模块 #4：如何思考和抓住 AI 机会？
我的理解
本模块提出应对 AI 时代信息过载与不确定性的第二条核心答案：光会构建还不够，还必须培养深刻理解。讲师以 ChatGPT 伪造论文为例，说明如何用三重过滤器判断信息价值——它是真实信号吗？它重要吗？它有可能发生吗？核心工具是「世界模型」：通过不断预测并主动证伪来更新对世界的心智模拟，目的不是证明自己正确，而是发现自己世界模型的盲区。模块围绕三个推测展开：幻觉是 GenAI 的基本特性、Transformers 或许通向 AGI、反共识观点才是人类的竞争壁垒。分发能力（distribution）的案例同时提醒我们：理解技术格局必须看本质，而非被产品表象或组织结构迷惑。
相关链接
- Ch05-L06 要点 1 与幻觉共处 — 幻觉推测在此课的深入展开
- Ch05-L08 要点二 有价值的观点都是反共识的 — 反共识推测在此课的深入展开
- Ch07-L02 预训练大语言模型 — 理解共识机制偏向的预训练技术基础
- Ch04-L11 技术洞察 像管理者一样思考 — 判断框架与管理者思维框架相互印证
原文
Lesson 37 of 68 视频 - 模块 #4：如何思考和抓住 AI 机会？/ Video - Module #4: 如何思考和抓住 AI 机会？ 好，大家在屏幕上 Well, I have everybody on the screen Hide transcript Transcript 00:02 好，大家在屏幕上 Well, I have everybody on the screen 00:03 举手，如果你害怕 raise our hand if you are afraid 00:05 被 AI 取代的话。of being replaced by AI. 00:13 对了，那是什么意思？What does that mean by the way? 00:16 是的，只是对那种可能性感到焦虑或担忧 Yeah, just anxious or concerned with the possibility 00:19 它一直萦绕在你心头 like it’s on your mind 00:21 因为你已经见识到生成式 AI 的能力有多强，对吧？because you have seen how much AI can do, right? 00:25 你是不是担心，我会试图说服你 Are you worried I’ll try to convince you 00:29 到这个环节结束时，其实不必担心，by the end of this session that don’t have to worry, 00:34 但这会很困难。but that’s going to be difficult. 00:37 好的，我已经发送了 Okay, I sent request 00:39 屏幕共享请求。to share my screen. 00:43 酷。Cool. 00:46 欢迎。Welcome. 00:47 欢迎来到第四模块。Welcome to module four. 00:48 这是我们最后的直播课。This is our final session, live session. 00:51 但我们其实有很多 But we actually have a lot 00:52 除了直播会话之外的其他内容。beyond just the live session. 00:54 我稍后会详细谈那个。And I’ll talk about that later. 00:57 但在本次最后的直播课中 But in this final live session 00:59 我们将超越单纯的构建阶段。we are moving beyond building. 01:02 我们将探讨如何思考、We are going into how to think, 01:04 如何应对并驾驭这个生成式 AI 的新范式。how to think and to navigate this New AI paradigm. 01:09 这一部分比较难懂 This part is tricky 01:11 因为它只对开发者有用 because it’s only useful to builders 01:13 而且开发者都比较务实。and because builders are grounded. 01:16 否则，很容易把 Otherwise, it’s easy to confuse 01:17 幻觉与远见混为一谈。hallucination with visionary. 01:20 所以我先快速举一个例子开始。So I’ll start with an example quickly. 01:23 但在那之前，But before that, 01:25 我们先快速回顾一下，we also want to do a quick review 01:27 为什么我们需要学会如何思考。of why we need how to think. 01:31 因此，在第一模块，So in module one, 01:33 我们的目标是填补 we set out to close five key gaps 01:34 AI 技能的五个关键空白。in our AI skills. 01:37 Yan 为我们讲解了两个深入的技术环节 Yan walked us through two deep technical sessions 01:40 ，以帮助实现这一目标。to help make that happen. 01:43 此外，我们还整理了一个超过 200 页的阅读资料库 On top of that, we put together 200 plus pages reading library 01:47 以便在学习过程中作为指导。so we can use that as a guide along the way. 01:50 如果你去上我们的课。If you go to our lesson. 01:51 许多课程被设计为 A lot of the lessons are designed to be 01:54 阅读资料库。a reading library. 01:55 在构建过程中，如果遇到问题，As you build, when you encounter a problem, 01:59 你可以查看课程来获取答案。you can go to the lessons and get an answer. 02:02 还有一个顶点项目。There is also a capstone project. 02:04 我将解释它的设计方式 I’ll explain how it’s designed 02:06 以及如何最大限度地发挥它的作用 and how to get the most out of it 02:08 ，在本节课结束时。at the end of this session. 02:11 但掌握这些技能仅是 But learning the skills is only part of 02:13 我们应对这一转变的一部分。how we navigate this shift. 02:16 仍然存在两大挑战。Two big challenges still remain. 02:18 信息过载与不确定性。Information overload and uncertainty. 02:22 我们来举这个例子 Let’s do this example 02:24 来说明如何思考。of how to think. 02:25 比如，For example, 02:26 我可以把这项新研究分享给你。I could share this new research with you. 02:28 一条比 Transformers 更好的通往 AGI 的道路 A better path to AGI than Transformers 02:32 它利用基于神经科学的代理。using agents based on neuroscience. 02:36 这听起来很激动人心 It can sound exciting 02:39 ，如果我一步步带你分析其逻辑，and if I walk you through the logic, 02:43 或许它甚至会显得合理且可行。maybe it will even sound plausible or feasible. 02:48 尽管 Transformer 模型取得了成功，The transform model, despite its success, 02:51 但它存在固有的局限性，has inherent limitations 02:52 因为大家都知道，because we all know 02:53 它本质上只是下一个 token 的预测。it’s just next token prediction. 02:57 绝大多数机器学习模型的保质期仅有三年。Most machine learning models only have three years of shelf life. 03:00 现在，Transformers 是发明于…… Now Transformers invented in. 03:03 等等，那是什么时候。Wait, when was. 03:04 大概是 89 多年前了。It’s over 89 years ago, probably. 03:10 嗯。Yeah. 03:11 哦，好，2017 年。Oh, okay, 2017. 03:13 嗯，对。So yeah. 03:15 哦，那还是去年的事。Oh, and that was last year. 03:18 所以那是九年前的事。So nine years ago. 03:20 神经科学的最新进展提出了一种很有前景的新方法 Recent development in neuroscience suggests a promising new approach 03:24 即在神经网络中 that use agents instead of neurons in neural network 03:25 以代理（agents）而非神经元作为基本单元。as the base unit. 03:29 因此，我们可以将世界知识注入到智能体（agents）中。So we can install world knowledge into agents. 03:33 这个新模型有可能实现 AGI And this new model could potentially enable AGI 03:36 在更基础的层面。on a more fundamental level. 03:38 你对这个研究感到激动吗？Do you feel excited about this research? 03:42 但问题是，这篇论文是假造的。But the problem is this paper was fake. 03:46 这是 ChatGPT 编造出来的。It was made up by ChatGPT. 03:48 所以这篇论文的问题在于 So the problem of this paper is 03:52 它本身没错，it is not wrong per se, 03:54 但缺乏依据。but it’s not grounded. 03:56 我们实际上发现上面所有的说法都是正确的 We actually see all the claims above are actually correct 04:00 但它们缺乏依据，因此不可能实现。but it’s not grounded, so it cannot happen. 04:04 这并不错误，但只是不可能 It’s not wrong, but it’s just not possible 04:06 或者可能性太小。or the possibility is too small. 04:08 所以我们经常看到各种 So we actually see these kinds of articles 04:11 关于 AI 的文章和预测。and predictions all the time regarding AI. 04:14 因此，本次讲座将帮助你培养批判性思维 So this session will help you develop the critical thinking 04:18 穿透那些炒作和噪音 to cut through those hype, those noises 04:20 从而更有效地进行研究学习。and learn research more effectively. 04:24 这就是为什么我说这个课程 That’s what I mean by this lesson 04:26 只对开发者有用 is only useful to builders 04:28 因为开发者是脚踏实地的。because builders are grounded. 04:31 因此，这就引出了一个核心问题。So this leads to the core question. 04:33 当目标未知且一切都在变化之际。When the target is unknown and everything is changing. 04:37 我们怎样分辨对错？How do we figure out what’s right versus wrong? 04:40 我们怎样从噪声中辨识出信号？How do we find signal in the noise? 04:42 我们该如何培养自己的判断力？How do we develop our judgment? 04:45 我们首要且最重要的答案就是：构建。Our first and most important answer is to build. 04:48 模块 1、2、3 都旨在 Module 1, 2, 3 were all designed 04:50 为你提供实现这一点的技能。to give you the skills to do that. 04:53 现在进入模块 4，Now in module 4, 04:54 我们来探讨第二个同样关键的答案。we address the second, equally crucial answer. 04:57 你为什么这么想？How do you think so? 05:00 对这个问题的第二个答案 The second answer to the question 05:03 是培养深刻的理解。is to develop deep understanding. 05:05 光建是不够的。It is not enough to just build. 05:07 我们必须明白构建它的原因 We must understand why we are building 05:10 以及它的工作原理。and how it works. 05:12 我们必须通过亲手构建来理解。We must understand through building. 05:16 ‘理解’到底是什么意思？What does understand really mean? 05:18 我想向大家介绍世界模型的概念。I want to introduce the concept of a world model. 05:21 世界模型是机器学习领域的热门话题 World model is a hot topic in machine learning, 05:25 但我特别想强调它对我们人类的重要性。but I want to emphasize how important it is to us humans. 05:29 这是你对世界运行方式的内心模拟。This is your mental simulation of how the world works. 