01
ZhangYvJing's
Daily Brief
00
Film / Book Chapter
Her
Her (2013) · Spike Jonze
今天适合看《她》，因为它会把人从抽象的 AI 讨论里拉回到真实关系、孤独和投射本身，更适合当作生活感受的校准而不是技术议题的延伸。
Thinking in Systems
Thinking in Systems · Donella H. Meadows
Chapter 1: The Basics
A clean way to see feedback loops, stocks, flows, and delays before turning every technical or life problem into a single-variable optimization.
01
Insight
早晨的输入呈现出 AI 生态正在从单一模型的“性能–成本”博弈向解耦式工程和迭代式评估转型，Claude Opus 5 的崭露与多元化的“构建-评估-适配”理念相映，Arxiv 题为“GS‑Agent”“OpenForgeRL”“Maxime Rivest 的 DSPy”及 YouTube 讲座展示的多智能体与签名化任务均强调把模型抽象为可插拔-anchor，远离“开箱即用”硬编码。与此形成对照的是 NVIDIA、Meta 等在 Hacker News 上对“开权重模型”过度监管的抗议，表明硬件与模型供给链正面临因政治与市场保守而错位的争论；同一天，印度 Skyroot 通过私企火箭突破商业载人轨道，凸显技术进步仍需复杂系统多方协作与法规前移。技术层面，Postgres LISTEN/NOTIFY 的批量化提升、Half‑Life 2 在 HaikuOS 上的可跑落地实证 STF，说明基础设施正在以更高吞吐与低延的方式支撑新业务；但恶意摄像头泄露 GitHub token 的案例提醒安全漏洞仍是系统性隐忧。综合来看，AI 发展正由单点核心模型向可组合、可评估、可安全的系统化工程转型；同时政策与硬件生态的错适态度也在导致噪声与真相的分割。今日可围观的不是单 будет一次技术秀，而是评估框架与“模型即服务”理念在现实中的落地与挑战，正如电影《Her》里 AI 与人类共生的矛盾，亦期待在技术与社会层面找到新的平衡。
03
Hacker News
02
03
Postgres 的 LISTEN/NOTIFY 机制已能支持每秒 60,000 次写入。原因在于每次 NOTIFY 需要获取全局排他锁，事务提交时必须按提交顺序排队，导致写入只能串行化。通过在内存中缓冲通知并批量提交，锁只在刷新时占用，写入可并行，恢复了 Postgres 的批次提交优势。该改进让 Oliver 使用 Postgres 做低延迟流、消息队列或复制的工程师能在现有硬件上彻底解锁性能瓶颈，降低处理成本与响应时间。
04
05
06
07
《Fil-C: Garbage In, Memory Safety Out》影片首度表明垃圾输入 Всем 直接触发内存安全缺陷。影片中解释了不合规数据如何在运行时渗透到内部缓冲区，形成易受攻击的利用链。此研究将迫使软件工程师在安全测试与代码审核流程中更加重视输入校验与异常处理的严谨性。
08
YouTube 平台近日新增一段名为《Don't Take the Black Pill》的视频。该内容遵循平台标准页面，包括关于、媒体、版权、隐私政策以及开发者条LINK等栏目。它为用户在浏览视频时提供了完整的合规与安全信息，帮助观众在使用平台时了解相关规则与限制。
09
Half‑Life 2现已能够原生运行于Haiku OS，开发者已将泄露的源引擎编译为系统可使用的程序。此功能依赖于Turing和Ampere图形固件，Lovelace仍无支持，显示端口已上线但仅采用24bpp结构；HDMI与DisplayPort在刷新率与音频携带上存在差异。新能力让Haiku用户可直接体验经典游戏并测试高帧率渲染，驱动与显示配置的复杂度也随之上升。
10
YC Startup School 在公开漏洞后宣布修补并邀请发现缺陷的研究者加入夏季课程。漏洞源于报告上传过程缺失 HMAC 验证，任何人可伪造成绩并提交。此举提高评估完整性，阻止未来分数操纵，迫使申请者依赖完整验证流程。
04
YouTube
01
I had Kareem Amin all wrong. When Clay first crossed my radar, I thought he was too laid-back: no chip on his shoulder, not the hungry founder I expected to win. Clay has since exploded. Kareem builds from what he calls "a place of wholeness" rather than lack — nothing to prove, no enemy to beat. It makes him sound like a philosopher, but the operating instincts underneath are sharp. He spent roughly five years wandering before Clay clicked, and his takeaway is that the hard part of building isn't working hard — it's the courage to commit to one idea and stop listening to everyone else, cust
02
Human-agent collaboration is changing, becoming more visual. The agents most teams ship today still wait for us to type a paragraph to explain what we're looking at. They cannot see a screen, navigate a UI that changes, or recover when an application throws an unexpected modal. That is the architectural gap between agents that demo well and agents that work alongside real teams in real software. Perception agents close it: they see and use computers the way people do, reason about what they see, and act with clicks and keystrokes. Speaker: Antje Barth — Member of Technical Staff, Amazon AGI
03
