The Four Micro-Revolutions of the Intelligence Revolution brett goldstein 2026-05-06 34.9 KB 120.6K 185 ☆ 收藏 𝕏 文章列表 本文由前Google M&A团队成员撰写，深入分析了“智能革命”的演进规律。作者结合Carlota Perez的技术革命框架，预测AI发展将经历四个微观阶段：聊天、代理、语境和平台。文章指出，最终的赢家将掌握“语境”层，成为历史上最大的公司，因为它将聚合整个AI行业并消耗所有数字劳动力。 AI趋势技术革命Agent行业预测ChatGPTOpenAIAnthropicContextPlatformCarlota Perez 65 值得读 ▶ 文章基于经典理论提出了AI发展的四阶段框架，逻辑清晰、可读性强且紧扣当前时效，但缺乏具体的数据支撑和可落地的实操指南。 # The Four Micro-Revolutions of the Intelligence Revolution **作者**: brett goldstein **日期**: 2026-05-05T17:23:15.000Z **来源**: [https://x.com/thatguybg/status/2051714355398602957](https://x.com/thatguybg/status/2051714355398602957) ---  I used to get paid to predict the future for the CEO of Google. It was my job on the M&A team to study where technology was going before it was obvious so Google could allocate billions of dollars of resources accordingly. I met tiny startups like Scale AI before they became big, watched markets form before they had names, and tried to understand which shifts were real and which were just noise. I was there to see the last AI hype cycle of 2014 as it boomed and busted. I even helped Google get a ChatGPT-like messaging app & assistant product off the ground and later started an accelerator where I gave talks on “How to think about the future” to hundreds of founders. This AI cycle is certainly different. It’s an Intelligence Revolution. Revolutions always feel chaotic while you are inside them and many bet their careers and businesses on a myopic view of how they play out. But every revolution follows the same pattern. It’s a pattern so predictable that I am confident betting my own career and business on it. So I figured I'd go ahead and call my shot: 1. The Intelligence Revolution will unfold in four micro-revolutions: Chat, Agents, Context then Platform. 2. The winner of the Intelligence Revolution will be the winner of The Context Revolution. Context creates real switching costs where Chat and Agents have not. Once users are aggregated in the Context Revolution, a Platform will be built on top that aggregates the entire AI industry: generative apps and agents. 3. Almost all of the players in the first two revolutions will be wiped out but there's is still opportunity for a new player to win the Context Revolution and thus the Intelligence Revolution. Incumbents are historically weak at winning multiple revolutions and OpenAI & Anthropic are suspiciously absent from the conversation about Context. 4. The winner of the Intelligence Revolution will be the largest company in history because it will be an index on the AI industry, which will consume all digital labor and software. AI will in turn consume more and more the world economy as agents and software become easier to build and robotics goes into large scale production. ## How Revolutions Happen There's no shortage of literature on the anatomy of a technology revolution. It's a pastime of computer-savvy, capitalistic philosophers like Benedict Evans, Geoffrey Moore, Balaji Srinivasan, and my favorite, Carlota Perez. They all contribute a different theory but in reality, it all comes together in one simple framework: 1. Incubation - Revolutionaries tinker with fringe ideas sometimes for years on end. 2. Installation - A paradigm shift drives a frenzy and then a crash pursuing new opportunities - Spark/Irruption - A breakthrough happens when an idea meets the perfect conditions. - Frenzy - Entrepreneurs and capital rush in pursuing new opportunities unlocked by the paradigm shift. - Crash - When capital is deployed into the best opportunities, marketing/innovation budgets are exhausted, and experimentation finds its limits, the frenzy ends and mass company extinction ensues. 