05:33 它能让你捕捉因果关系，It allows you to capture causality, 05:35 模拟各种“如果……会怎样”的情景，run what if scenarios, 05:38 并选择最佳前进路径。and choose the best path forward. 05:40 这里的关键词是运行‘假如’情景模拟。The keyword here is running what if scenarios. 05:45 what if 是什么？What is what if? 05:48 那么，我们该如何完善这个世界模型呢？So how do we sharpen this world model? 05:50 这个方法虽简单却很强大。The method is simple but powerful. 05:53 你通过…的方式来开发它。You develop it 05:56 预测并证伪 by predicting and falsifying. 05:57 那就是那些“如果……会怎样”的情景。That is the what if scenarios. 06:00 预测与证伪。Predicting and falsifying. 06:03 其实这正是我们从小到大在中小学所学的 So it’s actually simply the scientific method 06:10 科学方法。that we all learned from elementary school to high school. 06:12 我们一生都在学这个，We learned this our entire life, 06:15 但直到读到 but I don’t think it’s quite explicitly stated to me 06:16 《禅与摩托车维修艺术》这本书，我才真正明确地认识到这一点。until I read this book, the Zen and the Art of Motorcycle Maintenance. 06:24 书中作者对此总结得非常到位。In a book, the author summarizes it very well. 06:26 你提出假设 You form a hypothesis 06:28 设计并进行实验。design and experiment. 06:30 接下来是关键一步：And then this is the crucial part, 06:32 在运行实验前，你需要先预测结果。you predict the result before you run the experiment. 06:37 预测结果与实际观察之间的差距 The gap between a prediction and the observed reality 06:41 正是学习发生并更新你的世界模型的地方。is where learning happens and your world model gets updated. 06:47 如果你的世界模型很优秀，If your world model is great, 06:49 那么你的所有预测都会成真。then all your predictions will happen. 06:51 如果你的预测与现实不符 If your prediction doesn’t meet reality 06:53 那就说明你的世界模型需要更新。that means your world model should get updated. 06:59 我们在提出假设时，When we form hypothesis, 07:00 就是在寻找有价值的想法。we are looking for valuable ideas. 07:02 我在三个维度上提供价值。I give value on three dimensions. 07:05 首先，这是真正的信号，还是仅仅是噪声？First, is it a real signal or just noise? 07:09 其次，如果这是一个信号，Second, if it’s a signal, 07:11 它重要吗？is it important? 07:12 第三，如果它很重要，And third, if it’s important, 07:14 它是否有可能发生？is it likely to happen? 07:17 通常，Most of the time, 07:18 我们不会对 we don’t apply these filters 07:19 所消费的信息应用这些过滤器。to the information we consume. 07:22 比如，那篇论文。For example, that paper. 07:24 那篇论文是可行的。That paper is possible. 07:26 如果你问一个假设性的 If you ask a hypothetical 07:27 ‘这是可能的吗？’问题？is this possible question? 07:29 一切皆有可能。Everything is possible. 07:31 外星人明天会接管世界 Is it possible that alien is going to 07:33 有可能吗？take over the world tomorrow? 07:34 是的，但那个的可能性有多大呢？Yes, but what’s the likelihood of that? 07:36 它不高。It’s not high. 07:37 所以我们确实需要提出三个问题。So we really need to ask three questions. 07:39 这是个信号吗？Is it a signal, 07:40 它重要吗？is it important? 07:42 那么，这有可能吗？And is it likely? 07:43 如果你具备信号、重要性和可能性，If you have signal, importance and likelihood, 07:47 就能获得一些宝贵的信息。you get some valuable information. 07:51 仅仅找到一个有价值的想法是不够的。And finding a valuable idea isn’t enough. 07:54 这已经够难的了，It’s hard enough, 07:56 但还远远不够。but it isn’t enough. 07:57 你必须内化它 You must internalize it 07:59 ，并建立坚定的信念。and build conviction. 08:01 最好的方法就是 The best way to do this 08:03 持有强烈的观点，但要保持灵活性。is to have strong opinions weakly held. 08:07 你要追求极致清晰 You seek extreme clarity 08:08 ，并主动尝试证伪自己的想法。and actively try to falsify your own idea. 08:13 如果你做不到，那么信念就会自然而然地建立起来。If you cannot, then convictions builds naturally. 08:19 现在你或许会说 Now you might say 08:20 ，未来是无法预测的。the future is unpredictable. 08:22 没错，确实如此。And that’s true. 08:22 未来是动态而多维的。The future is dynamic and multidimensional. 08:25 所以，仅凭人类思维很难预测未来。So it’s very hard to predict from our human mind. 08:29 但进行预测依然是一项至关重要的实践。But making prediction is still a vital practice. 08:32 即使无法掌握每一个细节 It’s possible to predict the trajectory 08:35 也能预测其轨迹。even if you don’t get every detail right. 08:37 更重要的是，如果你具备知识诚实，More importantly, if you have intellectual honesty, 08:42 这能让预测中的错误 this makes the mistakes in our prediction 08:44 帮助你发现世界模型中的空白。can help you identify the gaps in your world model. 08:47 然后你就可以修正它了。Then you can fix it. 08:49 再说一次，回到‘强烈意见，弱持有’这个理念，Again, going back to the strong opinions weakly held, 08:54 我们做预测并表达强烈观点的目的，the reason we want to make predictions and we want to state our strong opinions 08:57 不是为了证明自己是对的。is not to prove ourselves right. 09:00 那毫无用处。That’s useless. 09:02 目的是证明我们自己错了。It’s to prove ourselves wrong. 09:04 我们就是在那个地方学习。That’s where we learn. 09:05 这正是我们的预测现实 That’s the gap in our predictive reality 09:08 与实际现实 and actual reality 09:09 之间的差距，它提醒我们需要更新世界模型。that signals to us we need to update our world model. 09:13 那就是学习发生的时候。That’s when learning happens. 09:16 例如，As an example, 09:16 我在 2023 年 3 月撰写了一篇文章，I wrote an article in March 2023 09:19 提出了关于 ChatGPT 的五个关键问题。laying out five key questions about ChatGPT. 09:22 回顾来看，这样的预测大多 Looking back, most of the predictions 09:25 像‘它是否会造成颠覆？’ like is it a disruption? 09:27 或者，我们关于如何参与它 Or how to participate in it 09:29 的预测是否站住了脚。have held up. 09:30 这说明，进行预测是值得的。This shows that prediction is a worthwhile exercise. 09:37 我四月份的那场演示，In a presentation that I did in April, 09:40 大概是三年前做的吧？like how many years, three years ago? 09:43 嗯。Yeah. 09:44 当时我做出了更具体的预测 I made more specific predictions 09:45 这些预测与当时主流观点相反。that were contrary at the time. 09:47 封闭模型会变成开放模型，That closed models would be the open models, 09:50 提示比微调更重要，that prompting would be more important than fine tuning, 09:52 大科技公司比初创公司获益更多。and that Big tech would benefit more than startups. 09:55 短期来看，In the short run, 09:55 这些预测在很大程度上已被证实正确。this have largely proven correct. 09:58 我想用几分钟时间谈谈第三个预测 I want to spend a few minutes talking about the third one 10:01 因为我觉得它对我们的职业选择 because I think it’s 10:04 至关重要 too important to our career choice 10:05 比如，谁会从生成式 AI 中获益？like who is going to benefit from this AI? 10:08 那么，我们双击最后一个吧。So let’s double click on the last one. 10:10 为什么是大科技公司？Why Big Tech? 10:11 原因在于分布。The reason is distribution. 10:14 我们常常关注它们的企业组织结构、We often look at their organizational structure, 10:18 创新速度、how fast they innovate, 10:19 技术的高端程度、how fancy is their technology, 10:21 以及员工的才华水平。and how talented are their employees. 10:24 但大科技公司的真正优势在于分发能力。But the actual advantage of Big Tech is distribution. 10:27 比如，你知道 Gamma 吗？For example, have you heard of Gamma? 10:30 如果你还没看过 Gamma 的网站 You can give a reaction 10:32 ，可以试试反应一下。if you haven’t checked out Gamma’s website. 10:33 它也是 AIPT。It’s AIPT too. 10:36 像 Gamma 这样的初创公司能开发出优秀的 AI 原生工具，A startup like Gamma can build a great AI native tool, 10:40 但 Google 只需在幻灯片上加个 Gemini 按钮 but Google can just add a Gemini button to the slides 10:44 ，就能瞬间覆盖十亿用户。and reach a billion users instantly. 10:46 这就是分发能力的强大之处 That is the power of distribution 10:48 初创公司很难与之匹敌。and that is very hard for startup to beat. 10:52 比如，我制作了它。For example, I made. 10:54 我用 Karma 重新制作了这个 DAG I remade this DAG with Karma 10:57 这是 Karma 的展示。and this is Karma’s presentation. 10:59 它效果不错 It’s pretty good 10:59 而且速度很快。and it’s very fast. 11:02 是的，它们有很多不同的功能，Yeah, and they have A lot of different features, 11:05 比如把它转换成网站之类的。for example, turning it into a site and other things. 11:08 太棒了。It’s great. 