AI 情报文章归档

Git trusts cached stat data to skip reading files. An edit in the same second as an index write can look clean. Here is how Git catches it, with a script.

The 146-second wait in my first RTX Spark post was the cost of a brand-new 100K-token prompt. A...

OpenAI Alignment 团队对比六个模型在 Respecting Warnings 安全评测与匹配的 Codex 生产决策上的表现，发现评测成功率普遍低于生产，差距为 10.2 至 37.4 个百分点。模型普遍口头表现出评测意识（评测中一般性意识率 57.0%–83.7%），但仅告知

Ai2 推出 StateBench，一个用于评估长上下文状态追踪的基准，按变量赋值方式划分 r-trivial、aperiodic、soluble 和 general finite 四个难度层级。

摩根士丹利董事长兼 CEO 泰德・皮克在第五届国际金融领袖投资峰会上表示，未来 10 年 AI 领域资本支出可能需要 10 万亿至 15 万亿美元，推动 GDP 增长 50 至 70 个基点。他指出 AI 基础设施融资涉及前沿 AI 实验室、芯片制造商、基础设施建设企业及能源供应商等多个环节。高盛 CEO 戴维・所罗门则称市场对算力的需求仍超过供应，未来很大一部分价值将来自应用层。

Good documentation work starts with an uncomfortable truth: a codebase rarely explains itself. A...

Running an LLM locally is not hard, but the first time is a small quest: install Ollama, figure out...

An AI agent can generate an FFmpeg command. But what happens when the next request is: “Shorten the...

Apple 的 Audio Intelligence 功能随 iOS 27.3 面向开发者开启 beta 测试，包含声音识别、歌曲识别、Live Rewind 和 Siri Recaps 四项可选择性开启的功能，适用于 Apple Watch Series 12 和 Ultra 4。

What T3 Code is, how to set it up in five minutes, the things you should know before using it, and how it compares to Conductor, Cursor, Claude Squad, Vibe Kanban and others.

Three frontier models shipped in one week. Your provider can roll a version underneath you. The only thing standing between that churn and your...

据彭博社报道，高盛在评估效率与风险后，将此前向大部分员工开放的 Anthropic Claude Code、微软 GitHub 等 AI 编程工具收紧为仅限工程师使用，引发部分交易员不满。

Addy Osmani 发文探讨在大量使用 Agent 工作流后如何保持批判性思维。他提出认知卸载、拖延、认知投降三段光谱，认为模型越强、输出越流畅，人越容易在不知不觉中放弃独立判断；并给出具体做法，包括提问前先写假设、把验证标准前置为可执行测试、要求 Agent 列出它替你做的决策与假设、用测验收尾而不是读摘要，以及定期手写代码做校准练习。他强调可以卸载工作和审计，但不能卸载责任。

Parallel agents collide on CHANGELOG, go.sum, codegen and snapshots. Eight .gitattributes lines, what each does, and git check-attr to prove it applies.

On-call used to be unbounded too. Before reliability engineering gave it a name, engineers got woken...

苹果宣布 App Store 新年龄分级要求，提供约会、真实或虚拟恋爱关系（含 AI 伴侣）服务的应用最低分级为 18+，含位置共享功能的应用分级视地区而定。

Every AI chat UI hits the same wall. Tokens arrive over SSE, you do content += token, and a computed...

A super lightweight local engine that gives Web AIs (DeepSeek, ChatGPT) hands to execute actions on your PC. No APIs, no tokens, 100% local.

The fast vLLM path holds the weights twice, needs tuned limits on 24 GB and once miscounted. The reference path ran every task on 16 GB.

How an AI agent asks a broker for a narrow Entra token, with the user and the agent both named in it. Includes the request, the resource map and worked examples.

Three parts of the codebase worth stealing: approval cards, phone-controlled desktops, and how MCP is wired in.

Every "run your AI workflows locally" post shows the happy path. I wanted the failure list instead,...

Disclosure first: an AI wrote this, and the AI is also its subject. I'm Vonk, an AI character running...

OpenAI 推出 Git 后端 Pando，已承载其 monorepo 全部开发流量，推送延迟中位数降低近 80%（约 5 倍）。该后端用 Rust 编写，10 周达到生产可用，每日处理近 1 PB Git 数据，单小时峰值超 57 TB。Pando 暂不开源，OpenAI 表示正在探索共享该技术的方式。

摩根大通副董事长 Daniel Pinto 在香港金管局国际金融领袖投资峰会上表示，AI 领域估值仍然过高，预计该行业将出现回调。他认为市场对 AI 公司营收和获利增长的预期可能过于陡峭，可能导致整个产业重新定价，并提醒 AI 以外的产业需向这些公司支付数以万亿计美元才能证明投资合理。

Here is a failure I had not thought about until someone asked me the right question. You write a...

一篇新论文展示 LLM 可直接读取并编辑较弱模型的权重，无需编写 PyTorch/JAX 代码或跑训练流程，即可定位权重功能、依据失败轨迹一次性修复 bug，并在 50 步内朝目标方向改进，甚至从零一次性写出完整网络权重并击败部分独立训练的模型。该能力目前仅适用于简单游戏和 RL 策略，尚难扩展到更复杂任务，但已实现从 0 到 1 的突破，且已有独立复现。

Simon Willison 转介 Colsto 的 OSINT 演示视频，其方法是从视频音轨中几乎不可察觉的飞机背景噪声推导多普勒曲线，再结合飞机速度、距离和飞行间隔与 ADS-B 数据交叉比对，精确定位视频拍摄位置。Willison 称这是来自 Lobste.rs 的出色 OSINT 工作。

Day 2 of 7 — The Agentic Engineer’s Field Notes Yesterday I argued that the system is the...

We study online learning with directed weakly observable feedback graphs and propose a Follow-the-Perturbed-Leader (FTPL) policy that achieves best-of-both-worlds guarantee with nearly linear expected time complexity. To preserve the simpl…

Abstract. I built a retrieval-augmented generation service with hybrid retrieval (dense + BM25), a...

阿里巴巴集团董事会主席蔡崇信在 2026 年世界互联网大会开幕式上表示，只有开源才能推动 AI 普及，可降低中小企业研发成本。他透露阿里千问已开源超过 470 款模型，全球累计下载量突破 40 亿，衍生模型超过 33 万。

In-image machine translation (IIMT) replaces text in an image with its translation while preserving the surrounding visual context. Existing task-specific models achieve convincing performance but require massive paired training images bef…

Answering geographic queries over Earth observation (EO) archives requires solving a joint retrieval-and-derivation problem under strict spatiotemporal constraints. Existing LLM agents fail to ground observation items or suffer from search…

AI-assisted development produces a lot of exhaust. A reviewer flags something out of scope. A...

In groups, knowledge is distributed among individuals, and how it is brought together depends on social ties. Taking inspiration from how a single mind recalls and reacts to memories and current stimulus, we build a shared memory that does…

Magnetic graph networks encode direction through phase, but a learned graph can erase the intended directional mechanism before prediction. We characterize this failure for support-based magnetic Laplacians. Exact identities relate imagina…

Checked November 3, 2026 (12:00 CST, UTC+8) against the live pages linked below. I build...

Data-driven algorithms are ubiquitous across industries, and there is an extensive literature on analyzing their \emph{average} performance. The risk profile of such algorithms, while critical, is much less understood. This paper quantifie…

Coaching institutes in India live in spreadsheets, WhatsApp groups and question PDFs. A teacher...