AI 情报文章归档

宇树科技（Unitree Robotics）发布上市后首份 Q3 财报，营收 9970 万美元，同比增长 70%；Q3 研发支出同比增长 250%，主要投向具身智能基础模型和算力基础设施。公司还与早期投资方 HSG（原红杉中国）合作推出孵化早期机器人开发者的「山雨（ShanYu）」计划，显示其正从硬件性能转向软件与生态建设。

Error-correction overhead fell by roughly 10x in five years. Here is what that curve does to drug discovery budgets, cryostat supply chains and the price of a quantum hour.

Why Forward Deployed Engineers Are Suddenly Hot: The Death of the Centralized AI...

Agents commit PHP that never met your formatter or PHPStan. A lefthook setup that blocks the commit, hands the agent the error, and survives new worktrees.

OpenAI 发布长文，详解其针对前沿模型失准行为的多层监测系统，包括激活探针、基于模型的审查、调查智能体和自动化调查四层架构，并对 GPT‑6 Astra 及以上模型的全部 RL 运行、评估和内部 Codex 流量进行监测。

A tool call can pass schema validation and still select the wrong product, skip a prerequisite, or...

Infini-AI-Lab 发布开源 harness 适配器框架 Hermit（v0.1.0，Apache-2.0），在模型和 harness 均冻结的情况下训练外部适配器，用 RL 式训练循环迭代适配器版本，可部署到 Codex、Claude Code 或 Pi 并导出为插件。

📌 Note: Interactive math rendering is currently being updated for this post. In the meantime, you...

IDC报告显示2026H1中国MaaS市场规模达106.3亿元，火山引擎以45.1%份额位居第一，MaaS市场营收同比近9倍增长。大语言模型占整体规模47.0%，视频生成模型贡献32.7%收入；火山引擎推出全托管AaaS平台Managed Agents，上线2个月产生近500万个Session。

Zero-width, bidi and Unicode tag characters hid instructions in AI rules files. How they work, plus a ripgrep check and a git hook that block them.

Anthropic 于当地时间周三发布 Claude 智能报告功能，现已面向 Claude Enterprise 付费套餐用户开放，可在组织设置页面开启。管理者可选定团队、最长 28 天的时间范围及纳入报告的 Claude 产品，系统读取对话记录并归类工作流与产出类型、统计成本，每月免费生成最多 10 份报告。

Summary: AI coding agents rarely lie. They do something worse for a one-person shop: they state...

# Build RAG in C# with ASP.NET Core and Qdrant Build RAG in C# with ASP.NET Core, Qdrant and Ollama:...

Block 发布研究预览版 Dessa，一款仅用人类对话音频训练的全双工语音到语音模型，不使用文本和预训练模型。它用 1 台 NVIDIA GB200 NVL72（72 块 GPU）训练，可在 MacBook Pro M4 Max 及以上本地运行且快于实时。目前仅有演示，未公布 benchmark 分数、参数量、权重、论文或 API，需申请 waitlist。

A few months ago my typical afternoon looked like this: four terminal tabs, four Claude Code...

据 X.PIN 报道，中国 CXMT 将在两个月内开始限量生产 HBM3E，以缓解国内 AI 内存紧缺。代价是良率仅 20-30%（SK Hynix 约 70%），速度接近较旧的 HBM3，且在韩国对手推进 HBM4 时落后一代以上，超过 70% 的产出会被废弃。

Deep dive into how function calling actually works under the hood — JSON Schema to CFG compilation, pushdown automata, logit masking, and why the model literally cannot produce invalid structure.

英伟达于 10 月 28 日发布开放权重语音转录模型 Nemotron 3.5 Transcribe，并公开训练数据管线 Granary v2。

Static analysis and LLMs are a good pairing. I'm not here to argue otherwise. But a lot of teams are...

PJM 提出市场规则改革，允许混合资源在能源市场和容量市场分别核算不同资源的贡献，从而让现有风电、光伏项目在不停原有购电协议的前提下加装电池，并把新增容量卖给不同买方。此前 FERC 已驳回 PJM 要求数据中心等大负荷自备容量的方案，而新规即便顺利推进，FERC 批准也可能要等到明年年底。

AssemblyAI 发布首个文本转语音模型 AssemblyAI-TTS-1，正式开放 API，面向实时电话语音 Agent，定价 $15/1M 字符且不限并发。

The second client on the hot path: net/http, encoding/json and a bufio scanner for Server-Sent Events, in a static binary from a scratch image. What the standard library gives you, and the one thing it does not.

