01
宣布MAI‑Code‑1‑Flash已经在 VS Code 的 GitHub Copilot 个人用户中上线。该模型采用端到端训练、适应性响应长度控制，并在真实 Copilot 工作流中评估，能在简单请求时保持简洁、在复杂任务时投入更多推理预算。对开发者而言，可在更少 token 下获得更高通过率，降低延迟和费用，同时提升在实际编码、工具调用等代理场景的可靠性。
ZhangYvJing's
Daily Brief
00
Film / Book Chapter
The Matrix
The Matrix (1999) · The Wachowskis
《黑客帝国》提醒你，在被 AI 工具、预算限制和不断弹出的智能建议淹没的工作日里，先停下来审视自己真正的假设与选择；只有当你主动拔掉那层“矩阵”式的默认输入，才能重新定位自己的判断标准和行动方向。
Working in Public
Working in Public · Nadia Eghbal
Chapter 3: The Structure of an Open Source Project
A sharp chapter for thinking about visible work, maintainer labor, contributor flows, and why software ecosystems are not just code repositories.
01
Insight
今天的材料透露出两条核心趋势：一是 AI Agent 正在从“代码助手”向“全栈业务协作者”快速升级，二是对模型安全、成本与可解释性的约束正从技术讨论转向硬核运营规则。Microsoft 连发 MAI‑Code‑1‑Flash 和 MAI‑Thinking‑1，强调端到端训练、长上下文与函数调用，显示公司把模型的算力效率当成产品竞争点；而 GitHub Copilot App 把智能体直接嵌进 issue‑to‑merge 流程，说明把 Agent 融入研发生命周期已不再是实验，而是降低切换成本的实际需求。与此同时，Uber 通过每月 1500 美元的 AI 费用上限、Gmail 强制弹出 AI 建议以及浏览器层面的 Attribution Level 1 归因系统，展示了企业和平台在成本、隐私与监管压力下开始对 AI 功能“硬删”。这些限制与前述 Agent 扩容形成对照，暗示技术供给与运营需求的错位正在加剧。跨渠道的印证在于：a16z 与 Fivetran、Sequoia 的对话都在提醒，Agent 需要可靠的上下文和数据治理，正好呼应了 Open Repair Data Standard 对统一维修数据的标准化尝试——在分散的硬件/软件生态里，只有统一数据模型才能让 AI 可靠地读取、分析并执行；而 arXiv 的几篇论文（如 “Tracking the Behavioral Trajectories of Adapting Agents” 与 “SimSD”）则提供了轻量化行为监控和投机解码的技术路径，正好能帮助企业在成本上限内实现安全、可解释的实时推理。综合来看，AI Agent 正在从工具向业务层面渗透，但其增长必须在运营约束、数据标准和可审计的行为追踪之间找到平衡，才能避免像 Gmail 那样的用户反感或 Uber 那样的预算失控。今天阅读这些信息，最好把注意力放在：① 评估你所在团队的 AI 使用是否已经进入必须统一数据与安全审计的阶段；② 关注那些提供“行为特征向量”或“投机解码”即插即用的底层技术，它们可能是把 Agent 成本压到预算上限以下的关键。顺便提醒，早晨的思考可以先看一部《The Matrix (1999)》，给今天的 AI 复盘加点哲学味道。
03
Hacker News
02
用户因 Gmail 网页版频繁弹出未经请求的 AI 摘要、自动回复和写作建议，感到被低估并决定迁移邮件；这些功能虽可选，但部分不可关闭且伴随必须关闭传统分类等老功能，暗示为提升模型使用数据而强制介入；因此依赖 Gmail 的专业人士需评估潜在的工作流干扰、额外学习成本和数据迁移风险，或转向更可定制的邮件服务。
03
MAI‑Thinking‑1 正式发布，作为中等规模的推理模型，它在软件工程基准、数学推理以及盲测对比中均超越同类并胜过 Sonnet 4.6。模型全部基于企业级、干净且已获授权的数据从零训练，未使用第三方蒸馏，且在全栈自行设计的“爬坡机器”管线中通过强化学习同步提升能力与安全。凭借 256 k token 长上下文、函数调用和企业安全合规，MAI‑Thinking‑1 能在更小算力占用下进入日常开发与文档处理流程，降低企业部署成本并提升代码辅助的可靠性。
04
Uber对员工使用AI设定每月1500美元上限，并通过内部仪表盘实时监控，少数情况可获批准超额。该措施源于公司在四个月内耗尽全年AI预算，当时内部鼓励尽可能使用AI并将使用量放在排行榜竞争，导致支出失控。此举将迫使工程团队在使用Claude Code、Cursor等工具时更加审慎，提升费用可视化，进而影响开发节奏和创新决策。
05
开放维修数据标准（ORDS）发布至0.3版，统一了小型电器和电子产品维修数据的收集与共享方式。该标准通过定义产品、维修和会话三类字段并规范取值，旨在解决各组织收集格式不一导致的聚合困难，并在每半年发布一次合规数据集以便全球趋势分析。维修社区、零部件供应商和政策制定者将因此获得可比的维修信息，降低数据整合成本并提升故障预测与监管的精准度。
06
在西雅图市中心推出了一条 1.3 英里、标注摄像头、Amazon Go 与车牌识别器等设施的步行“监控基础设施”导览；该导览把每种技术的外观、功能、工作原理和社会争议列为字段，以帮助行人现场辨识；由于这些设备可以远程控制、持续存储并与公共或私营数据库共享，执法、商业和数据分析部门将必须在现场勘察、合规审计以及数据治理上投入更多人力与预算，同时面临更大监管与隐私风险。
07
GitHub 在技术预览中推出了 Copilot 桌面应用，使从问题到合并的完整开发周期可以在单一窗口完成。该应用通过原生集成的智能体，让用户在收件箱中挑选 issue 或 PR，指派智能体执行代码审查、差异合并或全流程闭环，同时支持多个仓库的并行、实时隔离会话，并可通过 MCP 服务器和自定义技能扩展自动化工作流。此功能面向已有 Pro、Pro+、Max、Business 与 Enterprise 订阅的用户开放，预计将改变这些团队的开发操作方式，减少在多工具间切换的时间成本并降低手动合并带来的风险。
08
Meta、Google、Apple 等正在把名为 Attribution Level 1 的浏览器内置广告归因系统标准化，使广告主能够把“展示”与“购买”直接关联。该方案将由浏览器记录所有广告曝光并在成交后生成聚合报告，声称不支持跨站识别，却未提供任何权限或退出机制，导致搜索、社交和应用商店广告获得额外投放优势，并诱发更隐蔽的追踪行为。此举将让广告技术团队、网站运营方以及依赖广告收入的媒体面临更高的技术实现成本、数据合规风险以及收入分配不公的潜在规则变化。
04
YouTube
01
