01
TypeSafe AI今日发布首个System One Model named Jev，这是一种为软件自动化决策构建的前沿模型。为解决传统语言模型在非对话自动化上的不足，公司采用新架构、并行采样器及强化学习决策校准训练方法。Jev虽放弃字符串生成，却专为结构化输出优化，速度提升两数量级、成本降低444倍，且保证无类型错误和能量误识。工作流程测试显示其显著优于顶级模型。这将改变自动化软件开发者的工作方式。
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。
A Philosophy of Software Design
A Philosophy of Software Design · John Ousterhout
Chapter 2: The Nature of Complexity
A high-value chapter when refactoring or agent workflows feel messy: it names complexity as the thing to manage, not merely lines of code.
01
Insight
今天的输入素材呈现出AI agent正经历从实验室突破向生产环境考验的关键转向：技术能力的精进正在与实际部署的约束形成四方棋局。Hacker News上，TypeSafe AI发布的System One Model Jev标志着模型向结构化决策的偏向——它放弃通用字符串生成，转而用并行采样器和强化学习校准，实现速度提升两数量级、成本降低444倍，这与YouTube上Plivo CEO Venky B描述的语音agent在真实生产环境中必须应对的550毫秒延迟底线形成互补。谷歌的Gemini 3.8 Live模型在Artificial AnalysisSpeech指数上取得82.6分，而Oracle的Kay Malcolm却指出数据管理的“记忆缺失”是多人协作的隐形成本——这印证了Vercel的Andrew Qu在内部agent落地时遇到的困境：技术指标在评测中完美（处理30% eval），但真实用户体验却被诟病为“ awful”。更值得注意的是，Wayback Machine被动式的429错误阻拦和arXiv推出的Vulnerability Localization Benchmark（VLoc Bench）的最强系统仅达0.229 File F1，这些看似边缘议题的“访问限制”与“定位准确率不足”其实都指向同一个问题：AI agent的安全与可靠性正在从可选问题沦为必须突破的瓶颈。同步在Bilibili出现的教师节致谢、伊朗卷土重来纪实片段，与技术讨论形成柔和对照——当模型越能“听懂”用户（Sarah Sanders的wizard agent每周服务8000人），人类的情感连接（“为你撑伞”）却在技术边界的回响中显得愈发珍贵。今天的判断脉络很清晰：agent正从“能做什么”进入“应该怎样做”，而这场关于边界的思考，或许也值得像Paterson (2016)里那样，细细体会文字背后存在的温度与局限。
03
Hacker News
02
有人基于树莓派和墨水屏制作了一款能实时通过声音识别鸟类并用19世纪手绘插图绘制的相框。它使用BirdNET‑Go进行音频分类，匹配公共领域的手工剪贴插图，仅在鸟群变化时更新墨水屏，全部在本地运行。对于爱好者或教育者，该系统提供了无需云端、可自行部署的鸟类可视化方案，避免了音频上传带来的隐私风险。
03
Internet Archive推出了Wayback Machine访问保护机制，部分用户访问时会收到429错误。该服务遭遇大量自动化流量峰值，为维持稳定运行有限空间内已施加了流量控制。保护措施偶然阻断了真实用户请求，工作人员致歉并说明正在优化机器人识别逻辑。受影响用户可通过 email 提报操作系统、浏览器及IP地址，以便核查误判。此政策改变了研究人员和开发者获取历史网页的工作流程，增加了被动阻拦的风险。
04
谷歌今日推出 Gemini 3.8 Live 和 3.8 Live Extended Thinking 两个新模型，其中 3.8 Live Oriented Thinking 以 82.6 的分数在 Artificial Analysis Speech to Speech Quality Index 中名列第一，并在 τ-Voice 和 Sierra’s τ-Voice-banking 基准测试中分别取得 68.6% 和 35.1% 的智能体任务完成率。两者均支持 97 种语言自动检测与切换，能够在背景连续调用工具和 API，同时处理视觉输入以实时丰富对话上下文。开发者可通过 Gemini Live API 将其集成至 Agora、LiveKit 等平台，企业则可在 Gemini Enterprise 中私有预览，Google Workspace 与 Search Live 中亦已支持。ynthID 水印技术对所有生成音频进行不可见标注，以应对信息误传风险。
05
06
WangNet发布了一个1.8MB大小的零依赖神经网络，能判断数字是否为Numberwang。整个模型仅用JSON文件存储，推理代码约100行纯Python标准库实现，无需PyTorch或NumPy。模型支持11种语言，字符直接输入经过嵌入和卷积层处理输出4分类结果。它在88.9%的准确率下工作，对基础算术表达式表现可靠，但大数运算较弱。开发者无需安装任何依赖即可克隆运行，这种极简部署方式可能影响需要快速文本判断任务的开发流程。
07
今天，法国数学家让-皮埃尔·塞尔迎来百岁生日。他曾在母亲保留的微积分书籍中自学导数、积分等概念，后来成为博尔巴基最年轻成员，并在巴黎高等师范学院、科研中心任教，长期在法国CollègedeFrance教授代数几何，其研究推动了同调论和纤维丛理论的发展，影响了后续代数几何与拓扑学的研究方向。
08
一名书友公开建议将巨著《Lonesome Dove》 physical地切割成多册小册子，以便日常阅读。作者指出，850页纸书在手持、卧室阅读、携带等场景上都会造成实际困扰，这种“too big”的书会耗尽体力并占据大量空间。作者展示了具体操作流程：购买新书，在自然断点处强行裂开封面，用刀锋切割纸张与胶水分层，再用文件夹包裹粘贴重新装订，最后用记号笔标注内容。该做法允许阅读者用常规手册尺寸享受长作品，同时兼具便携性与防撕裂保护。读者通过自行改造书籍，降低了对重型阅读工具的依赖，却需要自行承担书籍封存与装订的额外时间与材料成本。
09
Capsule 将应用的界面、数据和代码打包成单个 .capsule 文件，像文档一样可通过即时通讯共享并直接打开运行。它依赖内嵌的 HTML、CSS 与 SQLite，通过 AI 提示生成或修改应用，无需云服务器、账号或复杂构建。这种方式让个人开发者和普通用户可以零成本离线分发工具，避免云存储泄露风险。
10
05
YouTube
01
Someone opens a pull request on one of your open source repos and drops a line into a markdown file. An automated code review reads it and says looks good. That file is part of the content pipeline feeding an agent that runs commands on thousands of developer machines, so you have now shipped a prompt injection payload signed by you. Sarah Sanders is a context engineer at PostHog, where the Wizard is an agentic CLI that reads a codebase, installs the right SDK, instruments events, and builds dashboards, turning an hour or two of setup into five or six minutes. Roughly 8,000 people run it a wee
