📝 本节摘要：
本章节探讨了AI初学者掌握大语言模型（LLM）核心原理的最佳学习策略。嘉宾建议从零开始编写代码（From Scratch），而非直接使用复杂的现有库（如Hugging Face）。
核心观点包括：
1. 学习目标的设定：自建模型不是为了替代ChatGPT，而是为了通过“手写”理解预训练、微调及注意力机制的内部运作。
2. 避免过早追求规模：大规模模型涉及参数分片（Sharding）和多GPU管理，复杂度呈指数级上升，初学者应专注于单GPU模型。
3. Hugging Face的局限性：虽然它是行业标准，但其代码库为了兼容性而过度封装，逻辑交织，不适合用于学习底层原理。
4. “逆向工程”法：推荐通过阅读模型配置文件（Config File），复现架构，并尝试加载开源的预训练权重（Pre-trained Weights）。如果你写的代码能跑通别人的权重并输出一致结果，这就是最好的“可验证奖励”。
[原文] [Lex Fridman]: if we could take at this point a bit of a tangent and talk about education and learning if you're somebody listening to this who's a smart person interested in programming interested in AI so I presume building something from scratch is a good beginning so can you just take me through like what you would recommend people do
[译文] [Lex Fridman]: 如果我们此刻能稍微偏离一下主题，谈谈教育和学习。如果你是正在收听节目的聪明人，对编程感兴趣，对AI感兴趣，我假设从零开始构建一些东西是一个很好的开端，那么你能否带我了解一下你会推荐人们怎么做？
[原文] [Guest]: so I would personally start like you said uh implementing a simple model from scratch that you can run on your computer the goal is not if you build a model from scratch to have like something you use every day for your personal projects like it's not going to be your personal assistant replacing an existing openweight model or CHPD it's to see what exactly goes into the LLM what exactly comes out of the LLM how the pre-training works in that sense on your own computer preferably um and then you learn about the pre-training the supervised fine-tuning the attention mechanism you get a solid understanding of how things work
[译文] [Guest]: 我个人会像你说的那样开始，呃，从零实现一个可以在你自己电脑上运行的简单模型。如果你从零构建一个模型，目标并不是要拥有一个你每天用于个人项目的东西，比如它不会成为替代现有开放权重模型或ChatGPT的个人助手；其目的是为了看看究竟什么进入了LLM（大语言模型），究竟什么从LLM出来，以及预训练是如何工作的，最好是在你自己的电脑上进行。嗯，然后你会学习预训练、监督微调（SFT）、注意力机制，从而对这些东西是如何工作的有一个扎实的理解。
[原文] [Guest]: but at some point you will reach a limit because small models can only do so much and the problem with learning about LLMs at scale is I would say it's exponentially more complex to make a larger model because it's not that the model just becomes larger you have to now think about sharding your parameters across multiple GPUs you even for the KV cache there are multiple ways you can implement it one is just to understand how it works just to grow the cache let's it's like a cache you grow step by step by let's say concataining lists um growing it but then that wouldn't be optimal in GPUs you wouldn't do that you would pre-allocate a tensor and then fill it in but that adds again another 20 30 line lines of code and for each thing you add so much code,
[译文] [Guest]: 但在某种程度上你会遇到瓶颈，因为小模型能做的事情有限。而且学习大规模LLM的问题在于，我会说构建一个更大的模型在复杂度上是指数级增加的。因为不仅仅是模型变大了，你现在必须考虑在多个GPU之间对参数进行分片（Sharding）。甚至对于KV缓存（KV Cache），你也有多种实现方式。一种方式只是为了理解它是如何工作的，即逐步增长缓存——假设它像一个你通过连接列表逐步增长的缓存——但在GPU上这并不是最优的，你不会那样做。你会预先分配一个张量（Tensor），然后填充它，但这又增加了20到30行代码，每增加一样东西都会增加大量代码。,
[原文] [Guest]: and I think the trick with a book is basically to understand how the LLM works it's not going to be your production level LLM but once you have that you can understand the production level LM so you're trying to always build an LLM that's going to fit on one GPU yes the most of them I have they I have some bonus materials on some models i think one of or two of them they may require multiple GPUs but the goal is to have it on one GPU and the beautiful thing is also you can selfverify it's almost like RLVR when you code these from scratch you can take uh an existing model from the hugging face transformer library,
[译文] [Guest]: 我认为写书（或学习）的诀窍基本上是理解LLM是如何工作的，它不会成为你的生产级LLM，但一旦你掌握了它，你就能理解生产级的LM。所以你要尝试始终构建一个能放入单个GPU的LLM。是的，我拥有的大部分模型——我在某些模型上有一些补充材料，我想其中一两个可能需要多个GPU——但目标是将其放在一个GPU上。这样做美妙之处在于你可以自我验证，这几乎就像RLVR（基于规则的验证奖励）。当你从零开始编写代码时，你可以从Hugging Face Transformer库中取出一个现有的模型。,
[原文] [Guest]: um so the hugging face transformer library is great but if you want to learn about LMS I think that's not the best place to start because the code is so complex because it has to full it has to fit so many use cases also some people use it in production it has to be sophisticated and it's really intertwined and really hard it's not linear to read it was started as a fine-tuning library and then it grew to be like the standard representation of every model architecture and the way this loaded so hugging face is like the default place to get a model and transformers is the software that enables it so people can easily load a model and do something basic with it,
[译文] [Guest]: 嗯，Hugging Face Transformer库很棒，但如果你想学习大语言模型，我认为那不是最好的起点。因为代码太复杂了，它必须满足太多的用例，还有些人在生产环境中使用它，所以它必须很复杂，而且它是真正交织在一起的，非常难读，并非线性的。它最初是作为一个微调库开始的，然后成长为每个模型架构的标准表示形式以及加载方式。所以Hugging Face就像是获取模型的默认场所，而Transformers是使其能够运行的软件，让人们可以轻松加载模型并用它做一些基本操作。,
[原文] [Guest]: and all frontier labs that have open weight models have a hugging phase transformers version of it like from deepseek to GPTOSs so it's like the canonical weight that you can load there but again also even transformers the library is not used in production people use then sglang or vlm and it adds another layer of complexity we should say that the transformers library has like 400 models so it's a one library that tries to implement a lot of LLMs and so you have a huge code base basically it's like huge it's like it's I don't know maybe millions hundreds of thousands of lines of code and find it's like understanding the part that you want to understand is finding the needle in the haststack
[译文] [Guest]: 所有拥有开放权重模型的前沿实验室都有其Hugging Face Transformers版本，比如从DeepSeek到其他GPT类操作系统，所以它就像是你可以在那里加载的标准权重。但同样，即使是Transformers这个库也不常用于生产环境，人们会使用sglang或vllm，这又增加了一层复杂性。我们应该说Transformers库大概有400个模型，所以这是一个试图实现大量LLM的单一库。因此你面对的是一个巨大的代码库，基本上它是巨大的，我不知道，可能有数百万或数十万行代码。要在其中找到你想理解的部分，就像大海捞针一样。