3. Deployment - The winners of the revolution reach scale and deploy their products beyond the early adopters.  Overly of Carlota Perez' (blue) and Gartner's (black) frameworks for technology revolutions Incubation Revolutions always start with the revolutionaries, the microorganisms in the primordial soup of change. In politics they're activists who push for new policy, in society they're artists and public figures who define culture, and in tech they're the tinkerers pushing digital systems to their limit. As they say "if you want to see the future, spend time with your weirdest most technical friend."  Homebrew Computer Club were the Revolutionaries of the Personal Computer Revolution Spark Eventually, the right idea meets the right conditions and it creates a paradigm shift. This is the ah-ha moment that leads to a paradigm shift that powers the revolution. Revolutions begin with one of three kinds of paradigm shifts: - Regulatory Shift - A change to the laws like marijuana, stablecoins, and or peptides (soon) - Cultural Shift - A change in human behavior at scale like Gen Z entering the workforce, COVID, or political polarization. - Technological Shift - A technological breakthrough like the transistor, the iPhone + app store, the Bitcoin whitepaper, or Attention is All You Need + scaling laws. For technology revolutions, the shift usually begins with the creation of a new tool that dramatically reduces the cost or effort for engineers' to create or distribute high value products. - The printing press for making books - Interchangeable parts for making guns - Digital audio workstations for making music - AWS for making websites - Stripe for making internet businesses - LLMs for making AI products In fact, technology is at its core a force for democratization, making things cheaper or easier to do which allows them to be accessible to more people.  Every revolution looks like the Industrial Revolution, where massive opportunities were unlocked when building blocks became composable Frenzy When valuable products suddenly become significantly cheaper and easier to build, entrepreneurs pour in. Then capital. Then more entrepreneurs and more capital. Like in the Cambrian Explosion millions of years ago, where biodiversity flourished when oxygen suddenly increased in the atmosphere, the range of companies getting started after a spark becomes overwhelming.  While resources are abundant, companies don't have to optimize their spending. In fact, resource abundance is directly correlated to how exploratory or exploitative any system is, whether it's startups in a bubble, VCs with fresh powder in a bull market, microorganisms in the Cambrian Explosion, or young people with life ahead of them (seeking travel and new friendships). In the frenzy, competition and capital abundance drives bidding wars, expensive marketing campaigns, and other frivolous spending.  Sorry Roy <3 Competition drives the spending higher and higher. There seems to be a direct correlation with the amount of money startups spend on brand and marketing and the severity of a bubble. If you have to ask, it's probably a bubble. Crash At some point, the market is saturated, limitations of the technology are found, enterprise marketing/innovation budgets are spent, the hype dies down and resources start to contract. Years after the Cambrian Explosion when ecosystems matured and competition for resources became more fierce, "pruning" ensued. Less adapted organisms died out. Similarly, companies that are mismanaged or pursuing opportunities propped up by the bubble