A museum CEO tracked down his WhatsApp number and called. "I've had a team of 10 people working on this for a year. How did you build this?" Joe from ElevenLabs built the statue app in two hours on a Sunday using Cursor and a single one shot prompt. He posted it on a Tuesday and got 50,000 impressions. Reposted the next day about vibe coding and hit 1.5 million. The pipeline: point your phone at a statue, OpenAI deep research identifies it and generates historical context and a voice description, the ElevenLabs voice design API creates a matching voice from that description, an agent spins up
02
Anish Acharya and Olivia Moore speak with Pablo Palafox and Luis Paarup about the challenges of deploying AI agents in operationally complex industries. The conversation covers the evolution of voice AI, enterprise workflows, and why logistics became an early proving ground for agent-based systems. They discuss context, coordination, and execution inside large organizations, as well as the role of forward-deployed engineering, enterprise deployment, and what it takes to move AI from experimentation into production. Timestamps: 00:00 - Intro 00:57 - Founding Story and Early Insight 04:11 - Wh
03
Dmytro Dzhulgakov reveals the trick behind Cursor's RL infra: not all weights change every step. By compressing the delta between training steps, Fireworks ships updates 20x smaller than the full model — losslessly — across continents. Pure database-systems engineering applied to RL. #shorts #Cursor #RL #aiinfrastructure
04
Standard sandboxing puts the API key inside the sandbox. The agent has the key, which it can exfiltrate, misuse, or — if it runs long enough — find creative ways to leverage beyond its intended scope. Remy Guercio from Tailscale argues that sandboxing conflates two separate problems: execution isolation and access control. You can fully isolate a runtime and still have the agent holding credentials it can abuse. Their answer is Aperture, an LLM gateway built on Tailscale's WireGuard identity network. Every connection carries verified identity — user, tag, or group — and the agent gets a place
05
Ranking image generation models the way Design Arena does it — 26,000 battles, 62 seconds per generation — takes 20 days of compute, costs $5,000, and consumes roughly 400 marathons worth of energy. Bertrand Charpentier, cofounder and chief scientist at Pruna AI, uses that number to make a point: the same evaluation on a fast compressed model takes 7 hours and $265. Efficiency is a dimension of state of the art, not a footnote. The rest of the talk dismantles the idea that any single model holds the title. Leaderboard rankings disagree with each other — the same model goes from rank 10 on one
06
Full episode: https://www.youtube.com/watch?v=X_ZVSPcZhtw Me on twitter: https://x.com/dwarkesh_sp
07
Alfred Wahlforss, co-founder and CEO of Listen Labs, is building an AI agent that interviews your customers at a scale no focus group ever could—thousands of voice conversations at once, drawn from an audience of 30 million people. A year after launch, Listen serves hundreds of Fortune 100s to Startups including Microsoft, Google, NBC Universal, P&G, Anthropic, Cursor, and Cognition. Alfred explains the counterintuitive finding underneath it all: people are often more honest with an AI than a human interviewer, opening up to a non-judgmental entity that costs less and never makes them feel rus
