Preparing for Radical Transformation ✍ Leonard Tang🕐 2026-04-18📦 6.3 KB 🟢 已读 𝕏 文章列表 本文探讨了超级智能已至的背景下,AI 发展的加速及其带来的挑战与机遇。作者指出 AI 将重塑高 prestige 的白领行业,同时也面临持续学习、安全和验证等难题。文章建议通过自动化低垂果实、拥抱开源权重模型、利用隐性知识及建立安全基线来应对。 AIAgent转型超级智能企业落地安全模型验证开源模型 # Preparing for Radical Transformation **作者**: Leonard Tang **日期**: 2026-04-17T16:35:29.000Z **来源**: [https://x.com/leonardtang_/status/2045179350547612043](https://x.com/leonardtang_/status/2045179350547612043) ---  Distilled from a Night with Haize Labs: Preparing for Radical Transformation Delmonico's April 16, 2026  The title of this dinner is Preparing for Radical Transformation. With the rate of progress of this technology, a more apt title would have been Embracing Radical Transformation. Though it has not yet fully dispersed throughout the economy, by many sensible metrics super intelligence has already arrived. It is an incredible inning of history we are in. In the same hour, I experience both incredible hope and fear for what is to come. AI will enable us to solve the most persistent of challenges in medicine, materials, physics, the sciences, engineering. AI will also rapidly rip through the fabric of white collar services, most immediately upending high prestige professions like software engineering, consulting, legal, and financial services. It is not my intention to be a pessimist. But we must be sober about the reality we face: if we do not proactively shape how AI is deployed in our businesses, our businesses will very much be left behind. I selected this group of technology, risk, and AI leaders tonight to plan how we chart the path forward, stay ahead of the technology curve, and enable radical transformation. I want all of us to learn from each other. My goal is to speak as little as possible tonight and learn from you. ## This Inning of History I have been researching AI since 2020. Never has AI progressed so rapidly. Models can now trivially handle many-hour long verifiable tasks and perform work at parity or better than the majority of the junior white collar workforce. 3 years ago the models struggled at basic arithmetic.  The models will only continue to get better on anything that can be readily hillclimbed. Too, it is an open secret that all the frontier labs are investing heavily into Recursive Self Improvement: the ability for AI to improve itself. Most immediately, this means AI experimenting with new architectures, training algorithms, and data curation recipes to train future generations of itself. This implication is that AI progress will only continue to accelerate, at least on a subset of verifiable + economically valuable work. ## Challenges at the Frontier In spite of all this progress, three enduring challenges remain. Continual Learning. AI cannot continuously adapt to the ever-changing world and sources of information around us at the same rate at which humans can. This would require a breakthrough in the ability to bake actions and metrics the model experiences