2026 年的大语言模型：推理、工具使用与后训练焦点

好的，各位。欢迎收听新一期的 Twimmel AI 播客。我是主持人 Sam Cherington。今天请到的是独立 LLM 研究员 Sebastian Raschka。在开始之前，别忘了在您收听节目的地方点一下订阅按钮。Sebastian，欢迎回到播客，好久不见。

All right, everyone. Welcome to another episode of the Twimmel AI podcast. I am your host Sam Cherington. Today I'm joined by Sebastian Raschka. Sebastian is an independent LLM researcher. Before we get going, be sure to take a moment to hit that subscribe button wherever you're listening to today's show. Sebastian, welcome back to the podcast. It's been a little bit.

谢谢你再次邀请我，Sam。很高兴回来聊聊 LLM、AI 以及你想到的任何话题。上次非常愉快，希望这次也能有趣又有料。

Yeah, thank you for inviting me back, Sam. I'm happy to be back and to chat about LLMs, AI, and whatever you have in mind. I had a lot of fun last time, so I hope we can make it fun and interesting again.

我这次的笑话可能有点老套了——上次我们聊还是三年前，没什么变化，对吧？

You know, my joke around this time, it's getting a bit old, but it's like, the last time we spoke was three years ago. Not much has changed, right?

好事成三嘛，有句俗话这么说，对吧？

Well, all good things come in threes. I think there's a saying, right?

实际上变化巨大。我们将聚焦于其中最新、最重要的变化，尤其是 LLM 的新进展以及 2026 年可以期待什么。这是你在研究和教育工作中长期关注的领域。也许我们可以从最直观的问题开始：如果从宏观角度对比现在和一年前，你对这个领域的演变有什么总体看法？

And in fact, a ton has changed. And we're going to be focusing on the most recent and most important of those changes. In particular, what's new with LLMs and what to expect with LLMs in 2026. This is an area that you spend a lot of time focusing on with your research and education work. Maybe we can start with just, you know, kind of top of mind like if you think about, very big picture where we are now compared to where we were a year ago, what is your broad reflection about the evolution of the space?

我会对比今天和一年前。差不多是 DeepSeek 的周年纪念，也就是 DeepSeek 版本 3 大模型加上 R1 模型。所谓的“推理革命”，打引号。它仍然是 LLM，仍然是同样的基础模型，但我们现在在它之上有了更多技术，让模型在解决更复杂问题方面变得更聪明。架构方面，LLM 架构仍然相对相似，但推理训练是相比去年新增的东西之一。另外，我认为对工具使用的关注也大大加强了。以前，ChatGPT 刚推出时，或者 LLM 的第一代版本，重点主要是通用任务，让 LLM 凭记忆回答我们好奇的所有问题。如果我们问一个数学问题或知识问题，LLM 基本上会从记忆中提取答案。但这并不总是最有效或最准确的做法，对我们人类来说也是如此。我是说，LLM 的思考方式与人类不同，但我们人类，如果你问我一个复杂的数学问题，比如两个大数相乘，我会掏出计算器来算。我不会在脑子里算。我或许能算，但会花很长时间，更容易出错，而且没必要。LLM 也一样。现在有了更现代的工具，让 LLM 也使用工具变得越来越流行。这需要训练 LLM 使用这些工具，但有了这个，我认为我们可以减少幻觉率，不是完全消除，而是减少，同时让答案更准确。至于推理能力，它本质上是给 LLM 更多时间（打引号）来思考问题。所以，如果对比去年和现在，这就是我们在过去一年可以调整并取得进展的两个主要旋钮。

I would look at today compared to one year ago. It's almost like the anniversary of DeepSeek, the big DeepSeek version 3 model accompanied by the R1 model. The reasoning revolution, in quotation marks. It's still an LLM. It's still the same base model, but we have now more techniques on top of that to make the models smarter in terms of solving more complex problems. Architecture-wise, LLM architectures are still relatively similar, but the reasoning training is one of the new things if we compare today to last year. Also, I think there's a more heavy focus on tool use. Back then, when ChatGPT was launched or also the first iteration of LLMs, the focus was mainly on general purpose tasks, but then also having the LLM answer all the things we are curious about from memory. If we ask it a math question or a knowledge question, the LLM would basically draw from its memory and then write the answer. But that's not always the most effective or accurate thing to do, similar for us humans. I mean, LLMs are different from how humans think, but we as humans, if you asked me a complicated math question like multiplying two large numbers, I would pull out my calculator and calculate that on a calculator. I wouldn't do that in my head. I maybe could, but it would take a long time. It's more error-prone and so forth, and there's no need to do that. The same with LLMs. Now with more modern tooling, it becomes more and more popular to use or to have the LLM use tools too. It requires training the LLM to use those tools, but with that, I think we can reduce hallucination rates, not completely getting rid of those, but reducing those, and then also making answers more accurate. And then with reasoning capabilities, it's essentially giving the LLM more time, in quotation marks, to think through a problem. So these are the two main knobs that we can tune and make progress on in the last year, if we look particularly like last year and now.

我们稍后会深入探讨推理和工具使用方面的技术演变。但在此之前，我想从实际角度聊聊：你认为我们今天所处的位置与之前相比有何不同和转变？这非常有趣。我们现在是 2026 年 2 月的第二周，今年已经有很多新消息、新模型：Opus 4.6、OpenAI 5.3，还有整个 OpenClaw Multibot。谈谈今年我们已经看到的情况，但要从你眼中 LLM 的实际应用角度出发。

We'll dig into the technical aspects of how we've evolved in reasoning and how we've evolved in tool use among other things. But before we do that, I was thinking it might be interesting to talk a little bit about from a practical perspective, how do you think where we are today is different and has shifted? It's super interesting. We're talking in the second week of February and already this year in 2026 there's been a ton of new news, new models: Opus 4.6, OpenAI 5.3. There's been the whole OpenClaw Multibot. Talk a little bit about what we've seen already this year, but in the context of where you see LLMs are from a practical perspective.

是的，说得好。我们才到 2 月的第二周，这意味着春节还没到，我认为届时还会有另一批发布，但主要是在开放权重方面。不过我认为这是另一回事：现在公司围绕 LLM 开发工具，这些工具越来越成熟，同时 LLM 本身也变得更好。我几乎要把这两者分开。我的假设是，如果你把最好的开放权重 LLM 放到 ChatGPT、Gemini 或 Claude 的界面里，你几乎会得到相同类型的质量表现。我认为现在很多用例都围绕 LLM 的工具包装器展开。这就是去年年底流行起来的“封装工程”概念。这也是我们使用 LLM 方式的变化，因为以前只是一个非常简单的宠物界面。后来它变得更复杂了，你可以上传文件和 PDF。就我个人而言，我使用 LLM 主要是为了——听起来可能有点奇怪——校对、检查之类的事情。就在我们录制之前，我刚写完一章，想更新目录，于是我把 PDF 上传到 ChatGPT 界面，说：“嘿，能给我标题吗？”这样我就不用自己提取了，然后还可以再核对一下是否正确。就是这类让工作更简单的小便利，这些繁琐的事情。但正如你所说，还有新的 Opus 模型，ChatGPT 也发布了 Codex 5.3 和对应的 Mac OS 应用，我认为这又是这些模型能力的一次飞跃。

Yeah, that's a good point. We are just in the second week of February, and that means the Chinese New Year hasn't even occurred, where I think there will also be another batch of releases, but I think on the open-weight front. But I think that is a separate thing where you have now companies developing the tooling around LLMs that is becoming more and more mature, and then you have better LLMs themselves. I would almost separate those two. My hypothesis is if you would take the best open-weight LLM and put it into, let's say, a ChatGPT or Gemini or Claude interface, you would almost get the same type of quality performance and everything. I think a lot of use cases evolve around the tool wrapper around the LLM nowadays. That's this idea that was popularized towards the end of last year around harness engineering. That is also something how we changed using LLMs, because before it was just a very simple pet interface. And then it became more sophisticated. You could upload files and PDFs. For my personal use case, I use LLMs mostly for, actually it sounds weird, but like proofreading, checking things, and these types of things. Just before we recording here, I was finishing writing a chapter and I wanted to update the table of contents, and then I just uploaded the PDF to the ChatGPT interface and said, "Hey, can you give me the headers?" so I don't have to pull that out myself, and then you can just double check also that it is correct. But little convenience tasks like making work a bit simpler, these tedious things. But then like you said, there was also the new Opus model, and then ChatGPT released Codex 5.3 and a Mac OS app with that, and I think that is also like yet another leap in terms of what these models are capable of.