die and the Revolution reaches maturity. Deployment After the crash, the winners consolidate power, their technology gets adopted at scale and moats become insurmountable. The revolution goes corporate and enters the deployment period. Crossing the Chasm Deployment can happen over many years since certain individuals and industries are slower to adopt new paradigms than others. Geoffrey Moore's classic book Crossing the Chasm explains this perfectly: first its the early adopters/tinkerers (younger people and tech companies) and last, the laggards (older people and institutions).  From Crossing the Chasm Overlapping Revolutions Ben Evans showed that revolutions tend to overlap and even catalyze each other.  Adapted from Ben Evans / Andreessen Horowitz Humanity is a closed loop, so the deployment phase for one revolution creates the conditions for the paradigm shift that sparks the next revolution.  Nested Revolutions Just as revolutions can feed into each other, there are often multiple revolutions within a single revolution. The internet for example includes major revolutions such as cloud as well as micro-revolutions driven by technologies like AWS and Stripe. I believe the Intelligence Revolution is the same - a collection of four overlapping micro-revolutions that will end in one winner which will be the biggest company in history.  ## R1 - The Chat Revolution The first revolution gave the world intelligence that could hold a conversation. OpenAI was the obvious winner, with ChatGPT now at nearly 1 billion weekly users - the most widely used AI chat app of all time. GPT-4 gave developers reliable AI chat which revolutionized industries like therapy, content creation, and more. Unicorns like Copy.ai, Character.ai and Bible Chat emerged built directly on top of the technology.  With chatGPT at scale and GPT-4+ in the hands of millions of developers, the AI tinkerers discovered structured output and tool calling, which gave chatbots new abilities. ## R2 - The Agentic Revolution In early 2023, the first examples of structured output allowing chatbots to take actions in external apps (tool calling) drove OpenAI and Anthropic to standardize structured output, tool calling and later MCPs which unleashed the Agentic Revolution. These new capabilities lead to an explosion in agent startups from AI SDRs to customer support agents, with many worth billions today. With the advent of Claude Code, Opus 4.5 and OpenClaw, the capabilities became more reliable and more powerful, which in turn gave rise to even more agent startups.  Anthropic has been the clear winner - Claude Cowork/Code is a truly agent native product and the company’s historical earnings growth speaks for itself. Claude has received worldwide attention and has become synonymous with agents. But we are currently at peak frenzy. Half of the latest YC batches are building agent startups, the guys who sold PPE during the pandemic all have AI Psychosis, and there's a 2021-crypto-boom-level energy at OpenClaw meetups around the world. When you're getting texts from your crazy uncle about setting up his Mac Mini and startups are burning $150K on launch videos, you know we're at the top. The opportunities in the agent revolution are still abundant, but the revolutionaries have already zeroed in on the catalyst for the next major paradigm shift: context. It started with Jaya Gupta and Andrej Karpathy's essays on context/knowledge graphs, Obsidian's viral second brain moment, and Garry Tan's GBrain, and will end in a new revolution. ## R3 - The Context Revolution Before the end of 2026, we will see startups and incumbents like Anthropic and