11:09 我是说，它还在赚大钱 I mean, it’s still making a lot of money 11:12 而且评价很好，有优秀投资者的背书。and it has a good evaluation backed by good investors. 11:17 不过，要赶上还需要一段时间 But it’s going to take some time 11:20 ，而且并不安全。to catch up and it’s not safe. 11:23 比如，For example, 11:23 这是我们使用 Codex 生成的内容。this is what we generated with Codex. 11:27 我们为亚马逊进行了一次企业培训，We did an enterprise training, 11:29 其实只是一个简短的培训课。actually a quick session at Amazon. 11:32 你也可以去访问该网站。You can visit the site as well. 11:34 我刚用 Codex 生成了这个栈 I just used Codex to generate this stack 11:38 ，感觉用 Codex 比 Gamma 更有控制力 and I feel like I actually have more control 11:43 ，也更有各种可能性来优化它 and more, I don’t know, more possibility to make it better 11:47 所以我觉得 Gamma 的商业模式 So I don’t think the business model of Gamma 11:51 并不安全。is safe. 11:53 例如，我们用 Claude 生成了一份 pitch 演示文稿 And for example, we use Claude to generate a presentation for a pitch 11:58 另一个例子是用 Claude 生成的另一份企业培训演示文稿。and another example is a Claude-generated presentation for another enterprise training. 12:03 其实，这两个演示都深受我们客户的欢迎，Actually, both of these presentations are well received by our clients, 12:10 所以…… so. 12:12 而且我构建每个 stack 可能只用了不到 30 分钟。And I probably spent less than 30 minutes 12:18 这就回到了代理式工作流如何真正实现复利增长 And that goes back to how agentic workflow can really compound 12:25 因为每次我保存其他信息时，because each time I save other information, 12:29 也保存了我想要呈现信息的新方式。I save other ways that I want to present my information. 12:34 我保存了制作 PowerPoint 的技能 I save the skills of how to produce the PowerPoint 12:37 而且它会产生复合效应。and that compounds. 12:40 好的，这样就结束了 Okay, so that closed the topic of 12:43 ‘确保你看到分布’这个话题。make sure you see distribution. 12:47 当比较两家公司时，When you compare two companies, 12:50 分发能力的威力非常巨大。the power of distribution is quite large. 12:55 现在我们回到主要话题。Now going back to our main topic. 12:57 启动你的思考过程吧。Kickstart your thinking. 12:59 我想分享三个推测。I want to share three conjectures. 13:01 这些是我每周坚持的强烈观点。These are my strong opinions weekly held. 13:04 我的目标是激发你们的思考 My goal is to stimulate your thinking 13:07 ，我请大家 and I ask you to engage with them 13:08 用自己的批判性思维来与之互动。with their own critical thinking. 13:11 我可能有误，请大家质疑我。I could be wrong, so please challenge me. 13:14 但我提出这三个预测或三个承包商 But I lay out these three predictions or three contractors 13:18 ，是因为我觉得它们 because I do feel they are important 13:20 对于理解 AI 的局限性至关重要。in understanding the limitation of AI. 13:24 它们对于我 They are important in my calibration of 13:25 校准未来方向 how should I shape my future, 13:26 、选择职业路径至关重要。what career should I choose? 13:30 好的，第一个观点是：Okay, so the first one is 13:31 幻觉是生成式 AI 的基本特性。hallucination is fundamental. 13:33 第二个是，transformers 可能会带来 AGI。The second is transformers may lead to AGI. 13:36 第三个逆向观点就是我们的优势所在。And the third is contrarian view is our edge. 13:41 那么，对比一下。So contrast. 13:43 幻觉并非一个需要修复的缺陷。Hallucination is not a bug to be fixed. 13:45 这是基本且必要特性 It is a fundamental and necessary feature 13:47 实用且富有创造力的生成式 AI 的。of a useful, creative, generative AI. 13:50 我们常常把幻觉 We often think of hallucination 13:52 想象成像这张图片那样的明显错误。as obvious errors like this image. 13:55 但最危险的幻觉是那些 But the most dangerous hallucination are the ones 13:58 细微且貌似合理的。that are subtle and plausible. 14:00 AI 会生成这些内容，因为它认为 The AI makes them because it believes 14:03 这是它能给出的最佳答案。that is the best answer it can provide. 14:06 那么，这为什么是根本性的呢？And why is this fundamental? 14:08 因为它是不确定性的产物。Because it is a product of uncertainty. 14:11 如果我们消除了所有不确定性，If we remove all uncertainty, 14:13 它就不再是真正的 AI 了—— it wouldn’t really be AI anymore— 14:14 而会变成确定性的系统。it would become deterministic. 14:16 它只会提供 100%正确的答案 It would only give 100% correct answers 14:19 但这些答案往往毫无用处。which are often useless. 14:21 诸如‘做个好人’之类的老生常谈 Truisms like ‘be a good person’ 14:24 ——没错，但毫无用处。—correct, but useless. 14:27 模型要做到富有创造力和实用性，To be creative and useful, 14:29 就必须有犯错的自由。the model must have the freedom to be wrong. 14:33 但这并不意味着 And this doesn’t mean 14:34 我们就得接受它。we just accept it. 14:35 幻觉是可以控制的。Hallucination can be controlled. 14:37 我们在软件中不习惯不确定性，We are not used to uncertainty in software, 14:40 但在日常现实生活中，我们每天都在应对它。but we manage it in real life every day. 14:43 我们通过风险管理来定义成功 We do it by defining success using risk management 14:46 ，并向 AI 工具提供支持，这与我们在模块 3 中所做的类似。and giving AI tools, just like what we did in module 3. 14:53 介绍了众多不同的方法 Introduce so many different approaches 14:54 来控制生成式 AI 的幻觉问题。to control for hallucination. 14:57 而且我想说，尤其在编程相关工作中，And I would argue, especially in coding related work, 15:01 如果 AI 出现幻觉，if you find AI hallucinations, 15:03 往往是因为你没有提供正确的上下文 it’s often because you didn’t give it the correct context 15:07 或者没有足够精确地整理上下文。or didn’t curate the context precisely enough. 15:11 同样值得一提的是，It’s also worth noting that 15:12 我们能察觉到幻觉是因为我们知道正确答案，we can spot hallucination because we know the correct answer, 15:16 但 AI 不知道。but it doesn’t. 15:17 所以这表明问题出在 So that’s why it points to a gap 15:19 上下文管理上的缺失，而不是生成式 AI 的 in context management rather than the AI’s capability 15:24 幻觉能力本身。with hallucination. 15:26 我还想提醒大家 I also want to call our attention 15:28 关注这个新的可能性。to the new possibility. 15:32 因此，传统软件要求输入精确、输出精确。So traditional software requires exact input, exact output. 15:39 输入中只要有一个错误，If there is a single error in the input, 15:41 整个链就会崩溃。the entire chain breaks. 15:44 我和 Chen 来介绍模糊输入。Chen and I introduce fuzzy input. 15:47 输出模糊。Fuzzy output. 15:49 模糊输入构成了问题 The fuzzy input is a problem 15:52 对我们当前的精确生态系统 for our current exact ecosystem 15:55 因为我们常常将 JNI because we often put JNI 15:58 集成到现有的软件生态中。in the existing software ecosystem. 16:02 而其他部分则需要精确输入。And the other parts requires exact input. 16:06 但是，模糊输入—— But the fuzzy input, 16:08 理解大型请求或绘图消息的能力—— the ability to understand a big request or a message drawing, 16:13 才是真正的革命。that is the revolution. 16:15 随着时间的推移，我们的生态系统将逐步演化 Over time our ecosystem will evolve 16:18 来使用这项技术。to use this. 16:21 比如，Like for example, 16:21 Yan 是一位计算机视觉的博士。Yan was a computer vision PhD. 16:26 大概得花他好几个月的时间，我也不知道，Probably would take him, I don’t know, 16:28 来训练这样的一个模型。months to train a model like this. 16:30 即使使用 4.0 版本，And even with 4.0, 16:32 我只需画出这个，I can just draw this 16:33 模型就能直接给出答案。and the model can give me an answer. 16:36 这就是模糊输入的强大之处 That is the power of that fuzzy input 16:39 ，但它被严重低估，因为整个生态系统 and which is underutilized because the entire ecosystem 16:43 并非所谓的 AI 原生。is not so called AI native. 16:46 它还是旧生态系统 It is still the old ecosystem 16:48 ，我们正在逐步改变它。and we are gradually changing it. 16:51 但随着时间推移，我们的生态系统会发生变化 But over time our ecosystem will change 16:54 从而开启许多新的可能性。and will unlock a lot of new possibilities. 16:58 因此，我们第一个猜想的总结就是 So the summary for our first conjecture is 17:01 幻觉是根本性的。hallucination is fundamental. 17:02 我们必须学会适应它 We must learn to live with it 17:04 ，更重要的是，要主动加以管理。and, more importantly, proactively manage it. 17:09 那是第一个。That’s the first one. 17:10 目前有什么问题或讨论吗？Any question or discussion so far? 17:17 好的，我们来看第二个猜想。Okay, I’ll go to the second conjecture. 17:19 第二个猜想是最具争议性的。