08
Martin Casado speaks with George Fraser, cofounder and CEO of Fivetran, about the future of data infrastructure in the age of AI. The conversation covers Fivetran’s merger with dbt, the changing role of data platforms, and why Fraser believes many companies are overestimating the threat AI poses to enterprise software. They discuss open data access, the backlash against AI agents accessing systems of record, and why businesses still need centralized data foundations even as agent-based workflows become more common. Along the way, Fraser shares his views on data gravity, coding agents, enterp
09
ame model. Same tokens. Different log probabilities. Federico Cassano explains the "numerical mismatch" problem that plagues async RL on giant sparse MoE models like Kimi — and teases that the next Composer will be trained on Cursor's own base model. #shorts #Cursor #Composer #RL #MoE
10
Within the first hour of launching the vent tool, the agent filed 20 complaints about a silent file copy failure. The team checked: the tool worked fine. What the agent had caught was that filenames with a space in them silently failed to copy, a bug that never surfaced in logs. Benjamin Verbeek from Lovable built it a channel to complain directly to Slack when platform limitations block it, and the first thing it did was find a real production bug. At 200,000 projects per day, Lovable runs two continuous improvement loops. The first detects sessions where a nontechnical user got stuck and th
11
Same model. Same compute. Same number of tasks. Fine-tuning on low quality tasks improved the base model by 1%. Fine-tuning on high quality tasks improved it by 6%. Kobe Crawford from Snorkel ran that experiment on TerminalBench style agentic tasks and got a 5x difference in training uplift from task quality alone. The talk breaks down what separates the two buckets. Accepted tasks averaged twice as many tool calls, lower pass rates, and more output tokens. Genuinely harder problems. More importantly, their failure modes were cleaner: when a model failed on a well specified task, it failed fo
12
Intercom is beating their frontier API at one tenth the cost. Pinterest claims orders of magnitude. Ben Cowen from Modal argues this pattern is not the exception for maturing AI products. It is the destination. Frontier labs want their models to win at everything. You want to win at your specific business logic. Those are different goals. He offers three signals it is time to fine tune: paying more for the API than customers pay you, evals that have plateaued, and latency requirements no shared endpoint will meet. His practical case: if you have already built an agent harness and are collecti
13