我的意思是，之前也有用于编程的 LLM，用 LLM 编程也变得更流行了，但它一直在变得更好。所以之前，我用的是 Visual Studio Code。我是说，因为我用 Visual Studio Code 这个代码编辑器已经好多年了，大概五年，十年，之前我用的是 Vim 和其他工具。但我对它的 UI 非常熟悉。所以我有我的 git 树。我知道终端在哪里之类的。所以我其实很喜欢把 LLM 作为一个插件放在里面，有时候你可以说，“好，能帮我看看吗？我有个 bug，你能帮我检查一下吗？”你知道，它就像是你工作流里新增的一层工具。所以 LLM 不一定要是主角。它也可以是一个小助手，就像之前——我的意思是，我还是会自己调试，但很多时候让 LLM 帮忙检查一下其实挺方便也挺快的。我喜欢它的一点是，它就像第二双眼睛，但又不是完全接管一切。它在某种意义上让你的工作变得更好，就像你有了额外的检查，你可以问，“嘿，你能建议一些改进，让我的代码，比如说，性能更好吗？”但你还是得问对问题，还得实际跑实验看看它是否真的让代码更快了。所以它不是说 LLM 帮你做所有事，而是它会建议一些有用的东西。我知道很多人也用 LLM 来写代码。所以新的，比如说，Codex 插件和 Codex 应用也能做到这一点。新的是，一两年之前，人们会把代码文件上传到 ChatGPT 或 Gemini 或 Claude，然后得到一些反馈，然后你得手动整合进去。而现在它更集成——

I mean, before there were also coding LLMs and it became more popular to use LLMs for coding, but it's always getting better and better. So before, I used Visual Studio Code. I mean, because I just used Visual Studio Code, the code editor, for like years, maybe like five years now, 10 years, and before that I was using Vim and other things. But I'm very familiar with the UI basically. So I have my git tree. I know where I have a terminal inside and that stuff. And so I actually liked having the LLM as a plugin there where you sometimes can say, 'Okay, can you? I have a bug. Can you just double check?' You know, it's just like another layer of tools you add to your workflow. So the LLM doesn't have to be front and center. It can be also this little helper, you know, like before you—I mean, I still debug things myself, but often it's actually quite nice and fast to ask the LLM to double check things. And what I like about it is when it's like a second pair of eyes, but it's also like it's not completely taking over and doing everything. It's making your work better in a sense, like you have additional checks and you can ask, 'Hey, can you suggest improvements to make my code, let's say, more performant?' Still, I mean, you as the person, you still have to kind of ask the right questions and you still have to run actually the experiments to see whether it actually makes the code faster. So it doesn't mean the LLM does everything for you, but it suggests useful things. I know a lot of people also use it for coding things. So that also works with the new, let's say, the Codex plugin but also the Codex app. What's new is, I mean, a year or two ago, people were uploading code files to ChatGPT or Gemini or Claude and then getting some feedback and then you had to manually incorporate that. And now it's more in line—

我觉得人们那样做已经有一阵子了。

I think it's been a while since folks have been doing that.

对，没错。所以现在更原生，你可以看到文件差异，不用离开你的编码环境。但除此之外，现在当你本地运行这些工具时，你可以让它访问你的整个文件夹，比如说整个 git 文件夹，然后它就能看到所有文件的上下文，你不需要手动上传任何东西。而且在此基础上，它现在还能自己使用工具。所以你可以允许 LLM 运行某些命令，比如自己运行单元测试之类的。这些加在一起——我不会说有什么单个东西是突破性的或者改变游戏规则的——但所有这些小东西加起来，让 LLM 变得更强大，因为它变得越来越复杂。我认为这就是我们最近几个月看到的，嗯，可能过去几个季度，人们开发这些类型的能力，而不是仅仅让模型变得更好。所以通过改进界面，我们可以从 LLM 那里获得很多性能提升。

Yeah. Right. So that is now more native where you can see the file diff, you don't have to leave your coding environment. But then on top of that, also now when you run these tools locally, you can give it access to your whole folder, let's say your whole git folder, and then it can see the context of all the files, you don't have to manually upload anything. And then on top of that, it can also nowadays use tools itself. So you can give permission for the LLM to run certain commands, to run a unit test by itself, and these types of things. And that together—I wouldn't say there is a single thing that is like groundbreaking or like a game changer—but all these little things, they add up to make the LLM more capable because it's more and more getting more sophisticated. And I think that's what we have been seeing in recent months and yeah, maybe the last few quarters, months, where people develop these types of capabilities instead of just making the model better. So there's a lot of performance we can get from the LLM by making the interface better basically.

你有没有发现这些新模型中的任何一个——你知道，你刚才说没有突破性的变化——但你是否对这两个模型中的某个新能力感到惊讶，还是说它很大程度上是对你已经在做的事情的增量改进？

Have you found that either of these new models—you know, you just said that there are no breakthrough changes there—but do you find yourself surprised with some new capability in either of these two models, or is it very much incremental to what you're already doing?

对我来说，个人感觉更多的是增量改进。更像是便利性。它们只是变得更稳健、更好，我不会说有什么东西是——对我来说没有那种“哇”的效果，比如“哦，我之前的模型做不到 XYZ”。只是好了一点，你知道？它变得更稳健、更好，然后我对结果也多了些信任。我认为这更像是一个渐进的改进。有一点是，我们仍然有不同推理努力程度的区别——就像一个滑块，决定 LLM 应该花多少时间来给你结果。有从低或零推理努力到高推理努力的设置，这改变了 LLM 生成结果所需的时间。我记得半年前或一年前，如果你想要好结果，你几乎总是得用最高设置，高推理模式，那要花很久。现在，即使是较低的模式，我觉得也相当不错，对大多数任务来说，用这些中高推理努力就足够了，而不是用超高的，然后你得到结果更快。我认为这也是这些模型生活质量上的改进，以前你可能偶尔才跑它们，因为不想等五分钟，但现在它们更常规地成为你工作流的一部分了。

For me, it's personally more incremental. It's just more like the convenience. They're just getting more robust and better, where I wouldn't say there is anything where it's like—there's no wow effect to me, like where it's like 'Oh, my previous model was not able to do XYZ.' It's just a bit better, you know? It's getting more robust, better, and then I also develop a bit more trust in the results. It's more like a gradual improvement, I think. The one thing is, we still have the distinction also between the different reasoning efforts in terms of—it's like a slider in terms of how much time the LLM should spend on getting you the results. And there are different settings from low or no reasoning effort to high reasoning effort, and that changes the time it takes for the LLM to generate results. And I remember like half a year ago or a year ago, if you wanted to have good results, you almost always had to use the highest settings, the high reasoning modes, which took forever. Nowadays, even the lower modes, I feel like, are pretty good, where for most tasks it's sufficient to use these medium-high reasoning efforts instead of the extra high ones, and then you get results faster. And I think that's also like a quality of life improvement for these models, where before you ran them maybe occasionally because you don't want to wait five minutes, but now it becomes more routine that they are part of your workflow basically.

我想补充一点，LLM 已经变得非常擅长自己知道需要多少努力才能对查询给出好的答案。所以我发现自己在绝大多数情况下，只是把提示输入到，比如说，ChatGPT，而不指定模型或思考级别，让它自己决定。如果我还想要更多，我会告诉它我想要更多。但它在判断什么时候只给我一个快速回答、什么时候使用搜索工具、什么时候做更多思考等方面做得相当不错。

I would expand on that and say that the LLMs have gotten really good at themselves knowing how much effort is required to provide a good answer to a query. And so I find myself, in the vast majority of times, just typing my prompt into, say, ChatGPT, for example, and not specifying a model or level of thinking and letting it figure it out. And if I want more, I'll tell it I want more. But it does a fairly good job of determining when to just give me a quick answer, when to use a search tool, when to do more thinking, that kind of thing.