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine. ...

据 Eurogamer 报道，Epic Games 开发执行副总裁马库斯·瓦斯默确认将扩大生成式 AI 应用范围，覆盖《堡垒之夜》和虚幻引擎 6，并承认公司采用了氛围编程，工程师指导 AI 工具而非亲自编写大部分代码。他表示 AI 可加快更新速度、减少程序错误，并称相比外界对 AI 的质疑，公司更担心开发进度落后的风险，此前 Epic 还用氛围编程开发了新版 Epic Games 启动器。

Static rules break on novel agent actions, and blind human approvals break on volume. A concept we are evaluating: AI at the meta level drafting rules, a deterministic pipeline deciding, and a human sealing.

n8n 内部 AI 与自动化负责人 Romina da Costa 撰文分析集中式自动化团队服务台模式的天花板，称公司已有超过 12,000 条内部工作流，任何团队都无法包揽构建与维护。她的团队转向赋能文化，提出四条原则：培养各团队的构建能力、立项前先摸清现状基线、按问题而非惯性选工具、让每个工作流有归属和文档，目标是让更多团队不再依赖他们。

An opinion on how IAM should treat AI agents, as a new kind of principal that starts every session with nothing and borrows exactly what the task needs, with the user's consent.

In the first article of this series we put four 2026 context-engineering principles on one map. This...

Eight checks before an unattended Claude Code run: sleep, idle compaction, usage-limit waits, turn caps, permissions, worktrees, a test gate, a ping.

This week's Hacktoberfest theme is Ghost in the Machine. Unfortunately, I don't have enough time to...

The vibe coding argument has been running for over a year and has settled nothing, because it's a...

OpenAI 发布 GPT-6 Luna 系统卡 10 月 29 日增补版，针对 10 月 28 日版本更新能力评估。该版本在生物与化学领域被列为 High capability，生物模型拒答评估中严重提示词表现提升、双重用途提示词出现退步；网络安全评估与 10 月 7 日版本持平，无统计显著退步。

Automated detection has become the essential safeguard against realistic synthetic media, but deploying it at scale requires training on sensitive data from multiple organizations, putting privacy and the transparency demanded by forensic …

彭博社援引美国全国学生信息交换中心的数据称，今年秋季美国计算机与信息科学专业本科入学人数下滑超 12%，研究生降幅超 10%。两年间计算机科学本科在读学生减少 18.4%，研究生减少 27.1%；美国劳工统计局 8 月报告认为计算机与数学相关岗位受 AI 冲击最明显。全美高校秋季本科入学总体上涨 0.6%，社区大学本科入学小幅上涨 1.2%，最终报告预计明年 1 月发布。

Graph analytics generates irregular, topology-dependent memory-access streams via indirect traversals, limiting data prefetching. Accurate prediction enables timely prefetches that hide memory latency. However, ML predictors couple short h…

Somewhere between faster technology and smarter machines, we must never forget the ordinary people...

Grok API 新增自定义审核功能：在任意文本 API 请求中加入 guidelines 参数，即可约束 Grok 对特定内容类别的处理方式，Grok 会在生成回复时以自己的话遵循这些指引。该功能面向所有 API 客户开放，文档见 https://docs.x.ai/developers/advanced-api-usage/custom-moderation-controls。

What PredictSphere taught me about orchestration, data quality, adaptive learning, and the...

小米 MiMo 宣布 Xiaomi MiMo Desktop 会员权益升级，会员总体额度显著上调且加量不加价，调整实时生效并适用于现有及新订阅会员。会员用户不再只能使用 MiMo 模型，还可选择 DeepSeek V4 Pro、DeepSeek V4.1 Flash、GLM 5.3、Kimi K3、Qwen 3.8 Max 等第三方模型，各模型标注消耗倍率。

AI coding agents now do long, multi-step and multi-session software work with little supervision. When a developer comes back to that work, the useful question is not what the model said but what was actually done, by whom, and whether it …