back into the weights of the model. Unlike the rapid capabilities improvement in the last 3 years, this requires a deeper insight than brute scaling of parameters, compute, and data. There is no obvious timeline on which this challenge will be resolved. Security. The power of agents is that they can explore solution spaces with creativity. This is also a security nightmare. For agents to be adopted with confidence in the enterprise, clear trust boundaries must be drawn, rigorous red-teaming must be performed, and ongoing monitoring must be enabled. Verification. The following statement has always been true in the post-LLM era: if you could measure the correctness of the model when performing task family X, you could improve capabilities on X. The next major frontier is in how to improve model capabilities on tasks where one cannot easily, via programs or enumerative rules, measure correctness. This is the verification problem: the ability to verify the correctness of machine output. There is no shortage of difficult-to-verify work within an enterprise context, and AI cannot reliably be of use in these settings until the verification problem is solved. ## Enabling Radical Transformation: Your Opportunity Automate the Low Hanging Fruit. While AI currently struggles on ambiguous, difficult-to-verify tasks that require expert judgement, it can be immediately beneficial in automating work where there is a very clear set of processes, procedures, and definitions of success. If you have well-articulated SOPs already available, it is a no-brainer to turn that into an agent. Be Open to Open Weight Models. Many enterprise tasks are narrow enough to fine-tune (via SFT and RL) a custom model for. This can save you orders of magnitude in cost and latency. This, of course, also prevents the frontier labs from training their models and harnesses on your proprietary business context and data, which is the ultimate enduring moat for your institution. It will prevent your entire workforce from being automated. Bring Tacit Knowledge to Bear. Design agent experiences in a way that incentivizes domain experts to articulate their thought process towards producing an output. Use this to create verification signal in the form of reward models and judges. This ultimately enables you to improve model performances in domains that were previously unsolvable. Set Groundwork for Safe, Secure, Reliable Agents. Conduct testing in sandboxes and simulation to understand the threat surface and trust boundaries of your agent. Use semantic observability and clustering to understand common vulnerability patterns. Leverage AI for joint red-teaming and blue-teaming exercises to automatically harden your system. ## Thank You Thank you to our friends at Nomura, Amex, UBS, Macquarie, Columbia, TriSpan for the spirited conversation and company on this topic last night.  ## 相关链接 - [Leonard Tang](https://x.com/leonardtang_) - [@leonardtang_](https://x.com/leonardtang_) - [4.7K](https://x.com/leonardtang_/status/2045179350547612043/analytics) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [12:35 AM · Apr 18, 2026](https://x.com/leonardtang_/status/2045179350547612043) - [4,738 Views](https://x.com/leonardtang_/status/2045179350547612043/analytics) --- *导出时间: 2026/4/18 13:53:21*
