与「manifest」相关的搜索结果

And yeah @fermah_xyz started with a problem that sits deep inside the ZK stack. Zero knowledge applications can prove that a computation was done correctly, but producing those proofs at scale takes serious infrastructure: specialised hardware, several proving systems, workload scheduling, and reliability and cost controls. A team building a ZK rollup, zkVM, bridge or coprocessor cannot simply ask for "a proof" and move on. Fermah's first answer was a universal proof generation layer. Provers with GPUs and other hardware supply capacity, applications that need proofs create the demand, and Fermah matches the two. That marketplace is now called Froben. The problem behind the proof Running this in production exposed something bigger. A proof is one step in a longer sequence. Something has to trigger the work, the work has to run, the result has to be verified, and then the result usually has to start the next action. Fermah's argument is that blockchains are good at recording state and weak at starting what happens next. Take a lending position that crosses its liquidation threshold. The chain records the fact, but nothing forces the liquidation. A keeper, bot or backend still has to notice the condition, build the transaction and submit it. Consensus settles what happened, and a human run service still decides what happens after. Fermah calls the missing piece protocol agency: a protocol's ability to act on a verifiable condition without depending on an outside operator. What Kernel does Kernel is Fermah's Protocol Agency Engine, and its pitch is short: Define. Execute. Attest. A developer defines a workflow, Kernel runs it, and the result is published onchain with a cryptographic proof. Three design choices matter here: -Typed nodes. Workflows are built from a library of verified building blocks such as HTTP calls, onchain reads, price feeds and math functions. The type system validates how they connect before anything runs. -Sandboxed execution. Workflows run in an isolated environment built for untrusted code, and the same inputs always produce the same output. - AI where it helps, code where it counts. Fermah says AI agents can compose workflows, but deterministic code executes them, with no LLM in the execution path. Kernel can work as an oracle too. A smart contract requests a workflow, Kernel runs it, and the contract receives the result with a proof that the workflow ran exactly as written. Froben and Flashcast Fermah did not walk away from ZK. Froben still runs, now on top of Kernel, and its paying customers include ZKsync Era and Abstract. In Froben, a workflow defines the proving pipeline and Kernel schedules the work across the prover fleet. Flashcast Social shows a different use of the same engine. A user posts a question, and within seconds it becomes a live prediction market. Kernel then runs the resolution workflow and settles the market automatically when the event happens, with no committee and no dispute window. Proof generation and market resolution have little in common on the surface. What they share is a condition that has to produce an action, and that is the workload Kernel is built around. Fermah also lists data aggregation, pricing oracles and cross chain coordination as things Kernel can run. Where agents fit Fermah's latest thinking extends this to autonomous software agents. An agent working onchain needs to find available actions, check whether it is allowed to run them, pay, execute, and use the output as the input to its next step. Fermah describes this as action manifests, which make actions machine readable across discovery, eligibility, execution, payment and output. The shift in one line The old question was how to generate ZK proofs efficiently at scale. The new one is how a protocol can turn a verifiable condition into a complete, executable sequence. Proving is still part of the system, but it is now one workload among several that Kernel coordinates.

they said to visualize what u want, so I did😅✨ 11 DAYS.
✍🏼TRACK 11: MANIFEST
hope. around this time I was really starting to lean into manifestation, my feminine energy, paying attention to the signs and synchronicities around me while again, getting intentional about what and who I wanted to attract into my life. and somewhere along the way I realized… aren’t fantasies basically manifestations too? 😂 you’re picturing something you want before it exists and letting yourself believe maybe one day it actually could. underneath the kitchen counters and staying up until the birds start chirping, this song really is about hope. letting yourself want something before you have any proof it’s coming. 10 more chapters to go 🤍
AWAKENED — OCTOBER 16 #singersongwriter# #newmusic# #womeninpop# #manifest# #popmusic#

Yesterday in Porto, I had the honour of interviewing the legendary Ai Weiwei about his book On Censorship, inside our permanent MANIFESTO library in partnership with Livraria Lello. What an honour to become a small part of this library’s rich history, and to have the opportunity to speak with one of the world’s most important artists and voices on censorship, freedom and the power of expression. A very special day, in a very special place. I can’t wait for you all to hear our conversation soon 📚🤞🏼 @service95

