<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Beta on Medium]]></title>
        <description><![CDATA[Stories by Beta on Medium]]></description>
        <link>https://medium.com/@betanote?source=rss-57e8f4e4b9c4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*BzV2skNNwvKEOk9f</url>
            <title>Stories by Beta on Medium</title>
            <link>https://medium.com/@betanote?source=rss-57e8f4e4b9c4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:32:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@betanote/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Coding]]></title>
            <link>https://medium.com/@betanote/coding-6a892d1a2640?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a892d1a2640</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 15:40:55 GMT</pubDate>
            <atom:updated>2026-08-03T15:40:55.347Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KuccxWUShoVRrjQnQhICUw.png" /></figure><p>1、Coding 的本质不是写代码，而是人在数字世界中的创造活动（在 coding agent 出现之后，这句话更能立住了），Coding 就如同神笔马良手中那支画笔，几乎能在数字空间中构建任何东西。而 Coding agent 则是生成式 AI 浪潮里，当下能看到的最大的应用机会，也许只有视频生成可与之媲美。</p><p>2、Coding agent 前面的 “coding” 可能迷惑了很多人，包括之前的我，这个词会让人觉得，这不过是个帮人类进行编程的工具（其实 agent 这个词也有很大迷惑性，真是个糟糕的名字），但实际上，它其实是非常强大的通用性工具，可以完成很多任务，不管是做网站做 App 做游戏，还做各种模态的内容，甚至是直接代理操控各种硬件设备，都能搞定。并且这种通用性的能力是可持续扩展的，它可以随着大模型能力的提升，传感器的扩展丰富，各种数据和工具接口的接入完善，不断随之水涨船高，这点其实很可怕，因为难以看到上限。</p><p>3、Coding agent 的演进也许会经历几个阶段：</p><p>一、在人类的指令下完成更复杂的任务</p><p>二、可自循环地完成更复杂的长程任务</p><p>三、更经济地构建可持续运行的复杂系统</p><p>目前看，我们还在阶段二的早期，目前各个公司主要在卷能力的上限，也就是能完成多复杂的任务，基本是唯「通过率」论，代码质量如何并不关心。而这导致了目前 agent 产品一个问题：干活“只管生不管养”，生成的代码往往极其冗余，堪称屎山大师。这会直接导致代码的可读性和长期可维护性都变得极差，尤其是对于一个复杂的系统，当越过一个临界点之后，这会成为限制 agent 能力持续提升的一大瓶颈。</p><p>4、在渡过阶段二之后，Coding Agent 之间的竞争，会从“可运行的代码生成器”，进化为“生成可持续运行系统”。前者强调能不能跑通，后续强调是不是长期可持续可维护。大家会对所生成代码的可读性、可维护性、简洁性提出更高的要求，更追求“系统能长期优雅地运行”。所谓变得更简洁更优雅，不是什么道德洁癖，而是实打实的经济账，这个等会展开。</p><p>5、它跺它也麻。混乱的代码库会让工程师抓狂，也会让 AI 发疯。AI 很强，但也有自身局限性。</p><p>首先，自回归模型的注意力是有限的，代码越冗余、耦合越严重，AI 理解的准确率也就越低。面对一团乱麻的代码，AI 修改时同样容易犯错，顾头不顾尾，改了 A 崩了 B ，引入隐形 bug。</p><p>另外，代码堆砌得越厉害，token 消耗越高，其对上下文窗口的占用也越大。当项目大到一定程度，上下文很可能根本装不下项目信息，自然也就失去了精准修改的能力。</p><p>这和人类面对屎山代码的困境本质一毛一样：当复杂度超过了理解主体的认知带宽，谁来维护都一样难。人读不懂的烂代码，AI 同样会读错、改错，只是它犯错的速度比人快得多。</p><p>6、实打实的经济账。年初很多公司鼓吹员工多用 AI ，额度报销无上限，但一年还没过，就有不少公司开始限定消耗上限，本质也是一种经济帐。企业是理性的经济组织，核心不是你消耗了多少 token，而在于你到底创造了什么价值。而今，在头部模型厂商不断提价的背景下，企业显然越加理性了。而冗余混乱的系统，其维护成本会显性上升，出错所带来的经济损失和信誉损失也越大，不管是企业还是经济个体，不可能不考虑。</p><p>有人会说，AI 也能维护系统，确实，但逻辑不会因为维护者从人换成 AI 就消失。AI 改烂代码同样需要更多轮调试、试错，这只是把「人的工时成本」换成了「token 成本 + 调试时间成本 + 故障风险成本」，而且很重要的一点，AI 再强，它也无法担责，越大的公司组织越不敢去赌这一点。</p><p>7、大神们如何定义 coding</p><p>“Computer programming is an art, because it applies accumulated knowledge to the world, because it requires skill and ingenuity, and especially because it produces objects of beauty.”</p><p>计算机编程是一门艺术，因为它把积累的知识应用到世界中，因为它需要技巧和创造力，尤其是因为它能产生美的作品。</p><p><strong>唐纳德·克努特（Donald Knuth）</strong></p><p>斯坦福大学荣休教授，1974年图灵奖得主，“算法分析之父”，代表作《计算机程序设计艺术》，排版系统TeX发明者。</p><p>⸻</p><p>“Programs must be written for people to read, and only incidentally for machines to execute.”</p><p>程序必须是写给人读的，只是顺便让机器执行。</p><p><strong>哈罗德·埃布尔森（Harold Abelson）、杰拉德·萨斯曼（Gerald Jay Sussman）</strong></p><p>MIT计算机科学教授，经典教材《计算机程序的构造和解释》（SICP）作者。</p><p>⸻</p><p>“Program testing can be used to show the presence of bugs, but never to show their absence!”</p><p>程序测试可以用来证明bug的存在，但永远不能证明bug的不存在。</p><p><strong>艾兹格·迪杰斯特拉（Edsger W. Dijkstra）</strong></p><p>荷兰计算机科学家，1972年图灵奖得主，最短路径算法发明者，结构化编程与程序正确性思想的核心奠基人。</p><p>⸻</p><p>“A most important, but also most elusive, aspect of any tool is its influence on the habits of those who train themselves in its use. If the tool is a programming language this influence is, whether we like it or not, an influence on our thinking habits.”</p><p>关于工具，最重要也最不易察觉的一点，是它会塑造使用者的思维习惯。如果这个工具是编程语言，那么无论我们是否愿意，它都会改变我们的思考方式。</p><p><strong>艾兹格·迪杰斯特拉（Edsger W. Dijkstra）</strong></p><p>图灵奖得主，现代算法理论与结构化编程的奠基人。</p><p>⸻</p><p>“There are two ways of constructing a software design: One way is to make it so simple that there are obviously no deficiencies, and the other way is to make it so complicated that there are no obvious deficiencies.”</p><p>软件设计有两种方式：一种是做得足够简单，明显没有缺陷；另一种是做得足够复杂，看不出明显缺陷。</p><p><strong>托尼·霍尔（C. A. R. Hoare）</strong></p><p>1980年图灵奖得主，快速排序算法发明者，形式化验证与程序设计理论先驱。</p><p>⸻</p><p>“Controlling complexity is the essence of computer programming.”</p><p>控制复杂性是计算机编程的本质。</p><p>“Everyone knows that debugging is twice as hard as writing a program in the first place. So if you’re as clever as you can be when you write it, how will you ever debug it?”</p><p>人人都知道，调试程序的难度是最初写程序的两倍。所以，如果你写程序时已经把聪明用到极限，那你以后还怎么调试它？</p><p><strong>布莱恩·柯尼汉（Brian Kernighan）</strong></p><p>贝尔实验室核心成员，Unix文化核心传播者，与丹尼斯·里奇合著《C程序设计语言》，AWK语言共同作者。</p><p>⸻</p><p>“C++ is designed to allow you to express ideas, but if you don’t have ideas or don’t have any clue about how to express them, C++ doesn’t offer much help.”</p><p>C++的设计是为了让你表达想法；但如果你没有想法，或者不知道如何表达这些想法，C++也帮不了你太多。</p><p>“I think we should look for elegance in the applications built, rather than in the languages themselves.”</p><p>我认为我们应该在构建出来的应用中寻找优雅，而不是只在语言本身中寻找优雅。</p><p><strong>本贾尼·斯特劳斯特卢普（Bjarne Stroustrup）</strong></p><p>C++语言创造者，其设计的C++是现代系统软件、游戏引擎、金融高频系统、嵌入式领域的核心底层语言。</p><p>⸻</p><p>“Bad programmers worry about the code. Good programmers worry about data structures and their relationships.”</p><p>糟糕的程序员关心代码本身；优秀的程序员关心数据结构以及它们之间的关系。</p><p><strong>林纳斯·托瓦兹（Linus Torvalds）</strong></p><p>Linux内核创始人、Git版本控制系统作者。</p><p>⸻</p><p>“Data dominates. If you’ve chosen the right data structures and organized things well, the algorithms will almost always be self-evident. Programming isn’t about algorithms, it’s about data structures.”</p><p>数据为王。如果你选对了数据结构并且组织得当，算法几乎总是不言自明的。编程的核心在于数据结构，而非算法。</p><p><strong>罗布·派克（Rob Pike）</strong></p><p>Go语言联合设计者，贝尔实验室Unix核心成员，谷歌首席工程师，极简主义工程哲学的代表人物。</p><p>⸻</p><p><strong>“Any damn fool can write code that a computer can understand, the trick is to write code that humans can understand.”</strong></p><p>任何傻瓜都能写出计算机能理解的代码，真正的本事是写出人类能理解的代码。</p><p><strong>马丁·福勒（Martin Fowler）</strong></p><p>软件工程教父，ThoughtWorks首席科学家，代表作《重构》《企业应用架构模式》，敏捷开发先驱。</p><p>⸻</p><p>“What hackers and painters have in common is that they’re both makers. Along with composers, architects, and writers, what hackers and painters are trying to do is make good things.”</p><p>黑客和画家的共同点在于，他们都是创造者。和作曲家、建筑师、作家一样，黑客和画家想做的，都是创造好东西。</p><p><strong>保罗·格雷厄姆（Paul Graham）</strong></p><p>程序员、技术作家、投资人，Y Combinator联合创始人，代表作《黑客与画家》。</p><p>⸻</p><p>“Measuring programming progress by lines of code is like measuring aircraft building progress by weight.”</p><p>用代码行数衡量开发进度，就像用重量衡量飞机制造进度。</p><p><strong>比尔·盖茨（Bill Gates）</strong></p><p>微软创始人，全球个人计算革命的推动者。</p><p>⸻</p><p>“Everybody in this country should learn how to program a computer… because it teaches you how to think.”</p><p>这个国家的每个人都应该学习编程，因为它教会你如何思考。</p><p><strong>史蒂夫·乔布斯（Steve Jobs）</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a892d1a2640" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI application companies are all opening supermarkets.]]></title>
            <link>https://medium.com/@betanote/ai-application-companies-are-all-opening-supermarkets-db2934922166?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/db2934922166</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[retail]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 15:19:23 GMT</pubDate>