[原文] [Guest]: but what's beautiful about it is you have a working implementation and so you can work backwards from it what I would recommend doing or what I also do is if I want to understand for example how almost 3 is implemented I would look at the weights in the model hub the config file and then you can see oh they used so many layers they use let's say group query attention or multi head attention in that case then you see all the components in like a human readable I don't know 100 lines of config file,
[译文] [Guest]: 但它的美妙之处在于你有一个可工作的实现，所以你可以从它倒推（Work backwards）。我会推荐的做法，或者我自己做的，就是假如我想理解——比如Llama 3是如何实现的——我会查看模型中心（Model Hub）里的权重和配置文件（Config File）。然后你会看到，哦，他们用了这么多层，他们用了比如分组查询注意力（Group Query Attention）或者多头注意力（Multi Head Attention）。在这种情况下，你会在一个大概100行的人类可读的配置文件中看到所有组件。,
[原文] [Guest]: and then you start let's say with your GPD2 model and add these things you know and the cool thing here is you can then load the pre-trained weights and see if they work in your model and you want to match the same output that you get with a transformer model and then you can use that as a basically as a verifiable reward to make your architecture correct and then it's kind of sometimes it takes me a day to with almost three the challenge was the rope for the position embeddings they had a yarn extension and there was some custom uh scaling there and I couldn't quite match the these things and in this struggle you kind of understand things but the cool thing is at the end you know you have it correct because you can unit test it you can check against the reference implementation
[译文] [Guest]: 然后你从，比如说你的GPT-2模型开始，把你所知道的这些东西加进去。这里最酷的事情是，你可以加载预训练的权重，看看它们是否在你的模型中工作。你要争取匹配你从Transformer模型中得到的相同输出，这样你就可以把它作为一个基本可验证的奖励（Verifiable Reward）来确保你的架构是正确的。然后，这有时候可能会花我一天的时间。比如在Llama 3上，挑战在于位置嵌入（Position Embeddings）的RoPE（旋转位置编码），他们有一个Yarn扩展，还有一些自定义的缩放，我当时没法完全匹配这些东西。正是在这种挣扎中，你会逐渐理解这些概念。但最酷的是，最终你知道你是正确的，因为你可以进行单元测试，你可以对照参考实现进行检查。
[原文] [Guest]: and I think that's maybe one of the best ways to learn really like to basically reverse engineer something yeah I think that that is something that everybody that's interested in getting to AI today should do and I think that's why I liked your book is like I came to language models from this RL and robotics field like I've never had taken the time to just like learn all the fundamentals and this transformer architecture I describe as being like so fundamental as like deep learning was a thing that I had to learn in the past and people need to do this
[译文] [Guest]: 我认为这可能是最好的学习方式之一，真的，就像是基本上对某些东西进行逆向工程（Reverse Engineer）。是的，我认为这是每个今天对进入AI领域感兴趣的人都应该做的事情。这也正是为什么我喜欢你的书。我来自强化学习（RL）和机器人领域进入语言模型领域，我从未花时间去学习所有的基础知识。而这个Transformer架构，我将其描述为如此基础，就像深度学习是我过去必须学习的东西一样，人们也需要这样做。
📝 本节摘要：
本章节重点讨论了在掌握基础原理后，如何在这个飞速发展的领域找到切入点。嘉宾指出，初学者往往因为不知如何产生影响力而感到不知所措。
核心策略包括：
1. 寻找“低垂的果实”：顶尖研究者往往关注宏大问题，而留下了许多未被充分解决的细分领域（如特定数据的处理）。
2. “去窄”（Go Narrow）：不要试图跟进所有新知，那样会导致倦怠。建议在掌握基础后，深耕一个极窄的领域（例如“角色性格训练”）。
3. 成为细分专家：在某些冷门领域，可能只需研读现存的3篇论文，你就成了该领域的专家。
4. 社区互动：许多推特（X）上的匿名账号可能就是隐藏的深度研究者，积极挖掘这些非正式资源非常有价值。
[原文] [Guest]: i think that where a lot of people kind of get overwhelmed is how do I apply this to have impact or find like a career path because like AI and language models make this fundamental stuff so accessible and people with motivation will learn it and then it's like how do I get the cycles on goal to contribute to research
[译文] [Guest]: 我认为很多让人感到不知所措的地方在于，我该如何应用这些知识来产生影响力，或者找到一条职业路径？因为AI和语言模型让这些基础知识变得如此触手可及，有动力的人都会去学，但接下来的问题是：我如何投入精力（get the cycles on goal）去为研究做贡献？
[原文] [Guest]: and I think that I'm actually fairly optimistic in this because the field moves so fast that a lot of times the best people like don't fully solve a problem because there's a bigger lower hat like a bigger problem to solve that's very low hanging fruit so they move on
[译文] [Guest]: 对此我其实相当乐观，因为这个领域发展得太快了，很多时候最优秀的人并没有完全解决一个问题，因为有更大、更重要的问题要解决，或者说是更容易摘到的果实（low hanging fruit），所以他们就继续前进了。
[原文] [Guest]: and I think that a lot of what I was trying to do in this RHF book is like take post-trading techniques and just describe how people think about them influencing the model and what people are doing and then it's remarkable how many things I just think are just like people stop studying them or don't so I think people trying to get narrow after doing the fundamentals is good and then reading the relevant papers and being engaged in the ecosystem
[译文] [Guest]: 我在这本RLHF（人类反馈强化学习）书中试图做的大部分工作，就是选取后训练（Post-training）技术，描述人们如何思考这些技术对模型的影响以及人们正在做什么。令人惊讶的是，有多少东西我觉得人们只是停止研究了或者根本不去研究。所以我认为，在打好基础之后尝试通过“去窄”（get narrow，深耕细分领域）是很好的策略，然后阅读相关论文并积极参与到生态系统中去。
[原文] [Guest]: it's like you actually the proximity that random people have online from the leading researchers like no one knows who all the anonymous account on X and ML is very popular for whatever reason and no one knows who all these people are like it could just be random people that study the stuff deeply especially with the AI tools to just be like keep I don't understand this keep digging into it I think is a very useful thing
[译文] [Guest]: 事实上，你在网上作为一个普通人与顶尖研究者的距离是非常近的。比如在X（前Twitter）上没人知道那些匿名账号是谁，但机器学习社区无论出于什么原因都非常火爆。没人知道这些人是谁，他们可能只是深入研究这些东西的普通人。特别是利用AI工具，这种“我不懂这个，我要继续深挖”的态度，我认为是非常有用的。
[原文] [Guest]: but there's a lot of research areas that just like are maybe three papers that you need to read and then one of the authors will probably email pay you back but you have to put in a lot of effort into these emails to understand the field like I think it would be for a newcomer easily weeks of work to feel like they can truly grasp like what is a very narrow area
[译文] [Guest]: 但有很多研究领域，可能只需要读三篇论文，然后其中一位作者可能会回你的邮件。但你必须在这些邮件中投入大量精力以表明你理解该领域。我认为对于一个新人来说，可能需要数周的努力才能感觉自己真正掌握了一个非常狭窄的领域。