我同意。我把 ChatGPT 的设置调成了自动模式，它会自动决定是否应该使用更多或更少的思考努力。同样的事情。我唯一还用专业模式的场景是——回到我提到的章节——当我写完一个章节，比如一个 40 页的 PDF，我会把它上传到那里，然后说，“嘿，你能检查一下有没有不一致的地方、错误的编号之类的吗？”然后我把它设成专业模式，就是那种要花 20 分钟的模式。我去吃午饭或晚饭，回来再看结果。那是我很少做的事情，我是说一个月一次，我写完一个章节之类的，或者如果我写了什么重要的东西，想要最大程度的质量检查。但就像你说的，对大多数任务来说，用轻量努力就足够了。对。或者用自动模式，它基本上自己决定。

I agree. I have my setting on ChatGPT on auto mode, where it automatically or by itself decides whether it should use more or less thinking efforts. The same thing. The only context where I still use the pro mode is when—coming back to the chapter I mentioned—when I have a chapter written like a 40-page PDF, I would upload it there and say, 'Hey, can you check for any inconsistencies, incorrect numbering, and all that type of stuff?' And then I set it to the pro mode, like the one that takes 20 minutes. I go have lunch or have dinner, come back, and look at the results. Where there, it's like a rare thing that I do, I mean once a month I finish like a chapter or something, or like if I write something important where I want the maximum quality check on that. But like you said, for most tasks, it's sufficient to use the light effort. Yeah. Or the automatic one where it decides by itself essentially.

我提到了 Multbot 和那个工具的发布。你花了很多时间研究它吗？

And I mentioned Multbot and the release of that tool. Have you spent much time digging into that?

嗯，是的。Multbot，我想现在叫 Open Claw 了。

Well, yeah. Multbot, I think now called Open Claw.

名字改了。对，它变了不少。

Open the name. Yeah. It changed a bit quite a bit.

它很有趣。就像一个本地智能体，人们现在可以在自己的电脑上运行。

It's interesting. It's like this local agent that people can now run on their own computers.

我觉得它有趣的地方在于，它能让人们对事物感到兴奋。这有点像当年 DeepMind 推出 AlphaGo（下围棋的模型）的时候。那真的很令人兴奋，因为在我认识的人里，之前没多少人下围棋，但当我家人和所有人看到它和世界冠军对弈时，都特别激动。我觉得 Moldbot 也类似：它让人们有兴趣去了解这些东西，并感到兴奋。我认为它也有很多实际用途，比如你可以用它来管理日历和邮件。就我个人而言，我还没这么做过。可能我有点信任问题。我不确定是否足够信任它来处理我的财务或日历。我对此还是有些犹豫，但我认为这是一个很酷的展示，让人们——尤其是不开发 LLM 的人——看到 LLM 能做什么以及它们的目的是什么。我觉得这真的很酷。

I think what I find interesting about it is it gets people excited about things. It's almost like back then when DeepMind had AlphaGo, the Go-playing model. It got really exciting because not many people in my circles played Go before, but it got my family and everyone excited to see this type of progress when it was playing against the world champion. I think with Moldbot, it's similar: it gets people interested in checking these things out and excited. I think there's also a lot of genuine use cases around it where you can run it to organize your calendar and emails. For me personally, that's something I have not done. Maybe I have a little bit of a trust issue. I don't know if I trust it enough to do my finances or my calendar. I'm a bit hesitant still to adopt something like that, but I think it's a cool demonstration of what LLMs can do and what their purpose is, especially for someone who is not developing LLMs. I think that's actually quite cool.

还有其他你依赖的、主要围绕 LLM 封装的服务或工具吗？还是说你主要直接使用模型本身或开发环境？

Any other tools or services that are largely wrappers around LLMs that you have come to depend on, or do you find yourself mostly turning to the models themselves or the dev environments?

我的工作流主要还是那样。我没有那种需要增量运行或智能体式设置的超级自动化工具。但我一直在做的是开发自己的应用，比如效率工具。我从小就是个程序员，用 bash、终端和 Python，给自己写脚本来自动化各种事情。现在有了 LLM，我有点转向开发原生 macOS 应用了。我一直想学 Swift 编程，但一直没时间，因为有太多其他重要的事要做。这成了一个机会，让我说，嘿，我想把这个脚本做成原生 macOS 应用，因为它更方便。比如前几天，我妻子也有一个播客——一个读书俱乐部播客——我帮她处理剧集，比如上传、编辑以及整个工作流，因为她不是技术背景的人。我原来有个脚本用来给播客添加章节标记。现在我做了一个原生 macOS 应用，你只需输入时间戳并点击按钮，它就会给音频文件添加章节标记。就是这种简单的东西。我可以和她分享，她现在就能用了。这些日常生活中的小便利：不用手动操作，而是自动化。这不是在运行 LLM，而是用 LLM 来开发一个确定性行为的工具。我更像那种人。比如，我读社交媒体时，主要是个对论文感兴趣的研究者。我经常收藏很多 arXiv 链接或摘要，然后我有一个 markdown 文件，里面有很多这些链接。现在我给自己写了一个原生 macOS 应用，只需输入这些链接，它就能提取标题、日期、作者姓名和链接，以漂亮的格式呈现，让我不用一个个点击。我得到一个漂亮的列表，可以看到标题。我觉得像这样的小事——我觉得 LLM 超级酷，可以用来开发这些我本来没时间开发的工具。

It's mostly still for my workflows. I don't have anything super automated where I need to run something incrementally or in an agentic setting. What I've been doing a lot though is developing my own apps, like productivity apps. I grew up as a coder using bash, the terminal, and Python, writing scripts for myself to automate things. Now with LLMs, I've shifted a bit towards developing native macOS apps. I always wanted to learn Swift coding, but never had the time because there were so many other important things to do. That became an opportunity to say, hey, I want this script as a native macOS app because it's more convenient. For example, just the other day, my wife also has a podcast—a book club podcast—and I help her with the episodes, like uploading everything, editing, and the workflow in general because she's not a tech person. I had a script to add chapter marks to the podcast. Now I made a native macOS app where you can just add the timestamps and click a button, and it adds the chapter marks to the audio file. Simple things like that. I can share it with her, and she can use it now. These little quality-of-life things in everyday life: instead of doing things manually, you can automate them. This is not running the LLM, but using the LLM to develop something that behaves deterministically. I'm more of a person who does that. For example, I also have, when I read social media feeds, I'm mostly a researcher interested in papers. I often end up bookmarking a lot of arXiv links or the abstracts, and then I have my markdown sheet with many of these links. Now I wrote myself a native macOS app where I just put in these links, and it pulls out the title, date, author names, and the link in a nice format, making my life easier so I don't have to click on them individually. I get a nice list and see the titles. I think little things like that—I feel like LLMs are super cool for developing these tools that I would not have time to develop otherwise.

这跟我的经历很相似。我觉得在过去一年左右，我从 LLM 中获得的最大好处就是编写自定义工作流工具。主要是围绕播客：我们和赞助商合作时，需要拉取分析报告，这很重复且耗时。所以我创建了一个基于网页的工具，它调用获取分析的 API，拉取剧集信息，你可以选择一集，然后把一堆数据导入 pandas，做一些分析，最后生成一个像 Google 文档那样的电子表格。这个应用本身没有用 LLM，但 LLM 被用来创建它。这只是大概六七个对工作流有重大影响的工具之一。

That parallels my experience quite a bit. I think some of the most benefit I've gotten out of LLMs in the past year or so has been writing custom workflow tools. Primarily around the podcast: one of the things we do when we work with sponsors is pull these analytics reports, and it was repetitive and time-consuming. So I created a web-based tool that hits the API where we get the analytics, pulls information about episodes, and you can choose an episode, then pull a bunch of data into pandas, do some analysis, and generate a spreadsheet like a Google Doc. The app isn't using an LLM, but an LLM was used to create it. That's one example of probably half a dozen fairly significant tools that have a big impact on our workflow.

说得好：在这些情况下，LLM 并不是在做常规工作或任务；它更多的是在开发完成任务的工具。我认为这一点很重要。LLM 非常有用且能力很强，但有些任务用 LLM 几乎是一种浪费。就像，如果你只有一把锤子，所有东西看起来都像钉子。我确实认为，如果任务是确定性的，那么开发一个确定性工具仍然是有意义的。你可以用 LLM 来做，但甚至问 LLM 一加一等于几都几乎是浪费。你可以用计算器。所以我认为，认识到问题的本质以及什么工具最适合那个问题，仍然很重要。

That's a good point: in these cases, the LLM is not doing the regular work or the task; it's more developing the tool to do the task. I think that's an important point. The LLM is very useful and capable, but there are tasks where it's almost wasteful to use an LLM for that. It's like, if all you have is a hammer, everything becomes a nail. I do think if you have a deterministic task, it still makes sense to develop a deterministic tool. You can use an LLM for that, but it is almost wasteful to even ask an LLM what is one plus one. You can use a calculator. So I think it's still important to recognize the nature of the problem and what is the best tool for that problem.