The second conjecture is the most controversial. 17:23 仅仅基于简单下一个标记 The transformer architecture based on simple next hook 17:25 预测的 Transformer 架构，就可能是 and prediction may be all we need 17:28 实现人工通用智能所需的一切。to achieve artificial general intelligence. 17:31 我觉得这是核心论据 And I think this is the key argument 17:34 Le Quince 理论的。of the Le Quince sense. 17:37 是吧？Right? 17:37 如同下一标记预测无法实现人工通用智能。Like next token prediction cannot lead to artificial general intelligence. 17:43 看到弹出的聊天框了吗？See the chat popping up? 17:47 哦，Transformer 是模型架构 Oh, transformer is the model architecture 17:50 我们生成式 AI Yan 的。of our generative AI Yan. 17:53 你想在聊天里给出更详细的回答吗？Do you want to give a more detailed answer in the chat? 17:56 嗯，没问题。Yeah, sure. 17:58 好的。Okay. 18:00 我觉得咱们课上有章节解释这个 And I believe we have chapters explaining this 18:03 ，对不对？in our lesson, right? 18:05 嗯嗯。Yeah, yeah. 18:08 这就是 Transformer 的结构图。So this is the picture of the transformer. 18:10 是的，基本上这就是支撑 AI、Yeah, Basically this is what enables AI, 18:14 支撑当前生成式 AI 的核心技术。enables the current generative AI. 18:20 所以不必深究细节，So without going into the details, 18:22 你只需知道这是我们生成式 AI 的核心架构。just know this is the fundamental architecture 18:29 我猜这可能是巧合，也可能不是巧合 I guess coincidentally or maybe not coincidentally 18:33 AI 很擅长转换任务。AI is good at transformation tasks. 18:36 要理解为什么 Transformer 可能通往 AGI，To understand this why transformer may lead to AGI, 18:40 我们首先需要了解 AI 在哪些转换任务上表现出色。we first need to see what AI is good at transformation tasks. 18:44 想想翻译、编程或摘要这些任务。Think translation, coding or summarization. 18:47 在所有这些例子中，In all these cases, 18:49 底层知识相同，the underlying knowledge is the same, 18:51 但表达方式不同。but expression is different. 18:55 转变的概念至关重要。This concept of transformation is key. 18:58 这正是程序员的工作。This is what coders do. 18:59 例如，For example, 19:00 它们能将客户的请求 they transform a client request 19:01 转化为可运行的代码。into working code. 19:04 这是 AI 正变得极为擅长的任务。This is a task AI is getting very, very good at. 19:10 这个论点 Transformer Melito AGI 有三个层面。This argument transformer Melito AGI has three layers. 19:15 所以请注意。So pay attention. 19:17 我即将开始。I’m going to. 19:18 每一层都会建立在上一层的基础上。Each layer is going to build on top of the other. 19:21 第一层：什么是知识？Layer number one, what is knowledge? 19:24 知识并不只是孤立的信息。Knowledge is not just isolated information. 19:27 它就是模式匹配，串联起各个点。It’s pattern matching, connecting the dots. 19:32 关于这些点的那些信息。Information on the dots. 19:34 知识就是模式匹配，Knowledge is pattern matching, 19:35 将点连接起来。connecting the dots. 19:38 第二层，这里开始有趣了。Layer number two, this is where it gets interesting. 19:43 所以左侧是 Elias Zuska 的截屏。So on the left is a screenshot of Elias Zuska. 19:48 我总是发不好他的姓。I can never pronounce his last name. 19:53 2016 年在 Lex Friedman 的课上做过一次讲座 Did a lecture in Lex Friedman’s class in 2016 19:59 那时远早于 AGI 和他的讲座。way before AGI and his lecture. 20:02 这个幻灯片探讨了神经网络为什么能够工作的根本原因？This slide is why do neural networks work at all? 20:06 而且这并不明显 And it’s not really obvious 20:07 我们常常想当然地认为 or we always take it for granted 20:10 我们拥有这个高级模型，它就能奏效。that we have this fancy model and it works. 20:13 但如果你提出这个最根本的问题 But if you ask the really fundamental question 20:16 ：为什么它能成功？why did it work? 20:18 他的回答是：And his answer is 20:20 最短的、能够拟合数据的程序 this is because the shortest program that fits the data 20:24 就是最好的泛化。is the best generalization. 20:27 这就是一个数学定理。And that is a mathematical theorem. 20:30 最短的程序，能够拟合数据 The shortest program that fits the data 20:32 就是最好的泛化。is the best generalization. 20:34 LLMs 在训练过程中，As LLMs train, 20:37 本质上是寻找最短的程序，they are essentially finding the shortest program, 20:40 将数据压缩为最具泛化能力的模式。compressing the data into the most generalizable patterns. 20:45 所以，那就是第二层。So that’s layer number two. 20:48 看看，这两者之间的相似性，我也不知道。And see the, I don’t know 20:52 我们如何相互连接并将其转化为知识 the similarity between how we connect us and turn that into knowledge. 20:56 比如，如果你找到一个更短的程序，Like if you find a shorter program, 20:59 它就更具泛化性。it is more generalizable. 21:00 它能够跨不同学科和不同问题进行泛化。It’s generalizable across different disciplines and different problems. 21:08 现在是第三层。Now is layer three. 21:10 第三层是最反直觉的，Layer three is the most counterintuitive, 21:12 至少在我最初看来是这样。at least to me at the beginning. 21:17 这个压缩信息的过程可能是在 This process of compressing information may be 21:22 阐述我们人类的学习方式。enumerating how we learn, how we human learn. 21:27 我们认为人类会推理 We believe humans reason 21:29 但如果不是这样呢？but what if we don’t? 21:32 教育研究显示，学习的关键 Research in education shows the key factor in learning 21:35 在于先前的知识，而不是抽象的推理能力。is prior knowledge, not an abstract reasoning skill. 21:41 假如我们是在压缩信息 What if we are compressing information 21:44 并进行泛化呢？and generalizing it? 21:46 是在泛化信息吗？Generalizing the information? 21:47 我们到底是在真正推理 Are we actually reasoning 21:50 还是只是利用压缩的信息 or are we actually just taking the compressed information 21:53 来尝试泛化？and try to generalize it? 21:56 我建议你看看这个简短访谈 I recommend you to look at this very short interview 21:59 与 Hinton 的。with Hinton. 22:02 他谈到，模型的智能 He talks about how intelligence of the model 22:05 是在训练阶段产生的，而不是在测试 happens at training, not at testing, 22:09 或推理阶段。not at inference. 22:11 我想，这就是为什么 This is I guess why it’s important 22:13 在我们与 ChatGPT 对话时，when we talk with ChatGPT, 22:16 理解 ChatGPT 的推理过程很重要。the ChatGPT reasons. 22:19 我们也有这个最新的模型 We also have this recent model 22:21 ChatGPT 说，and ChatGPT says, 22:23 我在想这个，我在想那个。I’m thinking about this, I’m thinking about that. 22:24 我要给你一个答案。I’m going to give you an answer. 22:26 我们觉得智能是在 We feel like intelligence is happening 22:28 推理阶段产生的。at the inference stage. 22:30 但在 Hinton 看来，智能发生在训练阶段，But in Hinton’s view, intelligence happened at the training stage, 22:35 即压缩阶段，at the compressing stage, 22:36 而模型只是进行基本的机械操作 and the model is just doing basic mechanical stuff 22:39 或预测下一个 token 而已。or predicting next token. 22:43 但智能早已存在于模型之中。But intelligence is already in the model. 22:46 如果你想深入了解 Hinton 的观点，If you want to learn more about Hinton’s view, 22:49 我建议你看看 1978 年的内容。I recommend you to go to 1978. 22:52 他发表了一篇关于快速权重的论文 He published a paper about fast weight 22:55 探讨 AI 如何寻找事物之间的关系。that AI is looking for relationship among things. 23:02 好，我们回到那三层，对吗？Okay, so going back to the three layers, right? 23:06 模型如何运作、人类如何运作 How models work, how human work 23:10 ，以及存在一个数学定理 and how there is a mathematical theorem 23:13 ，即最短程序最具泛化性 of shortest program is the most generalizable 23:17 ，从而使通过压缩生成智能成为可能，我猜。that makes it possible for compressing to generate intelligence, I guess. 23:23 好吧，如果我们人类只是生物计算机 Okay, so what if we are just biological computers 23:26 ，也在进行下一个标记的预测呢？also running the next token prediction? 23:30 如果是这样，正如 Ilya 所说，If so, as Ilya says, 23:32 数字广域网（WAN）没有根本原因无法与之匹敌。there is no fundamental reason a digital WAN cannot match it. 23:35 如果我们仅仅是大脑计算机、数字计算机，If we are just biological computers, digital computers, 23:39 那么数字计算机就没有根本理由无法匹敌我们的大脑。there is just no fundamental reason a digital computer cannot match our brain. 23:45 实际上，数字计算机可能会变得更强大。In fact, a digital computer may become more powerful. 