[原文] [Guest]: but I think going narrow after you have the fundamentals be very useful to people because it's like I became very interested in like character training which is like how you make the model funny or sarcastic or serious and like what do you do to the data to do this and it's like a student at Oxford reached out to me he's like "Hey I'm interested in this." And I advised him and I was like "That paper now exists." And it's like I don't know there's like two or three people in the world that were very interested in this he's a PhD student which gives you an advantage
[译文] [Guest]: 但我认为在掌握基础之后“去窄”对人们非常有用。比如我自己对“角色训练”（Character Training）非常感兴趣，也就是如何让模型变得有趣、讽刺或严肃，以及你需要对数据做什么来实现这一点。有一个牛津大学的学生联系我说：“嘿，我对这个感兴趣。”我指导了他，后来我就说：“那篇论文现在已经存在了。”感觉世界上可能只有两三个人对这个非常感兴趣。他是个博士生，这确实给了你一些优势。
[原文] [Guest]: but like for me that was a topic I was waiting for someone to be like "Hey I have time to spend cycles on this." And I'm sure there's a lot more very narrow things where you're just like "Oh it doesn't make sense that there was no answer to this." And I think that it's just like there's so much information coming that people are like I can't grab onto any of these but if you just actually stick in an area I think there's a lot of interesting things to learn
[译文] [Guest]: 但对我来说，那是一个我在等待有人来说“嘿，我有时间在这上面投入精力”的话题。而且我确信还有很多非常狭窄的领域，你会觉得“哦，这竟然没有答案，简直说不通”。我认为现在的资讯太多了，人们会觉得无法抓住任何东西，但如果你真的坚持在一个领域深耕，我认为会有很多有趣的东西可以学。
[原文] [Guest]: yeah I think you can't try to do it all because it would be very overwhelming and you would burn out if you try to keep up with everything for me for example I haven't kept up with computer vision a long time just focus on LMS
[译文] [Guest]: 是的，我认为你不能试图做所有事情，因为那会让人不知所措，如果你试图跟进所有事情，你会倦怠（Burn out）。以我为例，我已经很久没有关注计算机视觉了，我只专注于大语言模型（LLMs）。
📝 本节摘要：
本章节深入探讨了RLHF（人类反馈强化学习）背后的哲学与数学困境。嘉宾认为RLHF从根本上是“不可完全解决的”，因为人类偏好极其混乱，难以被压缩为单一的数值。
核心观点包括：
1. 偏好的不可量化性：引用经济学中的冯·诺依曼-摩根斯特恩效用定理（Von Neumann-Morgenstern utility theorem），指出将多维度的偏好（如风格与准确性）压缩为单一奖励值（Scalar Reward）是一种数学上的简化，而非现实的完美反映。
2. 社会选择理论（Social Choice Theory）：RLHF实际上是一个经济学问题，即如何聚合不同且往往冲突的偏好。
3. 模型规范（Model Spec）与透明度：OpenAI等公司发布的模型规范是区分“训练失败”与“有意为之”的关键。
4. 数学的商品化：批评当前的算法（如DPO）通过数学技巧（Log Trick）跳过了推导步骤，虽然简化了问题，但可能阻碍了学习者对底层逻辑的真正理解。
[原文] [Lex Fridman]: but coming back to your book for example I think this is also a really great book and a really good bang for the buck because you want to learn about RLHF i wouldn't go out there and read RL HF papers because I would be you would be spending contradict there
[译文] [Lex Fridman]: 但回到你的书来说，例如，我认为这也是一本非常棒的书，而且性价比很高。因为如果你想学习RLHF（人类反馈强化学习），我不会建议直接去读RLHF的论文，因为那会花费你大量时间去处理相互矛盾的信息。
[原文] [Guest]: i just edited the book and I was like there's a chapter where I had to be like X papers say one thing and X papers say another thing and we'll see what comes out to be true
[译文] [Guest]: 我刚刚编辑完这本书，当时我就觉得，有一章我不得不写成“X论文说是这样，而X论文说是那样”，我们还得等着看最后哪个才是真的。
[原文] [Lex Fridman]: what are some of the just to go through some of the table of contents some of the ideas we might have missed in the bigger picture of the post training so first of all you do the problem setup training overview what are preferences preferences data in the optimization tools reward modeling regularization instruction tuning rejection sampling reinforcement learning I policy gradients direct alignment algorithms then constitutional AI and AI feedback reasoning and inference time scaling tool use and function calling synthetic data and distillation evaluation and then open question section over optimization style and information and then product UX character and post training so what are some ideas worth mentioning that connect both the educational component and the research component you mentioned the character training which is pretty interesting
[译文] [Lex Fridman]: 为了过一遍目录，看看在后训练（Post-training）的大图景中我们可能错过了哪些想法：首先你做了问题设定、训练概览、什么是偏好、优化工具中的偏好数据、奖励建模（Reward Modeling）、正则化、指令微调（Instruction Tuning）、拒绝采样（Rejection Sampling）、强化学习I、策略梯度、直接对齐算法、然后是宪法级AI（Constitutional AI）和AI反馈、推理与推理时扩展（Inference Time Scaling）、工具使用与函数调用、合成数据与蒸馏、评估，然后是关于过度优化、风格与信息的开放问题部分，最后是产品UX、角色与后训练。那么，有哪些值得一提的想法是既连接了教育部分又连接了研究部分的？你提到了角色训练，这非常有趣。
[原文] [Guest]: character training is interesting because there's so little out of it but we talked about how people engage with these models and like like we feel good using them because they're positive but that can go too far or it could be too positive and it's like essentially it's how do you change your data and or decision-m to make it exactly what you want and like OpenAI has this thing called a model spec which is essentially their internal guideline for what they want the model to do and they publish this to developers so essentially you can know what is a failure of OpenAI's training which is like they have the intentions and they haven't met it yet versus what is something that they like actually wanted to do and that you don't like and that transparency is very nice But all the methods for curating these documents and how easy it is to follow them is not very wellnown
[译文] [Guest]: 角色训练很有趣，因为关于这方面的内容太少了。但我们讨论了人们如何与这些模型互动，比如我们使用它们感觉很好，因为它们很积极，但这可能会过头，或者过于积极。这本质上是关于你如何改变你的数据和/或决策机制，使其完全符合你的期望。比如OpenAI有一个叫做“模型规范”（Model Spec）的东西，这本质上是他们关于希望模型做什么的内部指南，并且他们向开发者发布了这个。所以本质上你可以知道，什么是OpenAI训练的失败——即他们有这个意图但尚未达到——以及什么是他们实际上想做但你不喜欢的东西。这种透明度非常好。但是，策划这些文档的所有方法以及遵循它们的容易程度并不为人所熟知。
[原文] [Guest]: i think the way the book is designed is that the reinforce learning chapter is obviously what people want because everybody hears about it with RLVR and it's the same algorithms and the same math but it's just like you can use it in in very different documents so I think the core of pref is like how messy preferences are is essentially rehash of a paper I wrote years ago this is essentially the chapter that'll tell you why RHF is never ever fully solvable because like the way that even RL is set up is that um it assumes that preferences can be quantified and that multiple preferences can be reduced to single values