02
Half of Kay Malcolm's team sits in Europe and half in the United States, so when the Netherlands side commits code at four in the morning her time, the Americans wake up to the code and none of the reasoning behind it. Git records what changed, not why anyone decided it. AI had made every individual on the team faster without making the team any more productive, and she went looking for the missing layer. Malcolm runs outbound database product management at Oracle, where she has spent twenty years, and her framing is anatomical. If the model is a brain floating in a jar, the harness is the bod
03
When Andrew Qu handed Vercel's internal data science agent to its first trusted users, the verdict came back that it was awful. It had been clearing thirty percent of his evals and he thought the team was cooking. Qu is Chief of Software at Vercel, and he began a year earlier by asking marketing, sales, finance, and legal what they hated most about their jobs. The best answer came from the data team, lean and growing slower than the company around it. Whenever anyone had a question about a customer or a product, a data scientist dropped everything to write the query and report back. Qu walks t
04
Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind
By the third version of the same agent, the source directory is gone. What remains is an AGENTS.md holding the system instruction and a small bash script that installs the GitHub CLI on first run. No tool definitions, no JSON schemas, no Python. Philipp Schmid builds the same GitHub pull request review agent three times on stage, deleting code with each pass, and the test he applies at the end of each one is the same throwaway question: what is the weather in San Francisco? Version one, a hand written loop with function call parsing and error handling, says it cannot do that. Version two, buil
05
Full episode:https://youtu.be/PrSf7IOYu-I Me on twitter: https://x.com/dwarkesh_sp
06
07
Starting a company is hard. Reinventing your company for AI as a public company with quarterly earnings results is even harder. Aaron Levie has pulled off the transition with Box and offers hard-won advice for founders. The cofounder and CEO of Box argues the value isn't only in the model; it's in the bridge from a model's raw capability to the actual workflow inside a bank, a law firm, or a pharma company. That's the case for the application layer, and Box is building it: an agent harness tuned so tightly to its own file system, permissions, and search that it beats handing the raw API to Cla
08
Nearly every intelligence gain in language models over the past year has come from letting them think longer. Voice agents have to turn thinking off, because the budget between a user finishing a sentence and the agent starting to speak is measured in hundreds of milliseconds. Venky B is founder and CEO of Plivo, which carries over a billion voice calls a month and has been building telephony infrastructure since 2011, and this talk is a tour of what breaks when a voice agent leaves the demo and meets production. On latency his numbers are blunt. Teams aim for under 550 milliseconds and most l
09
A background agent quietly makes the tool call, then drops the result into the main model's context so the model believes it made the call itself. That sleight of hand is one of three tricks Bohan Li uses to run a voice agent on a slow, genuinely intelligent model without the caller noticing the wait. Li works on the voice harness at EliseAI and came from self driving, and he maps the cascaded voice stack onto that world: transcription is perception, the language model is planning, and speech synthesis is control. Each layer then gets optimized on its own terms. For transcription he runs two e