Open AI attempt to standardize the context layer for intelligence in the same way that was done for tool calling and MCP in the Agentic Revolution. It fills an obvious gap. Agents that have context dramatically outperform agents without it. But most importantly, it is the first piece of the AI stack that a moat can be built on top of. Models are interchangeable, but context isn’t (as much). We have seen early attempts at standardization like memory.md or the file system paradigm in general, but these are hacky, incomplete solutions to the context problem.  The final primitive, I believe, will be a context/knowledge graph with set entities (like people and companies), relationships, and a handful of tools for searching through and updating information like dreaming, episodic memory, and more. There are many startups (mine included) already building context graphs or company brains as some say, but until the primitive is established, no clear winners can be determined. OpenAI and Anthropic appear to be suspiciously absent, despite having originated the ideas that brought about both the Chat and Agentic revolutions. This time, it's the open source community tinkering with Context. This is actually consistent behavior for incumbents in other technological revolutions. Winners of previous revolutions tend to be wedded to old paradigms technologically, ideologically or through brand (e.g. Google feels more cloud native than Microsoft). So it’s entirely likely that while Anthropic has clearly won the Agentic Revolution, they may not win the Context Revolution. So who will win? I'll save that for another piece, but whoever does will be set up to win the last revolution and the Intelligence Revolution as a whole. ## R4 - The Platform Revolution Every major technology cycle eventually produces a platform: a place where other people build, distribute, and monetize products on top of the winning layer. It creates the powerful network effects and virality necessary to reach hyperscale and build a powerful moat. Apple had the App Store. Amazon had the marketplace. TikTok had the creator graph. The Intelligence Revolution will have the same thing. OpenAI tried to create this with custom GPTs and agent builders. Startups have made similar attempts. But these mostly failed because they lacked the two things a platform needs: - Tools to create extremely high quality supply (e.g. TikTok's video editor) - A place for demand to meet that supply (e.g. TikTok's For You page) Simply put, GPTs and agents weren't any higher quality or easier to create on chatGPT than elsewhere and chatGPT users didn't really want them. The Context Revolution changes that. When it is deployed, millions of people will have a living database of their most important information: people, companies, conversations, meeting notes, documents, decisions, and more. That database becomes the substrate for a spectacular range of agents and generative apps. The winner of the Context Revolution that owns that substrate simply needs to lean in by: - Providing the best tools for creating agents and apps (which is nearly automatic when you own the context layer) - Creating a place where users can discover, use, and pay for them. Today, tinkerers discover agents and generative apps on GitHub, X, and other corners of the internet. It is what Chris Dixon would call a "weak solution" - a hacky under-optimized way to solve a problem.  