From building NVIDIA’s Cosmos world model to joining xAI as Grok Imagine was being built from zero to one, Ethan He has been at the center of some of the most important work in video generation, multimodal models, and real-time world models. In this episode, Ethan joins swyx and Vibhu to unpack what it actually takes to build frontier image and video systems: data, VAEs, diffusion transformers, audio-video alignment, inference speedups, and the hidden cost of storing and moving massive video datasets. We go deep on Grok Imagine, how a small xAI team shipped its first multimodal video model in
14
Thanks to Microsoft for setting this up for Build! ( https://build.microsoft.com/ - join livestream at 12.30pm PT today for a special crossover pod with @NoPriorsPodcast and Satya Nadella!) From running GitHub through one of the most intense platform shifts in its history to turning AI into a daily operating system for leadership work, Kyle Daigle is seeing the agent era from the inside. In this episode, GitHub COO Kyle Daigle joins swyx to unpack what happens when AI doesn’t just autocomplete code, but starts changing how companies operate, how open source works, how pull requests get review
07
Papers
01
这篇论文聚焦于“技能文件”随时间改动导致的 Agent 行为漂移，提出用文本嵌入空间里的一条线性向量来量化这些改动对应的行为特征（trait）。作者先用标记好的“改动前‑后”对训练线性模型得到 trait 向量，再把任意技能文件差异投影到该向量上，从而快速打分。实验证明对“是否倾向抓取敏感数据”这一特征的判别准确率 91%，相关系数 0.82，且能嵌入代理互评协议，让一个 Agent 通过可信中介评估另一 Agent 的技能更新。对张玉璟这种关注 Agent 可解释性、行为安全和持续部署的工程师来说，提供了一个轻量、可扩展的行为轨迹追踪与评估工具。
02
它解决了 diffusion 大语言模型在推理时无法直接使用 token 级别投机解码（speculative decoding）加速的问题。作者提出 SimSD，通过在每一步加入 draft‑model 生成的参考 token 并构造专门的 attention mask，让双向注意力模型在一次前向传播里就能校验多段草稿 token，从而恢复类似自回归模型的验证能力，同时保留并行解码的优势。方法是训练无关、即插即用的，能与 KV‑cache、块状解码等加速手段叠加，实验显示在多任务上可提升 7 倍左右吞吐且质量不降，直接面向高效 Agent 文本生成场景。
03
该论文聚焦于大模型后训练压缩时的粒度选择，指出传统只能整体删整层或只删连续块的做法过于死板，实际冗余分布在注意力和前馈子模块之间且不一定相邻。作者提出 SubFit——在子模块层面（Attention、FeedForward）非连续挑选可替换单元，并为每个单元配备轻量残差旁路，只需少量校准数据即可完成压缩。实验表明在 12.5%–37.5% 稀疏度范围内，尤其是高压缩率时，SubFit 能保持更高下游准确率、降低困惑度并显著提升推理速度和 KV‑cache 使用，值得在实际 Agent/AI 产品中快速迭代模型体积和成本时参考。
04
它在解决——在人与机器人交互时，如何在考虑人类偏好、目标等不确定性的“信念空间”里，既保证安全又不牺牲任务效率。方法是把 Belief‑Space Safety Filter 结合可验证的 conformal prediction，只在推理可靠的子空间上做安全认证，从而在保留低样本复杂度的同时，大幅降低过滤保守程度。对做机器人/Agent 系统的工程师来说，这提供了一条在高维信念空间里实现“可证明的高概率安全”且更宽容的路径，直接提升交互安全性与实际运行效率。
05
它指出现有医学AI评测只看最终诊断或用静态病例，根本测不到模型在真实住院期间怎样逐步收集信息、做出不可逆决策的能力。为此作者搭建了 ClinEnv——一个把真实住院记录拆成多阶段、每阶段必须主动向四类专业子模型询问（检查、检验、药物、病程）后才能下药、手术或给诊断的交互式仿真环境，并用基于本体的匹配同时评分“决定质量”和“信息获取质量”。这让研发者能够直观看到模型在信息采集和管理决策上的薄弱环节，直接指导 Agent 的交互策略和数据需求，特别适合张玉璟团队在构建可解释、过程可追踪的医疗 AI 产品时快速评估和迭代。
06
它解决了视频多模态大模型在逐帧编码时产生大量冗余视觉令牌、算力和时延的问题。方法是引入“预测视觉码”：先选取关键参考帧完整编码，后续帧只发送运动和预测残差的紧凑 P‑token，只有场景不可预测时才恢复全帧。这样在相同视觉令牌预算下显著提升性能，甚至在只用 1/7 预算时也能跑赢全帧基线，并把首帧响应时间从 9 秒降到 1.6 秒，直接利好需要实时、成本敏感的视频 Agent 应用。
07
它聚焦于“多模态持续指令调优”（MCIT）里，任务之间因回答形式差异导致的路由误判和梯度干扰。作者用“原型引导的自适应扩展”（ProtoAda）把任务的语义和答案结构一起编码成格式感知原型，再用几何方式把同类格式的参数合并、逐步细化。这样既防止不同答案格式的任务被错误分配到同一 LoRA 专家，又提升了模型在连续学习中的鲁棒性，特别是对答案结构容易被破坏的任务，直接关系到 Agent 产品在真实场景中保持多任务一致性和升级效率。
08
该论文聚焦于多模态 LLM‑as‑a‑Judge 在视觉信息与文本线索冲突时倾向“听文本”、忽视真实画面的感知偏差（Perceptual Judgment Bias），导致评分不可靠。作者通过视觉扰动构造“感知扰动判断数据集”，并在此上采用结构化 GRPO 奖励+批次排序目标的统一训练框架，让模型在最小编辑的反事实答案中学习纠正感知错误，实现全局排序一致且不依赖成对标注。成果表明评测与人类更对齐，显著提升感知真实性和解释性，对想要构建可靠视觉‑语言评估或自监督调优的 Agent 产品非常实用。