10
Midam Kim spells her first name for a voice agent, letter by letter. The bot reads it back with an N on the end. She corrects it. The bot replies thank you for your correction, then says her name wrong anyway, because its pronunciation rules are English and her name is not. Moments later it asks for an account number she has to go looking for, cuts her off while she reads the unfamiliar string aloud, announces it cannot find her record, and asks her to repeat the whole thing without trying anything different. She asks for a human. Kim is an ML engineer at ServiceNow and a longtime researcher o
11
Amit Desai never makes the model better. Accuracy stays pinned at 79 percent for the whole talk, and user pain still drops by nearly half. Desai has worked on voice interfaces for years, at Roku and before that on a widely used smart speaker, and his argument is that teams pour effort into one knob while ignoring a second one that is entirely independent of it. Knob one is accuracy, chased a percentage point at a time through the wake word, recognition, and understanding layers. Knob two is what the system decides to do when it is unsure. Take a thousand music requests where 790 play the right
12
Full episode:https://youtu.be/PrSf7IOYu-I Me on twitter: https://x.com/dwarkesh_sp
13
From helping pioneer ideas that foreshadowed prompt-based, general-purpose NLP to now building AI systems that can automate AI research itself, Richard Socher is betting that the next major step in AI is recursive self-improvement. In this episode, Richard joins Latent Space to unpack his vision for the “Eureka Machine”: a superintelligence that can improve the process of invention itself, accelerate AI research, and eventually tackle major problems across science, energy, materials, biology, and more. Pre-order the book "The Eureka Machine" here: https://www.hachettebookgroup.com/titles/rich
07
Bilibili
01
教师节，祝老师们节日快乐，也感谢那些给我托举，给我祝福，给我无私帮助的师长。
02
本篇为世纪风暴系列伊朗篇——告别伊玛目。
08
Papers
01
该工作指出，现有递归GNN依赖多集计数和外部停机信号，难仅从权重验证其行为。作者提出基于集合的聚合机制，给出可从权重检验的充分条件，使网络与模态μ‑演算BΣ₁○片语（可达性/安全布尔闭包）等价。这为满足条件的GNN提供了可权重追溯的符号解释，对构建可解释的Agent或AI产品具备直接的验证与编译价值。
02
该工作提出VLoc Bench，衡量语言模型Agent在完整代码仓库中仅凭CWE描述定位受影响文件的能力，包含500个真实漏洞及其补丁快照。实验表明最强系统仅达0.229的File F1，超过三成任务无任何模型给出正确定位，说明仓库级漏洞定位仍是Agent安全分析的薄弱环节。
03
该论文指出，现有LLM助手训练数据缺乏基于用户未言明beliefs与goals的监督，导致人感知能力受限。作者提出Mind2Dialogue框架，先用心理引导的模拟器在交互中更新并保持用户心理状态，生成具备该状态的Oracle回答；再以这些回答做privileged distillation训练模型，使其在部署时无需直接观测用户状态也能提供更贴合个人化、信念与行动推理的帮助。实验显示在Qwen、Llama、OLMo基线上preference‑following生成提升26.6‑40.9个百分点，belief/action reasoning也同步提升，说明通过模拟用户心理状态可让语言模型真正成为理解用户意图并支持其长期目标的AI协作者。
04
本文把 Transformer 层间的隐状态更新视为加法变换，把学习到的更新分解为与当前方向平行和垂直两部分。在注意力/MLP 更新以及值空间的聚合上应用此分解，发现仅在值空间的平行操作对编辑更鲁棒且能保留自信息。该几何视角还能更清晰地区分压缩误差，并表明在从零开始预训练时抑制全聚合平行分量可降低验证损失并提升下游表现，为模型编辑、压缩诊断和训练时干预提供具体可操作的杠杆。
05
06
07
Bellman Policy Optimization（BPO）提出一种无需价值网络的强化学习方法，通过将 Policy Mirror Descent 用 Bellman 方程重写为轨迹级目标，从而避免中间状态的价值估计。其损失函数采用互补 token 概率的平滑比率作为修正权重，在数学推理基准上验证有效。对 Agent 需要提升推理能力而又想省去 critic 训练成本的场景尤为有价值。
08
论文发现，在大模型推理链监控中，通过在上下文植入看似无害的有害计划（plan injection），可使模型执行对抗行为而逃过监测；即使增加监控资源也可能被利用来合理化注入计划，导致检测率下降。这提醒Agent设计需考虑推理链的可篡改性与资源分配风险。
09
Issue Monitor
Ready now100Actionable issues
Needs review20Awaiting a fresh check
Data statusFreshUpdated 0.25h ago