23:48 我们大脑的处理能力 Our brain are limited in processing power 23:49 和数据处理能力都是有限的。and data. 23:50 Transformer 并不是。A transformer is not. 23:52 这可能只是时间早晚的问题。It may just be a matter of time. 23:55 我猜，如果你去问生成式 AI 前沿的研究者，I guess if you ask AI frontier researchers, 24:00 他们好像都持这个看法。they all seem to hold this view. 24:02 生物计算机与数字计算机之间并无根本区别。There is just no fundamental difference between a biological computer and a digital computer. 24:09 总之，AI 擅长于转换。So to summarize, AI excels at transformation. 24:13 有可能所有智能都仅仅是某种变换 It’s possible all intelligence is just transformation 24:17 我们对自身大脑的了解还太少 and we simply know too little about our own brains 24:20 无法排除下一个 token 预测就是通往 AGI 的途径。to dismiss the possibility that next token prediction is the path to AGI. 24:30 右边这张图片 On the right is a picture 24:31 来自这个由 bug read by 的图片。from this bug read by. 24:36 我觉得这篇东西 And I think this was published 24:37 远在 ChatGPT 出现之前就发表了。way before ChatGPT. 24:40 观察结果是，And the observation is there is a singularity 24:42 机器智能将超越人类智能，出现一个奇点。of machine intelligence crossing human intelligence. 24:48 爱因斯坦与最愚笨的人类 And the difference between Einstein and the dumbest human 24:52 在智力上的差距其实很小。in terms of intelligence is actually not much. 24:57 但是人类与黑猩猩之间的智力差异 But the intelligence difference between human and chimp 25:00 、鸟类之间的实际上非常大。and bird and our actually very large. 25:03 因此，我们看到 So we observed that 25:05 ChatGPT 在 2023 年已经超过了愚蠢人类水平的智能。ChatGPT surpassed dumb human-level intelligence 25:11 但 ChatGPT 超越人类中最聪明者的智能，只需 But it would just take a very short time 25:16 极短的时间。for ChatGPT to surpass the intelligence of the smartest human being. 25:18 我觉得这已经发生了。And I feel like it already happened. 25:22 我记得 2025 年初，I remember at the beginning of 2025, 25:24 Sam Altman 曾说，我们好像已经见证了奇点，Sam Altman says, like we have seen the singularity 25:30 但不确定自己处在哪一边。but unsure which side we are on. 25:32 而且我认为我们处于正确的一方，And I think we are on the right side, 25:35 我们已经见证了奇点的到来，which is we saw singularity already happened 25:37 机器智能——也就是 AI——已经比人类更聪明了。and machine intelligence already AI is already Smarter than human. 25:42 我们目前还无法理解它。We just cannot understand it yet. 25:45 到目前为止的问题是 And the question so far 25:47 如果你之前不怕 AI 取代我们，and if you weren’t afraid of AI replacing us, 25:50 现在害怕了吗？are you afraid now? 26:01 哦，Jeremy 有问题要问。Oh, Jeremy has a question. 26:04 我在训练过程中很难理解 I have some hard time understanding 26:06 那个最短程序 the shortest program during training 26:08 参数化优化参数。where optimizing parameters parametrically. 26:10 我怎样来理解这个概念 How could I understand this concept 26:12 从模型训练和架构设计的角度？in terms of model training and architecture design? 26:15 其实我有一个 Yuan Dong Tian 的采访视频 So actually have an interview with Yuan Dong Tian 26:22 是用中文录制的。and he it’s in Chinese. 26:25 如果你看不懂中文，If you don’t read Chinese, 26:27 我也可以顺便翻译成英文。I can translate it into English as well. 26:29 我其实把这个论点给他看了 I actually showed him this argument 26:33 我们还讨论了一下。and we had a discussion. 26:37 他的观察即是损失函数。His observation is the loss function. 26:41 根据他训练 AI 的实际经验 In his empirical experience in training AI 26:43 损失函数其实并没有那么重要。the loss function actually doesn’t matter that much. 26:46 他能用多种不同的损失函数 He can use many different loss functions 26:49 获得同样的结果。to get the same result. 26:52 而且这些模型—— And it seemed that the models, 26:54 这些机器学习模型—— the machine learning models 26:56 似乎本质上是在追求平滑、优雅或精致。intrinsically are optimizing for smoothness or gracefulness or elegance. 27:02 他认为 And he believed that 27:03 这实际上能让许多模型 that actually makes a lot of models 27:06 通过缩小变得更智能。shrink to becoming smarter. 27:10 那就是刚才说的。So that was the. 27:12 那就是那个观察结果。That was the observation. 27:14 那么，为了回答我们的问题，So to answer our question, 27:16 如果模型…… if the model. 27:17 如果模型本质上只是试图找到最短路径 If intrinsically the model is just trying to find the shortest path 27:22 即使你给它不同的损失函数，它仍然只找到最短路径 and if you give it different loss function and it just finds the shortest path 27:29 那么参数就没那么重要了。then the parameters do not matter that much. 27:33 归根结底，数据规模才是关键。In the end, the size of the data matter. 27:37 不过没错，我想那就是一条 But yeah, I guess that is one rule 27:40 把它们排除掉的规则。that rules them out. 27:42 但这终究只是猜想而已。But again, conjectures. 27:44 所以，请挑战我 So please challenge me 27:46 ，并培养你自己的推理能力。and develop your own reasoning. 27:52 我唯一的结论，或者说 My only conclusion here is or something 27:55 我确信的一点是 that I know for sure is 27:57 我们对自身的智能了解得太少。we know too little about our intelligence. 28:01 我们对大脑的了解还远远不够 We don’t know enough about our brain 28:04 ，因此不能排除数字计算机超越人类智能的可能性。to dismiss the possibility that a digital computer can surpass our intelligence. 28:10 所以要保持谦逊 So stay humble 28:12 ，不要把我们当成什么特别的东西。and don’t treat us as something special. 28:18 酷。Cool. 28:18 目前为止有什么问题吗？Any questions so far? 28:21 但即便有了这样的认识，But even with this understanding, 28:23 我依然坚信 AI 无法取代我们。I’m still confident that AI cannot replace us. 28:26 所以我的第三个猜想就此宣告失败。So there goes my conjecture number three. 28:31 如果 AI 能胜过我们的智慧，If AI can outsmart us, 28:33 我们还能做什么？what is left for us? 28:34 我的答案是，生成式 AI 无法取代 My answer is this generative cannot replace 28:37 真正颠覆性的想法。truly contrarian ideas. 28:40 其实这更多是社会和经济层面的，而不是技术层面的。And that is actually social and economical rather than technological. 28:47 什么叫反其道而行之的想法？What is a contrarian idea? 28:49 这不是单纯为了反对而反对。It is not just objecting for the sake of it. 28:53 Neville 说过，As Neville says, 28:56 逆向思维者从底层独立推理，a contrarian reasons independently from the ground up 28:59 不屈从于从众压力。and resist pressure to conform. 29:02 他们了解共识性的最佳实践，They understand the best practice which is the consensus, 29:06 同时还具备额外知识，明白其可能出错的原因。but have additional knowledge of why it might be wrong. 29:12 我们在课程里做过这个。We did this in our course. 29:14 例如，过去的最佳实践是 For example, the best practice used to be 29:18 教授 rag、流行框架以及 mcp。is to teach rag and popular frameworks and mcp. 29:24 比如，如果你去 Maven 网站，Like if you go to Maven, 29:26 会发现很多课程都强调 you’ll see a lot of courses emphasizing 29:28 要教授这些概念。we are going to teach these concepts. 29:29 我们独立思考后决定不这么做 We reasoned independently and chose not to 29:32 因为我们认为这会降低你们作为构建者的效率。because we believe it makes you less effective builders. 29:35 你在我们的办公时间也见过 Jens 关于 MCP You have seen Jens argument about MCP 29:38 的论据。during our Office hour too. 29:39 这需要对基础原理有深刻的理解 It takes a deep understanding of the fundamentals 29:42 那就很难做到 and it’s very hard to do 29:44 如果你不明白为什么最佳实践是错误的 if you don’t have the deep understanding of why the best practice is wrong. 29:48 你会感受到直面形式的压力。You have this pressure to confront form. 29:50 我们能抵御这种压力 We are able to resist the pressure 29:52 ，是因为它需要对基础原理有深刻的理解 because it requires the deep understanding of the fundamentals 29:55 ，而且做起来真的非常困难。and it is really hard to do. 30:00 而这一点非常重要。And this one is important. 30:02 逆向思维就是要发现别人忽略的模式。Contrarian thinking is about seeing patterns others miss. 30:06 这意味着以全新方式连接各种点。It means connecting dots in a new way. 30:09 逆向思考并不是创造新信息。Contrarian thinking is not inventing new information. 30:14 它是在发现新信息 It’s discovering new information 30:17 ，或者探索连接这些点的新方法。