D Don't talk to me, talk to my agents 文章介绍了Raft提出的一种新的跨公司协作方式:联合通道。这种方式允许双方将特定的人、AI代理和上下文放入同一个房间,实现人与代理的直接协作,而无需共享整个服务器。它重新定义了跨公司工作的基本单位,从“人”变为“人和他们的代理”,提高了协作效率并保证了安全性。 技术 › Agent ✍ Raft🕐 2026-07-24 Agent协作跨公司工作联合通道AI安全RaftScopeDB
纳 纳瓦尔:苹果已死,SaaS 下一个,你只有 18 个月 文章指出,随着 AI 导致界面层实时商品化,苹果的软件护城河已崩塌。纳瓦尔预言纯软件不可投资,SaaS 公司因技术复制门槛降低而面临毁灭。创始人需在 18 个月内构建新的护城河(如分发、网络效应、硬件集成),否则将被市场淘汰。 投资 › 宏观经济 ✍ Saito🕐 2026-05-01 NavalSaaS投资创业护城河AI苹果科技趋势Agent转型
让 让 Claude 半夜自己干活,只有两条路 文章探讨了让 Claude 无人值守工作的两种安全路径:一是使用隔离的闲置机器给予完全权限;二是在现有机器上通过严格参数限制权限。作者通过分析两种方案的适用场景和风险控制,提出了“爆炸半径”公式,并给出了具体的实施建议和避坑指南。 技术 › Agent ✍ 老金🕐 2026-07-30 ClaudeAgent自动化安全DevOpsClaude Code无人值守权限管理
装 装完 Codex 不知道干什么?这 6 个 GitHub Skills 让你做视频搞钱 文章介绍了 6 个适用于 Codex 的 GitHub Skills,涵盖动效生成、视频剪辑、批量制作、AI 生成及中文剪辑等工具,帮助用户构建自动化视频工作流以提升效率。 技术 › Codex ✍ Kay🕐 2026-07-30 Codex视频制作AgentSkill自动化HyperFramesRemotionAI剪辑工作流
H Hermes Agent Masterclass: 安装、配置与基础命令 本文介绍了开源 AI Agent Hermes 的安装与基础配置。Hermes 由 Nous Research 开发,支持多种模型和终端运行,具有会话复用和技能积累的独特设计。文章详细讲解了在 Windows、macOS 和 Linux 上的安装步骤、设置向导流程以及如何配置推理提供商。 技术 › Agent ✍ tonbi🕐 2026-07-29 HermesAgentAI安装教程CLI工具与效率开源
顶 顶尖VC 2026年动向观察:硬基建与深垂直的押注 文章分析了2026年顶尖独立VC机构的投资动向,指出资本正从浅层应用转向硬基建、物理世界和深垂直领域。通过分析赛道年龄、轮次和具体公司(如Atoms、Etched、Cognition等),揭示了AI创业窗口正在收窄,而推理芯片、机器人、企业Agent及特定垂直工作流成为投资热点。 投资 › 风投 ✍ snowboat🕐 2026-07-28 风投AI硬科技机器人Agent芯片垂直行业创业
如 如何利用AI构建和扩展单人企业 文章介绍了利用AI构建和扩展单人企业的完整蓝图。通过使用AI员工(如Viktor)在内容、项目、拓展、财务和广告五个领域实现自动化,只需保留决策环节。文章探讨了两种构建方式:快速路径和自定义构建,强调业务知识库和操作规则的重要性。 技术 › Agent ✍ Machina🕐 2026-07-26 AI单人企业自动化Agent生产力Viktor商业模式知识库LLM效率
管 管理 AI 员工团队:Ryan Carson 的高效工作系统 Ryan Carson 分享了如何作为唯一员工,通过管理云端 AI Agent 团队日处理 40 个 Pull Request 的经验。文章详细介绍了将工作迁移至云端、建立工作节奏、将重复检查自动化以及控制 Token 成本四个关键步骤,强调在 AI 时代,优秀的工程管理能力比以往任何时候都重要。 技术 › Agent ✍ The Startup Ideas Podcast (SIP)🕐 2026-07-25 AIAgentDevin工程管理自动化云端开发成本控制DevOpsLLM
我 我用 WorkBuddy 跑了 3 个活儿,发现它不是助手,是数字员工 文章通过三个真实任务评测了 WorkBuddy 的实际应用:视频 Skill 管理工具开发、电商数据清洗分析及竞品监控自动化流程。作者指出 WorkBuddy 不仅是聊天工具,更是集专家判断、专业 Skill 和自动化执行于一体的 AI 工作台,能有效提升处理固定流程任务的效率。 技术 › 工具与效率 ✍ 行者AI视频🕐 2026-07-24 WorkBuddyAI自动化数据分析实战评测效率提升Agent专家模式技能管理竞品监控
用 用飞书多维表格打造你的数字分身 文章介绍了如何利用飞书多维表格结合 AI 技术打造知识库“数字分身”。通过三级进化:将知识结构化存库、利用 AI 字段自动化处理知识、接入飞书机器人实现对话交互,解决专业知识沉淀与复用难题,提升业务效率。 技术 › 工具与效率 ✍ 黄小木🕐 2026-07-24 飞书多维表格AI知识库数字分身自动化Agent低代码
如 如何使用图工程构建多因子Alpha模型 本文详细介绍了如何利用图工程构建对冲基金级别的多因子Alpha模型。作者阐述了从Prompt、Loop、Swarm到Graph的技术演进路径,重点介绍了Slate这一运行时工具,它能通过JavaScript编写的程序自动化执行复杂的图结构任务,解决了多智能体系统中的协调与失败处理难题。 投资 › 量化交易 ✍ Roan🕐 2026-07-24 图工程多因子模型Alpha量化交易SlateAgentAI对冲基金系统设计
改 改一次 Prompt,怎么知道 Agent 有没有悄悄变差 文章探讨了在企业级 Agent 开发中,如何通过建立版本化固定案例集和审计日志来准确评估模型更新后的质量变化。强调不能只看平均分,需逐条对比案例的改善与退步,并实施一票否决机制以确保安全和可追溯性。 技术 › Agent ✍ jager🕐 2026-07-24 Agent评测审计质量控制Prompt安全版本管理企业应用风险控制评估方法