书评
书评：投资AI:6个精神模型告诉Moats从陷阱由Rob May编写
对罗布·梅及其为投资者和企业家建设耐久AI企业的实用精神模型的个人回顾。
在 Google 上的首选来源中添加 Unite.AI
我非常兴奋地阅读了《投资AI：六种精神模式从陷阱中告诉Moats》，主要是因为我多年来一直关注Rob May，我认为，将罗布·梅与许多投资者分开的是，他愿意避免集团思维和理性与第一原则。 他看过去流行的叙述，想问一个企业是如何实际运作的，它的优势来自何处，这些优势是否能够在接下来的技术转变中存活下来。
这本书就是在这里发光的。 这是我遇到的第一本书，它接近AI投资相当如此。 它要求读者重新设计如何分析围绕它正在兴建的企业。 机会是巨大的，但承认有希望的AI能力和承认可投资的公司是两种截然不同的技能。
当AI的叙述性运行到现实
Rob May在任何关注人工智能生态系统的人所熟悉的故事中，贾斯珀以15亿美元估值成为媒体亲爱的融资1.25亿美元。 随后，ChatGPT直接向消费者带来了强大的文本生成，挑战了围绕获取类似基础能力而构建的产品的价值。 Rob May利用逆转来检查一个明显的优势能够变得脆弱的速度。
这个例子让我感兴趣的是它迫使投资者或创始人面对的问题：当基础技术被广泛利用时，客户仍然需要你的公司做些什么？有防護牆的生意和另一个人工智能包装有什么区别？早期的领先可以赢得时间，但它没有回答这个问题。 工作流程所有权、分配、专门知识和客户关系一旦进入模式成为商品，就变得更加重要。
DeepSeek呈现出这种不确定性的另一面。 书中讨论了GPT-4级模型的冲击，报告的培训费用为5576万美元。 投资教训是重大的：对竞争成本的假设可以比估值模式预期的要快得多。
SaaS 游戏本为何需要重新思考
Rob May重温了过去的AI冬天，认为今天部署的广度打破了充满希望的示威的旧模式，他的论点是，有用的AI现在被嵌入了太多真正的工作流程中，无法简单地退入实验室。 然而，即使继续收养，仍然难以确定下一个能够利用收养的独角兽。
他确定了四个有助于解释转变的基本区别：AI可以进行认知工作，学习数据，改变使用它的人的期望，成为一代人与它一起成长的默认生命的一部分。 第三点——弹性——特别有趣。 一旦客户体验到某种能力，他们就会开始在别处期待。 随着技术的采用，一家公司正在服务于变化的市场。
最大的游戏改变者是缩放速度。 AI企业可以在以前属于特殊外部的时间内达到相当规模。 但是，评估它们仍然需要了解商业模式。 每个任务或每个结果的充电，在价值，收入和成本之间建立了不同于出售固定数量软件座椅的关系。
萨阿斯的一些经验教训仍然十分重要。 领域专门知识仍然重要。 了解一个行业的真正问题的纵向工具可以建立有意义的优势。 专有数据循环可以复合，深度集成可以创造切换成本。 挑战在于如何区分这些持久的长处和客户可从其他几十家供应商获得的特点。
六种心理模式让问题变得更好
该书围绕6种精神模式组织其中心论点：将AI理解为技术集；挑战传统的SaaS假设；期待重大的市场混乱；承认市场和公司战略的变化；理解更快的尺度；以及仔细评价AI销售商品的成本。
两者合作良好，因为两者都暴露了表面投资论文中不同的弱点。 将一家公司称为“AI公司”，称对其经济学略微了解。 收入急剧增加，称的耐久性差一点。 当客户开始要求产品做更多困难的工作时，今天的强毛差称几乎没有发生什么。
我最珍视的是问第二个问题的习惯。 在询问该技术是否有效后，询问该技术是否对客户的实际工作流程有效。 在询问客户是否支付后，询问他们支付什么费用，以及这种支出是否可能继续。 这使得这本书在写出其例子的那一刻之后就变得有用。
Botkeeper 和成本轨迹的重要性
最有用的案例研究之一是Botkeeper，一家记账自动化公司Rob May投资。 他描述了一个最初严重依赖人类会计的企业，总利润约为20%至25%。 这对期望获得传统软件利润的投资者来说似乎没有吸引力。 但这些人所完成的工作也产生了可以帮助进程自动化的数据。 Rob May报告说，随着更多的工作转向软件，毛利润后来超过了70%。
我从中吸取的教训是研究人类参与正在建设什么。 如果每一项完成的任务都有助于系统更有效地处理下一个任务，早期服务费用可能为更可扩展的业务提供资金。 如果每个新客户都要求无限期地按比例增加人工工作，那么经济学会讲述不同的故事。
Rob May还研究了相反的轨迹：推论成本随着使用量的增长而变得更加苛刻。 AI系统可能需要更多的计算，重复尝试，或者人文审查来完成更艰巨的任务。 在小型飞行员中看起来有吸引力的产品在客户依赖它进行复杂工作时可能会变得昂贵。 这就是为什么费用的方向与目前的差值一样重要。
效率不是自动的模拟
我发现另一个有用的框架是罗布·梅的AI价值三合一：商品化AI，核心AI，扩展AI. 广义地说，这些描述竞争者可以随时采用的能力，加强公司现有优势的能力，以及开拓新的商业机会的能力。
他假想的航空公司例子使得第一类很容易理解。 一家推出航空公司的AI支持agent，并变得更加高效。 竞争者然后采用类似的工具。 短暂看来不同者变成了行业的期望，一些经济利益可能会被竞争掉。
这对将每个AI的公告作为企业更强大的证据来说是一种有益的解药。 我想知道，在竞争者获得可比工具之后，什么仍然是独特的。 一个将AI与专有反馈，深层操作知识，或难以操作的replicate客户工作流程相结合的公司，比一个宣布其他人可以购买的同样自动化的公司有更有说服力的答案。
寻找操作员谁可以让AI工作
这本书充满了实用的观察，包括支持能够实际吸收AI的操作者的重要性。 改变工作流程和责任的具体计划远比整个组织对大赦国际充满希望的介绍更有说服力。
Rob May关于收入质量的讨论证实了这一点。 付费的飞行员和概念证明可以产生令人印象深刻的早期数字，但是它们应与与正在进行的生产工作挂钩的收入分开。 有用的问题是，如果客户明天停止试验，业务会有多少剩余。
我认为这对创始人来说是特别宝贵的。 赢得一个飞行员是进步，但建立成为客户操作方式一部分的产品则是不同的成就。 了解及早区分可以决定产品的优先次序、雇用和公司衡量成功的方式。
最后判决
我最大的外卖是这本书不应该只留给投资者企业家从了解AI公司的可投资性、持久优势如何形成以及哪些启动陷阱在变得昂贵之前是可以避免的同样多。
最好把它作为判断的框架，而不是购买股票的清单。 它的最大贡献是帮助读者更加确切地了解他们为什么相信一个企业。 这些例子将不可避免地随着行业的前进而老化；质疑假设的学科就是如何使这本书具有持久效用。
投资AI：六种智商模式，对于试图将可辩护的企业与令人信服的故事分开的投资者，以及试图建立前者的企业家来说，这是一个深思熟虑和有价值的解读。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Book Reviews
