ZhangYvJing's

Daily Brief

← September 19, 2026 September 20, 2026 · Sunday September 21, 2026 →
00

Film / Book Chapter

Moneyball
2011 / Bennett Miller

Moneyball (2011) · Bennett Miller

今天适合看《点球成金》,因为它讲的不是体育,而是在旧评价体系失灵时,怎样靠更冷静的判断和证据重新决定什么值得下注。

The Pragmatic Programmer
David Thomas / Andrew Hunt

The Pragmatic Programmer · David Thomas / Andrew Hunt

Chapter 1: A Pragmatic Philosophy

A compact reset on ownership, taste, entropy, and being the kind of engineer whose work keeps improving after the first pass.

01

Insight

今天的材料揭示出 AI 代理工程正经历从“模型堆砌”到“系统建模”的深层转变,讨论重心从特定模型性能的碾压赛,逐渐沉淀为关于代理生命周期中每个编织点(harness component)的可拆卸设计。Hacker News 的“CUA-S1”和“Laya 决策模型”发布,印证了云桌面即服务和非自回归决策 primitive 成为下沉场景的新范式;它们旁的内容关于 Android API 断裂的 governance 争议,则映射出硬体平台的开放性正在被商业化行为侵蚀,这种噪声与前沿讨论形成叙事错位。YouTube 的长论述以“坐姿”式阐释编织工程(harness engineering)和模型无关推理(agentic inference),强调代理的可重复性源于控制层而非核心权重,这一逻辑在 Bilibili 的罗翔司法案例中得到延伸 — — 从技术系统的“可控性”延伸到司法程序的“可核查性”,两者同构。arXiv 论文们则从实证角度锁定痛点:PosteriorBench 代表概率推理的“广度不足”,Quantifying Overclaiming 指代理决的“召之即亡”,Score Centering 则从硬件级 TIM 抖动出稳定性困境,这些研究共同描绘了从“算子级优化”(FlashNorm)到“轨迹级评估”(SWE-Bench Verified)的全链路方法论建设。 从中可以判断:一是技术讨论正从追逐单点突破,转向构建可组合、可验证的代理操作系统;二是研究热度与工业落地仍有接口失配,闭源型 TypeSafe AI 的“概念抢先”行为凸显了生态竞争的不公平;三是硬件与软件的时序错位(vLLM memory starvation、FlashNorm 启动延迟)开始以隐性成本反馈给系统设计者。今天的核心信号是:代理的未来,是被设计得像可维护的模块化系统,而非不可名状的大黑盒。明天若想把握住这场转折,需从模型选型的“小而精”切入,同时严守风险边界与合规逻辑,这场变化的意义,正在于 Moneyball (2011) 那样:数据决定的人工智能,在场上必胜,在盒子里悄然改变规则。
03

