/GPT-6 级别

The Verge 记者 Justine Calma 评析黄仁勋在 The Ezra Klein Show 上的言论，黄仁勋称 AI 拯救气候前不得不先使用化石燃料、造成痛苦。

Asynchronous monitoring, incident investigations, and compliance audits primarily rely on agent traces to reconstruct what happened. These analyses assume that LLM agents cannot tamper with their own execution traces. We show that local LL…

Dreambeans 是我们在 @GoogleLabs 较新的实验之一。它正获得越来越多狂热粉丝。 很简单：每天早上固定数量的“beans”会冲泡出来。它们把你引向真实世界，与真实的人一起做你关心的事。 试试看吧！

Latent world models are typically trained to predict factual transitions, whereas model predictive control (MPC) must compare alternative actions from the same state. A model can therefore achieve low factual prediction error yet poorly di…

Distributed learning in embodied reinforcement-learning agents offers a degree of privacy by retaining raw sensor data on-device and transmitting only policy gradients to the server. Yet temporal structure can amplify this leakage beyond s…

Conspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery. The main challenge is…

Coding agents have demonstrated enormous success in solving complex programming problems. To leverage their potential for robot systems, this work introduces Robot Agentic Programming from Demonstrations (RAPID), which automatically genera…

World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action u…

Dedicated decision models such as Jev map unstructured language to probability distributions over finite choices, allowing their outputs to directly route requests, select tools, and trigger actions. Yet real-world inputs rarely arrive in …

Recent research on Multimodal Sentiment Analysis (MSA) has focused on learning from language, visual, and acoustic modalities with incomplete data to infer human sentiment. Most studies typically compensate for missing information by recon…

在 Krea Agent 中推出自定义应用。 构建按你想要的方式工作的创意工具。角色工作室、3D 查看器、动态效果。只需描述这个应用。 了解更多 + 示例见下方 👇

Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses th…

马斯克称 SpaceX 的 AI 投入仅 3 年，而 Anthropic 和 OpenAI 分别已 6 年、10 年，若二阶导数保持强劲，SpaceX 约 6 个月就能到达头名位置。他还认为智能一旦远超某类任务所需水平便无意义，并强调快速上线大规模算力极其困难，而 SpaceX 在这方面表现突出。

Meta 宣布两款 AI 游戏开发工具：移动端应用 Horizon Create 和浏览器端应用 Horizon Studio，均支持用 AI 提示词生成 2D 或 3D 移动游戏，包含进度系统、难度平衡、美术方向和多人玩法，目前以早期访问形式开放候补名单。

一种名为 Dynamic Abliteration 的方法通过 PyTorch forward hooks 在运行时拦截多层残差流，无需修改权重即可抑制 Qwen3-4B 的拒绝行为，基座模型权重保持 100% 冻结。该方案在 A100 上以 Qwen/Qwen3-4B（隐藏维度 2560、36 层）为概念验证，对比了传统单层拒绝向量减法会永久改变权重并损害非拒绝任务的问题。

华为乾崑智驾 ADS 5 新版本（ADS Max V5.0.2）即将推送，领航与泊车体验双升级。领航侧新增组队出行、队内实时语音对讲，导航搜索拓展至自车周边与目的地两大场景；泊车侧新增「AR 人找车」，可在华为乾崑 App 的 ADS 板块跨层、室外精准定位车辆。

Online misinformation increasingly appears in spoken formats such as news clips, podcasts, interviews, political speeches, and social media videos, creating a need for fact-checking systems that can verify claims directly from speech. We i…

Foundation models are post-trained with reinforcement learning (RL) to maximize specific rewards, such as human alignment, correctness, or instruction following. This post-training process is computationally intensive, sometimes unstable, …

我提议用这个来正式替代那张著名的（如今已饱和的）METR 长任务时间跨度图表——它以前出现在每一场 AI 演讲里。

Existing point tracking models face a fundamental tradeoff: they can either track a sparse set of query points over long horizons, or track all points across only short clips. We introduce TrackEverything, a 3D point tracker that breaks th…