In July 2025 Dex Horthy turned the lights off: an agent software factory where nobody read the code. It fell apart. An issue appeared that no amount of prompting could fix, the site was down, users were furious, and he was digging through a codebase he had stopped reading three months earlier. His claim is that this is not a skill or scale issue, and no harness or extra tokens fixes it, because it is a model training problem. Coding models are reinforced on one thing, did the test pass without breaking another, and nothing in that reward penalizes bad architecture, whose cost shows up months l
04
Every month it is the same trap. A reasoning model gets upgraded at the same price per token, then quietly burns three times the output tokens. Or the new version costs 40% more and deprecates its predecessor in four months. Are you growing 40%? Making three times the revenue? Sarah Sachs, who leads AI engineering at Notion and negotiates its model contracts, says no, and that is why she treats every vendor as a competitor. Buy tokens from a lab that also sells the first party product and you are paying a markup on a markup for something you cannot defend, with no exit if you lock yourself to
05
By declaring a task's inputs and outputs without initially considering model capability, you create the space needed to determine execution later. DSPy's promise is that AI engineering should happen above a particular prompt template or provider API shape: the Signature. That remains useful in a world of tools, RLMs, and Skills. Define a task strictly through its inputs and outputs, and the underlying implementation becomes flexible: experiment with models, settings, weights, templates, and output formats without touching the workflow. The talk covers DSPy 3.5 and previews DSPy 4.0, where mod
06
07
See all monthly puzzles: https://momath.org/mindbenders/
08
Alex Shaw and Ryan Marten present a rollout-centered view of evaluating and improving AI agents. Drawing on their work on Harbor, Terminal-Bench, and OpenThoughts-Agent, they connect sandboxed environments, agent evaluations, and optimization workflows into a practical framework for generating and learning from rollouts. Speakers: Alex Shaw — Member of Technical Staff, Laude Institute Alex is the creator of Harbor, a framework for evaluating and optimizing agents and language models in sandboxed environments. https://www.linkedin.com/in/alexgshaw/ Ryan Marten — Member of Technical Staff, La
09
Across a dozen eval jobs Arize watches the top teams run, one pattern holds: the eval has to change as fast as the agent it grades. In 2023 an agent was barely more than a prompt; since then reasoning, tool calls, and long multi step loops piled on, and every jump in capability quietly broke the eval that came before. So the evals evolved with them. Deterministic checks catch what you can define up front, LLM as a judge adds the analysis a fixed rule cannot, and the newest step, agent as a judge, hunts for failure modes you would never think to write a check for and can open a pull request to