EvoOntology：给数据智能体一个会成长的语义层 数据智能体（Data Agent）面对异构数据时，能“看见”的往往只有表名、列名和文件路径；指标如何定义、实体如何关联、业务上有何约束，这些语义都藏在数据之外。让智能体每次从零探索，代价高且易出语义错误；把人工维护的语义层全量注入提示词，又难以扩展、随数据漂移而失效。这就是论文中的 agent-data gap。 中国人民大学 ruc-datalab 提出的 EvoOntology（arXiv:2609.15779）给出了新解法：首个面向数据智能体的自进化本体层（Self-Evolving Ontology Layer）。核心洞察是把本体当作“可训练的智能体状态”：不更新模型权重，而让语义知识随真实使用持续积累、验证、版本化演进，每次变更可检查、可比较、可回滚。 本体以三层结构封装为 MCP 服务器，供智能体运行时主动查询： · Schema Layer 定义本体的表示边界（节点族、语义关系、引用模式）； · Content Layer 是类型化语义图，由 Terms / Mappings / Constraints / Evidence 四类节点构成，每个语义对象都有工作负载证据支撑，落地验证后才提交； · Tool Layer 仅暴露 browse_semantics 与 resolve_semantics 两个 MCP 工具加紧凑会话 manifest，智能体按需检索当前步骤所需语义，而非全量注入上下文。 本体遵循 Build → Use → Evolve → Evaluate → Publish 生命周期：构建器智能体从真实工作负载推导候选概念并对照原始数据验证；交互轨迹被持续记录；进化时诊断重复行为、归因到具体层级并生成局部补丁；候选版本只有在相同数据、相同智能体、相同解码设置与交互预算的配对评估中可复现地胜过父版本，才会发布。 在 BIRD、DDR-10K、InsightBench 三个基准、四个 LLM 骨干（GPT-5.5 / GPT-5.6-sol / Claude-Sonnet-5 / Claude-Opus-4.8）上，EvoOntology 一致优于无本体 ReAct 与静态初始本体。其中 DDR-Bench 上仅“自进化”环节就贡献了 +7.7（81.8 → 89.5），验证了持续演进相对静态语义层的增量价值。 论文： 代码：

