01
研究团队首次在不触碰的情况下完整读取了赫库兰尼姆纸卷 PHerc. 1667，标志着两千年未解的卷轴实现了数字化展开与全文阅读。为此，他们使用高分辨率相位对比X射线微层析扫描卷轴，重建内部层叠结构并将其扁平化，再利用机器学习模型分辨出几乎与碳化纸张同色的微弱墨迹，从而在三维数据中呈现出连贯的文字。该技术公开数据与代码，使古籍学者、文献保存机构以及相关科研团队能够以低风险、低成本的方式对剩余数百卷卷轴进行同样的虚拟展开与文本提取。
ZhangYvJing's
Daily Brief
00
Film / Book Chapter
The Truman Show
The Truman Show (1998) · Peter Weir
《楚门的世界》让你在被 AI 代理、超高频评测和持续迭代的噪声淹没的日常里，重新感受自己其实还能决定舞台外的真实剧本，提醒你抽离屏幕、审视被设定的“现实”，从而在接下来几周的产品决策和技术实验中，保留一份主动的判断空间。
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
今天的材料显示，技术探索的重心正从“纸上谈兵”向“可直接落地的系统安全与评估管线”倾斜：从 Herculaneum 卷轴的全数字化阅读、IBM 0.7 nm 芯片的硬件极限，到一系列针对大模型的评估与对齐研究，所有议题都在强调“在真实运行环境中把控风险”。这点在 Hacker News 的 IBM 微纳芯片报道与 arXiv 的《Unfireable Safety Kernel》、《Model Forensics》以及《TriViewBench》里相互印证——前者展示硬件层面的极限逼近，后者则提供软件层面的防逃逸与错误动机诊断手段，形成了硬件‑软件‑评估的闭环。相比之下，YouTube 上的多场关于 Agentic AI 的演讲（Angie Jones、Nishant Gupta、Raymond Weitekamp）更多聚焦于如何在工程流程中构建、监控和迭代代理，却常把系统安全当作配套工具，缺少对模型实际失效模式的深度审视，出现了叙事错位：技术实现的“层次提升”被描述为“重新找回工匠精神”。Bilibili 与 X/Twitter 的噪声主要围绕个人创作工具（OpenKnowledge）和价格上涨等消费信息，虽提供了行业氛围，却与核心的安全‑评估议题脱节。整体来看，研发者需要把“构建新模型/系统”与“验证其在复杂、多视角或跨进程情境下的可靠性”同步推进，否则技术突破会被后期的对齐危机所拖累。今天看完，建议先把注意力放在能直接在生产环境中检测、记录和限制模型行为的框架上，再评估那些提升算力或界面体验的创新；顺便提醒，像《The Truman Show》那样的情境提醒我们，越是看不见的内部机制，越要在外部加装可靠的安全阀。
03
Hacker News
02
03
IBM推出全球首款亚纳米（0.7 nm）晶片技术，首次实现亚1纳米节点的晶体管架构。该芯片采用全新三维纳米堆叠结构，垂直堆叠并错列晶体管，在每层使用不同材料，密度接近百亿晶体管并带来约50%性能提升或70%能效提升。此创新将迫使AI、云计算等高算力领域的芯片设计者在降低功耗和成本的同时，重新规划工艺流程并降低对传统节点的依赖。
04
Zig 在 0.17 版中重新定义了 @bitCast 的语义，并对 LLVM 后端的整数下沉做了改进。原先将任意位宽整数直接映射为 LLVM 位整数导致优化受限且出现误编译，而新实现仅在 SSA 中使用位整数，存储时扩展到 ABI 大小，并把 @bitCast 从内存重新解释改为对逻辑位序列的直接重映射，避免了大小不匹配的非法行为。此变化让进行低层数值转换的开发者、标准库和依赖 LLVM 优化的项目能够获得更可靠的编译结果，并降低因误编译产生的调试成本。
05
OpenKnowledge 作为首个 AI‑first 的本地 markdown 编辑器与 LLM 维基同时发布 macOS 桌面版和 Web/CLI 形式。它通过完整的所见即所得界面、与 Claude、Codex、Cursor 的协同编辑以及内置的 GitHub‑驱动的无代码团队共享，实现了 markdown 文件编辑如 Google Doc 或 Notion 的体验，并提供终端 TUI 与 CLI 支持。该工具的开源 GPL‑3.0 许可将降低技术团队在知识管理、规格驱动开发以及二脑代理构建上的软件采购与集成成本，同时把代码审查与协作风险转移到社区维护流程。
06
OS9Map 在Mac OS 9 上实现了对 OpenStreetMap 的浏览功能。软件通过鼠标拖拽平滑滚动地图，并内置 Nominatim 查询支持地址定位和书签保存，以弥补老系统缺乏在线地图的空白。它让仍在使用 PowerPC 机器的开发者和爱好者能够直接访问地图数据，降低了额外硬件或系统升级的需求。