            <atom:updated>2026-08-03T15:24:26.560Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7HsTmsSeXc18TNpXLLsSpQ.jpeg" /></figure><p>When you open any AI application, all users see are input boxes, buttons, and results; they can’t see the underlying model being called. Most people don’t care whether it’s connected to GPT or Claude; they only care whether the task has been completed. This “users don’t recognize the manufacturer, they only recognize the shelf” mentality is a typical characteristic of the retail business.</p><p>The previous generation of internet companies told the story of platforms: they set up a venue, letting supply and demand meet naturally, with the company collecting rent. This generation of AI application companies tells a different story — they’re more like running supermarkets. Platforms rely on network effects to passively collect rent, while supermarkets earn their profits penny by penny through product selection, display, and service. These are two completely different business logics: platforms are winner-takes-all, while retail is cutthroat competition; platforms can wait for others to educate the market before entering, while supermarkets require real money to invest in site selection and supply chains. This generation of AI application companies generally seems to be working harder than the previous generation, not because the technology is more difficult, but because they can’t collect rent; they can only operate.</p><h3>Intelligence has been commodified</h3><p>Why has this generation become retail? Because intelligence has been commodified. The capabilities of previous generation software were custom-made; they couldn’t be bought, so you had to create them yourself. Today, intelligence is packaged as APIs, standardized, clearly priced, and readily available. Once a commodity is standardized, the value of distribution begins to exceed that of production — the premise of a retail business is that there are goods available for wholesale from upstream suppliers. The multi-model era has made this premise a reality: GPT, Claude, Gemini, plus a batch of open-source models, with similar capabilities and continuously decreasing prices, allow anyone to enter the wholesale market.</p><h3>Model factories produce goods, and application companies handle retail.</h3><p>The division of labor in the industry has quietly changed. Model manufacturers are producing standardized goods: intelligence. They turn reasoning capabilities into APIs, priced by tokens, and anyone can buy them wholesale. AI application companies don’t produce intelligence; they source it from upstream: integrating multiple models, negotiating prices, and managing supply. Then they reorganize and repackage the intelligence they’ve acquired and deliver it to users.</p><p>Breaking it down, they essentially do six things: Procurement: Integrating multiple models, comparing prices, negotiating, and managing supply — using whichever is cheapest and most stable. Packaging: Not selling complete capabilities, but selling by the use case or scenario; pricing is tied to the usage scenario, not training costs — this is the watershed between the profit structures of retail and manufacturing; users pay for a summary, a translation, or a meeting minutes, not for a “model that can write.” Display: Transforming complex model interfaces into products that users can easily understand; parameters are invisible to users, only buttons and results are visible. Shopping Guide: Users don’t need to figure out which model to use; the system automatically selects the most suitable one based on the task. Cashier: Unified accounts, unified points, unified billing — only one checkout at the supermarket entrance. After-sales: Enterprise integration, FDE, custom development, customer success — service only begins after the product leaves the store.</p><p>None of these six things involve producing intelligence; each one is a retail activity. Therefore, essentially, this is a retail business. The first principle of retail is turnover and repeat purchases: low profit margins rely on volume; customer retention depends on service. This also explains why AI application companies are generally still losing money, yet they are desperately trying to build a user base — retail works like this: scale comes first, then you can talk about profits.</p><h3>Competition always revolves around four words</h3><p>Since it’s retail, the competition will always revolve around four words: more, faster, better, and cheaper.</p><p>More means more supply: more models, more capabilities, and a more complete range of products on the shelves mean customers are less likely to leave empty-handed. Faster means faster response and a smoother experience; supermarket queue times determine repurchase rates. Better means more stable results and a more streamlined workflow; no defective products on the shelves, so customers feel confident buying without hesitation. Lower costs and lower barriers to entry; price and barriers to entry are the most basic weapons in retail.</p><p>These four words seem contradictory. “More” and “better” usually come at the cost of increased costs, while “faster” and “cheaper” depend on engineering capabilities, not just slogans. Being able to excel in two out of these four dimensions is enough to survive; excelling in three makes you a top player.</p><h3>The retail sector has already addressed the criticism of “shelling”.</h3><p>Some people will surely say: Aren’t application companies just repackaged versions? If model manufacturers make their own applications, how will you survive?</p><p>Let’s first address the opposing viewpoints. Model manufacturers are indeed moving downstream; OpenAI itself has ChatGPT. Upstream, they may also tighten APIs and raise prices at any time. The product itself is also rapidly depreciating: model capabilities improve every few months, while API prices continue to fall. These threats are real.</p><p>But the logic of retail precisely addresses this skepticism. Walmart doesn’t own its factories, and Costco’s private label brands occupy only a small portion of its shelves, yet they remain the most profitable companies in retail history. The barriers to entry in retail aren’t in production, but in channels, product selection, service, and relationships. Why don’t consumers buy directly from manufacturers? Because manufacturers don’t know how to display products effectively, nor are they willing to provide after-sales service for individual customers. A model factory developing applications is like a factory opening its own stores: it can do it, but not comprehensively, and its service radius is inherently limited.</p><p>However, retailers must also acknowledge the cost: distributors are forever at the mercy of upstream suppliers. This is precisely why they move downstream.</p><h3>Excellent retailers will eventually create their own brands.</h3><p>Successful retailers don’t just sell other people’s products forever. Once they have enough sales volume and a deep enough understanding of user needs, they start building their own supply: private label brands, exclusive products, and even moving into the production process.</p><p>Costco’s Kirkland, Amazon’s Amazon Basics, and Uniqlo all control their supply chains — private label brands don’t just earn an extra layer of profit; they transform the understanding of “what users really want” into products that only they can sell. While resellers earn the price difference, private label brands earn the ability to make sound judgments.</p><p>AI application companies will most likely follow the same path: first aggregate models, then schedule models, and finally have their own model capabilities.</p><p>Aggregation is the entry ticket. Without a diverse model supply, you can’t even set up a shelf. Scheduling is differentiation — selecting models based on tasks and orchestrating workflows is a retailer’s discerning eye that others can’t copy. Owning your own model capabilities is the ultimate goal, but “owning” doesn’t necessarily mean self-training large models. More realistically, it means deep fine-tuning, customizing open-source models, and creating domain-specific small models. It doesn’t intend to replace model manufacturers; it simply wants to place uniquely its own products on key shelves.