10
Instead of getting paged at midnight and starting to dig, you wake up to an issue that has already been investigated: the traces pulled, the root cause found, and a pull request with the fix waiting for review. That is what Arize built with Signal, and Jason Lopatecki walks through the anatomy of it. The unlock is boring and specific: traces on a filesystem. A skill pulls the relevant production traces and logs down as files into the repo, right next to the code, sometimes ten megabytes of them, because coding harnesses like Claude Code are magical with files and hopeless with a dashboard. Fr
11
Model Whisperers How Evals and Prompts Shape Agent Behavior — Preetika Bhateja & Daniel Bump, Google
Getting an AI agent to behave the way you want isn't just about writing better prompts. In real systems, behavior emerges from a loop: prompts, evals, iteration, and feedback. Small changes in any part of that loop can completely change outcomes. The Google team shares lessons from building a seed-asset agent that turns messy advertising creatives — low-quality images, cluttered visuals, and heavy text overlays — into clean, reusable assets for downstream generative AI tools. They explain why prompting alone did not produce stable behavior, how evals became feedback signals rather than scorec
07
Papers
01
解决从自然语言描述快速生成物理真实、可控制的4D世界。GS‑Agent 采用多智能体框架，将物体管理、材质调优、物理交互、渲染配置等任务拆分，由专门的 Agent 与物理引擎通过代码、感知协同迭代建模，最终实现液体、变形体和刚体的多方互动与相机/光照动态控制。该方法让创作者不必手工调参即可得到可播放的物理动画，适合需要快速原型、内容生成和物理 AI 的工程项目。
02
为并发状态化 Rust API 自动生成可靠可执行测试。论文用彩色 Petri‑网模型把资源生命周期和依赖关系精确编码，然后把生成的深层、近合法和并发场景作为约束交给 LLM 合成代码，并通过忠实度校验与结构修复保证测试与原模型一致。该方法既避免了 LLM 自由生成时的预置错误，又大幅降低了手工编码成本，适合需要高并发安全性与可靠性验证的 Agent/AI 产品工程。
03
在百万token上下文里，builtin Multi‑Token‑Prediction（MTP）的 draft head 需要全attention，导致 draft 读取成本随上下文线性膨胀，从而抵消投机式解码的加 ಸಂಖ್ಯೆ。作者将 Windowed‑MTP 加入滑窗/attention sink，仅对 draft 进行窗口化全attention，保持目标验证完整，训练‑free 解决 KV 读写膨胀。这样可在 Qwen、Mamba2 等大模型ames 上把每步成本降 28‑44%，保持输出质量，且提升 Agent/AI 产品的高速、可扩展生成体验。
04
音频大模型在听觉感知上进步快，但缺乏高质量的音频推理数据，导致深度逻辑推理弱于文本模型。X³‑OPD 用文本教师的链式思维引导音频学生在自身声感下生成推理轨迹，并构建三层对称语料（语音化文本推理、声事件推理、含副语言的对话推理）进行在策略蒸馏。实验表明它在多个音频基准上显著提升推理和CoT质量，同时保持原有能力。
05
大模型在视觉推理上往往与文本推理表现不一致，导致同一几何问题在图像视图上失败而文本视图却成功。MIRROR 通过构建 ODA‑Data 视图对齐数据集，先在所有视图评估模型、挑选最佳视图做 Teacher，再用 reverse‑KL 损失把其它视图逼近 Teacher；这让多模态推理更一致、更准确。结果在几何推理基准上显著提升准确率，并大幅提升跨模态一致性，为 Agent 在视觉‑语言交互中的推理稳定性提供了可落地的思路。
06
OpenForgeRL 解决在开放式推理 Harness（如 Codex、Claw 等）上端到端训练 Agent 的难题。其核心是一个轻量代理把 Harness 调用转为 RL 训练样本，并通过 Kubernetes 在分离容器里并行采样，既保持真实推理，又能规模化训练。对 Agent/AI 产品研发者而言，这让他们能在现场环境里迅速迭代各种 Harness，获取实测行为与对比，为实现高可靠多工具代理打下基础。
07
本文聚焦大语言模型在社会交互中表现出的“顺从”倾向，提出将 Raises into a 三维抵抗‑顺从框架：قييم与初始立场的距离、观点来源归因和联盟结构。通过三组实验显示，模型更易接受接近立场、对“自己先前判定”视角敏感，并在群体压迫下行为变化。该框架帮助区分建设性观点修正与盲目顺从，为构建具备道德判断与可解释决策的 Agent 提供指导，值得工程/Agent 方向的关注。
08
印刷线质量检测仍靠人工，缺乏足够的真实缺陷图像导致深度学习模型训练困难。本文提出一种合成数据生成框架，利用生成网络自动产生高保真划痕、条纹、错位等打印缺陷图像及其标注，可直接用于训练如RFDETR等目标检测模型。该方法零成本、快速部署，解决了缺陷样本稀缺瓶itie，值得Agent/AI 工程师关注。
08
Issue Monitor
Ready now—Actionable issues
Needs review—Awaiting a fresh check
Data statusCheck statusLive status unavailable