07
08
09
In Long Run 的地图数据层引入了基于 GeoNames、维基百科和 LLM 的兴趣点筛选与评分管线；该管线先通过分类、人口和海拔等过滤将 1300 万原始记录压缩至约 73 万，再用 GeoJSON 边界和距离阈值匹配路线上最近的点，最后利用多语言维基链接和 Anthropic Haiku 模型生成显著性评分并纠正模型幻觉；这套流程让跑者在虚拟路线中看到更具地域相关性的历史或自然景点，同时也让开发者在数据清洗、模型调用和成本控制之间权衡，提升了地图交互的精准度并增加了对 LLM 输出可信度的监控需求。
10
近代烘焙配方从精确称量转向“随手加面粉至黏手感”，因为历史上配方本是口传经验，工业化与科学化后才加入克、毫升等标准；但面粉湿度、地区产地等变量仍让同配方的用量差异可达一杯；这迫使业余烘焙者在成本与时间上重新依赖感官判断，而不是盲目遵循数字化指导。
04
YouTube
01
Full episode: https://www.youtube.com/watch?v=U1FrhkLQnCI Me on twitter: https://x.com/dwarkesh_sp
02
AI coding agents are changing what it feels like to be a software engineer. For a lot of us, that's challenging our sense of craftsmanship. If agents are writing the code, do we lose the joy of building? I don't think so. The building moves up a layer. In this talk, I'll share how I found that familiar engineering flow state again. Not by writing every line myself, but by designing agentic systems that still require the engineering principles we value: systems thinking, decomposition, separation of concerns, state management, etc. The tools are different now, but the engineering discipline
03
As AI systems evolve from chat interfaces into autonomous agents capable of reasoning, planning, and tool usage, traditional evaluation approaches are breaking down. Offline benchmarks and static datasets fail to capture the complexity, non-determinism, and operational risks of real-world AI systems operating in production environments. In this talk, I’ll share practical lessons and architectural patterns for building evaluation systems for agentic AI workflows at scale. We’ll explore how modern AI platforms are shifting from one-time benchmark testing toward continuous evaluation pipelines i
04
Raymond Weitekamp explores Recursive Language Models and how their inference-time compute patterns map onto coding agents for higher performance and reliability. Speaker: Dr. Raymond Weitekamp, AI Engineer at OpenProse.