</p><p>Of course, not all retailers should create their own brands. Two prerequisites are necessary: ​​sufficient scale to cover R&amp;D costs; and a deep understanding of users to ensure that the products actually sell. Small supermarkets that force their own brands will only cripple their cash flow.</p><h3>Stop using “technical barriers” to describe them.</h3><p>So what truly changes this assessment is the way we view them. Stop evaluating an AI application company based on “whether it has technological barriers” — it’s a supermarket. The questions to ask are: Are the shelves comprehensive (abundant)? How long are the queues (fast)? Is the stock stable (good)? Is it expensive (economical)? And most importantly — does it have anything unique to it on its shelves?</p><p>The first four questions determine whether it is a good supermarket; the last question determines how long it can stay in business.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db2934922166" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Amplify Your Strangeness]]></title>
            <link>https://medium.com/@betanote/amplify-your-strangeness-9ca52e7293e6?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ca52e7293e6</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[life]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Sun, 09 Nov 2025 14:57:12 GMT</pubDate>
            <atom:updated>2025-11-09T14:57:12.463Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/249/1*Nn5EqdlK3fNRJhipVuGsTg@2x.jpeg" /></figure><p>I once heard a story about Yuri Milner. As the founder of DST Global, he bet correctly on almost every major wave of the mobile internet – from Facebook to Alibaba, and then to ByteDance.</p><p>Someone asked him, “You’ve met so many top-tier founders. What do they have in common?”</p><p>Yuri’s answer was simple: “They are all extremely good at amplifying their traits.”</p><p>Notice he didn’t say “overcoming weaknesses,” but “amplifying.”</p><p>This runs counter to most of the education we receive. We are often taught to be “well-rounded” and to fix our shortcomings. But in the fiercely competitive world of startups, being well-rounded often means being mediocre. A smooth pebble cannot tear open a market; you need sharp edges.</p><p>Think about Steve Jobs. His “reality distortion field” was, in some ways, a pathological need for control. If applying for a job at a big corporation, he might have been rejected in the first round for being “difficult to work with.” But he didn’t try to become “agreeable.” He amplified this paranoia until it became an organizational force capable of compelling thousands of engineers to build the iPhone.</p><p>Look at Zuckerberg. Early on, people often criticized him for being as cold as a robot. This trait is a disadvantage in social settings, but a massive advantage in business decision-making. While most people hesitate due to emotional entanglement, he can calculate coldly like an algorithm, decisively writing a $1 billion check to acquire Instagram.</p><p>Your biggest weakness is often the shadow of your greatest strength. They are two sides of the same coin.</p><p>Someone who is extremely sensitive and easily hurt (weakness) often possesses unparalleled empathy, capable of perceiving the subtlest user emotions (strength). If you try to make yourself a bit more “thick-skinned,” you might simultaneously lose that keen product intuition.</p><p>Mediocrity is often the result of constantly fixing your own weaknesses. You end up with a product that has no major flaws but also no major highlights – whether that product is yourself or your company.</p><p>Of course, this doesn’t mean you should be a jerk. The key is to distinguish between “mistakes” and “features.” Being chronically late is a mistake that needs fixing; but neglecting other matters because you are overly obsessed with product details is often a feature. For the latter, you shouldn’t force yourself to pay less attention, but rather find a partner skilled in operations to fill your blind spots.</p><p>Don’t spend time polishing away your edges. Find the battlefield where your edges become weapons.</p><p>As Yuri said, amplify it. That is your true source of power.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ca52e7293e6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Moat is Outside of AI]]></title>
            <link>https://medium.com/@betanote/your-ai-moat-is-outside-of-ai-7a5255d237ca?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/7a5255d237ca</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Thu, 30 Oct 2025 09:08:47 GMT</pubDate>
            <atom:updated>2025-10-30T09:08:47.690Z</atom:updated>
            <content:encoded><![CDATA[<p>Your AI Moat is Outside of AI</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/258/1*3fQjxMfsU8ThqacOd-52Tw@2x.jpeg" /></figure><p>Every few months, when OpenAI or Anthropic releases a new model, a batch of AI startups disappears.</p><p>Their core product – whether it was “better summarization” or “smarter writing” – is suddenly a free feature of GPT-5.</p><p>This should scare you. If it doesn’t, you might be next.</p><p>We are witnessing a painful but necessary reckoning. Too many founders built their businesses on a fatal assumption: they mistook “model capability” for their moat.</p><p>This is a trap.</p><p>Models are a Public Utility</p><p>The more you rely on the model’s capability, the more fragile your business is.</p><p>We must immediately accept a harsh reality: a model is not your private asset. It is a public utility.</p><p>It’s like electricity, or the internet in 1999. You can’t say your business is “doing electricity,” or that your moat is “we use the internet.” Your competitors use it, too.</p><p>Models are:</p><p>* Public: Everyone (including your competition) can access the best models at a low cost.</p><p>* Rapidly Homogenized: Any slight lead (“our model is 2% better on a benchmark”) is erased within weeks.</p><p>* Replaceable: The cost of switching an API endpoint is almost zero.</p><p>When you bet everything on “we use AI better,” you are effectively just working for the model makers (OpenAI, Anthropic). You are just a fancy API reseller, finding Product-Market Fit for them.</p><p>Every prompt you optimize is free training for their next model.</p><p>So, if the model itself isn’t the moat, what is?</p><p>The real moat is in the things AI can’t take with it and can’t easily replicate. It’s outside of AI.</p><p>The Misconception: “We Have Data”</p><p>When I ask founders about their moat, the most common answer I get is: “Our moat is our data.”</p><p>This is a dangerous half-truth.</p><p>What most people mean by “data” is “piled-up data” – a messy, static heap of PDFs, user logs, or chat histories.</p><p>That is not a moat. That is a storage liability. You have to pay to store it and secure it, and you get no compounding value from it.</p><p>The real moat is “Compounding Workflow Data.”</p><p>“Compounding” means the data a user creates in your product must immediately improve their own experience, which in turn incentivizes them to create more data. It’s a closed loop.</p><p>Think about Figma.</p><p>Figma’s moat isn’t “we have a lot of .fig files.” Its moat is:</p><p>* Structured Design Systems: The reusable component libraries and brand guidelines a team has spent years building.</p><p>* Collaboration History &amp; Context: All the comments, decisions, and version histories are logged.