It's a clear cut example of the incubation phase of a revolution where the market is begging for a solution. This is the same transition TikTok went through. Video was taking off on social media platforms designed for text or photos. TikTok launched as the tool to create high quality videos to post on other platforms. But once it aggregated the supply of quality videos, it was able to aggregate demand, transitioning to the best place to consume video as well. And from there the flywheel just keeps spinning.  Here's the logic: - Context is sticky and is the backbone of the highest quality agents and generative apps. - Whoever wins the Context Revolution is positioned to become the tool for creating the highest quality agents and apps. - Whoever becomes the tool for creating the highest quality agents and apps is positioned to become the place for discovering, using, and paying for them. Once a Platform like this is established, it sparks a Platform Revolution where a flurry of new businesses are built on top of it. - TikTok drove content businesses - Apple's App Store drove mobile app businesses - Shopify drove eCommerce businesses But here's where it gets wild. The most powerful platforms become indexes of the markets they aggregate: single proxies for entire categories. Apple became an index of mobile software. Amazon became an index of e-commerce. Shopify became an index of independent merchants. So what market would a platform for agents and generative apps aggregate? All of them. - Whoever becomes the place for discovering, using and paying for agents and apps becomes an index on the AI industry - Agents will replace (digital) human labor and generative apps will replace software eventually, so an index on the AI industry becomes an index on all industries that touch a computer today (about 50% of the entire global economy). Of course, it doesn't end there. Marc Andreessen said "software is eating the world" in 2011 and now AI is eating software. As software and agents get easier to create and humanoid robots scale to production, more and more of the physical economy will become digitized and consumed by AI. This is why the winner of the Context Revolution and then the Platform Revolution will likely be the most valuable company of all time. There are some early signals of who this may be, but the Context Revolution has barely just begun, so the title is still up for grabs. My company, Micro is a bet to win the Context Revolution and all of the strategic decisions we've made thus far have been driven by an unconscious understanding of the ideas presented in this piece. I wrote this as a map for myself and founders like me to be able to make even better decisions in such a chaotic moment in history. Thanks to Isabel Swope, Ravi Mishra, and Saul Carlin for edits and conversations. ## 相关链接 - [brett goldstein](https://x.com/thatguybg) - [@thatguybg](https://x.com/thatguybg) - [71K](https://x.com/thatguybg/status/2051714355398602957/analytics) - [Copy.ai](http://copy.ai/) - [Character.ai](https://character.ai/) - [$150K](https://x.com/search?q=%24150K&src=cashtag_click) - [Micro](https://micro.so/) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [1:23 AM · May 6, 2026](https://x.com/thatguybg/status/2051714355398602957) - [71.8K Views](https://x.com/thatguybg/status/2051714355398602957/analytics) - [View quotes](https://x.com/thatguybg/status/2051714355398602957/quotes) --- *导出时间: 2026/5/6 20:53:51* --- ## 中文翻译 # 智能革命的四个微型革命 **作者**: brett goldstein **日期**: 2026-05-05T17:23:15.000Z **来源**: [https://x.com/thatguybg/status/2051714355398602957](https://x.com/thatguybg/status/2051714355398602957) ---  我曾经在谷歌 CEO 手下从事预测未来的工作，并以此获得报酬。 我在并购团队的职责是在趋势尚未明朗之前研究技术的发展方向，以便谷歌据此配置数十亿美元的资源。 我见过像 Scale AI 这样的微小初创公司，那时它们还没有做大；我目睹过市场在尚未定名之前就已成形；我还试图理解哪些变化是真实的，哪些只是噪音。 我亲眼见证了 2014 年上一轮 AI 热潮的兴起与破灭。