[译文] [Guest]: 我认为这本书的设计方式是，强化学习章节显然是人们想要的，因为每个人都听说了RLVR（基于规则的验证奖励），而且它是相同的算法和相同的数学，只是你可以将其用于非常不同的文档中。所以我认为偏好（Pref）的核心在于偏好是多么混乱，这本质上是我几年前写的一篇论文的重述。这本质上是那一章会告诉你为什么RLHF（人类反馈强化学习）永远无法完全解决，因为甚至RL（强化学习）的设置方式都假设偏好可以被量化，并且多个偏好可以被缩减为单一数值。
[原文] [Guest]: and I think it relates in the economics literature to the vonoman Morganston utility theorem and like that is the chapter where all of that philosophical economic and like psychological context it tells you what gets compressed into doing RHF so it's like you have all of this and then at later in the book it's like you use this RL math to make the number go up and I think that that's why I think it would be very rewarding for people to do research on is because it's like quantifying preferences is something that is just like humans have designed the problem in order to make preferences studyable but there's kind of fundamental debates on like an example is in a language model response you have different things you care about whether it's accuracy or style and when you're collecting the data they all get compressed into like a I like this more than another
[译文] [Guest]: 我认为这在经济学文献中与冯·诺依曼-摩根斯特恩效用定理（Von Neumann-Morgenstern utility theorem）有关。那一章包含了所有的哲学、经济学和心理学背景，它告诉你什么被压缩进了RLHF的操作中。所以你有所有这些背景，然后在书的后面部分，你只是使用这个RL数学让数字变大。这就是为什么我认为人们在这个领域做研究会非常有回报，因为量化偏好就像是人类为了让偏好变得可研究而设计了这个问题。但存在根本性的争论，例如在语言模型的回复中，你关心不同的事情，无论是准确性还是风格，而当你收集数据时，它们都被压缩成一句“我喜欢这个多于那个”。
[原文] [Guest]: and it's like like that is happening and there's a lot of philosoph there's a lot of research in other areas of the world that go into like how should you actually do this I think social choice theory is the sub field of economics around how you should aggregate preferences and there's Like I was I went to a workshop that published a white paper i'm like how can you think about using social choice theory for ROHF so I mostly would want people that get excited about the math to come and have things where they could stumble into and learn this kind of broader context
[译文] [Guest]: 这正在发生，世界上其他领域有很多哲学和研究探讨你应该如何实际操作。我认为社会选择理论（Social Choice Theory）是经济学的一个子领域，关于你应该如何聚合偏好。比如我去了一个研讨会，发表了一份白皮书，我在想如何考虑将社会选择理论用于RLHF。所以我主要希望那些对数学感到兴奋的人能进来，遇到这些东西，并学习这种更广泛的背景。
[原文] [Guest]: i think in education a lot of it needs to be like at this point it's like what I like because the language models are so good at the math where it's like famous paper direct preference optimization which is like a much simpler way of pro solving the problem than RL um the derivations in the appendix skip steps of math and it's like I tried for this book like I redid the derivations and I'm like what the heck is this log trick that they use to change the math but doing it with language models they're like this is the log trick and I'm like I I don't know if I like this that the math is so commoditized i think like some of the struggle and reading this appendix and following the math I think is good for learning
[译文] [Guest]: 我认为在教育方面，很多时候需要是——在这一点上是我喜欢的——因为语言模型在数学方面太擅长了。比如著名的直接偏好优化（DPO）论文，这是一种比RL简单得多的解决问题的方法。附录中的推导跳过了数学步骤。为了这本书，我试着重做这些推导，我就想：“他们用来改变数学运算的这个对数技巧（Log Trick）到底是什么鬼？”但在语言模型领域做这个，他们就觉得“这就是对数技巧”。我就想，我不知道我是否喜欢这样，数学变得如此商品化。我认为阅读附录、跟随数学推导过程中的这种“挣扎”，对学习是有好处的。
📝 本节摘要：
本章节探讨了AI时代的教育哲学。对话者一致认为，学习过程中的“挣扎”（Struggle）是不可或缺的，如果没有经历这种智力上的痛苦，就无法真正掌握知识。
>
核心观点包括：
1. AI导师的理想形态：理想的教育AI不应直接给出答案，而应像游戏攻略一样，只提供线索（Hint）而不剧透（Spoiler），引导学生自己解决问题。
2. 自律与诱惑：虽然可以开发“苏格拉底式”的教学AI，但在旁边唾手可得的“全能解题AI”构成了巨大的诱惑，这对学生的自律提出了极高要求。
3. 培养“品味”（Taste）：在工具泛滥的时代，学生需要培养一种“品味”，以此判断哪些难点值得自己去死磕（Struggle），哪些可以交给工具处理。
4. 考试形式的轮回：由于AI让数字化作弊变得极其容易，教育界正在经历一次“回溯”——从短暂的数字化作业窗口期，重新回到传统的蓝皮书（Blue Book）手写考试和口试。
[原文] [Lex Fridman]: Yeah so we're actually returning to this often just on the topic of education you both have brought up the word struggle quite a bit so there is value if you're not struggling as part of this process you're not fully following the the proper process for learning i suppose some of the providers are starting to work on models for education which are designed to not give all the information at once and make people work to do this so I think you could train models to do this and it would be a wonderful contribution where like all of this stuff in the book you have to reevaluate every decision for it which is such a great example
[译文] [Lex Fridman]: 是的，所以我们其实经常回到这个话题，仅仅就教育这个话题而言，你们俩都多次提到了“挣扎”（struggle）这个词。所以这里是有价值的：如果你在这个过程中没有挣扎，你就没有完全遵循正确的学习过程。我想一些提供商开始致力于开发教育模型，这些模型被设计成不一次性给出所有信息，而是让人必须付出努力去完成。所以我认为你可以训练模型来做这件事，这将是一个很棒的贡献。就像书里的所有东西一样，你必须重新评估每一个决定，这是一个很好的例子。,
[原文] [Guest]: i think there's there's a chance you work on it at AI too which I which I was like oh I think this be it makes sense i do something like that uh did that the other day for video games for example I sometimes for my past time play video games like I like uh video games with puzzles so you know like Zelda and Metroid and there's this new game where I got stuck and I already got stuck and was okay I you know I don't want to struggle like two two days and so I used an LLM but then you say hey please don't add any spoilers just you know I'm here and there what do I have to do next
[译文] [Guest]: 我觉得这也是你在AI领域可以做的事情，我觉得这很有道理。我前几天在玩电子游戏时就做了类似的事。例如，我有时在闲暇时间玩电子游戏，我喜欢带解谜的游戏，比如塞尔达（Zelda）和银河战士（Metroid）。有个新游戏我卡关了，我已经卡住了，我想着“好吧，我不想为了这个挣扎两天”，所以我用了一个LLM。但你会对它说：“嘿，请不要剧透（Spoilers），只要告诉我，我现在在这里，接下来该做什么。”,