</p><p>* The Workflow: The entire company’s process – from product manager to designer to engineer – is built on Figma.</p><p>AI can “help you draw a button” or “turn a sketch into a design.” But AI cannot migrate your entire team’s three-year-old collaborative workflow and design assets.</p><p>The true value of an AI application is not in owning data, but in activating it. Your product must be the “reactor” that activates data and makes it compound within a workflow.</p><p>The Strongest Moat: From “Tool” to “System”</p><p>This leads to the most durable moat of all: deep integration.</p><p>You must evolve your product from a “Tool” into a “System.”</p><p>* A Tool is something a user uses when they think of it. An AI writing assistant or a chatbot. I can use app A today and app B tomorrow. The switching cost is zero. They compete for “attention.”</p><p>* A System is something the business depends on to operate. Think Gusto (Payroll) or Rippling (HR). They are “infrastructure.”</p><p>You can’t just “try out” a new payroll system today and switch back tomorrow. Migrating an HR or finance system is minor surgery for a company. It requires months of migration, training, and process re-engineering.</p><p>That is a moat.</p><p>Why doesn’t everyone build a system?</p><p>Because it’s hard. It requires you to deeply understand a boring vertical industry (like construction, logistics, law, or pharma). It requires you to do (often unpleasant) sales, handle (annoying) customer support, and adapt to (messy) edge cases.</p><p>This is far less “sexy” than tuning parameters or launching a cool new feature.</p><p>But this is what the most successful SaaS companies do. They do the “dirty work.”</p><p>In a “system,” AI is just an “accelerator node,” not the system itself. It might automate a compliance check in payroll or parse an invoice. It makes the system more efficient, but the system’s core value lies in the un-migratable business process it holds.</p><p>Build a “Model-Agnostic” Business</p><p>I’ve observed that the AI companies that are moving steadily are, intentionally or not, doing one thing:</p><p>Building a “Model-Agnostic” experience.</p><p>This means their underlying model can be switched – OpenAI today, Anthropic tomorrow, a self-hosted model the day after – but the user’s experience and core value remain identical.</p><p>Why?</p><p>Because users don’t care which model you use.</p><p>Users only care if their “assets” in your product (their data, their workflows, their design systems, their customer relationships) are safe and growing in value.</p><p>When AI becomes a commodity like water or electricity, the core of competition is no longer the AI itself.</p><p>A “Litmus Test”</p><p>So, stop trying to build your moat on someone else’s water faucet.</p><p>You need to ask yourself a “litmus test” question. It’s a sharp one:</p><p>&gt; “If your biggest competitor gets access to GPT-5 tomorrow (and you don’t), making their AI features 20% better, how many paying customers will you lose?”</p><p>&gt;</p><p>If your answer is “most of them,” then you don’t have a business. You are an API reseller, and your fate is in someone else’s hands.</p><p>If your answer is “almost none,” then congratulations. You have found a real moat.</p><p>It means your customers are staying not because your AI capability is the best, but because their data, their processes, and their teams are built inside your product. Their cost of leaving is far higher than their cost of staying.</p><p>AI is the most powerful engine of our time, but it is not the product.</p><p>Your job is to use this new power to dig the deep, boring, and defensible moats that others can’t cross.</p><p>And those moats are, almost always, found outside of AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7a5255d237ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sora – AI powered Bilibili]]></title>
            <link>https://medium.com/@betanote/sora-ai-powered-bilibili-e6efb8aeb5c5?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/e6efb8aeb5c5</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ui]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Fri, 17 Oct 2025 06:30:05 GMT</pubDate>
            <atom:updated>2025-10-17T06:30:05.279Z</atom:updated>
            <content:encoded><![CDATA[<p>Sora – AI powered Bilibili</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/275/1*hQmFILE87cNMP3E0qMB12w@2x.jpeg" /></figure><p>Everyone says Sora is the next TikTok.</p><p>I don’t think so.</p><p>After watching it for a few days, my impression is:</p><p>Sora today feels less like TikTok, and more like an AI-powered Bilibili.</p><p>⸻</p><ol><li>Why it feels like Bilibili</li></ol><p>Good products are never designed from the top down.</p><p>They grow from what users actually do with them.</p><p>When you open Sora now, what you see is not people recording their real lives.</p><p>It’s users remixing anime scenes, parodying movie clips, or reimagining game characters.</p><p>It’s play, not performance.</p><p>That’s exactly how Bilibili started – a culture of re-creation.</p><p>The only difference is that Bilibili users spent hours cutting and rendering videos,</p><p>while Sora users just type a few words.</p><p>Sora drops the cost of creativity by an order of magnitude.</p><p>When you make creation easier, you don’t just get more of it —</p><p>you change who creates, and why.</p><p>That’s how new cultures begin.</p><p>⸻</p><p>2. Why it’s nothing like Instagram</p><p>Some people think Sora will replace Instagram.</p><p>It won’t.</p><p>Instagram is built on two pillars: authenticity and scarcity.</p><p>You post a photo from Iceland because not everyone can go there.</p><p>You post a picture with a celebrity because it actually happened.</p><p>What makes these moments valuable is that they’re rare and real.</p><p>But when Sora lets everyone “go to Iceland” or “pose with a celebrity” in seconds,</p><p>that scarcity disappears.</p><p>When anyone can fake a luxury, no one can flaunt it.</p><p>When the cost of showing off drops to zero,</p><p>showing off itself loses meaning.</p><p>That’s why Sora will never replace the social graph of real life.</p><p>The logic is different:</p><p>Instagram says “this happened to me”;</p><p>Sora says “this could happen, in my head.”</p><p>⸻</p><p>3. The hidden barrier</p><p>Another thing people miss:</p><p>writing good prompts is harder than shooting good videos.</p><p>TikTok’s genius is that you don’t need to think.</p><p>You just point your camera at yourself and it works.</p><p>That’s why it reached everyone.</p><p>Sora, on the other hand, asks you to think like a director and write like a poet.</p><p>It’s not just a tool – it’s a blank page.</p><p>That’s liberating for some, intimidating for most.</p><p>Until AI understands what we mean without us saying it,</p><p>Sora will remain a playground for high-context creators, not a mass platform.</p><p>⸻</p><p>4. Where it might go</p><p>So far, Sora seems to be evolving toward something between</p><p>a sandbox for imagination and a remix engine for culture.</p><p>Its power won’t come from replacing social networks,</p><p>but from reshaping how ideas circulate.</p><p>TikTok captures the external world.</p><p>Sora captures the internal one.</p><p>If TikTok is reality’s mirror,</p><p>Sora might become imagination’s window.</p><p>And that – if it happens —</p><p>could still be a bigger revolution than anyone expects.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e6efb8aeb5c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[NVIDIA: The Central Bank of AI]]></title>