Book Review: Investing in AI: Six Mental Models to Tell Moats from Traps by Rob May
A personal review of Rob May and his practical mental models for investors and entrepreneurs building durable AI businesses.
Add Unite.AI to your preferred sources on Google
I was excited to read Investing in AI: Six Mental Models to Tell Moats from Traps, largely because I have been following Rob May for years through his Investing in AI newsletter. What separates Rob May from many investors, in my view, is his willingness to avoid groupthink and reason from first principles. He looks past the prevailing narrative to ask how a business actually works, where its advantages come from, and whether those advantages can survive the next technological shift.
That is where this book shines. It is the first book I have encountered that approaches AI investing quite this way. Rather than simply identifying an exciting technology or a growing market, it asks readers to re-engineer how they analyze the businesses being built around it. The opportunity is enormous, but recognizing a promising AI capability and recognizing an investable company are two very different skills.
When the AI Narrative Runs Into Reality
Rob May grounds the discussion in stories that will be familiar to anyone who has followed the AI ecosystem. Jasper became a media darling and raised $125 million at a $1.5 billion valuation. Then ChatGPT brought powerful text generation directly to consumers, challenging the value of a product built around access to similar underlying capabilities. Rob May uses the reversal to examine how quickly an apparent advantage can become vulnerable.
What interests me about this example is the question it forces an investor or founder to confront: what does the customer still need from your company when the underlying technology becomes widely accessible? What differentiates a business with a defensible moat from just another AI wrapper? An early lead can buy time, but it does not answer that question. Workflow ownership, distribution, specialized knowledge, and customer relationships all become much more important once access to the model becomes commoditized.
DeepSeek presents the other side of that uncertainty. The book discusses the shock of a GPT-4-class model emerging with a reported training cost of $5.576 million. The investment lesson is significant: an assumption about the cost of competing can change much faster than a valuation model anticipates.
Why the SaaS Playbook Needs Rethinking
Rob May revisits the AI winters of the past and argues that today’s breadth of deployment has broken the old pattern of promising demonstrations followed by widespread disappointment. His argument is that useful AI is now embedded in too many real workflows to simply retreat into the laboratory. Nonetheless, even with continued adoption, it can still be difficult to identify the next unicorn that can capitalize on that adoption.