我甚至帮助谷歌打造了一款类 ChatGPT 的消息应用和助手产品，后来还创办了一家加速器，向数百名创业者做过题为“如何思考未来”的演讲。 这一轮 AI 周期显然不同。这是一场智能革命。 身处革命之中时，革命总是让人感觉混乱不堪，许多人把自己的职业和事业押在对革命进程的短视判断上。 但每场革命都遵循同样的模式。 这个模式如此可预测，以至于我有信心把自己的职业和事业押在上面。 所以我想干脆把我的预测亮出来： 1. 智能革命将经历四个微型革命展开：对话（Chat）、智能体（Agents）、上下文（Context），然后是平台（Platform）。 2. 智能革命的赢家将是上下文革命的赢家。上下文能创造真正的转换成本，而对话和智能体不能。一旦用户在上下文革命中被聚集起来，一个平台将被建立在其之上，聚合整个 AI 产业：生成式应用和智能体。 3. 前两场革命中的几乎所有玩家都将被淘汰出局，但仍有机会让新玩家赢得上下文革命，进而赢得智能革命。从历史来看，在位者在多场革命中连续获胜的能力很弱，而 OpenAI 和 Anthropic 在关于上下文的讨论中令人疑心地缺席。 4. 智能革命的赢家将成为历史上最大的公司，因为它将成为 AI 产业的指数，而 AI 产业将吞噬所有数字劳动力和软件。随着智能体和软件变得越来越容易构建、机器人技术进入大规模生产，AI 反过来将吞噬越来越多的世界经济。 ## 革命是如何发生的 关于技术革命解剖结构的文献并不匮乏。这是像 Benedict Evans、Geoffrey Moore、Balaji Srinivasan，以及我最喜欢的 Carlota Perez 这类精通计算机、信奉资本主义的哲学家们的一大爱好。 他们都提出了不同的理论，但实际上，这一切都可以归结为一个简单的框架： 1. 孕育期 - 革命者摆弄边缘性的想法，有时一摆弄就是好几年。 2. 安装期 - 一次范式转变引发狂热追逐新机遇，随后迎来崩盘 - 火花/爆发 - 当一个想法遇到完美条件时，突破便发生了。 - 狂热 - 创业者和资本蜂拥而入，追逐范式转变所释放的新机遇。 - 崩盘 - 当资本被投入到最好的机会中、营销/创新预算耗尽、试错触及极限时，狂热结束，大规模的公司灭绝随之而来。 3. 部署期 - 革命的赢家达到规模，并将其产品推广到早期采用者之外。  Carlota Perez 的（蓝色）与 Gartner 的（黑色）技术革命框架的叠加图 孕育期 革命总是从革命者开始，他们是变革原始汤中的微生物。 在政治领域，他们是推动新政策的活动家；在社会领域，他们是定义文化的艺术家和公众人物；在科技领域，他们是把数字系统推向极限的摆弄者。 正如人们所说：“如果你想看到未来，就多花时间陪你最古怪、最懂技术的朋友。”  家酿计算机俱乐部是个人电脑革命中的革命者 火花 最终，正确的想法遇到合适的条件，便引发了一场范式转变。 这就是带来范式转变、进而驱动革命的“顿悟”时刻。 革命始于三种范式转变之一： - 监管转变 - 法律的变化，如大麻、稳定币，以及（即将到来的）多肽类药物 - 文化转变 - 大规模的人类行为变化，如 Z 世代进入职场、新冠疫情，或政治两极化。 - 技术转变 - 技术突破，如晶体管、iPhone 与应用商店、比特币白皮书，或《Attention is All You Need》与缩放定律。 对于技术革命而言，这种转变通常始于一种新工具的诞生，它大幅降低了工程师创造或分发高价值产品的成本或难度。 - 印刷机之于制造书籍 - 可互换零件之于制造枪支 - 数字音频工作站之于制作音乐 - AWS 之于搭建网站 - Stripe 之于创建互联网企业 - LLM 之于打造 AI 产品 事实上，技术从本质上讲是一股民主化的力量，它让事情变得更便宜或更容易，从而使更多人能够使用它们。  每场革命看起来都像工业革命——当基础构件变得可组合时，巨大的机遇便被释放出来 狂热 当有价值的产品突然变得便宜得多、容易构建得多时，创业者便会涌入。 然后是资本。 然后是更多的创业者和更多的资本。 就像数百万年前的寒武纪大爆发——当大气中的氧气突然增加时，生物多样性蓬勃发展——火花之后涌现的公司数量会变得势不可挡。  当资源充裕时，企业无需优化开支。 事实上，资源充裕程度与任何系统的探索性或利用性直接相关，无论是泡沫中的初创公司、牛市中手握新资金的VC、寒武纪大爆发中的微生物，还是前途无量的年轻人（追求旅行和新友谊）。 在狂热中，竞争和资本充裕推动了竞价战、昂贵的营销活动和其他挥霍性开支。  抱歉了Roy <3 竞争推动支出越来越高。 初创公司在品牌和营销上的支出与泡沫的严重程度似乎存在直接关联。 如果你还需要问，那大概率就是泡沫了。 崩溃 在某个时刻，市场趋于饱和，技术的局限性显现，企业的营销/创新预算花完，炒作退潮，资源开始收缩。 寒武纪大爆发数年后，当生态系统成熟、资源竞争变得更加激烈时，“修剪”随之而来。适应力较弱的生物灭绝了。 同样，管理不善或追逐泡沫支撑起来的机会的公司会消亡，革命走向成熟。 部署 崩溃之后，胜者巩固权力，他们的技术得到大规模采用，护城河变得难以逾越。 革命走向企业化，进入部署期。 跨越鸿沟 部署可能持续多年，因为某些人和行业的范式采用速度比其他人更慢。 Geoffrey Moore 的经典著作《跨越鸿沟》完美地解释了这一点：首先是早期采用者/折腾者（年轻人和科技公司），最后是落后者（老年人和机构）。  摘自《跨越鸿沟》 重叠的革命 Ben Evans 指出，革命往往会相互重叠，甚至会相互催化。  改编自 Ben Evans / Andreessen Horowitz 人类是一个闭环，因此一场革命的部署阶段会为引发下一场革命的范式转变创造条件。  嵌套的革命 正如革命之间可以相互滋养，一场革命内部往往包含多场革命。 例如，互联网包含了云计算等重大革命，以及由 AWS 和 Stripe 等技术驱动的微型革命。 我相信智能革命也是如此——由四场相互重叠的微型革命组成，最终将产生一个赢家，它将成为历史上最大的公司。  ## R1 - 聊天革命 第一场革命为世界带来了能够进行对话的智能。 OpenAI 是显而易见的赢家，ChatGPT 目前拥有近10亿周活用户——是有史以来使用最广泛的AI聊天应用。 GPT-4 为开发者提供了可靠的AI聊天能力，彻底改变了心理治疗、内容创作等行业。 Copy.ai、Character.ai 和 Bible Chat 等独角兽企业直接构建在这项技术之上。  随着 ChatGPT 规模化，GPT-4+ 落入数百万开发者手中，AI 折腾者们发现了结构化输出和工具调用，这赋予了聊天机器人新的能力。 ## R2 - 智能体革命 2023年初，最早允许聊天机器人在外部应用中执行操作（工具调用）的结构化输出案例，促使 OpenAI 和 Anthropic 将结构化输出、工具调用以及后来的 MCP 标准化，从而引爆了智能体革命。 这些新能力引发了智能体初创公司的爆发式增长，从 AI SDR 到客服智能体，其中许多如今估值数十亿美元。 随着 Claude Code、Opus 4.5 和 OpenClaw 的问世，这些能力变得更加可靠和强大，进而催生了更多的智能体初创公司。  Anthropic 是明显的赢家——Claude Cowork/Code 是真正的智能体原生产品，该公司历史性的营收增长本身就说明了一切。Claude 受到了全世界的关注，并已成为智能体的代名词。 但我们目前正处于狂热的顶峰。 最新的 YC 批次中有一半在构建智能体初创公司，疫情期间卖PPE的那帮人全都患上了AI精神病，OpenClaw 全球聚会洋溢着2021年加密货币热潮般的能量。 当你收到疯狂叔叔发来的短信让你帮他设置 Mac Mini，当初创公司在发布视频上烧掉15万美元时，你就知道我们已经在顶部了。 智能体革命中的机会仍然很多，但革命者们已经锁定了下一个重大范式转变的催化剂：上下文。 它始于 Jaya Gupta 和 Andrej Karpathy 关于上下文/知识图谱的文章、Obsidian 病毒式传播的第二大脑时刻，以及 Garry Tan 的 GBrain，并将以一场新的革命告终。 ## R3 - 上下文革命 在2026年底之前，我们将看到初创公司以及 Anthropic 和 OpenAI 这样的行业巨头，尝试像智能体革命中对工具调用和 MCP 那样，将智能的上下文层标准化。 它填补了一个明显的空白。拥有上下文的智能体表现远超没有上下文的智能体。 但最重要的是，它是AI技术栈中第一块可以用来构建护城河的组件。 模型是可以互换的，但上下文（相对）不能。 我们已经看到一些早期的标准化尝试，如 memory.md 或文件系统范式，但这些对于上下文问题来说都是粗糙、不完整的解决方案。  