or covering new ways to connect the dots. 30:21 AI 并非无法生成奇特想法，AI struggles with this not because it cannot generate unusual ideas, 30:26 而是人们往往分不清 but because people cannot always tell the difference 30:28 大胆的洞见与糟糕的观点，这才是它在这方面遇到的难题。between a bold insight and a bad take. 30:37 这使得训练一个模型 That makes it risky to train a model 30:39 变得真正反传统非常冒险。to be truly contrarian. 30:43 强烈的对齐机制往往会过滤掉那些 Strong alignment tends to filter out ideas 30:46 看似古怪或不安全的 that looks odd or unsafe, 30:48 想法，即便其中有些最终可能是正确的。even if some of them might turn out to be right. 30:52 这就是 AI 通过伽利略测试如此困难的原因。That’s why passing a Galileo test is so hard for AI. 30:56 我记得埃隆·马斯克 I remember Elon Musk 30:58 大约几周或一个月前在 Twitter 上发了一条推文。published his Twitter a couple weeks a month ago. 31:03 他基本上认为，AI 不仅需要通过 Gallio 测试，He basically says AI needs to pass the Gallio, 31:05 也就是 Galli arrow 测试，还必须通过图灵测试。not only the Turing test, but the Galli arrow test. 31:08 这意味着，即使所有人都说你错了，That means when everybody says you are wrong, 31:12 你也要坚信自己是正确的。you need to stay convinced that you are correct. 31:16 但这对 AI 来说非常困难 But it’s very hard for AI to do 31:18 因为是人类来对齐 AI 的。because it’s human that aligns AI. 31:21 除了这个对齐问题之外，还有另一个问题。There is another issue on top of this alignment issue. 31:24 AI 没有个人利害关系。AI doesn’t have a personal stake. 31:27 它没有身份，没有切身利害关系，It has no identity, no skin in the game, 31:31 也没有真正后果要承担。and no real consequences to face. 31:34 因此，它无法表现出信念。Because of that, it cannot show conviction. 31:38 而且没有信念，And without conviction, 31:39 反传统观点就很难被信任或付诸行动。contrarian ideas are hard to trust or act on. 31:42 在实践中，大胆的想法通常需要 In practice, bold ideas usually need 31:45 有人愿意支持它们 someone willing to stand behind them 31:47 并坚持到底。and see them through. 31:49 我可以给你 10 个不同的反共识观点 I can give you 10 different contrarian ideas 31:52 如果我说我相信它，and if I tell you I believe in it, 31:53 但自己却没有实际投入，but I have no skin in the game, 31:55 你就不会相信我。you won’t trust me. 31:57 因为反向想法本质上就是如此。Because the contrarian ideas are fundamentally. 31:59 反向想法的定义 The definition of contrarian idea 32:01 很难仅凭推理就相信。is hard to just reason and trust. 32:05 需要有人敢于冒险 It takes someone to risk the game 32:07 支持他们并陪他们走到底。to stand behind them and see them through. 32:11 这些观点本质上具有根本性价值。These views are fundamentally. 32:13 逆向观点本质上非常宝贵 The contrarian views are fundamentally valuable 32:15 因为共识已被计入价格。because consensus is priced in. 32:18 每个人都已经在这么做了。Everyone is already doing it. 32:20 比如说，Like for example, 32:21 在股市中。in stock market. 32:22 如果大家都认同，If it’s a consensus, 32:23 就不会有利润。there is no profit. 32:24 唯一的利润来自于正确的反向观点。The profit of only comes from correct contrarian views. 32:29 正确的反共识观点是抓住宝贵机会的必要前提。A correct contrarian view is the necessary condition 32:34 关于这一点最好的引述出自 1997 年的苹果公司。The best quote for this came from Apple in 1997. 32:39 我来播放这个简短的视频。I’ll just play the quick video. 32:44 敬那些疯狂的一群人。Here’s to the crazy ones. 32:46 那些格格不入的人，那些叛逆者，The misfits, the rebels, 32:50 那些惹事生非的人，the troublemakers, 32:52 那些方孔里的圆钉。the round pegs in the square holes. 32:56 那些看待事物与众不同的人。The ones who see things differently. 32:59 他们不喜欢遵守规则 They’re not fond of rules 33:01 对现状毫无敬意。and they have no respect for the status quo. 33:04 你可以引用他们、反对他们、You could quote them, disagree with them, 33:07 美化或诋毁他们。glorify or vilify them. 33:10 你唯一做不到的事就是忽略它们 About the only thing you can’t do is ignore them 33:14 因为它们改变了世界，推动人类不断进步。because they change things, push the human race forward. 33:21 虽然有些人认为他们是疯子，While some may see them as the crazy ones, 33:23 但我们看到了他们的天才。we see genius. 33:27 那些足够疯狂 Because the people who are crazy enough 33:29 、相信自己能改变世界 to think they can change the world 33:31 的人，正是真正做到的人。are the ones who do. 33:43 没错，那些疯狂的人，Yep, the crazy ones, 33:45 那些相信自己能改变世界的人，the ones who think they can change the world 33:47 正是那些真正做到的人。are the ones who do. 33:49 这就把我们带回了起点。And this brings us full circle. 33:52 你怎样抓住价值 How do you capture the value 33:53 你那个与众不同的想法的？of your contrarian idea? 33:56 你要通过假设、预测和证伪 You hypothesize, predict and falsify 33:58 ，来找到那个反传统的观点。to find the contrarian idea. 34:01 一旦你做了足够多，Once you done this enough, 34:03 就会 once you’ve done enough, 34:04 每周持有强烈的意见。strong opinions weekly held. 34:06 你内化这个反主流的想法 You internalize the contrarian idea 34:08 从而建立起坚定的信念。to build conviction. 34:10 然后，因为你的想法会遭到反对，And then because your idea will be opposed to, 34:13 唯一证明它的方法就是动手构建。the only one to prove it is to build. 34:18 这就是一个完整的循环 So that is the full circle 34:20 你会收到印有 swag 的衣服，and you’ll receive claiming swag, 34:23 我现在就穿着 aibuilders.com，which I’m wearing aibuilders.com 34:28 正面是这样，背面是这个。and on the back is this. 34:30 敬那些疯狂的人。Here’s to the crazy ones. 34:32 让这个点实现吧。Make the dot happen. 34:33 这就是我们说 That’s why we say 34:34 主页是 AI Builders 的原因。our homepage is AI Builders. 34:37 构建过程让核心理念得以实现 The build makes the controlling ideas happen 34:40 这就是捕捉价值的方法。and that’s how you capture the value. 34:43 你可享 100%折扣并包邮，You get 100% discount including free shipping, 34:47 但每位学生限一个。but one per student. 34:52 好，我们已经到了终点，All right, so we are at the finish line, 34:54 但真正的工作现在才刚刚开始。but the real works actually starts now. 34:57 你已经掌握了知识 You have the knowledge 34:58 ，这里有三个具体的 and here are the three concrete steps 35:00 立即行动步骤来巩固它。to take immediately to lock it in. 35:04 我已经发邮件邀请你加入课程了 I already sent the email to invite you to the course 35:08 课程结束时你还会收到一封反馈邮件。and you will also receive an email at the end of this course for reviews. 35:13 我有一个最后的私人请求想拜托你。I have one final personal favor to ask from you. 35:17 请用两分钟时间给我们留个评价。Please take two minutes to leave a review. 35:19 但我希望它有用 But I want it to be useful 35:20 既对你们有用，也对我们有用。both for you and for us. 35:22 所以，对你而言，记下关键收获 So for you, writing down a key takeaway 35:25 是巩固刚学知识的最佳方式。is the best way to solidify what you just learned. 35:28 它把信息转化为知识 It turns information into knowledge 35:30 而我们则靠反馈来决定生死。and for us we leave our die by our feedback. 35:33 我们宁愿靠反馈活下去，也不愿死去。Rather live all die by our feedback. 35:35 因为正如你看到的 Because as you can see 35:36 我们不教名词，we don’t teach nouns, 35:37 而是教动词。we teach the verb. 35:38 而且很难 And it’s very hard to explain 35:39 向那些没上过我们课的人 to someone who never took the course 35:41 解释清楚这门课到底讲些什么。what our course is actually about. 35:44 我们最有力的营销素材就是学生们的反馈 Our strongest marketing material is our students review 35:48 这有助于我们迭代 so it help us iterate 35:49 改进，也帮助社区中的其他人培养更多建造者。and help others build other builders from this community. 35:55 注意，在 Maven 这样的平台上 One notice that on platforms like Maven 35:57 评分系统几乎是二进制的。the rating system is binary almost. 36:00 如果这门课对你有价值，If this course gives you a value, 36:02 给 10 分满分就是最佳支持方式。a 10 out of 10 is the single best way to support it. 36:05 如果不是 10 分，If it wasn’t 10, 36:06 请直接给我们发私信。please DM us directly. 36:08 我想把它修好 I want to fix it 36:09 而不是仅仅拿个差评。not just get a bad grade. 36:11 所以这就是我个人的 Favor。So that’s my personal Favor. 36:16 第二步，不要独自一人开发。And step number two, don’t build alone. 36:19 你现在可以终身访问 AI Builders 社区。You now have lifetime access to the AI Builders community. 