He identifies four fundamental differences that help explain the shift: AI can perform cognitive work, learn from data, change the expectations of the people using it, and become a default part of life for a generation growing up with it. The third point—reflexivity—is especially interesting. Once customers experience a capability, they begin expecting it elsewhere. The market a company is serving changes as the technology is adopted.
The biggest game changer is the speed of scaling. AI businesses can reach substantial size on timelines that previously belonged to exceptional outliers. But evaluating them still requires understanding the business model. Charging per task or per outcome creates a different relationship between value, revenue, and cost than selling a fixed number of software seats.
Some SaaS lessons remain highly relevant. Domain expertise still matters. Vertical-specific tools that understand an industry’s real problems can build meaningful advantages. Proprietary data loops can compound, and deep integrations can create switching costs. The challenge is distinguishing those durable strengths from features a customer could obtain from dozens of other vendors.
Six Mental Models That Make the Questions Better
The book organizes its central argument around six mental models: understanding AI as a collection of technologies; challenging traditional SaaS assumptions; expecting major market dislocations; recognizing changes to markets and corporate strategy; understanding faster scaling; and carefully evaluating AI’s cost of goods sold.
These work well together because each exposes a different weakness in a superficial investment thesis. Calling a company an “AI business” says little about its economics. A dramatic revenue increase says little about durability. A strong gross margin today says little about what happens when customers start asking the product to do more difficult work.
What I value most is the habit of asking a second question. After asking whether the technology works, ask whether it works reliably enough for the customer’s actual workflow. After asking whether customers are paying, ask what they are paying for and whether that spending is likely to continue. This makes the book useful beyond the moment in which its examples were written.
Botkeeper and the Importance of the Cost Trajectory
One of the most useful case studies is Botkeeper, a bookkeeping automation company Rob May invested in. He describes a business that initially relied heavily on human accountants and had gross margins of roughly 20 to 25 percent. That looked unappealing to investors expecting conventional software margins. But the work those people performed also generated data that could help automate the process. Rob May reports that gross margins later exceeded 70 percent as more work shifted to software.
The lesson I take from this is to examine what human involvement is building. If each completed task helps the system handle the next one more efficiently, early service costs may be financing a more scalable business. If every new customer requires a proportional increase in manual work indefinitely, the economics tell a different story.
Rob May also examines the opposite trajectory: inference costs that become more demanding as usage grows. An AI system may need more computation, repeated attempts, or human review to complete harder tasks. A product that looks attractive in a small pilot can become expensive when customers rely on it for complex work. This is why the direction of costs matters as much as the current margin.
Efficiency Is Not Automatically a Moat
Another framework I found useful is Rob May’s triad of AI value: commoditized AI, core AI, and expansion AI. Broadly, these describe capabilities competitors can readily adopt, capabilities that strengthen a company’s existing advantage, and capabilities that open up new business opportunities.
His hypothetical airline example makes the first category easy to understand. An airline introduces an AI support agent and becomes more efficient. Competitors then adopt similar tools. What briefly looked like a differentiator becomes an industry expectation, and some of the economic benefit may be competed away.
This is a helpful antidote to treating every AI announcement as evidence of a stronger business. I want to know what remains distinctive after competitors have access to comparable tools. A company that combines AI with proprietary feedback, deep operational knowledge, or a difficult-to-replicate customer workflow has a more persuasive answer than one announcing the same automation everyone else can buy.
Look for Operators Who Can Put AI to Work
The book is full of practical observations, including the importance of backing operators who can actually absorb AI. A concrete plan for changing workflows and responsibilities is far more compelling than a presentation promising AI throughout the organization.
Rob May’s discussion of revenue quality reinforces this. Paid pilots and proofs of concept can generate impressive early numbers, but they should be separated from revenue tied to ongoing production work. The useful question is how much of the business would remain if customers stopped experimenting tomorrow.
I see this as particularly valuable for founders. Winning a pilot is progress, but building a product that becomes part of how a customer operates is a different achievement. Understanding that distinction early can shape product priorities, hiring, and the way a company measures success.
Final Verdict
My biggest takeaway is that this book should not be reserved exclusively for investors. Entrepreneurs have just as much to gain from understanding what makes an AI company investable, how durable advantages form, and which startup pitfalls can be avoided before they become expensive.
It is best approached as a framework for judgment rather than a list of stocks to buy. Its strongest contribution is helping readers become more precise about why they believe in a business. The examples will inevitably age as the industry moves forward; the discipline of questioning assumptions is what gives the book its lasting usefulness.
Investing in AI: Six Mental Models to Tell Moats from Traps delivers the first-principles thinking that made me a reader of Rob May’s newsletter in the first place. For investors trying to separate a defensible business from a compelling story, and entrepreneurs trying to build the former, it is a thoughtful and worthwhile read.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.