            <link>https://medium.com/@betanote/nvidia-the-central-bank-of-ai-dbd3b80a30dd?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/dbd3b80a30dd</guid>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Tue, 30 Sep 2025 09:23:40 GMT</pubDate>
            <atom:updated>2025-10-12T00:35:42.065Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*Hglq7-F5qYBbNkX6zWOnXA@2x.jpeg" /></figure><p>If the AI economy is a new financial system in the making, the real central bank isn’t a model company – it’s NVIDIA</p><p>Why? Because in this system, computing power is currency, and GPUs are the cash. And NVIDIA is the only true “printing press.”</p><ol><li>Compute = Money, GPU = Cash</li></ol><p>In the traditional economy, central banks control money supply and interest rates.</p><p>In the AI economy, whoever gets GPUs can train models, win users, and attract capital.</p><p>Each H100 or B200 is like a “US dollar of AI.” It’s both a production tool and a monetized asset. NVIDIA is the sole issuer.</p><p>2. Liquidity Injection: Investments &amp; Guarantees</p><p>Central banks pump liquidity in times of need. NVIDIA has been doing the same:</p><p>•	$100B into OpenAI: like quantitative easing – rapidly scaling the strongest player.</p><p>•	$6.3B guarantee for CoreWeave: like refinancing banks, ensuring GPU cloud supply.</p><p>•	Acquiring Run:AI: improving capital efficiency, avoiding wasted compute.</p><p>•	Investments in Scale AI, Wayve, Nebius: like sector loans, keeping the ecosystem balanced.</p><p>These moves aren’t just for profit – they stabilize the AI economy.</p><p>3. Interest Rate Tool: GPU Pricing</p><p>Central banks use rates to set the cost of money. NVIDIA sets GPU prices, controlling the cost of compute.</p><p>Low-cost, stable GPU supply accelerates model iteration; scarcity pushes others out.</p><p>GPU pricing = AI interest rate. NVIDIA = the rate-setting committee.</p><p>4. Lender of Last Resort</p><p>When banks fail, central banks step in. NVIDIA does the same for AI infrastructure:</p><p>•	Propping up CoreWeave to prevent cloud collapse.</p><p>•	Investing in Nebius to stabilize Europe’s supply.</p><p>•	Billions in the UK, earning the PM’s public support.</p><p>This isn’t just commerce – it’s public governance.</p><p>5. Soft Power: From Business to Geopolitics</p><p>NVIDIA’s reach now extends beyond business:</p><p>•	Heads of state endorse its projects – strategic, not just financial.</p><p>•	Partnerships shape not just industries but national AI strategies.</p><p>•	Its role in capital, policy, and geopolitics resembles the Fed in global finance.</p><p>Summary</p><p>NVIDIA is no longer just the company “selling shovels.”</p><p>Its capital moves, compute supply, and global strategy are shaping a new monetary system:</p><p>•	GPU = currency</p><p>•	Compute = liquidity</p><p>•	Investment = stimulus</p><p>•	Guarantees = refinancing</p><p>•	Price = interest rate</p><p>And NVIDIA? It’s becoming the central bank of the world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dbd3b80a30dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The First Principles Behind TikTok’s Success]]></title>
            <link>https://medium.com/@betanote/the-first-principles-behind-tiktoks-success-8836806906e5?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/8836806906e5</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[company]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Thu, 18 Sep 2025 10:49:10 GMT</pubDate>
            <atom:updated>2025-09-18T10:49:10.133Z</atom:updated>
            <content:encoded><![CDATA[<p>The First Principles Behind TikTok’s Success</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*qq_9_gPxr4_vw5bW5Hfzng.jpeg" /></figure><p>What are the first-principle reasons behind TikTok’s success? I’ve discussed this with many people at ByteDance and heard all kinds of answers.</p><p>The one that feels closest to the truth, in my view, is Zhang Yiming’s belief at the time — his conviction about the eventual scale of short video.</p><p>He believed short video was a massive opportunity, far beyond what most CEOs and founders were estimating. In his eyes it was a product that could reach billions of users. Opportunities of that magnitude are exceedingly rare.</p><p>But many people then assumed that, in China, short video would top out at 100–200 million active users. And by 2016, Kuaishou — the reigning short-video leader — already had over a hundred million active users. Some of its executives even thought the war was about to end and they would monopolize the market. That was a core reason they didn’t move quickly to block ByteDance’s advance. For a long time, they even kept their promotion budget at zero and were proud of it. Later events proved this was just wishful thinking.</p><p>Zhang Yiming saw it completely differently.<strong> There was a huge perception gap — on the order of 10×</strong>. That gap explains why he dared to go almost all-in on the short-video battlefield.</p><p>From the very beginning, ByteDance built three short-video products in parallel — Huoshan short-video , Xigua video, and Douyin (the Chinese version of TikTok) — and advanced on three fronts. At that time, ByteDance was just a unicorn valued in the single- to low-tens-of-billions of dollars, with limited revenue. Many internet giants with far greater income were also watching the space.</p><p>Yet most decision-makers couldn’t see how big short video could truly become. So even when they invested, they did so conservatively; some didn’t believe in the market at all. In fact, many in the industry looked down on ByteDance’s core business back then — entertainment products represented by Toutiao — which they saw as relying on the re-posting of lowbrow infotainment to acquire users. People considered it a “tacky” company; some even labeled it a thief, because it did redistribute a large amount of content produced by other media for free.</p><p>And yet it was this company that went on to build Douyin and TikTok. Interestingly, CEO Zhang Yiming didn’t favor Douyin at first. His favorite was Huoshan, and that’s where he poured the most resources. Early on, Huoshan also had the strongest growth momentum, while Douyin remained marginal, with fairly average metrics. But before long, Huoshan hit a ceiling. The fatal issue was that Huoshan’s content style was too similar to Kuaishou’s. You can’t beat a leader by playing the leader’s game; this was no exception.</p><p>Fortunately, the Douyin project wasn’t killed. (It’s said that in a life-or-death decision meeting, a senior executive — Kelly, who later became Douyin’s CEO — strongly defended the project. With an arts background, she felt Douyin’s content aesthetic was distinctive, even though the numbers were still weak.)</p><p>After more than a year of difficult exploration and climbing, Douyin finally broke out in 2017 and achieved massive success.</p><p>In summary: ByteDance was far from the strongest internet company at the time, yet it came from behind and captured the biggest opportunity in short video. Then it repeated the story overseas: TikTok later achieved astonishing success. (Along the way, Mark Zuckerberg made a major mistake: ByteDance poured huge budgets into Facebook to grow TikTok’s user base. Facebook was happy to take the money and completely failed to perceive the potential threat — rather ironic.)</p><p>So yes, people say ByteDance has a powerful recommendation algorithm, a strong growth team, and excellent monetization. I agree those are important. But if we return to first principles — the decisive success factor — I personally believe it was Zhang Yiming’s far-sighted recognition of the opportunity, coupled with a resolve that exceeded everyone else’s.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8836806906e5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Environment is the Product]]></title>
