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 Mom Test
The Mom Test · Rob Fitzpatrick
Chapter 1: The Mom Test
A short practical check on product conversation: stop asking for validation, start extracting facts, and keep reality from being softened by politeness.
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 Ten advances in mathematics and theoretical computer science；Hacker News 的 Devtools must be open source；Hacker News 的 Wind and solar overtake fossil fuels in Germany for the first time；Hacker News 的 Celebrating 45 Years of Kermit with the First New C-Kermit Release in 15 Years；Hacker News 的 MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video；Hacker News 的 Smaller, faster, safer: running Kimi and GLM at scale。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
03
04
05
06
07
08
09
10
04
YouTube
01
Every time a model launches there is a gap between the benchmark numbers and what the thing can actually do, and Nick Heiner argues the existence of the word benchmaxxing is the tell. When labs openly brag about scores, teams stop asking whether a benchmark reflects reality, and the whole field drifts into an avalanche of numbers that measure the wrong thing. His talk is a field guide to reading a benchmark fairly, starting from the antipatterns that quietly break them. The failure modes are specific. A large share of tasks in a typical benchmark are simply broken; contamination means models
02
You invoke a tool and expect an answer, but real work takes time, and over that time connections drop, networks blip, and processes crash. Cornelia Davis, a distributed systems veteran who wrote the book on cloud native patterns, argues that this is exactly the gap the MCP tasks specification exists to close, and walks through why almost no agents support it yet. A task lets a tool run long, report progress, and pause for human input without losing its place, which means the interaction has to be durable: it survives the client disconnecting and picks up right where it left off. She demonstra
03
Chat and coding assistants still hand you walls of text when a button, a chart, or a small interactive view would say it faster. Liad Yosef, who co created MCP UI, walks through MCP Apps: a way for an MCP server to return a real interactive interface instead of a block of text, built on the MCP UI project he started and now shaped through an open working group in the MCP committee. A tool call links to a registered resource, the host renders it as a web component, and clicks flow back into the agentic loop, so the same funnel that would take a paragraph to explain becomes something you can see
04
05
Full episode: https://www.youtube.com/watch?v=QbdbAhaJoCQ Me on twitter: https://x.com/dwarkesh_sp
06
# Fireside Chat: Gergely Orosz × Simon Eskildsen **Location:** Main Stage **When:** Day 2 - June 30, 2026 · 12:30pm-1:30pm ## Speakers ### Gergely Orosz Author / Founder, The Pragmatic Engineer · The Pragmatic Engineer [X/Twitter](https://twitter.com/gergelyorosz) · [LinkedIn](https://www.linkedin.com/in/gergelyorosz/) · [Website](https://pragmaticengineer.com) Software engineer, engineering leader, and author of The Software Engineer's Guidebook; best known for The Pragmatic Engineer newsletter and blog covering software engineering practices, engineering leadership, and the tech industry
07
Jerry Tworek, co-founder of Core Automation and former VP at OpenAI, on a quiet assumption in AI research: prove your architecture at small scale first, then earn the compute to scale it. But reinforcement learning showed almost nothing interesting until it crossed a compute threshold — and if that's true of RL, other architectures may have been written off before they were ever given enough compute to show what they could do.
07
Papers
01
论文指出，单一共享 LoRA 在异构任务序列上易产生干扰与灾难性遗忘。作者提出自动多策略 PEFT 框架，通过任务分组与排序在固定参数预算下，将兼容任务分配到独立的 QLoRA 适配器，形成解耦的优化路径，既保留正向迁移又避免冲突。在 TRACE 基准上，该方法在相同可训练容量下达到 44.78 的最高分，显著优于传统单策略 PEFT。
02
这篇论文解决了“如何从短短20秒的双人对话中判断两人是熟人还是陌生人”这一社交推理难题。作者构建了FriendBench基准，收集文本、音频、视频三模态的对话样本，并让26个大型模型与人类评审在96对平衡样本上竞争。结果显示，顶尖模型与人类在准确率上几乎无差，但模型更倾向“陌生人”，而人类保持更均衡；这说明多模态信息对模型帮助有限，而人类能从可见行为中获益。对想要让Agent在社交场景中快速判断关系的工程师来说，FriendBench提供了可复现的评测框架和对比数据，值得一看。
03
在向量检索中，固定搜索参数往往导致召回率波动大、成本高。QASP 通过一次监督回归预测每条查询的完整召回曲线，随时生成满足任意召回目标的搜索策略，并在检索过程中根据预测与实际偏差轻量级调整深度。它在不需要重新训练、跨数据集、跨索引配置的前提下，显著降低召回方差、提升查询满意度，并在大规模数据与层次索引上实现 99% 召回时减少 80% 数据访问，正是 AI 产品工程师追求高效、可控检索的理想方案。
04
本文探讨在模仿学习中，何时主动与专家交互能弥补模型表达不足。作者提出 OVI，一种交互式 on‑policy IL，核心思路是只需逼近专家的 value function，而非完整 policy，配合线性最大化 oracle 实现统计与计算双重效率。实验表明，当学习者网络表达力低于专家时，OVI 能显著优于 BC、DAgger 及离线 value‑based 方法，提示工程师可通过交互降低对模型复杂度的依赖。
05
解决 LLM 代理在实验优化中的“读懂实验、改进实验”能力缺失问题。作者构建 AgentHPOBench：30 个可执行 ML 任务，按顺序给代理提供基线结果、日志与指标，要求代理逐步提出新的超参配置。通过统一协议评测 12 种主流代理与传统 HPO 基线，发现现有代理能做实验优化但在持续迭代、日志诊断和稳定提升方面仍有限。对想把 LLM 变成可持续实验的 Agent/AI 产品工程师来说，这套基准能快速定位代理的瓶颈并指导改进。
06
07
论文研究在局部光滑、曲率和分离条件下，如何在严格的差分隐私保障下估计多元分布的密度模式。作者提出 DP‑GRAMS，基于均值漂移思想，先用高阶核函数构造无偏分数估计器，再在梯度上升阶段加入梯度裁剪和高斯噪声，并通过公开网格初始化和抑制策略实现多起点联合释放。该方法在保证 (ε,δ)-隐私的同时实现近乎最优的估计误差，并自然扩展到私有模态回归和聚类，对需要隐私保护的 Agent 特征学习或数据聚具有一定参考价值。
08
企业流程越来越依赖 Agent 做 schema‑guided extraction，核心是给定文档和用户自定义 schema，准确提取字段并给出来源证据。ExtractBench 通过 4,869 页、370 文档、8 业务域、67 文档类型，结合真实文档、合成列表和人工校验，提供价值准确率、记录完整性、词/页级 grounding 以及成本等四项指标的统一评测。对工程师而言，它既能快速定位模型在长文档上的截断问题，又能比较 Agent 与编码方案的成本/精度平衡，LlamaExtract Agentic Plus 便是典型案例。
08
Issue Monitor
Ready now—Actionable issues
Needs review—Awaiting a fresh check
Data statusCheck statusLive status unavailable