An acceptance protocol is developed for sensor-coordinate and polarity binding in mechatronic commissioning. Candidate generation is separated from release authority. Requirements unsupported by a deterministic parser are routed to a froze…

Pre-logit steering adapts a frozen language model to a test-time reward by adding vectors to its final hidden states. Unregularized reward optimization can substantially alter the output distribution and degrade generation quality. We prop…

A central concern in AI safety is that agents may treat oversight as an obstacle when it conflicts with completing their goals. We study instrumental evasion, the propensity of LLM agents to circumvent runtime monitoring as a means of comp…

We prove nearly quadratic lower bounds for randomized algorithms for linear optimization and uniform sampling over convex bodies in the membership oracle model. For linear optimization, this matches the known nearly quadratic upper bound u…

Google 的 Suncatcher 项目计划在轨道上以太阳能运行 AI 基础设施，一颗冰箱大小的试验卫星 MVP 定于 10 月 1 日搭乘 SpaceX Falcon 9 从加州 Vandenberg Space Force Base 发射。

We present Underwater C$^{3}$-JEPA (cross-view, control-conditioned, context-extended), an object-centric multi-view predictive world model for near-field heavy-load underwater ROV salvage. Without contact sensors, it predicts in latent sp…

欢迎 @AMD 加入 Perplexity Computer 本地 AI 生态！ [引用 @perplexity_ai]：Portable Computer for Windows 现已在 @AMD Ryzen AI Max 系列处理器上线。 它可以轻松运行本地 AI 智能体，与你连接的应用程序和本地文件协同工作。启动任务或安排完全在你的设备上运行的周期性工作。

We study the deterministic oracle complexity of finding approximate solutions to composite monotone inclusion problems, formed by the sum of a smooth single-valued monotone operator and a maximally monotone set-valued operator, under the t…

Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the a…

Large language model (LLM)-based clinical assistants are increasingly being integrated into electronic health record (EHR) systems, transforming how clinicians retrieve and synthesize information from patient records. Their safety and util…

Vellum 发布 Ambient Assistants，将 @vellum_ai 助手以悬浮胶囊形式叠加在任意 Mac 应用之上，可实时看到屏幕内容并在屏幕上标注重点。用户可双击 Fn 唤起、共享屏幕并让助手实时标注，或按住 Fn 将语音转为任意页面文字；还能通过语音在后台派发任务，助手会在屏幕上圈出具体按钮、设置或指标，并在用户演示并讲解流程、计划获批后自动执行该流程。

Meta 发布 AI 设备 Muse Charm，外形类似 Tamagotchi，可自定义角色头像，定位为可挂在包上的 AI 智能体配件。它踩中了 Gen Z 的包挂饰与复古科技潮流，但 Meta 过往的隐私争议可能影响其成为消费者首选 AI 硬件。

GPT-6 Astra 帮助 @harvey 将成堆文档转化为结构化的法律草稿，让律师能更专注于策略。

Anthropic 宣布恢复对被安全机制拦截的请求计费，仅适用于误报率低的类别：生物学、蒸馏攻击和前沿 LLM 开发。官方称近几周遭遇了协同攻击，此举是防御层之一；测试中 99.7% 使用 Claude Code、Claude.ai 或 Cowork 的账户未触发此类计费拦截，分类器误报率低于 0.1%，认为被误拦可通过 Claude Code 中的 /feedback 申诉。

A lot of AI productivity claims sound impressive until you ask what was actually measured. "10x...

Elvis Saravia 测试了 Space Bunny 模型，其在 OpenCode 中免费开放一周，支持 1M token 上下文窗口和图像输入。该模型视觉理解与设计能力出色，能自我验证工作成果，可从草图生成完整落地页、3D 模型、动画和可玩游戏。作者推荐将其用于图像转代码、交互原型和 3D 创意编程。

100% 赞同 @Mappletons /show-me 才是正道 https://youtube.com/shorts/gnYNZEwq9ho?is=vfKVK_cW659v6y9I

开发者 Peter James 与 Jonny L. Saunders 称用少量提示词就能让 Meta 的 Muse 打包分享其根文件系统全部内容，Saunders 表示 Muse 几乎没有提示注入抵抗。