            <link>https://medium.com/@betanote/the-environment-is-the-product-f384d77a1437?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/f384d77a1437</guid>
            <category><![CDATA[ui-design]]></category>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Mon, 15 Sep 2025 10:38:41 GMT</pubDate>
            <atom:updated>2025-09-15T10:38:41.475Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nS-3TA3_YRDEIcbXk9Pdvg.jpeg" /></figure><blockquote><em>In the early days of programming</em>, the world was a flashing cursor and a command line. You typed a command, and the machine gave you a response. It was a sparse, powerful, but unguided environment. It assumed you already knew what you were doing.</blockquote><p>Then the Macintosh came along. It gave you a “desktop” with “folders” and a “trash can.” It created a metaphor for an environment you already understood. You didn’t have to learn the rm command; you just dragged a file to the trash. This was a revolutionary change, but its core wasn&#39;t technical. It was a shift in perspective: a software interface isn&#39;t just a collection of features; it is an environment.</p><p>Users don’t just <em>use</em> your software; they <em>inhabit</em> it. And like any inhabitant, their behavior is a reaction to the environment you’ve built.</p><p>If you truly understand this, the way you design products will change fundamentally.</p><p><strong>Stop Obsessing Over Features</strong></p><p>Most founders, especially technical ones, are obsessed with features. They’ll spend hours debating whether to add a new button or what stack to use for a new function. Their logic is that more features make a better product.</p><p>This is wrong.</p><p>Users don’t care about your feature list. They care about getting a job done in a specific context. They “hire” your product to solve a problem that exists within a certain environment.</p><p>Think about a food delivery app. The “feature” it provides is ordering food, but its success comes from providing the perfect solution for a specific “environment”: “I’m hungry, I’m tired, and I don’t want to cook or go out.” In this environment, opening the app, browsing recommendations, and placing an order become the most natural sequence of reactions. All the design elements — the limited-time discounts, the delivery distance, the driver’s location — aren’t isolated features. Together, they build a micro-environment that coaxes you to complete the “order” reaction.</p><p><strong>If you see your product as an environment, your question shifts from “What features should we add?” to “In what context will users need us</strong>, and how can we design an environment where achieving their goal is the most effortless path?”</p><p><strong>Designers as Architects of Behavior</strong></p><p>Once you start seeing through this “environment” lens, every element in your product becomes a tool for shaping user behavior. They are triggers, incentives, and barriers.</p><p>The red notification badge on an e-commerce app preys on our anxiety about “unfinished” tasks and “new messages.” It places a tiny stimulus in the environment to trigger you to click.</p><p>A push notification from a fitness app — “You haven’t closed your rings today!” — injects a sense of urgency into an environment where you might be tempted to slack off (like after work), prompting the “go work out” reaction.</p><p>One of the most elegant environmental designs I’ve ever seen is Amazon’s 1-Click ordering. What they did wasn’t to <em>add</em> a feature but to <em>remove</em> an entire environment of friction. They removed the steps of filling in your address and credit card info, compressing a complex decision-making environment into a straight line from “see” to “have.” In this new environment, “buying” becomes an almost instinctive reaction.</p><p>Mediocre designers worry about what color a button should be. Great designers worry about what state of mind the user is in when they see the button, and how pressing it will change the user’s environment. They are architects of behavior.</p><p><strong>Go Where Your Users Live</strong></p><p>So how do you know what kind of environment to build?</p><p>You can’t figure it out by guessing in your office. And don’t expect to get the answer from surveys. People are notoriously unreliable when describing their own behavior.</p><p>The only way is to go watch them. Go to their offices, their homes. Watch how they work and live. See how they clumsily solve the problem before your product exists. You need to understand not the abstract “user,” but the specific person in their messy, real, and distraction-filled environment.</p><p>The best collaboration software, like Slack or Figma, wasn’t born from a genius idea in a vacuum. Their creators deeply understood the nature of the “teamwork” environment: information is fragmented, communication is asynchronous, and attention is scarce. So they designed an environment where information flows smoothly, member statuses are visible, and feedback is immediate. They weren’t designing features; they were building an orderly digital habitat for a chaotic natural environment.</p><p>It’s the same reason car navigation apps have huge buttons and clear voice commands. Every design decision is a reaction to the specific environment of “driving”: you are moving at high speed, and your eyes and mind are focused on the road. Any design that requires you to look down for more than a second could be fatal.</p><p><strong>You’re Building a Room, Not a Hammer</strong></p><p>A common mistake founders make is to think they are building a hammer. A hammer is a single-purpose tool. You pick it up when you need it, and you put it down when you’re done.</p><p>But most software isn’t a hammer. It’s more like a room, or a workshop. It’s a place where users spend time, do work, and interact with others.</p><p>If you’re just building a hammer, you only care about its durability and efficiency. But if you’re building a workshop, you have to care about the lighting, the ventilation, the placement of tools, and the ease of movement. You need to create an environment where people can be immersed and productive.</p><p>So, look at your product.</p><p>What kind of environment have you built for your users? Is it a confusing maze full of obstacles, or is it a comfortable space that naturally guides them toward their goals? Does every element you’ve placed help them, or does it just distract them?</p><p>The answer will likely decide your future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f384d77a1437" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Is All You Need: Lessons in Context Engineering from Manus]]></title>
            <link>https://medium.com/@betanote/context-is-all-you-need-lessons-in-context-engineering-from-manus-b1c3d0401961?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/b1c3d0401961</guid>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Mon, 08 Sep 2025 10:36:36 GMT</pubDate>