我还做过一些工具，把 LLM 几乎当作分类器来用，一个非常简单的用例。我有一个工具，输入嘉宾的名字，然后我从 Google Docs API 拉取一堆最近的目录，让它找到对应这位嘉宾项目的目录。正则表达式或文本模式匹配并不总是有效，因为它们有时会不同，但 LLM 可以很容易地做到，重复性非常高，错误率很低。

I've also done some tools where I'll use LLMs almost like a classifier, a very simple use case. I have one where there's the name of the guest, and then I pull a bunch of recent directories from the Google Docs API and say, find the directory that corresponds to the project for this particular guest. A regex or text pattern match doesn't always work because they can be different sometimes, but an LLM can do it pretty easily with a very high level of repeatability and a low error rate.

嗯，这就像当你需要近乎人类或某种不那么结构化的方法时，LLM 就很适合。我大学时其实也做过一个类似的项目。

Well, it's like where you need almost a human or some less structured approach, then LLMs are great for that. I had actually a similar project as a college student.

我做过一个体育预测的副项目，纯粹为了好玩，比如每日梦幻体育，预测周末英超联赛中哪个球员会进球。为此，我还开发了一个非常复杂的系统，从不同网站抓取球员信息，查看谁受伤了、谁状态好等等。现在我用 LLM 重新启动了那个项目，纯粹为了好玩。这跟你提到的名字问题类似，因为有些球员的名字拼写略有不同，某些字母上有重音符号，但有些数据库里没有，然后要把它们对齐到数据库里。用正则表达式或确定性方法很难处理。所以这其实是 LLM 的一个很好的用例，用来解析这些非结构化的、模糊的数据集，这些任务也依赖一些上下文。

I was doing a sports prediction as a side project for fun, like daily fantasy sports, predicting outcomes like which player scores a goal in the Premier League soccer on the weekend. For that, I was also developing a very sophisticated thing which was pulling information about the players from different websites and looking at who's injured, who's in good form, and these types of things. Now I kind of revived that project just for fun using an LLM. It's kind of the same problem you mentioned with the names because some players have the spelling of the name slightly different. There are these accents over certain letters and sometimes not in certain databases, and then just getting them lined up in the database. It's really hard with regex or just deterministic things. So that is actually a great use case for an LLM to kind of parse these unstructured, vague data sets, things that depend also a bit on context.

那么，让我们回到 LLM 的实际应用层面。我觉得 Sam 和 Seb 都在用它们。主要得出两点。如果你有开发思维，但我们也看到，通过“氛围编码”，即使技术不太强或非技术人员也能通过创建自定义工具来自动化工作流中的特定部分，从而获得巨大价值。这是过去一年左右对我们俩影响很大的事情。另外，就是利用模型本身的改进。对我来说，我无法明确说出一套规则，但面对特定任务时，我内心有一个模糊的模型，知道该先用 ChatGPT 还是 Claude。关键是，我们俩都没有规律地使用 OpenClaw 或任何特别花哨的 LLM 封装智能体工具。可能对我来说，例外是像 Circleback 或 Granola 这样的会议摘要工具，但除此之外，主要还是通过原生聊天界面和面向开发的用例。

So maybe pulling back to where we are with LLMs from a practical perspective. I think both Sam and Seb are using them. Two main things came out of this. If you are in a development mindset, but we've seen with vibe coding that even less technical or non-technical people can get a lot of value by creating custom tools to automate specific parts of their workflows. That's a huge thing that has been very impactful for both of us over the past year or so. Otherwise, taking advantage of the improvements in models. For me, it's like I can't really articulate a rule set, but if I'm confronted with a particular thing, I have a soft mental model for whether I'll start with ChatGPT here or Claude for that. The takeaway is that neither of us are using open claw or any particularly slick LLM wrapper agentic tools with any regularity. Maybe the caveat for me would be something like Circleback or Granola to do meeting summaries, but beyond that, it's mostly use cases through the native chat interfaces and development-oriented use cases.

我想再补充一点。这其实是一个滑块：你可以完全不使用 LLM，仍然手动完成所有事情；也可以只用 LLM。我认识一些人，他们甚至仅基于 LLM 代码创办了一家公司，人们称之为“氛围编码”，但我认为“氛围编码”甚至不足以形容它，比如不再做任何手动编码，只用 LLM 来构建网站、产品和一切。这两个极端都存在，而我认为我们处于中间地带，我们采用 LLM，但不会完全依赖 LLM。对我来说，学习编程仍然有价值。对于现在学习编程的人来说，这值得吗？我认为即使有 LLM 能做这些，学习数学和编码仍然值得，因为这会让你的生活更高效，并让你更擅长使用这些 LLM。例如，我用 LLM 给我的网站添加了深色模式。网站是我 12 年前自己写的，但那时我对 HTML、CSS 和 JavaScript 的了解比现在好得多。我一直拖延添加深色模式按钮，因为要花一个月才能做好，而且这不是我的主要工作。所以我想，试试用 LLM 来做吧。它做得很好，但并不完美。按钮没有对齐。然后我让它调高、调低、左移，我觉得这样效率很低。为什么不直接进入 HTML 或 CSS 文件调整设置呢？因为我仍然懂一点 CSS，我自己做这些调整比让 LLM 暴力重做更有效。从这个意义上说，理解这些东西的工作原理确实有意义，因为有些情况下自己动手比提示 LLM 重做所有事情更高效。所以存在一个中间地带，学习工作原理仍然有价值。

I would add one more thing. It's also a slider: you can use LLMs not at all, you can still do everything manually, or you can only use LLMs. I know some people who develop even a company just based on LLM code, people call it vibe coding, but I think vibe coding doesn't even do it justice anymore, like not doing any manual coding anymore, just using LLMs to build the website, the product, and everything. These two extremes exist, and I think we are more in the middle, where we adopt LLMs but are not going full LLM. For me, I think there is still value in learning how to program. For people who are learning how to program nowadays, is it worthwhile? I think it is still worthwhile to learn math and coding even though there are LLMs that can do that, because it makes your life more efficient and makes you better at using these LLMs. For example, I used an LLM for my website to add a dark mode. I wrote the website myself 12 years ago, but I knew HTML, CSS, and JavaScript much better back then than I do now. I always procrastinated on adding a dark mode button because it would take me a month to do it well, and it's not my main job. So I thought, let me try using an LLM for that. It did a really good job adding it, but it was not perfect. The button was misaligned. Then I told it to make it higher, lower, move it left, and I thought this was very inefficient. Why don't I just go into the HTML or CSS file and adjust the settings there? Because I still knew a bit about CSS, I was more effective making those adjustments myself instead of having the LLM brute force everything. In that sense, it does make sense to have an understanding of how these things work, because there are cases where it is more efficient to do things yourself than to prompt the LLM to redo everything. So there is a middle ground where there is still value in learning how things work.

我想知道你的体验。我经常在社交媒体上看到新模型发布时有人说：“哦，我一次就搞定了这个。我一次就搞定了那个。”我试着回忆上一次我有这种体验是什么时候。然后我去尝试一次搞定同样的事情，得到的结果却很糟糕，跟社交媒体上报道的完全不一样。是我有问题，还是人们只是为了互动而报道这些成功，其实并不真实，或者是假的？你怎么看？你有类似的经历吗？

I wonder what your experience is. I often see around these new model releases on social media: "Oh, I one-shotted this. I one-shotted that." I try to remember the last time I had that experience. Then I go and try to one-shot the same thing, and the results I get are horrible, nothing like what is reported on social media. Is it me, or are people just reporting these successes for engagement and they're not really there, or they're fake? What's your sense? Do you experience similar things?

是的，我也这么觉得。我提到过我的原生 Mac 应用，甚至是一个简单的 Mac 应用，只需放入 PDF 就能导出特定分辨率的 PNG、WEBP 和 PDF 版本。即使是用当时的 Claude 5.2，也花了多次尝试才让所有按钮正常工作。就像你说的，根本不是一次搞定的。即使这么简单的东西，也经过多次迭代才成功。我有时会想：是不是我的指令不好？也许我没说清楚。也许你必须说“请彻底测试所有内容，确保一切正常”，并且要非常明确地说明，而我们没有这么明确，因为我们假设它会确保一切正常。或者，我们看到的情况只是运气好，有时候某些事情碰巧就运行得很好。

Yeah, I would say so. I mentioned my native Mac app, even something like a Mac app where I just put in a PDF and it exports PNG, WEBP, and PDF versions in a certain resolution. It took multiple tries even with Claude 5.2 back then to get all the buttons working correctly. Like you said, it was not one-shotted at all. It was multiple iterations to get it to work, even for something simple like that. I sometimes wonder: are my instructions maybe bad? Maybe I wasn't clear. Maybe you have to say "please test everything thoroughly and make sure everything works" and be super explicit about that, and we are not that explicit because we kind of assume it will make sure everything works. Or maybe these cases we see are just lucky, sometimes on certain things it just happens to work very well.