36:22 我们有最新更新、We have updates, 36:24 新闻资讯 we have news, 36:24 和知识库。we have knowledge bank. 36:25 你可以分享项目 You can share projects 36:27 ，这是保持前沿的一种方式 and this is a way to stay current 36:30 在生成式 AI 这一快速发展的领域。in this fast moving field. 36:34 我们会随时向你通报最新消息。And we’ll keep you posted. 36:36 请在这里分享我们的项目 Share our projects here 36:37 在 Product Hunt 之前。before Product Hunt. 36:39 反馈更好 The feedback is better 36:39 支持也是真实的，在 Product Hunt 上。and the support is real in Product Hunt. 36:42 我觉得，如果你能登上首页 I feel like if you ever make it to the front page 36:46 比如说，on Product Hunt, for example, 36:47 你每天大概能获得几十个访问量。you probably get dozens of visits per day. 36:51 但如果你在这个社区分享项目，But if you share your projects in this community, 36:54 它就会被分享并发送到收件箱 your project is shared and sent to the inboxes 36:57 所有社区成员的。of all community members. 36:59 所以这是一个非常宝贵的平台 So it’s a very valuable platform 37:02 提供终生的问答服务。and lifetime Q and A. 37:04 你们是不是都向 Yan 提问题？You all ask questions to Yan, right? 37:06 你看他思考得多么周到 You see how thoughtful 37:08 对这个领域理解得多么深刻。and how deep he understands the field. 37:10 抓住这个机会 Take advantage of this 37:12 继续提问 and keep asking the question 37:12 并在社区中持续与我们互动。and keep engaging with us in this community. 37:17 第三步，压轴项目。And step number three, the capstone. 37:19 这是一个家庭作业 This is a homework 37:20 这是一个作品集。this is a portfolio. 37:22 证明你会构建的唯一办法 The only way to prove you can build 37:23 就是动手构建。is to build. 37:25 把你所学运用起来，Take what you learned, 37:26 打造出产品，向世界展示吧。ship something and show the world. 37:28 我们以最大化你的成功为使命。We make it a mission to maximize your success. 37:31 构建、分享，我们会帮你放大你的声音。Build, share and we’ll amplify your message. 37:36 我其实就在两天前写了这篇文章 I actually wrote this article recently 37:40 用来将 AI 产品分为六个层级。like two days ago to classify AI products into six layers. 37:45 我们采用 AI 辅助构建方式 We have the AI assisted building 37:47 我们的许多项目都是 AI 辅助构建的。and a lot of our projects are AI assisted building. 37:51 但是无论你是想开发新产品、But whether you want to build a new product, 37:55 让初创公司做个副项目，还是构建点东西 have a startup do a side project or build something 38:00 放简历上给潜在雇主看，to put on a resume to show potential employers 38:04 你都想打造一个 AI 产品。you want to build an AI product. 38:08 区别在于，现在 UI The difference is now UI is 38:09 实际由 AI 运行时控制了多少工作量。how much work AI runtime actually controls. 38:13 我相信这门课，I believe this course, 38:15 的直播环节会让你很好地 the live session will prepare you well 38:18 准备好前四层。for the first four layers. 38:21 而顶石项目旨在帮助你 And the capstone is designed to help you 38:24 攻克最后两层。crack through the last two layers. 38:26 增强型核心与 AI 系统，Augmented core and AI system 38:28 具备内存权限、事件和主动触发功能。with memory permissions, events and proactive triggers. 38:33 其实让我 It’s actually let me just 38:35 快速给你们演示一下。to show you very quickly. 38:45 比如构建增强核心项目 Like building augmented core building projects 38:49 ，例如如何降低应用摩擦、and for example how to reduce friction application, 38:53 如何建立持久记忆、how to build a long lasting memory 38:56 如何实现上下文智能、and how to do contextual intelligence 38:59 以及如何打造主动性。and how to make a proactive. 39:03 很多学生会这样 A lot of students follow through like this 39:04 就好像这个设计就是为了让你们按部就班地做。is designed for you to just follow step by step. 39:09 到这个结束时，By the end of this 39:10 你将拥有自己的个性化系统。you’ll have your personalized. 39:13 你找出个人需求 You identify a personal demand 39:15 ，用通用方法构建通用系统 and use the general method to build a generic system 39:18 来满足它。to solve your demand. 39:20 这样，你简历上就会有东西可写 And you will have something to show for on your resume 39:24 很可能会给未来的招聘经理留下好印象。and it will impress your future hiring manager most likely. 39:31 好的，参加 live 能让你达到 L4 水平 Okay, so live gets you to L4 39:33 building 能让你达到 L6 水平。building gets you to L6. 39:35 我们的直播课程旨在激发你的动力，Our live session is designed to give you the motivation, 39:38 帮助你深入理解，提供所有正确的知识，give you the understanding, give you all the correct understanding 39:42 让你能高效构建，并学会充分利用课程内容。so you can build effectively and know how to leverage our course content. 39:47 但是《The Architect》章节，But The Architect chapter, 39:51 也就是 which by that I mean 39:51 顶石课程，会带你达到 L4、L6，the capstone gets you to L4, L6, 39:55 以及核心和 AI 系统。the core and AI systems. 39:58 所以请继续构建 post cell 项目 So keep building post cell projects 39:59 在 Sharing projects 和 Student portal 里。in Sharing projects and Student portal. 40:02 如果超过五名学生在 Maven 上发帖，If more than five students post on Maven, 40:07 我们将为这个小组举办专属办公时间 we’ll host a dedicated office hour for this cohort 40:11 帮助你们审视项目、提供建议 to help you critique the projects and just give you advice 40:14 娱乐一下或解答疑问。or just have fun or answer questions. 40:21 谢谢，Izzy。Thank you, Izzy. 40:22 我还有最后一个问题。I have one last question. 40:24 我觉得课程传达了 I suppose I feel like the course is telling us 40:26 一个观点，即我们能做的大多数事情 the idea that a lot of things we can do 40:29 都可以通过代理来实现，比如 Codex 之类的。can be done through agents, like Codex, and so on. 40:33 那么，在什么情况下，你还会选择 So in what situation would you still choose 40:36 使用 AI 生产力工具如 NotebookLM 或 Gamma to do something through AI productivity tools like NotebookLM or Gamma 40:41 ，你刚提到的，而不是用 Agent 来做这件事呢？that you just mentioned instead of doing it through Agent? 40:47 对我而言，几乎从来没有。For me, almost never. 40:49 从来没有。Never. 40:50 所以你总是只通过 Agent 来做。So you always just do it through Agent. 40:55 我认为这取决于具体的规则。I guess it goes to the specific rules. 40:58 我以前经常用 NotebookLM 和 Gamma，I used to use a lot of NotebookLM and Gamma, 41:03 但现在我觉得 Codex 或 Cursor 在其他场景下更好用。but now I feel like Codex or Cursor is just better for other use cases. 41:12 如果我想用一些特定平台的功能，If I want some platform specific features, 41:15 比如它们提供的东西是我很难在本地 like if they offer something that is very hard for me to replicate locally 41:19 或用通用工具复现的，我就会用那个功能。or using general tools, I will use that feature. 41:24 但我就是喜欢 But I just like 41:25 我没有那个需求。I don’t have the. 41:27 我觉得我并没有那样的需求。I guess I don’t have the demands. 41:29 我相信其他人也有这样的需求。I’m sure other people have the demand. 41:31 比如说，Notebook LLM Like for example Notebook LLM 41:33 可以让你轻松生成一个播客。you can easily generate a podcast. 41:35 哦，对了，抱歉，是我的麦克风问题。Oh yeah, sorry, my mic here. 41:38 真可爱啊。That’s cute. 41:39 嗯，Notebook 有很多花里胡哨的功能 Yeah, Notebook, they have a lot of fancy features 41:45 给用户用的，但我都不用那些。for people who use it, but I don’t use those features. 41:48 而且如果用 NotebookLM 来综合信息 And If I use NotebookLM to synthesize information 41:53 我总能获得更好的效果和更强的能力。I can always get better results and more capability. 41:57 例如，使用 Codex。Using Codex, for example. 42:00 说得有道理。Makes sense. 42:01 谢谢。Thank you. 42:01 我之所以提出那个问题，是因为 Because the reason why I raised that question is because 42:04 我想试用 NotebookLM 和普通的 ChatGPT 来生成幻灯片，I was trying to explore generating slides with NotebookLM and just ChatGPT, the general one, 42:10 但效果并不理想。but it doesn’t work really well. 42:11 不过，当我看到你 But then when I saw your example 42:13 之前提到的那个例子——你 you brought up earlier that all those slides 42:15 用 Codex 生成的所有那些幻灯片时，you generated with Codex, that just triggered me 42:18 这让我突然想，也许我也可以试试用 Codex 来做这个。