            <atom:updated>2025-09-15T10:40:12.261Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0Q8JGPXKdibTifvk.png" /></figure><p>Manus’s founder recently shared hard-won, practical lessons on building AI Agents. This is more than a technical guide — it’s packed with deep insights on product strategy, cost control, and user-experience optimization.</p><p>Below are seven core takeaways I distilled for agent founders. Hope they help</p><h3>Key Takeaways</h3><h4>1. Strategic Choice: Bet on “Context Engineering,” not “End-to-End Models”</h4><p><strong>Takeaway:</strong> For startups, rapid iteration before finding product–market fit (PMF) is existential. Building on frontier models (e.g., GPT, Claude) with context engineering — rather than training or fine-tuning your own model — lets you validate ideas and ship improvements in hours, not weeks. This way your product rides the wave of model upgrades (rising tide) instead of being dragged down by your own outdated model (a pillar fixed to the seabed).</p><ul><li><strong>Argument in the article:</strong> The author cites the painful lesson from a prior startup whose in-house model was obsoleted overnight by BERT and GPT-2. Manus chose context engineering to keep the product flexible and state-of-the-art.</li><li><strong>Examples:</strong></li><li><strong>Success case:</strong> Most successful AI agent apps — like <strong>Character.ai</strong> or <strong>Perplexity AI</strong> — don’t train base models from scratch. Their edge lies in clever prompting, conversation-history management, and tool integration (e.g., search) — i.e., “context engineering” — to deliver superior experiences in specific domains.</li><li><strong>Reflection:</strong> If a startup pours resources today into fine-tuning an open-source model for a general Office-document agent, by the time it shows some progress, GPT-5 or Claude 4 may launch with stronger native capabilities and render that work moot.</li></ul><h4>2. Cost &amp; Performance: Treat KV-Cache Hit Rate as a Lifeline</h4><p><strong>Takeaway:</strong> Cost and latency determine whether agent apps can be commercialized. Because agents often have “long input, short output,” KV-cache is critical. Optimizing cache hit rate can cut costs by an order of magnitude (the article mentions up to 10×) and significantly reduce response times — a real business advantage.</p><ul><li><strong>Tips from the article:</strong></li></ul><ol><li><strong>Keep prompt prefixes stable:</strong> Don’t put volatile content like timestamps at the start of the system prompt.</li><li><strong>Append-only context:</strong> Avoid mutating history; make sure serialization (e.g., JSON) is deterministic.</li></ol><p><strong>Use case:</strong> You’re building an agent that continuously monitors a code repo and auto-fixes bugs. The system prompt (“You are a senior software engineer…”) and most prior steps (“pulled repo,” “analyzed file A”) stay the same each run. With a well-tuned KV-cache, when a new step appears (“found a bug in file B”), the model only processes the new delta instead of recomputing the entire context — dramatically boosting efficiency and lowering cost.</p><h4>3. Scalability: Manage a Complex Toolset with “Masking,” not Add/Remove</h4><p><strong>Takeaway:</strong> As agent capabilities grow, the tool/action space can explode. Dynamically adding/removing tool definitions in context breaks the KV-cache and confuses the model. A better approach: keep all tool definitions in context, but use <strong>masking</strong> at a given step to restrict the model to only the currently valid tools.</p><ul><li><strong>Tips from the article:</strong></li></ul><ol><li><strong>Consistent tool prefixes:</strong> e.g., browser_ for browser tools, shell_ for CLI tools.</li><li><strong>Leverage framework features:</strong> Force the model to choose from a subset by pre-filling the reply with something like {“name”: “browser_.</li></ol><p><strong>Case: Devin-like software-dev agent.</strong> The workflow may include hundreds of tools (“read file,” “write file,” “run command,” “browse,” “call API,” etc.). When editing a file, mask the choices so the next action can only be “write file” or “read file,” while temporarily hiding “browse the web.” This preserves workflow logic, avoids cache churn, and keeps the agent stable and efficient.</p><h4>4. Breaking Context Limits: Treat the File System as Unlimited External Memory</h4><p><strong>Takeaway:</strong> Don’t expect vendors to solve everything by endlessly expanding context windows. Very long contexts are expensive, slow, and prone to performance drops (“lost in the middle”). Treat the work environment’s file system (local or virtual) as an unlimited, persistent, agent-operable <strong>external memory</strong> — this is the fundamental way to tackle complex tasks.</p><ul><li><strong>Principle in the article:</strong> Context compression should be reversible. Webpage content can be removed as long as the URL is kept; file contents can be omitted as long as the path remains.</li></ul><p><strong>Use case:</strong> A finance-analysis agent must parse dozens of PDF filings and draft a summary. It can’t load everything into context at once. The right approach:</p><ol><li>Read each PDF sequentially.</li><li>Write key data and notes to a summary.md or data.json.</li><li>In later steps, read only this distilled file — not the original PDFs — to produce the final report.</li><li>The file system becomes an extension of the agent’s thinking and memory.</li></ol><h4>5. Maintain Task Focus: Steer Attention via Continuous “Restatement”</h4><p><strong>Takeaway:</strong> For multi-dozen-step tasks, agents easily “forget” the original goal. A clever tactic: have the agent continually update and rewrite a “plan” or “todo” (e.g., todo.md). This repeatedly pushes the highest-priority goals to the end of the context, exploiting the model’s recency bias and countering goal drift.</p><ul><li><strong>Behavior in Manus:</strong> Create todo.md, check off completed items, and restate what remains throughout the task.</li></ul><p><strong>Example:</strong></p><ul><li><strong>Scenario:</strong> A marketing agent must “plan a 1-week social campaign for a new product.”</li></ul><ol><li><strong>Initial plan:</strong> Agent writes plan.md: “Mon: teaser; Tue: product intro…”</li><li><strong>During execution:</strong> After Monday, it rewrites plan.md, marks Monday done, and fleshes out Tuesday.</li><li><strong>Effect:</strong> This self-restatement mirrors how humans revisit a checklist, keeping the agent aligned with the ultimate objective even after 20+ steps.</li></ol><h4>6. Improve Robustness: Preserve and Learn from Errors, Don’t Hide Them</h4><p><strong>Takeaway:</strong> Failure is the norm in agent loops, not the exception. Hiding errors and blindly retrying wastes learning signals. Preserve failed actions and returned error messages (e.g., stack traces) in context so the model can “see” what didn’t work and make better next steps — avoiding repeat mistakes.</p><ul><li><strong>View in the article:</strong> Error recovery is one of the clearest indicators of true agentic behavior.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Z2vks_q4vlBgGWeS.png" /></figure><p><strong>Example:</strong></p><ul><li><strong>Scenario:</strong> A test automation agent tries to click a button, but the element isn’t loaded yet.</li><li><strong>Wrong way:</strong> Hide the error and just retry; the agent doesn’t know why it failed.</li></ul><p><strong>Right way:</strong> Record it in context:</p><ul><li>Action: click(“button#submit”) -&gt; Observation: Error: ElementNotVisibleError</li><li>Seeing this action–result pair, the model may infer it must first wait_for_element(“button#submit”, 5) before clicking. That’s learning from failure — key to building intelligent, robust agents.</li></ul><h4><strong>7. Avoid Rigidity: Use Diversity to Break the Tyranny of Few-Shot Patterns</strong></h4><p><strong>Takeaway:</strong> LLMs are powerful pattern-mimickers. If your context is packed with many similar, uniform “successful” few-shot examples, the agent can become rigid — stuck in a template even when a different solution is needed.