所以，我不太确定，但我同意你的看法，当有人展示说“哦，我一次性搞定了”，这并不像表面看起来那样。我不认为这反映了今天的工作方式。那么，我们稍微换个话题，聊聊你认为在未来一年里，LLM 会在哪些关键领域持续创新。然后针对每个领域，我们深入探讨一下近期的历史以及你预期的走向。

So, I don't know for sure, but I agree with you that it's not all what it seems when someone shows you, oh, I one-shotted this. I don't think that's reflective of how things work today. So, let's switch gears a little bit and talk through some of the key areas that you expect to see continued innovation around with LLMs in the upcoming year. And then for each of them we'll dig in and talk a little bit about the recent history and where you expect to see things going.

我认为仍然是推理——我们可以深入探讨，因为这是一个非常宽泛的话题。所以，在 后训练 中进一步推动推理。第二个是 推理 时的 算力 扩展，比如更复杂的技术，部分与训练相关，但主要是如何在训练后使用 LLM。然后，我认为我们还会看到更多这种 智能体 式的使用方式，因为目前大多数 LLM 都专注于逐轮交互，而人们会加倍投入这种循环，基本上像 mold bot 那样将 LLM 作为循环运行并优化。我认为这三件事将是公司最主要的关注领域。

I would say it's still going to be the reasoning — we can maybe go into more detail there because it's a very broad topic. So, pushing more on the reasoning front in post-training. The second one I would say is also inference scaling, like more sophisticated techniques that are partly related to training but mostly how to use the LLM after training. And then I also think we will see more of this agentic type of use, like how — because right now mostly LLMs are focused on a turn-by-turn basis, and how people will double down on this loop, basically running LLMs as a loop like mold bot and optimizing for that. I think these three things will be mainly the biggest focus areas for companies.

太棒了。那我们深入聊聊推理。为了铺垫你认为 2026 年的方向，你觉得 2025 年在推理方面有哪些重大进展？

Awesome. So let's dig into reasoning. To set the stage for where you think we'll be heading in 2026, what do you think were the big advancements in 2025 around reasoning?

最大的进展是，首先 OpenAI o1 让所有人都很兴奋。OpenAI o1 同时使用了 推理 时的 算力 扩展——没人确切知道，因为没有论文，但可能也用了训练技术。然后 DeepSeek R1 发布了他们的推理流程——抱歉，预训练——推理流程。我认为这确实是一个突破，很多其他公司也纷纷跟进。但总体来看，这仍然非常新，才一年左右。我最近在写一章关于推理的内容，算法有很多改进。就在前几天，我整理了一份清单，列出了 15 种不同的调整和改进，从基本的东西比如将序列级别的对数概率改为 token 级别，到 NVIDIA 的 GDPO。这方面进展很大，我认为我们还会看到更多。但还有一个原因是，预训练 基本上仍然有效，而且我认为它仍然是整个训练流程中最大的部分，因为数据量巨大且非常昂贵。但研究团队的研发重点如今更多地放在 后训练 上，比如从中获得更多性能，因为它更像是一个新范式，而且还有低垂的果实可以摘取，而 预训练 已经相当成熟了。你仍然需要大量数据和算力，但在算法调整以获取更多性能方面，与 后训练 相比，你能做的确实不多。当然，你仍然可以这样做，如果使用更多数据、优化数据混合、可能还有多 token 预测之类的方法，你仍然会得到更好的结果。但大多数有趣的事情现在都发生在 后训练 的推理领域。所以我认为我们会在那里看到更多进展。

So, the biggest advancement was, I mean, first OpenAI o1, which got everyone excited about it. And then OpenAI o1 was using both inference scaling and — I mean, no one knows for sure because there's no paper, but likely also training techniques. But then DeepSeek R1, they published their reasoning pipeline — sorry, pre-train — the reasoning pipeline. And I think that was really something that took off, where a lot of other companies also doubled down on that. But it's still very new in the grand scheme of things. It's just like a year old, and there have been — I was recently working on a chapter on reasoning. There were so many improvements to the algorithm. I mean, just the other day I compiled a list of 15 different tweaks and improvements, from basic things like changing sequence-level log probabilities to token-level, but also GDPO by NVIDIA. Lots of progress there, and I think we will see more of that. But also one reason is that with pre-training, we have seen basically that it still works, and I think it's still the biggest part of the whole training pipeline because it's just so much data and very expensive. But the R&D focus of the research team, I think, is more focused nowadays on post-training, like getting more performance out of that because it's more like the newer paradigm and there are still low-hanging fruits to be picked, whereas in pre-training it's already pretty sophisticated. You still need a lot of data, you still need a lot of compute, but there is nothing you can really do much there compared to post-training in terms of changing up the algorithms to get more performance. Of course, you can still do that and you will still get better results if you use more data, optimize the data mix, maybe multi-token prediction and these types of things. But most of the interesting things are happening now on the post-training front in the reasoning realm basically. So I think we will see more there.

在推理方面，去年我听到一个经常被提及的话题是“可验证奖励”的概念。我认为这促成了我们在编码模型方面看到的许多进展。你能谈谈这个范式以及过去一年我们看到的一些重要里程碑吗？

On the reasoning front, one topic that I heard come up quite a bit last year is the idea of verifiable rewards. And I think that led to a lot of the advancements, or contributed to a lot of the advancements, that we saw in terms of coding models. Can you talk about that as a paradigm and some of the big milestones that we've seen there over the past year?

是的，谢谢你的问题。这确实是一个非常关键的点。推理训练本质上主要基于可验证奖励，这意味着有些任务你可以验证答案。例如，在 DeepSeek R1 中，可验证奖励是编码和数学。以数学为例，你让模型以 LaTeX 的方框格式输出最终答案，比如 LaTeX 命令 \boxed{}，然后你可以用确定性的正则表达式或代码提取答案，再用 SymPy 或简单比较参考答案。如果 2/3 匹配 2/3，或者 4/6 匹配 2/3，本质上是相同的答案，但你可以符号化地双重检查答案，并获得一个奖励信号来判断是否正确。这实际上很棒，因为你可以无限地评估无数个答案。以前，使用基于人类反馈的强化学习（RLHF），你基本上需要人类反馈；你可以训练一个奖励模型来近似，这是训练的一部分，你会为每个答案得到一个分数，但它不如一个真正正确的答案那么准确——比如数学，要么对要么错，是绝对的。如果你有这种可以确定性地、低成本地验证答案的方法，你可以让 LLM 生成无数个答案。你可以说，好吧，为这个问题生成 6 万个答案，然后在很短的时间内计算所有答案的奖励。生成这些答案仍然很昂贵，但你没有模糊性，也不需要人类来评估这些答案。所以我认为这有助于扩展这些东西。代码也一样：在 DeepSeek R1 论文中，最初的方法是让代码基本能编译——如果编译正确，你也可以使用代码解释器。我认为两者都很棒，但这只是开始。我们可能会看到它扩展到不仅仅是正确性奖励。已经有其他类型的奖励被加入。例如，格式奖励，你希望模型使用——不是必须的，但有些公司更喜欢将思考放在 think 标签中。所以他们有一个 token <think> 和一个结束 token </think>，就像 HTML 一样。这不是必须的，但有助于解析中间内容并加以利用，你可以训练模型输出这种结构。这被称为格式奖励。

Yeah, thank you for the question. That's actually a really important point. So the reasoning training is essentially mainly based on verifiable rewards, which means there are tasks where you can verify the answer. For example, in DeepSeek R1, the verifiable rewards were coding and math. With math, for example, you ask the model to output the final answer in a boxed format in LaTeX, like a LaTeX command \boxed{}, and then you can have a deterministic regex or code to extract the answer, and then you can use something like SymPy or simply compare to a reference answer. If 2/3 matches 2/3 or 4/6 matches 2/3, it's essentially the same answer, but you can symbolically double-check the answer and get a reward signal whether it's correct or not. This is actually great because you can kind of infinitely evaluate infinite numbers of answers. Before, with reinforcement learning from human feedback, you need human feedback essentially; you can train a reward model to approximate that, and it's part of the training where you get a score for each answer, but it's not quite as accurate as a really correct answer where you can verify the answer — like there's an absolute, it's math, it's either correct or not. And if you have something like that where you can verify the answer deterministically and cheaply, you can have the LLM generate infinite answers. You can say, okay, generate 60,000 answers for this problem, and then you can calculate the reward on all of them in a very short time. It's still expensive to generate these answers, but you don't have vagueness and you don't have to have human evaluating these answers. So I think that helps scale these things. And the same with code: in the DeepSeek R1 paper, the original approach was to have the code and make sure that the code compiles basically — if it compiles correctly, and you can also use a code interpreter for that. I think both are great, but I think this is just the beginning. We will probably see this being extended to have more than just the correctness of reward. There are already other types of rewards being added. For example, a formatting reward where you want the model to use — it's not required, but some companies prefer to have the thinking in the think tag. So they have like a token <think> and then a closing token </think> like in HTML. It's not required, but it can be helpful to have it because then you can parse out the intermediate stuff and do something with it, where you can train the model to output this structure. This is called a format reward.

