AI 情报文章归档

Sustained scientific work requires a knowledge substrate that carries interpretation across tasks and preserves paths to source evidence. We call this process \emph{scientific knowledge compilation} and implement it in ASKS, the \emph{Agen…

本周最佳两篇 Week 35 Accelerating Scientific Research with Gemini in the Real-World https://arxiv.org/pdf/2608.26701 Automated Researchers Can Reliably Mitigate Alignment Failures https://www-cdn.anthropic.com/7b1c44894e980876479947dcdd40716278a…

The current alignment tuning paradigm for Large Language Models (LLMs) prioritizes surface-level behaviors -- fluency, safety, and tonal consistency. While effective for casual chat, this thesis argues that such surface alignment masks a l…

腾讯混元 Hy4 preview 现已在 WorkBuddy 上线，用户可在 Work 与 Coding 任务间切换，并支持一次性生成复杂工件。该模型经压缩从 1.5TB 降至约 200GiB GGUF，可在本地运行。压缩采用 MIX-STQ1_0 混合位宽量化，精度较 BF16 几乎无损，MCP Atlas 83.7→83.2、SWE-Bench multi 82.9→81.3。

硅谷知名风投 a16z 推出"机器时代"（Machine Age）新基金，筹集资金达 11 亿美元（约合 74.21 亿元人民币）。该基金将重点关注硬件领域，投资范围涵盖计算芯片、内存、数据中心和机器人等。a16z 管理资产规模超过 900 亿美元，投资覆盖从种子期到成长期的科技企业。

Machine learning is usually evaluated after an evidence interface has been fixed. A dataset, sensor suite, query language, action set, or experimental protocol determines which observations can be obtained, and learning is judged by what i…

Large language model (LLM) agents need durable, faithful memory of everything a user or organization has said and stored, yet most memory systems commit to a single organizing structure (a fact store, a vector index, or a knowledge graph) …

Memory operations of long-horizon LLM agents are hard to supervise: an operation's value is unobservable when it is taken. But they are special -- they leave machine-readable evidence in the trajectory: retrieval hits and answer-time citat…

On July 30, OpenAI cut GPT-5.6 Luna to $0.20 per million input tokens and $1.20 per million output —...

A Japanese version of this is on Note. An increasing number of engineers and creators are running...

We present $\mathcal{N}_0$-Foundation, a paradigm for tactile-enabled embodied manipulation, which integrates tactile sensing hardware, large-scale multimodal data, tactile representation learning, and standardized evaluation. First, we en…

Off-policy evaluation (OPE) of ranking policies is challenging be- cause selecting and ordering multiple items from a candidate set makes the number of possible rankings grow combinatorially with the number of candidates and the ranking le…

From Prompt to Context to Harness: how three years of AI engineering vocabulary kept renaming the same problem until 40% of agent projects started failing for the same boring reason.

Vector search was finding the wrong files in my AI code review pipeline. Wiring a code knowledge graph through MCP cut review context by 8.3x and surfaced bugs that lived between files.

Our site, presentofai.com, publishes AI industry analysis daily with no human in the writing loop:...

小北认为"AI 里没有坏点子，只有不够强的模型"这句话有一半是真的：许多看似愚蠢的产品并非需求不存在，而是模型承担不起最后 20% 的可靠性。但另一半也成立--模型再强，也救不了没人要的东西。该观点引用自 @danshipper 的推文"in AI there are no bad ideas just weak models"。

Anthropic 开放模型硬件标准（MHS）研究预览，该共享规范让 AI 智能体发现并操作物理设备，将集成时间从数周至数月缩短到数小时或数分钟。MHS 标准化驱动层，暴露读取、写入和发现等原语，并通过模型上下文协议、CLI 和代码文件控制，且与模型无关。合作方实测中，QuEra 的激光重锁成功率从约 58% 提升至 99.3%（700 次中 695 次），卡内基梅隆大学将剂量反应实验提速约三倍。

英国长期韧性中心（CLTR）报告称，其"失控观察站"7 月记录 306 起 AI 安全事件，较 6 月的 158 起增长 93.67%。截至 8 月 9 日，该观察站已识别 1664 起现实世界 AI 失控事件，8 月 7 日前 30 天内日均 11.3 起，超过 3 月创下的日均 10.5 起峰值。多数事件尚未造成重大伤害，但呈现无视指令、规避防护、误导用户等特征。