阿里云开源「企业级 Agent 白皮书」 2026 年最新发布，是 2025 年 9 月「AI 原生应用架构白皮书」的升级续作。全书按 架构 → 构建 → 运行 → 治理 → 调优 的全生命周期组织，共 7 篇 30 章，由阿里云数十位一线工程师分工撰写，并纳入吉利、塔斯汀、MiniMax、哔哩哔哩、信永中和等外部企业案例。 它的写作动机很明确：过去一年市场重心已经从 “如何快速搭出一个 Agent” 转移到三个新挑战，工程化（从概率智能到可靠生产力）、规模化（从单点试验到智能基础设施）、组织化（从 Agent 孤岛到进入核心业务流程）。现有的框架文档和教程基本不回答这些问题，这本白皮书填补的正是这个空白。 开源地址 # 各篇核心内容 架构篇（1–2 章） 建立认知框架。给出 Agentic Application 的六个判定特征（以任务结果为中心、运行时决定部分执行路径、能作用于环境、维持跨请求状态、受确定性机制约束、可观测可评估）和成熟度四级模型（L1 辅助生成 → L2 受控自动化 → L3 Agentic Execution → L4 规模运营）。两个重要的解耦判断：用哪种形态取决于任务结构，处于哪级成熟度取决于治理完备程度；单 Agent / Long-Horizon / 多 Agent 是沿时间跨度和协作结构两个正交维度的扩展，不存在“必须升级到多 Agent”的路径。贯穿的原则是“最低充分架构”，为任务选择成本与风险可接受的最低复杂度。 构建篇（3–6 章） 是方法浓度最高的部分，按“范式—任务—信息—行动”还原构建过程： · 任务：Agent Loop 五阶段（Prepare→Model→Act→Observe→Verify）+ 十态任务状态机，要害是“消息历史不应是任务状态的唯一来源”；完成判定的核心原则是“模型只能申请完成，Harness 依据环境证据提交完成”，验证分五级并与风险匹配。 · 信息：Context 是动态“编译”而非静态字符串。本章的独创设计是 Context Manifest，每次调用记录上下文每个片段的来源、作用域、版本、信任级别、选中理由和内容哈希，使“模型看见了什么”变得可解释、可回放、可审计。信息被五分为 Context/State/Memory/Knowledge/Skill，其中 Memory（个人经验）与 Knowledge（组织内容）必须分列，因为治理责任不同，“放进同一个向量库会同时失去两类治理能力”。 · 行动：统一 Action Plane（意图→Schema 校验→身份绑定→策略决策→执行→观测），关键三分：“模型看见工具 ≠ Harness 注册了工具 ≠ 获得执行授权”。协议定位清晰：Function Calling 是模型-Harness 意图接口，MCP 是 Harness-能力提供方连接协议，A2A 面向拥有独立任务循环的远程 Agent，“协议选择由能力是否拥有独立任务循环决定，而非新旧或流行度”。 运行篇（7–12 章） 处理规模化后的工程问题，大量内容达到了分布式系统的专业深度：沙箱后端选型判据（容器/gVisor/MicroVM 按代码可信度与租户边界取舍）；状态外置后 Event Log / Checkpoint / 工作区快照三者不可互相替代，且“Durable Execution ≠ 外部动作恰好执行一次”；AI 网关对 LLM/MCP/Agent 三类流量按不同粒度治理，其中“严格预算需要原子预留而非阈值检查”的数学化分析（余额 100、两笔 80 的并发请求都会通过）是真实的并发工程细节；多 Agent 编排强调“最小充分共享”，共享的是上下文来源而非同一个 Context 窗口。 治理篇（13–16 章） 让自主运行的系统变得可信。可观测性的判据是“请求成功 ≠ 任务成功”；安全章同时把 Agent 当被攻击对象和行为主体来防护（身份是全章最扎实的部分：数字工牌、Token Exchange 权限收敛、On-Behalf-Of 且 Agent 权限 ≤ 用户权限）；资产管理把 Prompt/Skill/MCP/Agent 当作运行时依赖做注册与版本治理。第 16 章 Agent Simulation 是全书原创性最强的一章：Agent 行为之所以不可验证，是缺制度前提（角色无外部标准、失败无自然代价、身份不连续），模拟是当下唯一可做的事，本质是“用可靠 Harness 约束不可靠内核”。它甚至给出诚实的统计学提醒：n 次零违规的 95% 置信上界约为 3/n 而非零。 调优篇（17–24 章） 的组织原则是“归因决定方法”：先排除环境故障、再修 Harness、最后才动模型，“把本应由上下文或工具协议解决的问题当成模型不行，是代价最高的一类误判”。主线是数据飞轮：Trace→Trajectory→黄金数据集（输入/轨迹/结果/判据四要素）→Badcase 闭环→受控自进化（模型生成的改进一律是候选变更，须回流构建、过门禁、可回滚）。模型调优章对 SFT/Agentic RL/蒸馏的适用边界、奖励投机的三套机制分离（训练奖励、独立评测、系统硬约束）论述相当严谨，广引 DeepSeek-R1、Tulu 3、FrugalGPT 等外部工作。 总结篇（第 30 章） 是全书思想密度最高的总结。当企业同时运行多 Agent、多框架、多租户时，同样的工程要求在每个应用里被重复且不一致地实现，这本质上是缺一个共享的系统层。Agentic OS 被给出“窄定义 + 三条否定”：为 Agent 任务提供公共运行对象、能力接入、可强制边界与统一证据的系统层，它不持有任务语义、不是又一个框架、不必然改内核。能力下沉有三条判据（复用性 + 强制性或可验证性），九类管理对象（其中 Budget Lease 预算租约最易被忽略），并提出“自治上限由可撤销范围与可证明范围决定，而非模型能力”。 调研报告 的 1906 份问卷给出一个关键发现：已开发或开发中 Agent 的企业占 46%，但真正上生产的仅 18%；有评估体系的企业任务成功率是无评估者的约两倍，卡点不是模型能力，是 Harness 层的工程配套。这与全书立意互为印证。

I just published a manifesto for all the developers out there who use an AI coding tool, but feel like they're not shipping much faster I wrote 10 principles for changing how you build software with AI, based on what I've seen work for teams across Amazon