所以除了正确性奖励之外，你还可以向这个东西添加多种类型的奖励。我认为我们可能还会看到一些有趣的事情，人们会提出格式化奖励或辅助奖励，帮助整个模型学习。还有一件事是，他们在 DeepSeek R1 论文中尝试评估答案的解释，而不仅仅是看最终分数或是否正确。确保或评估推理（即解释）是否正确。

So you can have multiple types of rewards added to this thing in addition to the correctness reward. And I think we will maybe also see interesting things there where people will come up with formatting rewards or auxiliary rewards that help the overall model to learn. And one thing is also they tried that in the DeepSeek R1 paper to evaluate the answer explanation instead of just looking at the final score or if it's correct or not. Making sure or evaluating if the reasoning like the explanation is correct or not.

过程奖励。

Process reward.

没错。是的。他们使用了一种叫做过程奖励模型的东西，基本上是另一个你训练的模型，用来给这个解释打分。但我记得，R1 论文发布已经有一段时间了，他们有一个章节将其列为失败的尝试或不成功的尝试。所以他们试过了，但认为这增加了奖励黑客攻击的风险，然后就不值得了。它更昂贵，导致了奖励黑客攻击，模型利用漏洞，因为这样模型更容易作弊，误导评估模型的模型。所以做这件事仍然很棘手。但我认为在最近几个月里，也有一些更有趣的成功案例，比如 DeepSeek Math 3.2 版本。他们使用了类似的方法，用评分标准评估整个答案，用另一个模型来做，然后还有另一个模型来评估那个有评分标准的模型，依此类推。这就像多个层次，似乎有效。他们还有消融研究表明这确实有帮助。我认为我们还会看到更多这样的做法。这基本上是一个全新的范式，让推理训练变得更加复杂。

Exactly. Yeah. They used a called process reward model that is basically another model that you train to give a score for this explanation. But I remember, I mean it's been a while since R1 came out in the paper, they had a section that they listed that as a failed attempt or unsuccessful attempt. So they tried it but they thought okay this increases the chance of reward hacking and then it was just not worth it. It's more expensive, it resulted in reward hacking, the model exploiting thing because it's easier that way for the model to kind of cheat, to mislead the model that evaluates the model basically. And so it is still tricky to do that. But I think also in the recent months there were some more interesting success stories like DeepSeek Math version 3.2 too. They used something like that where they evaluate also the whole answer with a rubric, have another model for that, and then they have another model that evaluates that model that has the rubric and so forth. It's like multiple levels and that seems to work. And they had ablation studies that showed that this is actually helping. And I think we will see also more of that. It's just like a very new paradigm, making the reasoning training more sophisticated essentially.

目前验证器主要关注数学和编码，这是因为对于给定的响应，有具体的验证能力。你认为这种验证范式会扩展到数学和编码之外吗？我认为部分原因是对数学和编码的关注之所以成功，是因为尽管并非所有 LLM 响应都涉及数学和编码，但这些事情本身具有内在的逻辑或推理能力，因此模型的推理能力可以泛化到非数学和编码问题。但你是否看到将验证概念扩展到数学和编码问题类型之外的关注点？

Right now the verifiers are focused on math and coding, and that's because for a given response there's a concrete ability to verify. Do you see this verification paradigm expanding beyond math and code? And I think in part the focus on math and code is successful because even though not all LLM responses are about math and code, those things kind of have an inherent logic or reasoning capability in them, and so the ability for the model to reason generalizes to non-math and coding problems. But do you see a focus on expanding this idea of verification beyond math and code types of problems?

是的。这实际上是一个非常有趣且重要的观点。我们如何做到？我的意思是，你提到如果你在数学问题上训练模型，进行数学推理，它也会在一般推理上变得更好。但如果你有一个目标领域，专门在该领域训练模型进行推理，那会更好。我认为你是对的，会有更多这样的情况。目前我只是缺乏创造力，想不出可以验证的问题的例子。但我会说，也许像生物学相关的东西，比如药物设计（如制药药物设计）或蛋白质结构建模，其中有物理约束。比如原子之间的角度，它们只能有特定的角度等等，你可以有一个物理类型的方程来双重检查生成的分子是否符合这些特定格式，然后将其作为训练模型时的一种奖励形式。我的意思是，这可能不是典型的推理案例，因为当你生成模型时，推理解释是什么？但总的来说，其他领域也可以有类似的东西。在最坏的情况下，你总是可以——这更像是一个粗略的近似——训练另一个提供正确性奖励的模型。但我认为这更具挑战性，因为它容易受到奖励黑客攻击的影响。甚至回到过去，生成对抗网络，生成器很容易崩溃。你有一个判别器，它说这张图像是真实的还是生成的，然后设置是训练生成器来欺骗判别器，判别器变得更好于区分，你几乎有一个类似的设置。你可以用它来给予奖励或不给予，但模型可能会也可能不会利用它。在某个时刻，它学会了一个技巧：如果我只生成这一个词或类似的东西，那么我就骗过了那个评估器。但我认为我们可能还会看到更多这样的做法，比如开发基于 AI 的奖励模型，然后可以在其他领域用于训练更好的推理模型。

Yes. So it's actually a very interesting and important point. How can we? I mean you mentioned that if you train the model on math problems, reasoning on math, it will also become better at reasoning in general. But then it would be even better if you have a target domain to train the model specifically on that target domain on reasoning in that target domain. I think you're right, there will be more of that. For me right now I just lack the creativity to come up with examples of problems that can be verified. But I would say maybe something even like biology related where for a drug design like a pharmaceutical drug design or protein structure modeling where you have physical constraints. So there are like the angles between atoms, they can only have a certain angle and so forth, where you could probably have a physics type of equation that double checks whether the generated molecule adheres to these certain types of formats and then have that as a form of reward when you're training the model. I mean this is maybe not a typical case of reasoning because what is the reasoning explanation when you're generating a model? But in general, something like that for other fields. And in the worst case you can always, I mean this is more like a rough approximation, but you can always train another model that provides the correctness reward. I think this is more challenging though because it's susceptible to reward hacking. Even going back to back in the day, generative adversarial networks where it's easy for the generator to collapse. You have the discriminator which says is this image real or generated, and then it was like the setup where you train a generator to fool the discriminator and the discriminator gets better at distinguishing, and you have almost a similar setup. You can use it to say give a reward or not, but then the model may or may not exploit it. At some point it learns a trick: if I only generate this one word or something like that, then I fool that evaluator. But I think maybe we'll see also more of that, like developing AI-based reward models essentially that can be used then in other fields to train better reasoning models.

除了对验证模型的更多关注和调整之外，你是否看到其他领域有助于未来更强的推理？

Beyond increased focus and tweaks to the verification models, are there other areas that you see as contributing to stronger reasoning going forward?