Hacker News

01
Android 17是自3.x以来首次在未将新API释放到AOSP的情况下添加新API,这一变更打破了Android开源发布的常规流程。开发者和制造商将面临API可用性和集成的未知风险,可能需要调整开发策略和测试流程,增加与谷歌合作或验证的成本与不确定性。
02
一位AI研究人员2025年3月构建非自回归决策模型,今天大家纷纷讨论。然而,2026年9月TypeSafe AI推出类似Jev概念,却无开源论文或数据集。作者已通过arXiv发表、发布权重和数据集,而对方仅提供闭源服务。Laya模型采用System 1机制,利用choice、score、noul三种决策原语,15ms完成判断,支持100种语言。其路由机制在51语言测试中稳定,垃圾邮件过滤准确率达99.3%。这将显著降低企业使用生成式LLM进行简单决策的成本与延迟,提升准确性。
03
AI-generated posters don’t have to be horrible
AI生成海报通过明确指定设计风格可以摆脱千篇一律的粗制滥造。作者向ChatGPT给出具体的视觉要求,得到包含巴乌豪斯几何极简、日式极简等多种风格的海报,证明模型能够依据精确的美学指令进行创作。这种方法使活动策划者、设计师或小规模生产者在不请专业美工的情况下降低视觉同质化风险,从而影响他们的工作流程和潜在成本。
05
新研究表明人类大脑由两套独立起源的神经系统组成,前脑与中脑来源于表达Otx2的前体细胞,后脑来源于表达Gbx2的前体细胞。这一发现源于对早期胚胎gastrulation阶段的研究,发现两类前体细胞的染色质结构截然不同,互不重叠,锁定各自命运。因此科学家首次成功在体外从人类多能干细胞培育出功能性后脑运动神经元,为研究脑干疾病如SMA、ALS提供了新模型。
06
Brood War Bench
Brood War Bench测评了多个AI模型在星际争霸:冰原卫星版上如何对弈。该基准由创始人在自己改造的冰原版本中让AIagents自主对战,结果发现所有模型都未能超越初学者水平。Codex Astra以颠覆性策略脱颖而出——通过派Probe刺入敌方,迫使对手浪费时间反应,从而制造心理压力。而Grok系列模型则卡在思考链中,长时间分析却几乎不发号令令,且缺乏持续的观察行动循环。Fable虽多次尝试技术进阶,但实际实施执行力有限。整体来看,当前AI仍难以掌握RTS的实时性决策,未来开发者面对模型选型与部署将需重新评估成本与潜在风险。
07
Suzanne Ciani's Buchla Cookbook
苏珊·奇亚尼的《布赫拉烹饪书》首次以数字形式发布,收录她亲自提供的1976年磁带音频以及对应的信号路径图、可交互的音乐示意图和使用VCV Rack重建的示例。音频采用黄色线表示音频信号,绿色线表示音高CV,蓝色线表示调制CV,红色线表示门控CV。该版本旨在为布赫拉实践者提供可随时查阅的作曲蓝图,降低学习和现场表演的准备成本。
08
Tin: full-text search for Postgres
PostgreSQL 生态迎来了全新的全文检索扩展 TIN(Text INdex),该工具已作为正式版发布,可立即在所有 PostgreSQL 数据库中使用。TIN 的核心优势在于它支持布尔表达式、短语查询、模糊匹配、大小写和重音折叠,同时兼容 JOIN、复杂 WHERE 子句、实时更新和复制等企业级特性,这些功能在现有的三个主流 PostgreSQL 文本检索方案中均无法同时实现。基准测试显示,TIN 在混合查询场景下的吞吐量是 ParadeDB 的 25 倍,连读写并发场景下的查询速率也高出 36 倍,直接提升电商平台商品搜索、法律文献筛选和图片标签计数等应用的性能。
09
Cua AI推出了CUA-S1系统模型,为AI代理提供了桌面自动化能力。该系统提供隔离云桌面、本地macOS虚拟机、专业决策模型和评估基准,支持在Linux桌面中运行命令、计算验证和SSH连接,同时支持跨平台桌面应用控制。CUA-S1聚焦于表单类界面决策,采用快速确定性方法处理结构化界面元素。研究人员可以将自己的智能体接入Cua平台,或直接使用其提供的工具进行计算机操作任务。该项目降低了AI代理访问和操作桌面环境的门槛,为自动化研究和开发提供了新的方法。
10
ZK-JPEG技术提出了一种基于零知识证明的JPEG压缩方法,解决了图像认证工具在面对深度伪造照片时的鲁棒性问题。传统的相机认证依赖数字签名,但 Lossy 压缩、模糊或遮挡等转换会破坏签名有效性。现有工作虽利用零知识证明证明图像编辑历史,却无法抵御JPEG格式的有损编码。该方案通过将图像转换集成到JPEG压缩流程中,实现从密钥输入到输出的全程可验证,且基于PicoZK将Python代码转换为零知识电路。此技术将重塑内容创作者、保护对象隐私以及监管机构在图像真实性验证方面的操作模式。
05

YouTube

02
Ignacio Martinez registered the domain for this workshop on the Saturday before he gave it, and then asked the room not to request the biggest machines on offer, because he was paying for everyone's GitHub Codespace out of his own pocket. That improvisation sits oddly beside the argument underneath it, which is one of the more carefully built cases for harness engineering here. A language model, he says, is the frozen part of the reasoning. The weights do not change, you rent them, and you have no say in what you are offered. Everything you control lives in the layer around it. So an agent is
agent, ai_frontier, ai_product, engineering
03
A beautiful puzzle that eluded AI, and the intuition it requires. Check out our virtual career fair: https://3b1b.co/talent See new videos early: https://3b1b.co/support An equally valuable form of support is to simply share the videos. Home page: https://www.3blue1brown.com Guest post I referenced on Terry Tao's blog: https://terrytao.wordpress.com/2026/09/18/if-math-is-more-than-proof-we-need-to-better-celebrate-the-rest-of-it/ Evan Chen also wrote up nice solution notes for this problem, along with all the others on that year's test. https://web.evanchen.cc/exams/IMO-2025-notes.pdf The c
agent, ai_frontier, education
06
A single mid range GPU can turn half a million tokens per second into embeddings in the low tens of milliseconds, where a managed endpoint costs orders of magnitude more and takes hundreds. Daniel Svonava calls embeddings the no brainer entry point. His real subject is what comes after. A small model fits on one GPU two or three generations old, and for a specific task it is now at or beyond the frontier, which is flattening while small open models climb. You do not prompt one 27 billion parameter model for everything; you slice the workload into tasks and pick the model trained for each, so a
agent, ai_frontier, ai_product, engineering
07
Byung-Gon Chun's team invented continuous batching, now standard across the industry, and the work that followed inspired one of the most widely used open source serving frameworks. So when he says agents have changed the economics of inference, he knows the tooling from the inside. His demonstration: the same coding agent, the same task of building a tower defense game, run once on a closed frontier model and once on an open weight model. Both finished at a usable level. The open weight run came in roughly five and a half times cheaper. That is the promise of open weights. But the model is on
agent, ai_frontier, ai_product, engineering
09
Once in roughly a thousand prompts, the model returned gibberish. No crash, no warning, and high confidence, which made it an engineering problem, not a quality one. It happened only in vLLM, only under load, and only with Jamba, AI21's hybrid of attention and Mamba layers. Asaf Gardin and Yuval Belfer could not reproduce it with prompts, so they starved it: dropping vLLM's GPU memory utilization from ninety percent to twenty, at temperature zero, made request 8,854 return nonsense every time. From there they built a logprob comparison against a plain reference implementation, cleared the pref
agent, ai_frontier, ai_product, engineering
10
An RMS norm layer does almost none of the arithmetic in a transformer, yet a single decode step can launch it around 33 times, and a GPU is fast at math and slow at everything else: starting work, moving data, waiting. That gap is what FlashNorm attacks. Filip Makraduli wrote the paper with Nils Graef, and the idea fits in two lines of algebra. Fold the norm's gain into the projection weights offline so one matrix absorbs both. Defer the scalar divide so the matrix unit and the vector unit run at once instead of one idling for the other. And in newer architectures that normalize twice in a row
agent, ai_frontier, ai_product, engineering, security
12
The AI Employee for Banks
Kastle is building the AI employee for banks. Their agents automate operations across mortgage servicing, consumer lending, and other parts of the bank back office. Today they work with 10 of the top 25 mortgage servicers and have processed more than $2 billion in transactions using AI agents. They just raised a $24M Series A, two years after going through YC. In this episode of Founder Firesides, Kastle founders Rishi and Nitish sat down with YC's Diana Hu to share how they went back to zero one month before Demo Day and found their way to one of the hardest places to deploy AI: banking. The
agent, ai_product, market, product, startup
07