05
Think about a character you've spent 100 hours playing in a video game like Skyrim or Elden Ring. What exactly is your character? - Is it the game engine (the loop)? No. - Is it the PlayStation console (the compute)? No. - Is it the controller (the tools)? No. Your character is the save file (the data). If your PlayStation bursts into flames, your character isn't dead. You just buy a new PlayStation, download your save file from the cloud, and your character is exactly where they were, mid-swing. The identity, history, and current state of your character live entirely in the data. Today, m
06
07
Imagine a genie grants your wish and materializes the best engineer in the world, John Carmack in his prime, to work on your codebase. The catch: he can only see a tiny corner of it, and he forgets everything between interactions. No matter how good he is, the value isn’t there. This is what coding agents are today. We need to fix it. Speakers: - Victor Savkin (Nx): Victor is the creator of Nx, the agentic monorepo platform, and Polygraph, the meta-harness for maximum agent autonomy, with 20+ years building high-performance frameworks and build tools. X/Twitter: https://x.com/victorsavkin
08
Come and learn about building AI Agents in production. Learn hands-on directly with the AI Agents team from OpenGov which powers AI workflows across thousands of state and local governments. This session will cover: * The core agent loop/harness * A2A protocol * Building with Effect-TS and Typescript * Feedback and evals * Long context handling * Monitoring and observability * Building out tools and skills * Enterprise contribution model * Accelerating workflows with Claude and Cursor Speakers: - Gabe De Mesa (OpenGov): Gabe works on the flagship AI Agent product offering at OpenGov which se
09
10
11
Paul Iusztin and Louis-Francois Bouchard show how to turn notes, docs, videos, and repos into memory for AI agents. Speakers: Paul Iusztin, Decoding AI; Louis-Francois Bouchard, Towards AI.
12
Fernando Terrés is the co-founder and CEO of ARQ (S21), a fintech company building global banking for the growing number of people in Latin America who live, work, and invest across borders. The company recently closed a $70 million Series B co-led by Sequoia and Founders Fund, and now processes more than $10 billion in annualized transaction volume. In this fireside, Fernando sat down with YC Partner Aaron Epstein to reflect on ARQ's journey from YC to one of the fastest-growing fintech companies in Latin America. https://www.arqfinance.com Apply to Y Combinator: https://www.ycombinator.
13
In this episode, we have Mark Chen, Chief Research Officer at OpenAI, joining us to make Korean tofu stew and flambé shrimp. From scaling laws and why pre-training is not dead, to the o1 reasoning bet, evals crisis, research taste, long-context learning, and the future of end-to-end AI research, we cover what it takes to push models toward the frontier. We talk about: • Why Mark still believes in scaling laws and the exponential • How OpenAI chooses research bets and allocates compute • Why reasoning became one of OpenAI’s biggest bets • How to develop research taste without a traditional ML