是的，我确实认为训练是一部分，但另一部分是推理时 Scaling（规模扩张），如果你使用简单的——我是说带引号的“简单”——但如果在训练后你花费更多的算力，你可以获得更好的性能。推理时 Scaling（规模扩张）就是在训练后、推理时花费更多算力，当有人使用模型生成答案时，你可以通过多种方式做到这一点。我的意思是，推理模型本身已经是一种推理时 Scaling（规模扩张）的形式，因为它们比常规模型生成更多的 token。解释比常规模型提供的更长，但它通常有助于 LLM 达到正确答案。但这更像是顺序推理时 Scaling（规模扩张）。你还可以有并行形式的推理时 Scaling（规模扩张），即生成多个答案，这被称为自一致性。例如，如果你有一个数学问题，你可以让 LLM 使用不同的温度设置多次回答问题，然后进行多数投票或类似操作。有不同的方法可以做到这一点。还有不同的评分方法或其他 LLM 查看所有答案并给出最可能正确的答案。这样你也可以提升模型的性能。不过它更昂贵。所以总是这样——没有一刀切的方案。你不会一直使用它。你在需要的时候使用它。但我认为有趣的是改进判断何时需要的方法。我认为当 ChatGPT 5.1 或 5 发布时，他们有我们一开始讨论的那个自动设置。

Yeah, I do think it's also the training is one part, but the other one is the inference scaling that you can get much better performance if you use simple — I mean simple in quotation marks — but if after training you spend more compute essentially. Inference scaling is all about spending more compute after training during inference when someone uses the model to generate the answer, and you can do it in multiple ways. I mean reasoning models themselves are already kind of a form of inference scaling because they generate more tokens than regular models. The explanation is longer than a regular model provides, but it helps often the LLM to reach the correct answer. But this is more like a sequential inference scaling. You can also have parallel forms of inference scaling where you just generate multiple answers, and that's called self-consistency. So for example, if you have a math problem, you can have the LLM with different temperature settings, answer the question multiple times, and then you take a majority vote or something like that. There are different ways you can do it. There are also different scoring methods or other LLMs that look at all the answers and give you the most likely correct answer. And with that you can also boost the performance of the model. It's more expensive though. So it's always like this — it's not a one-size-fits-all. You don't want to use it all the time. You use it when you need it. But I think what will be interesting is to improve the way to tell when it's needed. I think when ChatGPT 5.1 or 5 launched, they had that automatic setting that we talked about in the beginning.

一开始效果很差，但我觉得几个月、几年下来已经好多了。我不太确定开源开放权重生态里有没有类似的东西——也许听众可以纠正我——但我认为这类东西会变得越来越重要。一方面，我们在开发非常昂贵的模型，它们能解决像数学奥赛那样极难的问题，但我们不想一直用它们，因为它们又慢又贵。与此同时，也会更关注更便宜的模型。比如，就在前几周，Qwen 3 Coder Next 发布了。Qwen 3 是使用最广泛的开放权重模型之一，因为他们有各种尺寸的高质量模型。但 Next 模型本质上是个混合体，不再是纯粹的 Transformer 了，它借鉴了状态空间模型来降低成本。但这始终是个权衡：人们在开发更高精度的模型，也在开发更便宜的模型。一种方法是改变架构来控制质量和价格，另一种是推理扩展。但目前开放权重生态里，推理扩展还不那么流行。所以我认为我们会在本地工具等地方看到更多这类技术。我不知道有任何开源项目或模型集成了这个，但从交流中我感觉到，很多围绕 Qwen 等开放权重模型构建的公司，通常会在架构里放一个路由组件，用来评估提示的复杂度或类别，然后路由到最经济或可能经过后训练以获得更好响应的模型和提示。我觉得这是应对你描述的这个挑战的常见方法。

It was very bad at the beginning, but I think it got much better over the months and years. I'm not quite sure we have anything like that in the open source open weight ecosystem. Maybe listeners may correct me here, but I think something like that I can see also being more important. On the one hand, we are developing these very expensive models that can solve very hard problems like in this math olympiad, but we don't want to use them all the time because they are slower and more expensive. There's also going to be more focus at the same time on cheaper models. For example, just the other week, Qwen 3 Coder Next came out. Qwen 3 is one of the most widely used open weight models because they have a lot of really high quality models in all different types of sizes. But the Next model is essentially a hybrid; it's not a pure transformer anymore. It's inspired by state space models to make things cheaper. But it's always this trade-off. People are developing higher accuracy models, and people develop cheaper models. One way would be changing the architecture to control the quality and price. The other one is inference scaling. But right now in the open weight ecosystem, it's not quite as popular yet. So I think we will also see more of that in local tools and so forth. I don't know of any open-source project or model that incorporates this, but from conversations, I do get the sense that a lot of companies building around the Qwen models and these open weight models commonly have a router component in their architecture that tries to assess the complexity or category of a prompt and routes it to the right model and prompt that is either most economical or maybe post-trained for better responses. My sense is that that's the common approach to addressing this challenge that you're describing.

你提到这个，让我想起另一个例子：GPT OSS 模型，OpenAI 去年夏天发布的开源模型。在那个模型里，即使你用像 Ollama 这样简单的推理工具，也可以在系统提示中设置推理努力程度。你可以说“无”、“轻度”、“中等”、“高度”推理努力，然后根据推理努力来扩展推理。但我不认为还有其他像自洽性或自我改进这样的技术被自动集成进去。大多数时候，研究人员得自己动手。

Now that you mentioned that, another example came to mind. It's the GPT OSS model, the open-source model by OpenAI, which came out last summer. In that model, even if you use a very simple inference tool like Ollama or any comparable tool, you can set the reasoning effort in the system prompt. So you can say no, mild, medium, high reasoning effort, and then scale inference based on the reasoning effort. But I don't think there's any other technique really automatically incorporated like self-consistency or self-refinement. It's mainly you have to as the researcher do it yourself most of the time.

你能多谈谈自我改进和自洽性，以及人们如何使用这些技术吗？

Can you talk a little bit more about the self-refinement and self-consistency and how folks use those techniques?

是的。自洽性和自我改进是推理扩展的两个例子。我认为两者最大的区别在于，一个是并行技术。自洽性作为并行技术会生成多个答案，然后你通过多数投票选择正确答案，或者用一个评分器来评估答案。人们把这种技术称为 best-of-n，几乎就像经典的集成方法。另一个是自我改进，你让大语言模型生成答案，然后把答案喂给另一个大语言模型或它自己，说：“这是答案，这是问题。写一个摘要，说明答案是否可能正确，以及有哪些弱点。”你提供一个包含某些检查项的评分标准，然后它返回一份报告，说：“这可以改进，这可能不正确，解释与最终答案不匹配。”然后你把那个输出反馈给原始大语言模型，说：“看看那份报告，根据报告改进你的原始答案。”这常常能让大语言模型改进自己的答案。这几乎就像一种现象：有时你问 ChatGPT 某个问题，它给出一个答案，你想：“等等，这不对。”你问某个模型是什么时候发布的，你知道年份完全错了。你告诉它：“你错了，你犯了个错误。”它说：“哦，你是对的，我犯了个错误。”然后它再试一次，下次就好了。这几乎是同样的机制。它有时也会让答案变差，比如它会过度思考，或者原本正确但反馈很奇怪或很差，然后答案就错了。所以这不是一个万无一失的技术，它有注意事项。但在 DeepSeek Math 3.2 版本的论文中，他们以更复杂的方式使用了自我改进，用第三个模型来评估评估者，他们展示了一张漂亮的图表，显示了准确率能提升多少。基本上，如果他们加大自我改进和自洽性的力度，就能在某些数学竞赛中达到金牌级别的表现，这非常令人印象深刻，因为用的还是同一个模型，只是加大了推理扩展的力度。

Yeah. Self-consistency and self-refinement are two examples of inference scaling. I would say the biggest difference between the two is one is a parallel technique. Self-consistency as a parallel technique generates multiple answers, and you choose the correct answer based on majority vote, or you can have a scorer who assesses the answers. People call that technique best-of-n, like classic ensembling almost. The other one is self-refinement, where you have the LLM generate the answer, and then you feed the answer to another LLM or to itself and say, 'Here's the answer. This is the question. Write a summary if the answer is likely answered correctly and what are weaknesses.' You provide a rubric with certain things that the LM should check, and then it gives you back a report and says, 'This could be better. This is likely incorrect. The explanation doesn't match the final answer.' Then you feed that output back to the original LLM and say, 'Look at that report and refine your original answer based on the report here.' Often this can lead to the LLM improving its own answer. It's almost like this phenomenon. Sometimes you ask ChatGPT something and it gives you something and you think, 'Wait, that can't be right.' You ask something like when a certain model was released, and you know the year is totally wrong. You tell it, 'You are incorrect, you made a mistake,' and it says, 'Oh yeah, you are right, I made a mistake,' and then it tries again and it's better next time. It's almost like that same mechanism. It can also sometimes make answers worse, like it will overthink, or it was originally correct but then the feedback is weird or bad, and then it makes the answer incorrect. So it's not a foolproof technique; it has caveats. But in the DeepSeek Math version 3.2 paper, where they had a self-refinement in a more sophisticated way with a third model evaluating the evaluator, they showed a nice graphic of how much the accuracy can improve. Basically, if they cranked up the self-refinement and self-consistency, they were able to achieve gold-level performance in certain math competitions, which was very impressive given it was still the same model as before, but they just cranked up the inference scaling.

