01
ZhangYvJing's
Daily Brief
00
Film / Book Chapter
Paterson
Paterson (2016) · Jim Jarmusch
今天适合看《Paterson》,因为它更像一次生活和判断方式的校准,能把注意力从持续输入里稍微抽出来,重新放回你真正想怎样生活和做事上。
The Staff Engineer's Path
The Staff Engineer's Path · Tanya Reilly
Chapter 2: Three Maps
Good for calibrating work beyond code: where influence actually travels, which systems matter, and how to avoid mistaking activity for leverage.
01
Insight
今天的输入更像几股不同语气的材料同时挤在一起:社区链接在暴露工程和产品环境里的真实焦点,长视频在把这些焦点放回更完整的语境里,研究材料则提醒人热度和可落地性并不总是同一件事。如果先不急着做结论,至少可以把这几条线索放在一起看:Hacker News 的 Claude Fable 5.1 and Claude Mythos 5.1;Hacker News 的 The ChatGPT/Codex app bundles a full copy of LibreOffice;Hacker News 的 Refurbishing a Tektronix TDS7104 Oscilloscope;Hacker News 的 Launch HN: Nori Robotics (YC S26) – A low-cost humanoid robot for development;Hacker News 的 AnkiDroid: Google Play no longer allowing Open Collective donation link;Hacker News 的 Ambient CSS v3 – Blender meets CSS。真正值得注意的不是单条内容本身,而是它们共同指向了什么、彼此漏掉了什么。
03
Hacker News
02
OpenAI Codex 桌面客户端(现称为 ChatGPT)在本地缓存中包含 LibreOffice 办公套件及其依赖的 Python、Node.js、Poppler、git 二进制文件。这些文件位于 /.cache/codex-runtimes/codex-primary-runtime/plugins/openai-primary-runtime/plugins/documents 目录下,随附的技能插件指导模型如何定位和调用这些程序处理文档。这样使得依赖该运行时的用户可在离线环境下直接使用办公功能,但也会增加磁盘占用和安全更新维护成本。
03
04
Nori Robotics 发布了一款低成本人形机器人,面向开发者提供可在家中训练并共享技能的平台。该机器人通过可教学的技能市场,让用户自行教授日常家务操作并在全球技能库中发布。面向个人用户与开发者,可在此基础上降低机器人开发门槛,同时需要注意安全使用及可能的数据隐私规则。
05
06
07
08
09
Hacker News 定期更新「Ask HN: Who is hiring?」招聘帖,9 月 2026 版已发布。帖子要求明确标明工作地点,REMOTE 表示远程,REMOTE (US) 表示仅限美国,ONSITE 表示必须到岗。发布者必须是公司方主动发帖,禁止招聘机构或工作板块,单家公司限发一条,若公司不为大众所熟知需说明业务。同时要求申请者积极回复,评论区禁止抱怨。读者可通过 nthesis.ai、dheerajck.github.io、nchelluri.github.io、emilburzo.com 等站点检索,还留有「Who wants to be hired?」求职线程供参考。该规则影响求职者筛选成本,雇主招聘流程合规性,增强平台信息纯净度。
10
04
YouTube
01
Full episode: https://www.youtube.com/watch?v=aV26V1UvkJw Me on twitter: https://x.com/dwarkesh_sp
02
03
04
Most public safety technology companies grow by collecting more data. Peregrine inverted the model: no sensors, no new data, a business built on connecting the data and information cities already own. Co-founders Nick Noone and Ben Rudolph received more than two dozen no's before San Pablo PD let them in the door in February 2018. Today, Peregrine powers law enforcement, emergency medical services, fire and rescue, and other services in more than 400 cities and communities globally. Nick and Ben explain their north star for data sovereignty, and discuss how Peregrine's philosophy and privacy-f
05
Between the moment an x402 server verifies a payment signature and the moment it settles on the blockchain, nothing stops the buyer spending that money somewhere else. Jan Curn's point is that a client can mint a thousand signatures against one wallet, so any seller starting real work on the strength of a verification is exposed. The workaround is to do the work only after settlement, fine for a fixed price API call and awkward for anything slower. Curn is bullish enough on x402 to have shipped it two days before this talk, adding 20,000 Apify tools to a marketplace that previously carried abo
06
Pranav Maheshwari ran one prompt in two terminals: find the email address of the head of crypto and blockchain at Mastercard. The terminal without a payment skill file returned the company's email format and an invitation to work the rest out. The one with it paid a fraction of a cent to a metered endpoint and came back with the address, the person's location and their handle. That gap is the whole argument. Most MCP servers are free today, and his claim is that the useful ones will not stay free, which leaves an agent only as capable as the tools it can pay for. Rodrigo Coelho sets up the lon
07
Gus Iwanaga opens by disowning his own session title, then shows the demo that did not work. His team asked their system for a sales report for Q1 four times over. It returned four different layouts: different KPI cards, different charts, different amounts of text, and copy that drifted from Q1 in one run to January and March in another. He would not ship it. None of that was a model failure. It was the consequence of handing a model a component catalog and asking it to compose the experience, which is precisely what his team had done. The rest is what they built instead, at a company carrying
08
Full episode: https://youtu.be/X50zezLFWWI Me on twitter: https://x.com/dwarkesh_sp
09
Two hours before David Levine took the closing slot of the conference, the West Virginia Secretary of State wrote back to confirm organization number 62847. The paperwork had gone out by FedEx. What it registered was a DUNA, a decentralized unincorporated nonprofit association, under a law that had taken effect the day before, and Levine's claim is that an organization composed of agents now has legal standing. It can own property, enter agreements, open bank accounts, raise capital and be held responsible in court, with the catch that it cannot distribute profits to members without turning th
10
Wafer (S25) is building a fast AI inference cloud that runs open-source models at industry-leading speeds by using agents to optimize GPUs at every level of the stack, from custom kernels to speculative decoding. The result is the low cost of open source with the latency of a much smaller model. In this episode of Founder Firesides, YC Managing Director Jared Friedman sat down with Wafer co-founders Emilio Andere and Steven Arellano to talk about the college side project that became the company, how they got GLM 5.2 running two to three times faster than anyone else in the market, and why spe
07
Papers
01
02
03
工业级LLM后训练像维护遗留代码:在固定算力和混合预算下,只能用有限混合补丁改进而不退化旧行为。论文把后训练混合视为dataware,通过提升teacher distillation转化率(2.84×)和yield导向的补丁,使CodeForces pass@1提升+2.59、LiveCodeBench v6提升+6.11,表明需要一种编程dataware的工程纪律而非一次性配方,对Agent产品的持续改进有参考价值。
04
05
06
传统神经网络混合效应模型需手动推导似然和梯度,限制模型复杂度。本文利用TMB自动求导和Laplace近似,仅需提供联合对数似然即可自动求边缘似然及精确梯度,提升灵活性和精度。对构建Agent的层次模型有启发。
07
本文提出OntoAligner-Ensemble,通过可配置的两阶段投票融合与后融合选择,统一整合lexical、结构、KGE、LLM等不同范式的对齐器。实验表明融合在精确率和召回率间取得更好平衡,常优于单一对齐器;异质集合提升精度,同质LLM集合提升F1,为Agent系统选择对齐策略提供指导。
08
08
Issue Monitor
Ready now—Actionable issues
Needs review—Awaiting a fresh check
Data statusCheck statusLive status unavailable