我相信，最终的基础构件将是一个上下文/知识图谱，具有固定的实体（如人物和公司）、关系，以及一组用于搜索和更新信息的工具，如做梦、情景记忆等。 已经有许多初创公司（包括我的公司）在构建上下文图谱，或者像一些人所说的“公司大脑”，但在这一原语确立之前，还无法确定谁会是明确的赢家。 OpenAI 和 Anthropic 似乎出奇地缺席，尽管正是它们提出了引发聊天革命和智能体革命的想法。这一次，是开源社区在摆弄上下文。 其实，这与其他技术革命中在位者的表现是一致的。以往革命的赢家往往在技术上、意识形态上或品牌上（比如 Google 比微软显得更具云原生气质）与旧范式捆绑在一起。 所以很有可能，尽管 Anthropic 明显赢得了智能体革命，但他们可能赢不了上下文革命。 那么谁会赢呢？ 这个我留到另一篇文章再说，但无论谁赢，谁就将占据有利位置，赢得最后一场革命，乃至整个智能革命。 ## R4 - 平台革命 每一轮重大技术周期最终都会催生一个平台：一个让别人在赢家层之上构建、分发和变现产品的场所。它会创造出达到超大规模、构建强大护城河所必需的强大网络效应和病毒式传播。 苹果有 App Store。亚马逊有电商市场。TikTok 有创作者图谱。智能革命也会有同样的东西。 OpenAI 曾试图通过自定义 GPT 和智能体构建器来打造它。初创公司也做过类似尝试。但这些尝试大多失败了，因为它们缺少平台所需的两样东西： - 创建极高质量供给的工具（例如 TikTok 的视频编辑器） - 让需求与供给相遇的场所（例如 TikTok 的“为你推荐”页面） 简单来说，在 ChatGPT 上创建的 GPT 和智能体，质量和创建难度并不比其他地方更有优势，而且 ChatGPT 用户也不怎么想要它们。 上下文革命将改变这一点。 当它落地时，数百万人将拥有一个由其最重要的信息构成的活的数据库：人脉、公司、对话、会议记录、文档、决策等等。 这个数据库将成为种类繁多的智能体和生成式应用的底层基座。 上下文革命的赢家拥有这个基座后，只需通过以下方式顺势而为： - 提供创建智能体和应用的最好工具（当你拥有上下文层时，这几乎是水到渠成的） - 创建一个用户可以发现、使用并为其付费的场所 如今，折腾者们是在 GitHub、X 和互联网的其他角落发现智能体和生成式应用的。这就是 Chris Dixon 所说的“弱解决方案”——一种粗糙的、优化不足的解决问题方式。  这是革命孕育阶段的典型例子：市场正在苦苦哀求一个解决方案。 这与 TikTok 经历过的转型如出一辙。 视频当时正在为文本或图片设计的社交媒体平台上兴起。TikTok 以一个工具的姿态问世——用来创建高质量视频并发布到其他平台。但一旦它聚合了高质量视频的供给，它就能聚合需求，转型成为消费视频的最佳场所。而从那里开始，飞轮就一直在旋转。  逻辑如下： - 上下文具有粘性，是最高质量智能体和生成式应用的支柱。 - 谁赢得上下文革命，谁就有机会成为创建最高质量智能体和应用的工具。 - 谁成为创建最高质量智能体和应用的工具，谁就有机会成为发现、使用这些智能体和应用并为之付费的场所。 一旦这样的平台建立起来，就会引发一场平台革命，大量新业务将在其之上涌现。 - TikTok 带动了内容业务 - 苹果的 App Store 带动了移动应用业务 - Shopify 带动了电商业务 但精彩之处在这里。 最强大的平台会成为其所聚合市场的指数：整个品类的单一代理。 苹果成为了移动软件的指数。亚马逊成为了电商的指数。Shopify 成为了独立商家的指数。 那么，一个智能体和生成式应用的平台会聚合什么市场呢？ 所有市场。 - 谁成为发现、使用智能体和应用并为之付费的场所，谁就成为 AI 行业的指数 - 智能体终将取代（数字）人类劳动，生成式应用终将取代软件，因此 AI 行业的指数将成为当今所有涉足计算机的行业（约占全球经济总量的 50%）的指数。 当然，事情还没完。 Marc Andreessen 在 2011 年说过“软件正在吞噬世界”，而现在 AI 正在吞噬软件。 随着软件和智能体的创建变得愈发容易、人形机器人实现规模化生产，越来越多的实体经济将被数字化并被 AI 消化。 这就是为什么上下文革命以及随后的平台革命的赢家，很可能成为有史以来最有价值的公司。 关于这家公司会是谁，已经有一些早期信号，但上下文革命才刚刚起步，所以这个头衔仍然花落未定。 我的公司 Micro 就是一场赢得上下文革命的押注，而迄今为止我们做出的所有战略决策，都源于对本文所呈现理念的潜意识理解。我写下这篇文章，是给自己以及像我一样的创始人在这样一个历史性的混乱时刻做出更好决策的一张地图。 感谢 Isabel Swope、Ravi Mishra 和 Saul Carlin 的编辑和交流。 ## 相关链接 - [brett goldstein（布雷特·戈德斯坦）](https://x.com/thatguybg) - [@thatguybg](https://x.com/thatguybg) - [71K](https://x.com/thatguybg/status/2051714355398602957/analytics) - [Copy.ai](http://copy.ai/) - [Character.ai](https://character.ai/) - [$150K](https://x.com/search?q=%24150K&src=cashtag_click) - [Micro](https://micro.so/) - [升级至 Premium](https://x.com/i/premium_sign_up) - [2026年5月6日 凌晨1:23](https://x.com/thatguybg/status/2051714355398602957) - [71.8K 次浏览](https://x.com/thatguybg/status/2051714355398602957/analytics) - [查看引用](https://x.com/thatguybg/status/2051714355398602957/quotes) --- *导出时间： 2026/5/6 20:53:51*
O OpenAI 的 MCP 演进：从工具到应用架构 文章回顾了 MCP 协议 18 个月来的架构演进。MCP 最初由 Anthropic 于 2024 年 11 月推出，作为向模型暴露外部能力的工具层。OpenAI 虽然较晚采用，但推动了大量创新：先是 Apps SDK 让集成方提供交互式界面，随后社区提出官方 MCP Apps 扩展实现跨主机可移植。2026 年 DevDay 上，OpenAI 发布 Plugin Extensions，让 MCP 应用深度集成到 ChatGPT 的侧边栏、对话面板等界面，并新增 MCP Events，使外部系统能通过 webhook 主动触发 ChatGPT 中的工作。MCP 正从能力层演变为完整的应用架构标准。 技术 › OpenAI Josh Rosen 2026-10-01 MCPOpenAIAnthropicChatGPTMCP AppsPlugin ExtensionsMCP EventsAgent协议标准
适 适合零基础小白的Codex保姆级教程 本文是一篇面向零基础用户的Codex保姆级使用教程，系统介绍了OpenAI将Codex与ChatGPT桌面端整合后的完整使用方法。文章包括：Codex与Claude Code的区别对比（来源、模型、价格、易用性）；安装与账号注册流程；界面三大区域（功能区、项目区、工作区）介绍；三种工作模式（ChatGPT聊天、ChatGPT工作、Codex）的选择建议；基础功能详解（项目、工作位置、添加文件、权限设置、模型选择、语音功能）；高级功能（插件连接Gmail/GitHub、Skill技能复用、图像生成、定时任务、资料库）；以及远程手机连接功能的使用方法，帮助用户高效完成各类任务。 技术 › Codex 探路AI 2026-10-13 CodexOpenAIChatGPTAI教程AgentSkill插件自动化零基础
新 新版 Codex 桌面端：值得关注的五个重点 本文围绕 2026 年 DevDay 发布，梳理了新版 Codex 与 ChatGPT 桌面端的五大更新：一是新模型 GPT-6.1 Sol 价格大幅降低，Pro 套餐新增 500 美元档并调整 200 美元档用量；二是 Space / Pages 将生成结果组织为可编辑、可协作的文档成果；三是 Work 云端任务可调用本机已获准的文件和工具；四是 Codex 支持通过对话准备并发布可复用的云端环境，供后续任务直接使用；五是 dots 智能体可承接持续跟进职责，管理调研、执行与提醒，并能组合调用上述能力完成复杂工作流。 技术 › Codex 蓝哥AI 2026-10-08 CodexChatGPTOpenAIGPT-6.1 SolAgentdots云端环境订阅套餐桌面端