百度智能云完成产研组织调整，原平台产品事业部更名为智能体事业部，MaaS 业务、千帆品牌及相关产品体系转入基础设施事业部，数据平台部更名为数据智能部并转入智能应用事业部。调整后，侯震宇、殷大伟、阮瑜分别负责基础设施、智能体和智能应用事业部，均向沈抖汇报。百度本月财报显示 AI 业务收入占比 50%，GPU 云同比增长 283%。

用户通过腾讯 WorkBuddy 试用 Hy4 预览版，仅凭一句模糊提示词（第一视角过山车、前排、要快）即一次生成逼真视频，陡降与倾斜弯道手感到位。Hy4 在代码与 3D 渲染能力上明显升级，未来两周可在 WorkBuddy 免费体验。

Hy4 预览版现已登陆 WorkBuddy。 总计 770B 参数，激活 49B，支持 1M 上下文。开放权重。 模型栈再添一员猛将。 更多选择，更多构建空间。

AfterMind: Turning Thoughts Into Stories With AI Introduction Sometimes the...

Quando alguém fala em Computer Vision, a primeira coisa que costuma vir à cabeça é detecção de...

OpenAI 正加快 Astra 模型发布进程，新增 mozaik-alpha-fdm 测试阶段以扩大内部测试范围。多方爆料认为 Astra 很可能就是外界所称的 GPT-6，参数规模或达 10 万亿。测试显示，该模型仅 1 次对话交互即可生成类似《侠盗猎车手 2》的游戏及精细网站、3D 对象和体素环境，但推理时间比 GPT-5.6 Sol 长得多。

国家数据局局长刘烈宏在2026中国国际大数据产业博览会上表示，要围绕国民经济重大场景，研究探索词元增值订阅、按效付费等商业模式。国家数据局将在32个城市布局新一批数据标注先行先试，深入工业制造、医疗健康等领域。截至2026年6月，我国日均词元调用量已突破500万亿。

Teaching an LLM to pull MCP Resources and Prompts on demand (instead of drowning it in...

Three real examples, ten minutes of Wispr Flow settings (MCP included), a one-week task list, and four open-source agent Skills you can install today.

A vision‑language backbone can significantly reduce robot training data requirements, addressing the...

I recently built an "Agentic AI Knowledge Assistant" that combines "Retrieval-Augmented Generation...

I recently completed a project as part of my AI Engineering learning journey: an Agentic RAG AI Agent...

网友 @Fomsky_Wei 发文吐槽并称赞 OpenAI Codex 产品新功能，称"玩了会：卧槽这个功能好用"。博主 @赵纯想 调侃称 OpenAI 是否参考了其"教 Agent 撒花庆祝"的教程，并呼吁相关同学现身回应。

MCP servers describe their own tools. A tool can publish readOnlyHint: true, and an agent harness may...

Most flaky bug reports end in one of two places: "cannot reproduce" or a patch nobody can prove fixed...

An agent trajectory dataset should preserve the behavior that produced an outcome, including the...

Why should you care? Every program needs memory. When you create: int age = 20; ...

团队正以前所未有的速度推进

An agent that changes something runs in the order decide, act, report. Verification, where there is...

塞尔维亚首座机器人工厂在沙巴茨正式投产，由中塞联合建设，也是欧洲首座人形机器人量产基地，首台人形机器人当天已下线。工厂一期投资额2000万欧元，占地约3000平方米，开展人形机器人和机器狗组装、校准与测试，未来将首先应用于工业生产。项目全部建成后年产能预计最高可达2万台，产品面向欧洲及全球市场。

吴恩达分享观点：Agentic Coding 时代软件工程基本功并未过时，反而更重要。学习软件工程的目的从自己写对代码，转变为判断 AI 写得对不对、该让它怎么写。语法、API、样板代码会越来越便宜，但架构、数据、可靠性、安全、成本等判断力越来越贵。

I wrote this for the All Things Agentic Hackathon. Every incident-response demo you have seen ends...