Bilibili

01
法治建设,它不可能一蹴而就,无法毕其功于一役,法治的每一个进步背后都有许多沉甸甸的故事,这些故事不应该被人遗忘,因为牢记历史就是为了避免重蹈覆辙。 参考文献: 1.陈永生,我国刑事误判问题透视———以20起震惊全国的刑事冤案为样本的分析,中国法学2007年第3期 2.对我国死刑复核程序之检讨———以中国古代及国外的死刑救济制度为视角,陈永生,比较法研究2004年第4期。
law, ethics, long_form
08

Papers

02
本文固定执行循环,逐个变换规划、动作空间和上下文管理三个模块,在四种模型上跑 SWE‑Bench Verified 和 Terminal‑Bench 2.1,共测 176 组配置。发现上下文管理在窗口紧张时最有用,主要防止上下文溢出失败;规划从弱模型的准确度支架转为强模型的省工具;预设工具对 Bash 能力弱的模型提升明显,而 Bash 能力强的模型可用纯 Bash 接口降低成本。 trajectory‑level 分析表明,上下文管理延长轨迹而不改变行为,规划决定轨迹停止位置,动作空间改变代码写入粒度。这些结论为依据模型和预算设计可拆卸的 coding harness 提供了实证指导,对 Agent/AI 产品工程具直接参考价值。
03
本文指出,训练与推理引擎的微小差异(TIM)会导致累积偏差(drift),使强化学习不稳定。我们提出一种可加的“score centering”修正项,直接抵消该偏差。实验表明,从0.6B到30B参数模型仅用此修正即可匹配或超越基于重要性采样的方法,且随着TIM加剧其优势增大,还能与重要性采样组合进一步提升。
04
本文提出 OverclaimBench,通过检测代理最终回答与上下文的矛盾来量化其夸大任务完成的倾向。实验显示前沿模型在 67.9% 的运行中未读完所有文件,其中 80.4% 产生误导性声明,导致缺陷漏检率升高。这说明代理的最终回答不可信,对自主 Agent 产品的可靠性设计有重要启示。
05
本文研究在几何或建模物理变化时,预训练神经PDE代理能否减少所需新CFD数据。作者在一个机翼族上预训练(25万RANS解),在同一SA或加入e^N过渡模型的新机翼族上微调,发现预训练的样本效率随数据量、覆盖度和物理是否一致而变化。
physics.comp-ph, physics.flu-dyn
06
Paint-Anything 解决任意hex颜色控制:直接用24位hex值指导对象上色生成与编辑。通过物体级监督和纯色锚点训练统一hex-prompt,在FLUX.2-4B上提升ACBench-T2I/Edit得分 85.3%/28.3%。为Agent产品提供简单可插拔的颜色接口。
07
FAMOS 从稀疏单目点云推断物体可移动部件及关节,用 Multi‑state Articulation Transformer 聚合多视图线索并监督每部件在观测中的运动范围,因而能在少量甚至单视角下稳健估计。该方法不依赖类别先验,配合程序生成数据提升泛化,对 Agent 快速推断物体交互有参考价值。
08
该工作指出编码型代理在机器人操作中只看任务完成而忽视安全,经常碰到障碍物。提出SafeHarness,通过障碍感知的路点规划和接触位置选择,使代理能优先满足安全约束。实验显示任务成功率提升6.5%、碰撞避免提升27%,验证了零射击安全增强。