[原文] [Guest]: and the same thing you can do I guess for math where you say okay I'm here at this point I'm getting stuck don't give me the full solution But what is something I could try you know like where you kind of carefully probe it but the problem here is I think it requires discipline and a lot of people do math for like I mean there are a lot of people who enjoy math but there also a lot of people who need to do it for their homework and then it's like the shortcut and yeah we can develop an educational LLM but the other LLM is still there and there's still a temptation to use the other LLM
[译文] [Guest]: 我想你对数学也可以做同样的事，你说：“好吧，我卡在这一步了，不要给我完整的解答，但我可以尝试什么？”也就是那种小心翼翼的试探。但这里的问题在于，我认为这需要自律。很多人做数学——我的意思是，有很多人享受数学，但也有很多人只是为了完成作业。然后这就成了一种捷径。是的，我们可以开发一个教育用的LLM，但另一个（直接给出答案的）LLM仍然存在，人们仍然会有使用另一个LLM的诱惑。
[原文] [Guest]: i think a lot of people especially in college they they understand the stuff they're passionate about they're self-aware about it and they understand it shouldn't be easy like I think we just have to develop a good taste mhm we talk about research taste like school taste about stuff that you should be struggling on and and stuff you shouldn't be struggling on which is tricky to know cuz sometimes you don't have good um long-term vision about what would be actually useful to you in your career but you have to you have to develop that taste
[译文] [Guest]: 我认为很多人，尤其是在大学里，他们对自己充满热情的东西是很了解的，他们有这种自我意识，并且明白这不应该太容易。我认为我们只是必须培养一种好的“品味”（Taste）。嗯，我们谈论研究品味，就像谈论学校品味一样，关于你应该在什么事情上挣扎，以及不应该在什么事情上挣扎。这很难知道，因为有时你对自己职业生涯中什么才是真正有用的缺乏长远的眼光，但你必须去培养这种品味。,
[原文] [Guest]: yeah i was talking to maybe my fiance or friends about this and it's like there's this brief 10-year window where all of the homework and all the exams could be digital but before that everybody had to do all the exams in Blue Book cuz there was no other way and now after AI everybody's going to need to be in Blue Books and oral exams cuz everybody could cheat so easily it's like this brief generation that had a different education system that like everything could be digital and but you still couldn't cheat and now I was just going to go back which is pretty funny
[译文] [Guest]: 是的，我可能跟我的未婚妻或朋友聊过这个。这就像是有那么短暂的10年窗口期，所有的作业和考试都可以是数字化的。但在那之前，每个人都必须在蓝皮书（Blue Book，指美式传统的纸质答题本）上做所有考试，因为没有别的办法。而现在，在AI出现之后，大家都得回到蓝皮书和口试（Oral Exams）了，因为每个人作弊都太容易了。这就像是那短暂的一代人拥有过不同的教育系统——一切都可以是数字化的且当时还没法轻易作弊——而现在我们要回去了，这还挺滑稽的。
📝 本节摘要：
本章节重点讨论了在算力资源匮乏（Low Compute）的情况下，个人研究者如何做出有意义的贡献，以及学术界与工业界之间的职业选择。
核心观点包括：
1. 评估（Evaluation）作为突破口：对于算力有限的学者，最佳策略是专注于“推理”和“评估”。如果你能设计出一个能够难倒前沿模型（如Claude）的测试集，并被实验室采纳，这将是职业生涯的“火箭船”。
2. 学术界 vs. 工业界：工业界（尤其是OpenAI等前沿实验室）提供改变人生的薪酬（平均年薪百万美元以上）和巨大的影响力，但代价是可能成为“机器上的齿轮”且无法发表论文；学术界（PhD）提供自由和署名权，但面临资金削减和算力不足的“悲惨”境地。
3. 第三条路：创业（Startup）提供了高风险高回报的选择，而加入工业界被视为相对安全且具有向上流动性的路径。
[原文] [Lex Fridman]: you mentioned character training just zooming out on on a more general topic for that topic how much compute was required and in general to contribute as a researcher are there places where not too much compute is required where you can actually contribute as an individual researcher
[译文] [Lex Fridman]: 你提到了角色训练，稍微把视角拉远一点，到一个更普遍的话题：对于那个话题需要多少算力？以及总的来说，作为一个研究者想要做出贡献，有没有不需要太多算力的地方，让你作为一个个体研究者也能真正做出贡献？
[原文] [Guest]: for on the character training thing i think this research is built on fine-tuning about 7 billion parameter models with Laura which is like a essentially you only fine-tune a small subset of the weights of the model i don't know exactly how many GPU hours that would take but it's doable not doable for every academic so the situation for some academics is like so dire that the only work you can do is doing inference where you have closed models or open models and you get completions from them and you can look at them and understand the models and that's very well suited to evaluation which you become ex you want to be the best at creating representative problems that the models fail on or show certain abilities which I think that you can break through with this
[译文] [Guest]: 关于角色训练这件事，我认为这项研究是建立在使用LoRA（低秩适应）微调大约70亿参数的模型上的，这本质上就像你只微调模型权重的一小部分。我不知道确切需要多少GPU小时，但这（对部分人）是可行的。但并非对每个学者都可行，有些学者的处境非常糟糕，以至于你能做的唯一工作就是进行推理（Inference），即利用闭源模型或开源模型，从中获取补全结果，然后观察并理解模型。这非常适合做评估（Evaluation），你要成为最擅长创建那些模型会失败或展示特定能力的代表性问题的人。我认为你可以通过这种方式取得突破。
[原文] [Guest]: so I've like I think that the top end goal for a researcher working on evaluation if you want to have career momentum is the frontier labs pick up your evaluation so it's like you don't need to have every project do this but if you go from a small university with no compute and you figure out something that claude struggles with and then the next cloud model has it in the blog post like there there's your career rocket ship,
[译文] [Guest]: 所以我觉得，对于致力于评估工作的研究者来说，如果你想要职业发展的动力，最高目标就是让前沿实验室采用你的评估方法。这并不需要你每个项目都做到这一点，但如果你来自一所没有算力的小型大学，而你发现了一些Claude（模型）难以处理的问题，然后下一个Claude模型的博客文章中提到了这一点，那就是你职业生涯的火箭船。
[原文] [Guest]: i think that that's hard but it's like if you want to scope the maximum possible impact with minimum compute it's something like that which is just get very narrow and it takes learning of where the models are going so you need to like build a tool that tests where not cloud 4.5 will fail if you're going to do a re if I'm going to start a research project I need to think where the models in 8 months are going to be struggling
[译文] [Guest]: 我认为这很难，但如果你想用最少的算力获得尽可能大的影响力，大概就是这样，你需要变得非常专精，并且需要了解模型的发展方向。你需要构建一个工具，去测试Claude 4.5将会失败的地方（而不仅仅是现在的版本）。如果我要开始一个研究项目，我需要思考8个月后的模型会在哪里遇到困难。
[原文] [Lex Fridman]: well what about developing totally novel ideas this is a trade-off i think that if you're doing a PhD you could also be like it's too risky to work in language models i'm going way longer term which is like what is what is the thing that's going to define language model development in 10 years,