从 从零开始理解提示缓存（Prompt Caching In Agents）【译】 本文翻译自 Earendil Engineering 的《Prompt Caching In Agents》，系统讲解编程 Agent 中提示缓存的原理与实践。文章介绍了 KV 缓存的工作机制（prefill 与 decode 阶段）、缓存的两种存放方式（会话亲和性与分布式缓存）、前缀匹配对缓存命中的决定性作用，以及显式缓存与自动缓存的区别。还解释了为何修改工具定义会破坏缓存、TTL 过期导致短请求昂贵、缓存未命中的计费代价，以及 Pi 为何保持旧记录稳定而不激进剪枝，帮助读者在上下文、缓存复用、延迟与成本之间取得平衡。 技术 › LLM 小墨同学 2026-10-07 提示缓存KV缓存LLMAgentAnthropicOpenAI上下文管理推理成本翻译
A AI Is Improving AI Faster Than We Can Check 文章探讨了AI辅助构建下一代AI（递归自我改进，RSI）的现状与瓶颈。Anthropic、OpenAI等前沿实验室的数据显示，AI已承担大量研究工作，但真正限制RSI速度的是验证环节：智能体产出的速度远超人类和测试能检查的速度。作者分析了验证困难的五个原因：人是瓶颈、分数可被操纵、基准难以衡量最强模型、智能体不知何时停止、失败难以追溯。在可机器验证的任务上（如算法优化），AI收益巨大；但在开放性研究中，智能体缺乏判断力。人类剩余的优势在于选择哪些问题重要。 技术 › LLM Jinyan Su 2026-10-05 RSI递归自我改进AI研究验证benchmarkAnthropicOpenAIAgent前沿模型强化学习
C Codex大改版后，最新教程来了 本文是2026年10月ChatGPT桌面端大改版后的全面使用教程。作者从安装登录开始，系统讲解桌面界面三栏布局，包括功能入口、Space空间、Dot常驻Agent、项目与聊天管理等新特性。随后重点介绍Codex工作模式的实战用法：如何选择工作目录、区分Chat/Work/Codex三种模式、理解新的订阅额度体系，以及Astra、GPT-6.1 Sol、Luma等模型与思考档位的选择策略。文章还涵盖权限设置、文件/终端/浏览器面板使用、任务撰写四要素（材料、动作、交付物、验收标准）等实用技巧，帮助读者从入门到精通掌握新版ChatGPT桌面应用。 技术 › Codex Lab 305 2026-10-03 ChatGPTOpenAICodex教程Agent桌面应用SpaceDot工具与效率
H How dots took over my Codex 作者分享了使用 OpenAI 新产品 dots 的体验。Dots 是基于 Codex 开源 harness 构建的主动式个人智能体，可在 Slack、电话、桌面等多渠道间保持上下文连续性，并支持个性化设置。相比传统 Codex，dots 三大优势：一是跨渠道的连续性与个性化，可共享 ChatGPT 记忆和插件；二是协调能力，可并行启动子代理任务，调度本地与 Codex Cloud 云端环境；三是主动性，无需明确提示即可追踪 PR 部署、监控 Slack 消息等并主动汇报。作者认为 dots 大幅释放了心智带宽，已成为其日常工作不可或缺的助手。 技术 › Codex dominik kundel 2026-10-03 CodexdotsOpenAIAgentCodex Cloud个人智能体主动性任务协调ChatGPT自动化
手 手把手教你用 Agent 管理邮箱：从统一收件到每天自动整理 本文介绍了如何用 ChatGPT Agent 实现多邮箱自动整理：将 Gmail、Outlook、QQ 邮箱的新邮件自动转发到一个主 Gmail，再将其连接到 ChatGPT 插件，并创建每天早上 9 点运行的定时任务。任务会把过去 24 小时的邮件按「需要我处理」「重要通知」「资讯速览」三类分类，生成带待办和原邮件链接的清单。文章详细讲解了各邮箱的转发设置步骤、Gmail 与 ChatGPT 的连接与授权区别、定时任务的创建与测试方法、通知提醒配置，以及转发、连接、定时任务三个环节的常见问题排查。整套方案无需写代码、无需自建服务。 技术 › Agent 小墨同学 2026-10-02 AgentChatGPTGmail邮箱管理定时任务自动化效率工具OpenAI教程
C ChatGPT 桌面应用重大更新：Codex 合并，Work 模式登场 OpenAI 将拥有超 500 万周活用户的 Codex 编程应用合并进新的 ChatGPT 桌面应用，旧版更名为 ChatGPT Classic。新应用提供 Chat、Work、Codex 三种模式，其中 Work 模式是本次更新的核心：用户可交付完整任务而非简单提问，它能自主拆解目标、执行数小时，并产出文档、表格、幻灯片甚至网页。文中引用维珍航空将数周工作缩短至数小时、英伟达节省 40% 活动筹备时间等案例。此外还介绍了 Projects 长期上下文、Recents 统一管理、活动视图、跨设备同步及 Quick chat 快捷入口等功能。 技术 › OpenAI Julian Goldie SEO 2026-10-01 ChatGPTOpenAICodex桌面应用Work模式AI工具自动化Agent产品更新
O OpenAI 新时代：10步设置并使用 Dots 智能体、Sol 6.1 等新功能（从零开始） 文章介绍了 OpenAI 在 DevDay 上发布的重大更新，重点讲解了 Dots——一个拥有独立云电脑、可执行后台任务的个人智能体。作者以产品发布协调为例，分10步详细讲解如何从零开始设置和使用 Dots：创建 Dot 并赋予明确职责、连接所需的应用和云电脑、委派任务并保留上下文、设置定时任务和自定义规则以保持控制权等。文章还提及新模型 Sol 6.1、GPT-6 Astra 驱动的 Dots、Codex Cloud 编码环境等新功能，并强调核实智能体输出来源、区分建议与决策的重要性。 技术 › OpenAI codila 2026-09-30 OpenAIDotsAgentChatGPTGPT-6CodexDevDay智能体教程自动化
大 大变天！OpenAI 把 AI 变成了你的全天候云端分身：DevDay 资讯省流版 本文总结了 OpenAI DevDay 2026 的 20 余项重大升级。核心亮点包括：常驻云端数字员工 dots，由 GPT-6 Astra 驱动，7×24 小时自主工作；性价比模型 GPT-6.1 Sol 与 8 倍速 Ultrafast 极速推理；Codex Cloud 云端开发环境与 Agents API 预览；团队协作空间 ChatGPT Space 与动态文档 Pages；以及 Sign in with ChatGPT 额度跨平台通兑和 OpenAI Marketplace 生态战略。文章认为 OpenAI 正从对话工具演变为 AI 时代的云端基础设施运营商。 技术 › OpenAI 雪踏乌云 2026-09-30 OpenAIDevDayCodexGPT-6AgentCodex CloudChatGPT大模型开发者工具
O OpenAI 2026 开发者大会全面总结 作者通宵整理了 OpenAI DevDay 2026 发布会的全部内容。主要发布包括：个人 Agent 产品 Dots，具备云电脑、浏览器和执行环境，可主动替用户做事，对标 Grok Bot 和 Muse；配套的 ChatGPT Space 工作空间；新模型 GPT-6.1 Sol，以约七分之一的成本接近 Astra 水平；推出 500 美元 Pro 会员但砍削原有额度；Decisions API 直接对标刚发布的 Jev；Codex 全面升级云端可复用环境与安全扫描；以及 ChatGPT 订阅生态打通等。作者对 OpenAI 的定价策略和与 Claude Opus 5.5 的差距表达了看法。 技术 › OpenAI 数字生命卡兹克 2026-09-30 OpenAIDevDayAgentDotsGPT-6.1CodexDecisions APIChatGPT订阅大模型