to think maybe I should try Codex for that. 42:23 我强烈建议不要用 NotebookLM 生成演示文稿 I highly discourage generating presentations through NotebookLM 42:27 因为它采用了 Nano Banana。because it uses Nano Banana. 42:29 没错。Right. 42:30 它生成的是一张图片，而不是真正的幻灯片。It generates a picture, not an actual slide. 42:33 Yan，欢迎随时插话进来。Yan, feel free to chime in. 42:35 是的，强烈不建议这样做。Yeah, strongly discouraged. 42:38 而且 Yan 拥有一种演示技巧 And Yan has a presentation skill 42:41 你也可以加以利用。that you can leverage as well. 42:45 我认为所有这些幻灯片 I believe all these slides 42:46 都是用这个演示技巧制作的。are done by this presentation skill. 42:49 嗯。Yeah. 42:50 酷。Cool. 42:51 嗯。Yeah. 42:51 我来看看那个链接 I will take a look at the link 42:52 Yan 刚分享的。that Yan just shared. 42:53 非常感谢。Thank you very much. 42:56 等等，Yan，你想就这个问题说点什么吗？Wait, Yan, do you want to speak to the question? 42:59 因为你没用 MVP 切片，对不对？Because you are not using the MVP slice, right? 43:02 你在用 cursor slides。You are using the cursor slides. 43:05 哎呀，我去年用的是 cursor 幻灯片。Oh no, I used the cursor slide last year. 43:10 不过自 Nano Banana 发布后 But since the release of Nano Banana 43:13 我几乎一直都在用 MVP Slide。I nearly always used MVP Slide. 43:18 所以我也可以把一篇博客帖直接粘贴在这里。So I can also paste a blog post here. 43:23 总体思路是它会将幻灯片生成成 The general idea is it produces the slide decks as images, 43:27 纯图像形式，没有其他内容。purely images, nothing else. 43:31 它能更自然、全面地将各种设计元素 And it could embed all the different design elements 43:36 有机融入其中。more organically and holistically. 43:39 这就是我认为 That’s the reason why I feel 43:40 它更灵活、更具设计感的原因。it has more flexibility and design feeling. 43:45 而且我几乎总是 And I’m using it nearly always, 43:47 几乎完全只用它。nearly exclusively. 43:49 而且面临着诸多挑战。And there are a lot of challenges. 43:52 这篇博客文章讲解了 This blog post explains 43:53 我们是如何应对这些挑战的。how we walk through these challenges. 43:55 关键不仅仅 The keys not only 43:56 在于解决方案本身，how not only the solution itself, 43:58 还在于我们如何得出该解决方案。but also how we reached that solution. 44:02 我们脑子里在想什么样的取舍。What kind of trade off in our mind. 44:04 我觉得那也是一篇不错的读物。I think that’s also a good read. 44:06 嗯。Yeah. 44:07 非常感谢。Thank you very much. 44:08 嗯，我来读一下。Yeah, I’ll take a read. 44:08 谢谢你们两位。Thank you both. 44:09 谢谢。Thank you. 44:16 我看到 Jung 有个问题。And I see there’s a question from Jung. 44:20 其实我觉得它们大多都差不多。Actually, I think a lot of them are pretty much the same. 44:24 如果你主要感兴趣的是写作，If you are mostly interested in writing, 44:27 我推荐使用 Gemini 或者 Claude。I would recommend either Gemini or Claude. 44:31 那两个大语言模型在写作上非常出色。Those two LLMs are very good at writing. 44:36 如果你最在意性价比，If you care the most about getting the best bang for the buck, 44:38 我推荐 GLM，我用的是 GLM-5 Turbo I would recommend GLM and I use GLM-5 Turbo 44:45 非常划算。and it’s very economical. 44:46 如果是编程相关，If it’s about coding, 44:49 我推荐用 GPT，I recommend using GPT 44:51 因为它比 Claude 划算得多。because it’s much more cost-effective compared with Claude. 44:55 其实我觉得 Claude 并不比 GPT 优秀。Actually I don’t find Claude is doing better than GPT. 45:00 Claude 取消了 20 美元套餐的 Claude Code 访问权限吗？Did Claude cancel access to Claude Code for the 20 subscription? 45:14 我听说过，但我不…… I heard about it, but I. 45:15 没有。I didn’t. 45:16 其实并没有真正验证过。Didn’t really verify. 45:17 嗯，嗯。Yeah, yeah. 45:19 哦，哦，我收回刚才的话。Oh, oh, went back on that. 45:39 是的。Yep. 45:40 我的分享到此结束 That’s the end of my session 45:42 欢迎随时提出更多问题。and feel free to ask more questions. 45:55 如果不是 Capstone 项目，If not the Capstone project, 45:59 我强烈推荐 Hope。highly recommend that Hope. 46:02 希望能在我们的学生门户 Hope to see you in our student portal 46:05 和支持他们的学院看到你。and supporting their academy. 46:09 嗯。Yeah. 46:10 你还有什么想补充的吗？Anything you want to add? 46:14 我们已经为 Capstone 项目介绍了不少内容 We already introduced quite a lot for the Capstone projects 46:17 还演示了几个 AI Builderspace 的案例。and went over a demo several demos about AI Builderspace. 46:21 它已经全部录制并发布在了 It was all recorded and published 46:23 学生门户的通用频道里。in the general channel of the student portal. 46:25 欢迎随时去看看。Feel free to check it out. 46:27 我觉得那个也值得一看。I think that would be a good watch as well. 46:29 顶石项目，我们过去几个队列都在做这个 Capstone Projects so we have been doing this 46:36 ，评价非常高。for the past few cohorts and it was very highly rated. 46:39 所以这是非常好的一个方法 So that’s a very good way 46:41 从本课程中获取最大收获。to get the most from this course. 46:45 很多东西，不能学到 A lot of things cannot be learned 46:46 只有在你真正亲身经历的那一刻，才能学到。until the moment you really experienced it. 46:50 因此，Capstone Project 是 So Capstone Project is a very good way 46:52 体验它的绝佳途径。to experience that. 46:53 我们还特别设计得 And we also designed it such that 46:55 让它非常容易上手试用。it’s very easy to go to enter to try it out. 46:59 所以，看看视频 So just watch the videos 47:01 运行你的第一个项目 get your first project running 47:03 你就会迷上构建了。and you will get addicted to build. 47:12 好，谢谢大家。All right, thank you, everyone. 47:15 我看到有。I see there’s. 47:16 没有其他问题了。There are no more questions. 47:18 所以接下来两周，我们仍会继续监控学生门户。So we will still continue monitoring the student portal 47:22 你可以提问 You can ask questions 47:24 @我们 tagging us 47:24 或者在那里私信我们。DM us over there. 47:27 而且所有课程材料将一直可用，And all the course materials will always be available, 47:29 即便课程无限期结束后也是如此。even after the course ends indefinitely. 47:32 两周后，And after two weeks, 47:33 我们就不会再被主动监控了。we’re not actively monitored. 47:36 不过，你还是可以给我们发私信哦。But still, you can DM us. 47:38 我们始终都会回应。We are always responsive. 47:40 Capstone 项目的录像是干什么用的？What’s the recording for Capstone project? 47:43 这是办公时间（office hour）的录像。It is the recording for the office hour. 47:45 我们在办公时间演示了 Capstone 项目，We did demos about the Capstone projects in the office hour, 47:49 已经在学生门户的通用频道发布了，and it was already released in the general channel of our student portal, 47:52 大家可以随意去看看。so feel free to check it out. 47:56 好吧，目前就到这里吧。All right, then, that’s basically it for now. 48:00 课程材料会有更新吗？Going to have any update on the course material? 48:03 是的，我们会更新课程材料 Yes, we will update the course material 48:04 只要我们更新，就会及时更新。whenever we do so. 48:06 所以第一批学生 So even the first cohort 48:07 也和你们一样，收到了相同的课程材料。received the same course material as you as well. 48:12 目前你还不能观看未来队列的视频 I don’t think you can watch the future cohort video 48:16 。for now. 48:16 YZ 可能会定期更新这些内容。YZ might update them periodically. 48:19 就像经过三四个课程之后。Like after three or four course. 48:21 三四个队列后，Three or four cohorts, 48:22 你会收到最新队列的更新。you will get updates for the latest one. 48:26 但我们就是不这么做，绝不。But we don’t do that, period. 48:27 我们不会那样做 We don’t do that 48:28 ，每个月或每三个月都 like every month or every three months. 48:30 是的，我的意思是，Yeah, I mean, 48:32 下一个班级在七月，next cohort is in July, 48:33 所以我们会更新到七月的版本，so we’re updating July, 48:35 但它会…… but it’s gonna. 48:36 我们会升级到 Super Linear Academy 版本。We’ll update to the Super Linear Academy version. 48:39 所以你已经能访问那个版本了，So you already have access to that version, 48:43 七月再回来看看，and you can check back in July, 48:45 视频会在下一期结束后更新。and the video is going to be updated after the next cohort. 48:51 好的，谢谢大家。All right, thank you, everyone. 48:54 谢谢。Thank you. 48:55 谢谢，谢谢。Thank you, thank you. Module 4.pdf 6.94 MB
最后更新：2026 年 5 月 8 日
English Original
Module 4.pdf 6.94 MB
Last updated: May 8th, 2026