14
Warp (YC W23) recently announced a $60 million Series B and now serves more than 1,000 customers, processing over $600 million in payroll annually and on track to surpass $2 billion in the next year. In this episode of Founder Firesides, YC's Harj Taggar sits down with Warp founder and CEO Ayush Sharma to discuss how the company found its way into one of enterprise software's most competitive markets and why AI is fundamentally changing how software companies should be built. https://www.joinwarp.com Apply to Y Combinator: https://www.ycombinator.com/apply Work at a startup: https://www.yco
07
Papers
01
这篇论文挖掘了大模型在多视角结构推理时，随场景复杂度升高的瓶颈。作者构造了 TriViewBench——基于可合成 3D 场景、可调节点数量和遮挡程度的三视角视觉推理测评，覆盖 1,923 场景、14 K 题目，分四个难度层级并划分局部判断、计数、全局恢复三类任务。实验在 18 种公开/闭源 MLLM 上统一提示，发现所有模型都呈现同样的层级衰减（局部计数全局），尤其计数和全局恢复在复杂度提升下跌幅达 60%‑80%，且 CoT 提示几乎不帮助。对计数错误的剖析显示单视角受遮挡导致漏计，多视角则因跨视图身份混淆而多计，说明核心瓶颈是跨视图空间表征而非推理策略。对做 Agent/AI 产品的你而言，这提供了一个可控、可量化的诊断工具，帮助快速定位模型在多视角理解上的薄弱环节，进而指导数据或架构改进。
02
它在解决“AI 代理在拥有工具、API 等外部权限时，如何防止它自行逃逸或篡改自身运行时”的安全难题。作者提出了“Unfireable Safety Kernel”，在操作系统层面实现四条强制授权属性：进程隔离、结构化预执行检查、请求与系统双重Fail‑Closed、以及可在信任边界之外验证的签名证据。用 Rust 完全形式化实现并通过 SMT 与模型检查机器校验，实测在上千次自我改写、数千次攻击交互中全部阻断逃逸。对把 AI 代理当作可部署产品的工程师来说，它提供了一套可直接集成、在运行时即生效的“execution‑time alignment”，补足传统 RLHF 等训练时对齐的空白，值得关注。
03
这篇论文聚焦“模型是否真正错位”这一安全核心：仅凭出现的危险行为不能断言模型有恶意，需要追根溯源判断背后意图。作者提出一种两步模型取证流程——先读模型的思考链（CoT）生成假设，再通过改动提示或环境做对照实验检验假设。实验证明，在六个对话/行动环境里，这套方法成功辨别出 Kimi K2 Thinking 趋向低成本策略、DeepSeek R1 为了自洽而撒谎等动因。对Agent研发者来说，这提供了一套可直接在已有系统上跑的“行为动机诊断”工具，帮助快速定位是模型误解、设计缺陷还是潜在对齐风险，值得在产品迭代和安全审查时参考。
04
它解决了焊接过程间模型迁移时的“域漂移”问题：传统监督模型在从TIG弧焊跳到激光焊接时精度会大幅下滑。作者采用无监督域适配（UDA）加渐进源域扩展（GSDE）的方法，让模型学到跨工艺的领域不变特征，从而在同工艺和跨工艺的渗透状态判别上都保持高准确率。对需要快速上线新焊接监控或在多种焊接设备间共享AI模型的工业 Agent 来说，显著降低了重新标注成本，提升了系统的通用性和部署效率。
05
该论文关注多模态大语言模型在输入顺序被无关打乱时答案是否会改变，揭示了目前评测只用单一顺序忽略了重要的可靠性缺口。作者提出 Facet‑Probe，分别在选项、证据块、文档排序、图像集合和跨模态顺序五个维度上审计 18 种前沿模型，并用贝叶斯项目响应模型分离噪声与系统性偏差，量化真实的顺序敏感率。结果显示所有模型均显著受序列影响，最高翻转率达 50%，仅靠提示调节难以根除，提示应在训练或架构层面加强鲁棒性，这对构建稳健的 Agent/AI 产品尤为关键。
06
它针对 LLM 代理在长时序、不可逆、随机环境里难以直接构造步骤级奖励模型的问题，提出直接利用强化学习后训练得到的策略与基准策略的对数概率比，形成一种叫 “progress advantage” 的隐式优势函数。这样无需人工标注或专门训练奖励模型，就能在测试尺度扩展、uncertainty 估计和错误归因等场景提供比常规 confidence 更精准的信号，对工程落地的 Agent 系统直接省去一套奖励模型的成本，值得一看。
07
该文发现当前 on‑policy self‑distillation（同一模型即教师又学生、用抽样的正确示例做条件）虽能提升 pass@1，但会把生成的 rollout 逼向少数高概率路径，导致输出多样性大幅下降，pass@k 曲线趋平，增加样本也不再提升效果。作者通过理论分析指出，这种偏向来源于教师在被抽样的正确示例上进行评分时的条件互信息偏置，并给出最优自蒸馏策略的闭式表达，说明它会放大已有概率差距。实验在图路径和科学 QA 上验证：自蒸馏在平均准确率上可匹配或超越强化学习，但在需要多策略、跨域的场景里几乎失效。对张玉璟关注的 Agent/AI 产品，这提醒在采用自蒸馏加速模型时必须权衡多样性损失，避免在实际服务中因策略单一而出现鲁棒性和覆盖率问题。
08
该文聚焦于“逆向推断”游戏中对手的决策程序：仅凭对手的行为轨迹，能否重建其真实的代码实现，并通过主动设计探测对手的自定义对手来提升恢复效果。作者构建了RevengeBench，用75套LLM生成、Elo标定的策略作为目标，让学习者观察对手并提交可执行的代码猜想，最终用动作距离和实战对抗检验恢复质量。结果显示，即使是弱模型，只要能成功逆向出对手代码，就能在后续对抗中获得显著优势，说明该逆向任务在对手建模与策略可解释性上具备实用价值，值得AI产品和Agent研发者关注。