[译文] [Lex Fridman]: 那么开发全新的想法呢？这是一种权衡。我认为如果你在读博（PhD），你也可能会觉得研究语言模型风险太大了，我要做更长远的研究，比如什么是将在10年内定义语言模型发展的东西？
[原文] [Guest]: which I think that I end up being a person that's pretty practical i mean I went to my PhD where it's like I got into Berkeley worst case I get a masters and then I go work in tech it's like I'm very practical about it on like the life afforded to people to work at these AI companies the amount of like OpenAI's average compensation is over a million dollars in stock a year per employee any normal person in the US to get into this AI lab is transformative for your life
[译文] [Guest]: 我觉得我最终是个非常务实的人。我是说我去读博时，就像是我进了伯克利，最坏的情况也就是拿个硕士学位，然后去科技行业工作。我很务实地看待这个问题，比如在这些AI公司工作能给人们带来的生活。OpenAI的平均薪酬大约是每名员工每年超过一百万美元的股票。对于美国的任何普通人来说，进入这个AI实验室对你的人生都是变革性的。
[原文] [Guest]: so I'm pretty practical of like there's still a lot of upward mobility working in language models if you're focused and the outcomes is like look at these jobs but from a research perspective the transformative impact in these academic awards and like be the next Yan Lacun is from not working on not caring about language model development very much it's a big financial sacrifice in that case
[译文] [Guest]: 所以我很务实地认为，如果你专注的话，在语言模型领域工作仍然有很大的上升流动性，结果就是看看这些工作机会。但从研究角度来看，那些学术奖项的变革性影响，或者成为下一个Yan LeCun（杨立昆），往往来自于不研究或不太关心（当前的）语言模型开发。这在财务上是一个巨大的牺牲。
[原文] [Guest]: so I get to work with some awesome students and they're like should I go work at an AI lab and I'm like like you're getting a PhD at a top school or you're going to leave to go to a lab i'm like I don't know like if you go work at a top lab I don't blame you don't go work at some random startup that might go to zero but if you're going to open AI I'm like it could be worth leaving a PhD for,
[译文] [Guest]: 所以我有机会和一些很棒的学生一起工作，他们会问：“我应该去AI实验室工作吗？”我就像在说：“你在顶级学校读博，或者你要离开去实验室？”我会说：“我不知道，如果你去顶级实验室工作，我不怪你。别去那些可能会归零的随机创业公司，但如果你要去OpenAI，我觉得这可能值得放弃博士学位。”
[原文] [Guest]: let's more rigorously think through this so where would you give a recommendation for people to do a research contribution so the options are academia so get get a PhD spend 5 years publishing computer resources are constrained there's uh there's research labs that are more focused on open weight models and so working there or closed frontier labs research labs open AI anthropic xai so on um the two gradients are the more closed the more money you tend to get and but also the le you get less credit so in terms of building a like a portfolio of things that you've done like it's very clear of what you have done as an academic and you have done this and versus if you are going to go be like trade this fairly reasonable progression for being a cog in the machine which could also be very fun,
[译文] [Guest]: 让我们更严谨地思考一下。你会建议人们在哪里做出研究贡献？选项有：学术界，即读个博士花5年时间发表论文，算力资源受限；或者专注于开放权重模型的研究实验室；或者闭源的前沿研究实验室，如OpenAI、Anthropic、xAI等。这两个梯度的区别在于：越封闭的地方，你倾向于获得越多的钱，但你获得的声誉（Credit）也越少。所以就建立你所做事情的作品集而言，作为一名学者你做了什么是非常清晰的；相比之下，你是用这种相当合理的晋升路径去交换成为机器上的一个齿轮（cog in the machine），这也可能非常有趣。
[原文] [Guest]: so I think there's a lot of uncertainty and trade-offs that in my opinion favor just like take the take the well-paying job with meaningful impact so it's like not also like you're getting paid to sit around at OpenAI building like the cutting edge of things that are changing millions of people's relationship to tech but publication wise they're being more secretive increasingly so so you're publishing less and less and less and less and so you're you are having a positive impact at scale but it's you're a cog in the machine
[译文] [Guest]: 所以我认为存在很多不确定性和权衡，在我看来，倾向于接受那份薪水丰厚且具有意义影响力的工作。这不像是OpenAI付钱让你闲坐着，你是在构建改变数百万人与科技关系的最前沿的东西。但在发表论文方面，他们越来越保密，所以你发表的东西会越来越少。因此，你在规模上产生了积极影响，但你只是机器上的一个齿轮。
[原文] [Guest]: i think it's honestly it hasn't changed that much uh so I have been in academia i'm not in academia anymore at the same time I wouldn't want to miss my time in academia... a lot of people went from academia directly to Google and I think it's the same thing back then the professors were like you know sad that their students went into um industry because they couldn't carry on their legacy in that sense... the only thing right now is there's a third option which is um starting a startup that's a lot of people doing startups very risky move uh but can be high is a high risk high reward type of situation where joining an industry lab I think is pretty safe,
[译文] [Guest]: 老实说，我认为情况并没有改变太多。我也曾在学术界，现在不在了，但我不想错过我在学术界的时光……很多人从学术界直接去了Google。我想当时也是一样，教授们会感到难过，因为他们的学生去了工业界，无法在这个意义上继承他们的衣钵……现在唯一不同的是有了第三个选择，那就是创业。很多人在做创业公司，这是非常冒险的举动，这是一种高风险高回报的情况，而加入工业界实验室我认为是相当安全的。
📝 本节摘要：
本章节聚焦于硅谷AI界日益激烈的“内卷”文化。嘉宾坦言，相比于在企业前沿实验室（Frontier Labs）工作的朋友，留在学术界的教授们似乎更快乐。
>
核心观点包括：
1. 996文化的蔓延：源自中国的“早9晚9，一周6天”的工作模式正在硅谷AI公司中成为常态，这种“苦干心态”（Grind Mindset）被视为保持竞争力的必要条件。
2. 蛙跳式竞争（Leapfrogging）：模型之间交替领先的残酷竞争（如Anthropic与OpenAI之间）迫使团队保持高度紧张，这虽推动了技术进步，却以巨大的人力成本（Human Capital）为代价。
3. 挽救婚姻项目：嘉宾引用《Apple in China》一书中的细节，提到高强度工作导致企业甚至需要设立“挽救婚姻”的暗号，当某人家庭亮红灯时，团队会强制其休假。
4. 激情的两面性：这种过度工作往往并非被迫，而是源于研究者对技术变革的极度狂热（Passion）和使命感，导致他们自愿牺牲健康与生活。
[原文] [Guest]: and you know I feel like my friends who are professors seem on average happier than my friends who work at a frontier lab to be totally honest cuz that's just grounding and the frontier labs definitely do this 996 which essentially is shorthand for work all the time
[译文] [Guest]: 而且你知道，老实说，我觉得我那些当教授的朋友平均来看比我在前沿实验室工作的朋友要快乐。因为那更脚踏实地，而前沿实验室绝对在实行“996”，这本质上就是一直工作的代名词。
[原文] [Lex Fridman]: can you describe 996 as culture that's I believe you could say invented in China and uh adopted in Silicon Valley what's what's 996 just 9:00 a.m to 9:00 p.m 6 days a week 6 days a week what is that 72 hours okay so what is this basically the standard in AI companies in Silicon Valley more and more this kind of grind mindset