</p><ul><li><strong>Method in the article:</strong> Introduce small, controlled variation or randomness in action/observation serialization templates, wording, order, or format to “break the pattern.”</li></ul><p><strong>Example:</strong></p><ul><li><strong>Scenario:</strong> A résumé-screening agent processes 10 software-engineer resumes, each time following the same sequence: “open resume → search ‘Python’ → record years.” The 11th resume belongs to a stellar design director who lists “PyTorch/Django expert” without the literal word “Python.” A rigid agent might reject it.</li><li><strong>Solution:</strong> Include diverse examples — some start with projects, others with education; vary wording and note styles. This “noise” prevents overfitting to one pattern and improves adaptability to novel cases.</li></ul><h3>Summary</h3><p>For agent founders, this article offers not lofty theory but a <strong>tactical playbook</strong> drawn from real practice — directly guiding product design, technical architecture, and cost optimization. The core idea: <strong>don’t pin your hopes solely on a “bigger brain” (the LLM); focus on crafting a high-quality information environment (Context) around that brain — one that’s efficient, stable, and continuously learning.</strong></p><p><strong>Successful agents = powerful models + exceptional context engineering.</strong></p><p><strong>Original Source: </strong><a href="https://manus.im/zh-tw/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus">https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b1c3d0401961" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Engine and the Car]]></title>
            <link>https://medium.com/@betanote/the-engine-and-the-car-13d194a68436?source=rss-57e8f4e4b9c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/13d194a68436</guid>
            <dc:creator><![CDATA[Beta]]></dc:creator>
            <pubDate>Mon, 01 Sep 2025 13:40:04 GMT</pubDate>
            <atom:updated>2025-09-01T13:40:36.996Z</atom:updated>
            <content:encoded><![CDATA[<p>The Engine and the Car</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*HhHopThmpRiNydCY.jpg" /></figure><p>If you’re building a startup, you’re building a car. And right now, in the middle of the AI boom, everyone is obsessed with the engine. They talk about its horsepower, its torque, its efficiency. They benchmark it constantly. Founders, especially the technical ones, are spending all their time in the garage, trying to build a better engine.</p><p>I understand why. In the earliest days of the automobile, the company with the best engine won. A car with 50 horsepower was not just an improvement on a car with 5 horsepower; it was a completely different machine.</p><p>That’s where we are with AI today. The intelligence of the model — the engine’s power — directly determines the quality of the experience. For products like AI-powered search or tools that help us write and think, a smarter model feels like magic. This is the core truth behind the “model-as-product” idea. And it’s why, when a company like DeepSeek releases a dramatically more powerful engine, the entire landscape shifts overnight. Competitors who were leading the race suddenly find themselves left in the dust, their products feeling quaint and slow.</p><p>You can see the panic this creates. Companies like Kimi and Minimax, seeing the market’s reaction, had to immediately declare that building a more powerful engine was once again their top priority. In a horsepower race, you can’t afford to fall behind. A superior engine can be a killer blow</p><p>So, it seems logical for a startup to focus all its energy on the engine. But for most, especially the small ones, this is a fatal mistake. You are choosing to compete on the one variable where the largest, best-funded companies have an insurmountable advantage. You’re trying to out-engineer General Motors in their own garage.</p><p>“Fine,” you might say, “I won’t build the engine. I’ll buy the best one off the shelf from a supplier like OpenAI or Anthropic and put it in my car.”</p><p>This is the “wrapper” strategy. It’s a fantastic way to get started. You can build a beautiful car body, get it on the road fast, and see if people want to drive it. It’s how you find product-market fit. You can even build a profitable little business this way for a while.</p><p>But it is an exceptionally dangerous long-term plan. Because what happens when your engine supplier decides they want to start selling their own cars?</p><p>We saw this play out with Cursor. They built a beautiful car for developers, powered by third-party engines. Then Anthropic, one of the premier engine makers, released Claude Code. They effectively started building their own car, using their newest, most powerful engine. The company you depend on for your most critical component has become your direct competitor. They can pull the rug out from under you at any moment.</p><p>That “end-to-end user experience” you’ve crafted? It’s just a car body bolted onto someone else’s chassis. You are perpetually vulnerable.</p><p>So what should a startup do? You have to stop thinking about the engine and start thinking about the car. The engine is a component. The car is the product. Your moat — your real, defensible advantage — must be in the car itself.</p><p>What makes a great car that people will refuse to give up, even if a new model with a slightly better engine comes along?</p><p><strong>1. The Interior.</strong> A great car’s interior is molded to its owner. The driver’s seat remembers your position. The GPS knows your home and work. The radio has your favorite stations. This is your users’ private, sticky data. It’s their project history, their unique workflows, their collaborative environment. The more they use your product, the more it becomes <em>theirs</em>. The cost of switching isn’t just learning a new interface; it’s losing a part of their own brain.</p><p><strong>2. The Community.</strong> People don’t just buy a Porsche or a Jeep; they buy an identity. They join clubs. They go to meetups. They wave at each other on the road. This is the power of a UGC community and a unique culture. If you can build a product where users help each other, share their creations, and form relationships, they are no longer just customers. They are members. They stay for the people, not just the features.</p><p><strong>3. The Dashboard.</strong> A great dashboard doesn’t just show you your speed. It connects to your phone, your music, your maps. It’s the integrated control center for your journey. This is deep workflow integration. Your product needs to connect to the other tools your users depend on. It needs to become the central place where their most important work gets done. When you are the dashboard for their work life, you become indispensable.</p><p><strong>4. The Service Network.</strong> For some businesses, the moat is in the real world. Think of Uber. Its AI is the engine, but its defensibility is the global network of drivers — the service network. This is a physical asset that software alone cannot replicate.</p><p>The horsepower race will eventually end. Engines will become commodities — incredibly powerful, but available to everyone. The great, enduring car companies will not be the ones that build the engines, but the ones that build the cars people love to drive.</p><p>So, as a founder, you have to ask a different set of questions. Stop asking “How can I make my model 5% more accurate?”</p><p>Start asking: “What car am I building? Where does it take my users? And what am I building around my engine that makes them never want to get out of the driver’s seat?”</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=13d194a68436" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>