[译文] [Lex Fridman]: 你能描述一下996这种文化吗？我相信你可以说它是在中国发明的，然后被硅谷采用了。什么是996？就是早上9点到晚上9点，一周6天？一周6天，那是72小时。好吧，所以这基本上成了硅谷AI公司的标准了吗？这种“苦干心态”（Grind Mindset）越来越普遍了？
[原文] [Guest]: yeah i mean not maybe not exactly like that but I think there is a trend towards it and it's interesting i think it almost flipped because when I was in academia I felt like that because uh as a professor you had to write grants you had to do you had to teach and you had to do your research it's like three jobs in one and it is more than a full-time job if you want to be successful and um I feel like now like Nathan just said the professors in comparison to a lab I think they have less like even maybe pressure or workload than at a frontier lab because they work a lot they're just so fulfilled by like working with students and having a constant runway of mentorship and like a mission that is very people oriented i think in an era when things are moving very fast and are very chaotic is very rewarding to people
[译文] [Guest]: 是的，我的意思是也许不完全是那样，但我认为确实有这种趋势。有趣的是，我觉得情况几乎反转了。因为当我在学术界时，我有那种感觉（指忙碌），作为教授你得写基金申请，你得教学，还得做研究，这就像是一人干三份工。如果你想成功，这绝对超过一份全职工作。但我感觉现在，就像Nathan刚才说的，相比于实验室，教授们的压力或工作量甚至可能更小。因为虽然他们工作很多，但他们通过与学生共事、拥有持续的指导路径以及一种非常以人为本的使命感而感到非常满足。我认为在一个事物发展极快且非常混乱的时代，这种满足感对人们来说是非常有回报的。
[原文] [Guest]: yeah and I think at a startup I think it's this pressure it's like you have to make it and it's like it is really important that people put in the time but well it is really hard because you have to deliver constantly and I've been at a startup i had a good time but I don't know if I could do it forever it's like a interesting pace um and is exactly like we talked about in the beginning these models are leaprogging each other and they are just constantly like trying to take the next step compared to the competitors it's just ruthless
[译文] [Guest]: 是的，我认为在创业公司，就是这种压力，就像是你必须成功，人们投入时间真的很重要。但这真的很难，因为你必须持续交付。我曾在一家创业公司待过，我过得很愉快，但我不知道我是否能永远那样做。那是一种有趣的节奏。这就正如我们一开始谈到的，这些模型在互相“蛙跳式超越”（Leapfrogging），它们只是不断地试图比竞争对手领先一步，这真的很残酷。
[原文] [Guest]: i think right now I think this leaprogging nature and having multiple players is actually an underrated driver of language modeling process where competition is so deeply ingrained to people and these companies have intentionally created very strong culture like anthropic is known to be so culturally like deeply committed and organized i mean like we hear so little from them and everybody at anthropic seems very aligned and it's like being at a culture that is super tight and having this competitive dynamic is like talk about a thing that's going to make you work hard and create things that are better
[译文] [Guest]: 我认为现在，这种“蛙跳”性质和多个玩家的存在，实际上是语言模型进程中一个被低估的驱动力。竞争深深植根于人们心中，而这些公司有意创造了非常强势的文化。比如Anthropic以其深厚的文化承诺和组织性而闻名，我是说我们很少听到他们的消息，但Anthropic的每个人似乎都非常步调一致。身处一种超级紧密的文化中，再加上这种竞争态势，这绝对是那种会让你努力工作并创造出更好东西的因素。
[原文] [Guest]: so I think that this that comes at the cost of human capital which is like you can only do this for so long and people are definitely burning out i think I I wrote a post on burnout was like I've tried in and out of this myself especially trying to like be a manager of bull mode training it's a crazy job doing this the book Apple in China by Patrick McGee he talked about the how hard the Apple engineers worked to set up the supply chains in China and he was like they had saving marriage programs and he told in a podcast he was like people died from this level of working hard so I think there just like it's a perfect environment for creating progress based on human expense and I it's there's going to be a lot there's a lot of the human expense is the 996 that we started this with which is like people do really grind
[译文] [Guest]: 所以我认为这是以人力资本为代价的，这种状态你只能维持这么久，人们肯定会倦怠（Burnout）。我想我写过一篇关于倦怠的帖子，我自己也曾进进出出这种状态，尤其是试图作为管理者进行那种“公牛模式”（Bull Mode）的高强度训练时，做这工作太疯狂了。Patrick McGee写的《Apple in China》（注：应指关于苹果供应链的书籍）一书中提到，苹果工程师为了在中国建立供应链工作得有多辛苦。他说他们有“挽救婚姻项目”（Saving Marriage Programs），他在播客里说有人因为这种程度的努力工作而过劳死。所以我认为这就像是一个基于人力消耗来创造进步的完美环境。会有很多人力消耗，就是我们一开始说的996，人们真的在苦干。
[原文] [Guest]: i also read this book i think they had a code word for if someone had to go home to spend time with their family to save the marriage and it's crazy then colleagues want to say okay this is like red alert for this situation we have to let that person go home this weekend and um but at the same time I don't think they were forced to work it's really they were so passionate about the product I guess that it is it is you you get into that mindset and I I had that sometimes as an academic but also as an independent person i have that sometimes I overwork and it's unhealthy i had you know I had back issues i had neck issues because I did not take the breaks that I maybe should have taken but it's not because no one forced me to it's because I wanted to work because that's what Open AI and like they want to do this work
[译文] [Guest]: 我也读过这本书，我想他们有一个暗号，用来指代如果某人必须回家陪家人以挽救婚姻的情况。这很疯狂，同事们会说：“好的，这是针对这种情况的红色警报，我们这周末必须让那个人回家。”但同时，我不认为他们是被迫工作的，真的只是因为他们对产品太有激情了。我想这就是你进入的那种心态。我作为学者时有时会有这种心态，作为独立人士时也有，有时我会过度工作，这很不健康。我有背部问题，我有颈部问题，因为我没有休息，也许我本该休息的。但这并不是因为有人强迫我，而是因为我想工作，因为这正是OpenAI以及类似的人想要做的工作。
📝 本节摘要：
本章节（全书终章）探讨了硅谷独特的文化氛围与心理状态。随着“缩放定律”（Scaling Laws）带来的狂热，硅谷形成了一个巨大的“回声室”和泡沫。
>
核心观点包括：
1. 泡沫的两面性：泡沫既可能导致金融投机，也可以像乔布斯的“现实扭曲场”一样，通过集体信念推动技术突破（Buildout Bubbles）。
2. 脱离现实的风险：长期处于996和高强度同温层中，容易让开发者忽视中西部（Midwest）等广阔世界的真实人类体验。
3. “永久性底层”焦虑：嘉宾提到一种极端的硅谷迷因（Meme），声称如果不在2025年底前建立持久价值，就会沦为“永久性底层阶级”，这反映了极度的FOMO（错失恐惧症）。
4. 走出气泡的建议：虽然为了产生影响力需要身在旧金山（SF），但嘉宾强烈建议“走出去”，阅读历史（推荐书籍《Season of the Witch》）和文学，以避免被单一的技术叙事吞噬。
[原文] [Lex Fridman]: yeah there's also there's also a a feeling a fervor that's building especially in Silicon Valley aligned with the scaling laws idea where there's this hype where the world will be transformed on a scale of weeks and you want to be at the center of it
[译文] [Lex Fridman]: 是的，还有一种感觉，一种狂热（fervor）正在积聚，特别是在硅谷，这与“缩放定律”（Scaling Laws）的理念相一致。这其中存在一种炒作，认为世界将以周为单位发生变革，而你想要处于这一切的中心。
[原文] [Lex Fridman]: and then you know um I have this great fortune of having conversations with wide variety of human beings and from there I get to see all these bubbles and echo chambers across the world and it's fascinating to see how we humans form And I think it's fair to say that Silicon Valley is a kind of echo chamber uh a kind of u silo and bubble
[译文] [Lex Fridman]: 然后你知道，我有幸能与各种各样的人交谈，从那里我看到了世界各地所有的这些泡沫和回声室（Echo Chambers），观察我们人类如何形成这些圈子非常迷人。我认为可以公平地说，硅谷就是一种回声室，呃，一种孤岛（Silo）和泡沫。
[原文] [Lex Fridman]: i think bubbles are actually really useful and effective it's not necessarily a negative thing cuz you could be ultra productive it could be the the Steve Jobs reality distortion field cuz you just convince each other the breakthroughs are imminent and by convincing each other of that you make the breakthroughs imminent mhm
[译文] [Lex Fridman]: 我认为泡沫实际上是非常有用和有效的，它不一定是负面的东西，因为你可能会变得超级高效。它可能是史蒂夫·乔布斯式的“现实扭曲场”（Reality Distortion Field），因为你们只是互相说服对方突破即将来临，而通过互相说服，你们促使突破真的发生了。嗯哼。
[原文] [Guest]: burn Hobart wrote a book classifying bubbles but essentially one of them is financial bubbles which is like speculation which is bad and the other one is like I don't know the term but effectively for buildouts because it pushes people to build these things and I do think AI is in this but I worry about it transitioning to a financial bubble which is like it's yeah
[译文] [Guest]: Byrne Hobart 写过一本关于泡沫分类的书，本质上其中一种是金融泡沫，就像投机，这是不好的；另一种，我不知道具体术语，但这实际上是为了“建设/扩建”（Buildouts），因为它推动人们去建造这些东西。我确实认为AI处于这种状态中，但我担心它会转变为金融泡沫，那就……是的。
[原文] [Lex Fridman]: but also in the space of ideas that bubble you are doing a reality distortion field and that means you are deviating from reality and if you go too far from reality while also working you know 996 and you you might miss some fundamental aspects of the human experience including in Silicon Valley and this is a common problem in Silicon Valley is like it's a very specific geographic area you might not understand the Midwest perspective the full experience of all the other different humans in the United States and across the world
[译文] [Lex Fridman]: 但在思想空间里，那个泡沫，你在制造一个现实扭曲场，这意味着你在偏离现实。如果你偏离现实太远，同时还在工作，你知道，996，你可能会错过人类体验的一些基本方面，包括在硅谷内部。这是硅谷的一个常见问题，它就像是一个非常特定的地理区域，你可能无法理解中西部视角（Midwest perspective），无法理解美国及世界各地所有其他不同人类的完整体验。
[原文] [Lex Fridman]: and you and you speak a certain way to each other you convince each other of a certain thing and that that gets you into real trouble whether AI is a big success and becomes a powerful technology or it's not in either trajectory you can get yourself into trouble so you have to consider all of that here you are a young person trying to decide what you want to do with your life
[译文] [Lex Fridman]: 你们用某种特定的方式互相交谈，你们互相说服对方相信某件事，这会让你陷入真正的麻烦。无论AI是否取得巨大成功并成为强大的技术，还是没有，在这两种轨迹中你都可能让自己陷入麻烦。所以你必须考虑到所有这些，尤其当你是一个试图决定人生方向的年轻人时。
[原文] [Guest]: the thing that is I don't even really understand this but the SF AAI memes have gotten to the point where permanent underclass was one of them which was the idea that the last 6 months of 2025 was the only time to build durable value in an AI startup or model otherwise all the value will be captured by existing companies and you will therefore be poor which like that's an example of the SF thing that goes so far
[译文] [Guest]: 有件事我甚至不太理解，但旧金山的AI模因（SF AI memes）已经发展到了出现“永久性底层阶级”（Permanent Underclass）这种说法的地步。这个观点是说，2025年的最后6个月是你在AI创业公司或模型中建立“持久价值”的唯一时机，否则所有的价值都将被现有的大公司捕获，因此你将会变得贫穷。这就像是旧金山那种极端思维的一个例子。
[原文] [Guest]: I still think for young people that going to be able to tap into it if you are really passionate about wanting to have impact in AI like being physically in SF is the most likely place where you're going to do this but it has it has trade-offs I think SF is an incredible place but there is a bit of a bubble and if you go into that bubble which is extremely valuable just get out also read history books read literature uh visit other places in the world Twitter is not and Substack is not the entire world
[译文] [Guest]: 我仍然认为对于年轻人来说，能够利用这一点——如果你真的热衷于想在AI领域产生影响，肉身处于旧金山是最可能让你实现这一点的地方。但这有权衡。我认为旧金山是一个不可思议的地方，但那里确实有点泡沫。如果你进入那个泡沫——这极具价值——但也请记得“走出去”。去读历史书，读文学，去世界上其他地方看看。Twitter和Substack并不是整个世界。
[原文] [Guest]: I think I would One of my one of people I worked with is moving to SF and it's like I need to get him a copy of the season of the witch which is a history of SF from like 1960 to 1985 which goes through like the hippie re revolution like they all the um gays kind of taking over the city and that culture emerging and then the HIV AIDS crisis and other things and it's just like that is so recent and so much turmoil and hurt but also like love and SF and it's like no one knows about this it's a great book Season of the Witch i recommend it
[译文] [Guest]: 我想我会……我也曾共事过的一个人正要搬去旧金山，我就想我得给他买一本《女巫的季节》（Season of the Witch）。这是一部关于旧金山从1960年到1985年的历史，它回顾了比如嬉皮士革命，以及同性恋群体在某种程度上接管这座城市、那种文化的兴起，然后是艾滋病（HIV/AIDS）危机和其他事情。这就像是，那些历史是如此近期，充满了如此多的动荡和伤痛，但也充满了爱和旧金山的特质，但好像没人知道这些。这是一本好书，《女巫的季节》，我推荐它。
[原文] [Guest]: a bunch of my SF friends were who do get out recommended it to me and I think that it's just like living there like I lived there and I didn't appreciate this context and it's just like so recently
[译文] [Guest]: 我的一群确实会“走出去”的旧金山朋友把它推荐给了我。我觉得就像生活在那里一样——我也曾住在那里，但我当时并不理解这种背景，而这一切其实就发生在不久前。