反思我们讨论的这些主题，有趣的一点是它们都紧密相关。推理是一个关键主题，推理由推理扩展实现。我们在讨论中听到很多像循环和递归这样的概念，这些都是关键思想。你提到的第三个关键主题是大语言模型的智能体式使用。以此为引子，谈谈你到目前为止在智能体方面看到了什么，以及你认为这个领域有什么令人兴奋的地方。

One thing that's interesting reflecting on these themes we're discussing is just how they're all very interrelated. Reasoning is a key theme. Reasoning is enabled by inference scaling. A lot of what we're hearing as we talk about this is like loops and recursion, and those are key ideas. The third key theme you mentioned is kind of agentic uses of LLM. With that as a segue, talk a little bit about what you've seen thus far around agentic and what you think is exciting in that space.

智能体式用例，甚至像 Cursor 或 Claude Code 这样简单的东西，它们会进行多次迭代来解决问题。这不是一次性完成，而是更像执行一个任务，而不仅仅是提供一个答案。Multi-port 是另一个智能体系统的例子。智能体这个词定义得也不太好，因为人们用法不同。对于这个播客，也许我们可以把智能体理解为在循环中运行的东西。

The agentic use cases, even simple things like Cursor or Claude Code where it does multiple iterations to solve a problem. It's not just one shot; it's more like doing a task rather than just providing an answer. Multi-port would be another example of agentic systems. Agentic is also a not well-defined term because people use it differently. For this podcast, maybe we can think of agentic as something that runs in a loop.

在过去一两年里，我们在 LLM 的智能体式使用场景中经常听到的一个概念是多智能体系统，以及将问题分解为具有各自角色的独立智能体。我看到很多“我创建了我的 AI 团队”，这些“员工”通过 Slack 或 Notebook 互相交流。从具体的构建者或技术角度来看，你看到了什么？人们是否从中获得了很大价值？

One of the things we heard a lot about in the context of agentic uses of LLMs over the past year or two is the idea of multi-agent systems and decomposing a problem into independent agents with their own personas. I'm seeing a lot of 'I created my AI team' with employees that talk to each other using Slack or Notebook. What have you seen from a concrete builder or technical perspective? Are folks getting a lot of value out of that?

说实话，我希望我能给出一个很好的答案，但这并不是我亲自探索过的领域。我的大部分经验都是单一用例，即一个 LLM 提供解决方案或处理特定任务，但通常不与其他智能体交互。我认为这更像是一个上下文工程问题——如何将结果提供给另一个 LLM。这有点像图像或视频生成，其中一个 LLM 解析文本并将其传递给扩散模型部分。它可能涉及基本数据库或 Slack，其中一个模型输出内容，另一个通过 API 摄取。我认为这刚刚起步，比如 Moldbot 和 Open Claw，我们将会看到更多这样的应用。但我还没有具体的经验。

To be honest, I wish I had a really good answer, but this is something I've not explored personally. Most of my experiences are single use cases where one LLM provides solutions or tackles a specific task, but it's mostly not interacting with other agents. I think it's more of a context engineering problem—how to provide the result to another LLM. It's almost like a form of image or video generation where one LLM parses the text and passes it to the diffusion model part. It could involve basic databases or Slack, where one model outputs something and the other ingests it via API. I think this is just getting started with Moldbot and Open Claw, and we'll see a lot more of it. But I don't have concrete experience yet.

你感觉在未来一年里，我们会在这些智能体式使用场景上看到哪些重点和创新？有哪些差距需要弥补？

Do you have a sense for where we'll see focus and innovation around these agentic uses in the upcoming year? What are the gaps that need to be worked on?

每个 LLM 都有自己的失败率。进展通常通过它们能自主工作多久而不失败来衡量。你添加的模型越多，如果它们相互依赖，其中一个失败的风险就越高。改进模型本身将有助于改进整个系统。目前，API 中的大多数普通 LLM 并没有专门针对多智能体设置进行训练。如果你用多智能体交互的数据对它们进行微调，你可以获得更好的性能。我们在编解码器上看到了这一点——GPT-5.2 或 5.3 Codex 与基础模型不同；它们被分叉并专门针对 Codex 应用进行了训练。我认为我们也会在智能体模型上看到类似的情况。消费者很难做到这一点，因为他们无法访问这些模型。像 OpenAI 和 Claude 这样的公司可能会开发自己的版本，针对多智能体环境进行微调。

Each LLM has its own failure rate. Progress is usually measured by how long they can work autonomously before failing. The more models you add, the higher the risk that one fails if they depend on each other. Improving the model itself will help improve the whole system. Right now, most vanilla LLMs in APIs are not specifically trained for multi-agent settings. If you fine-tune them with data for multi-agent interactions, you can get more performance. We've seen this with Codex—GPT-5.2 or 5.3 Codex is not the same as the base models; they were forked and specifically trained for the Codex app. I think we'll see similar for agent models. It's harder for consumers to do that because they don't have access to the models. Companies like OpenAI and Claude might develop their own versions, fine-tuning for multi-agent environments.

回顾过去一两年，从架构角度来看，许多重大进展都相对渐进。核心架构相当稳定。你同意吗？你如何看待 LLM 架构的未来？

Looking back, many big advancements over the past year or two from an architecture perspective have been relatively incremental. The fundamental core architecture has been fairly stable. Do you agree? How do you think about the future of LLM architecture?

这是一个有趣的问题。我要加个星号来说一切，因为 DeepSeek 版本 4 还没有发布；它可能会改变一切。但如果我们看 2025 年截至二月第二周，我认为在最新架构方面没有任何根本性变化。

That's an interesting question. I'll say everything with an asterisk because DeepSeek version 4 is not out yet; it might change everything. But if we look at 2025 up to the second week of February, I don't think there were any fundamental changes in terms of state-of-the-art architecture.

所以我认为，我们需要区分两类架构：一类更侧重于让现有功能更高效，另一类则旨在提升模型的建模性能准确率。如果看那些推动建模性能达到前沿的模型，最近并没有太多变化。2025 年，混合专家模型卷土重来。之前也有 Mixtral 和 DeepSeek 等模型，但它们真正流行起来是在 DeepSeek V3 发布之后。DeepSeek V3 之所以流行，是因为 DeepSeek R1——它基本上是 DeepSeek V3 的微调或后训练版本。很多公司采用了这种架构。我认为 Kimi 直接用了这个架构，并将其从 6700 亿参数扩展到 1 万亿参数。甚至欧洲公司 Mistral AI 也用了 DeepSeek V3 的架构。所以很多人并不是在赌尝试不同的东西；他们采用行之有效的方法，然后在数据和算法上寻求进步。但这并不意味着没有新想法。DeepSeek V3 除了混合专家之外，还有多头潜在注意力。它之前也在某篇论文中出现过，但多头潜在注意力本质上是对注意力机制的一种调整：你有一个中间更小的压缩状态，用于键和值。压缩键和值很重要，因为这样你的 KV 缓存会变小。你不存储完整的键和值，而是存储压缩形式，然后在推理时从压缩形式重建键和值。所以你基本上是在用算力换内存。为了更好地解释，你可以把它想象成类似 LoRA 的低秩适应。基本上，你将其投影到压缩空间，然后再投影回来。这就是多头潜在注意力。这是一个有趣的调整。在 2025-2026 年，人们采用了它。然后 DeepSeek V3.2 有了另一个调整：稀疏注意力。稀疏注意力也不是新东西；一直有研究如何让注意力更便宜，因为它随序列长度呈二次方缩放。已经有成百上千篇论文。但对于论文，我总是有点谨慎；我等着看它在生产环境中、在旗舰模型中的表现。这个想法可能对小模型有效，但一旦扩展到 5000 亿、6000 亿或 1 万亿参数，可能就会出问题。DeepSeek 是一个很好的案例，因为他们有旗舰模型，如果他们在旗舰模型中用了某个东西，你就知道它在规模上有效。他们有自己的稀疏注意力版本，叫做 Deep Sparse Attention。它不是让一个 token 关注所有之前的 token，而是更有选择性；它选择要关注哪些 token。这有点像掩码。你计算一个覆盖所有 token 的掩码，以选择一个子集，从而降低成本，实现次二次方缩放。这些都是调整，但并没有从根本上改变注意力机制的工作原理。仍然是同样的注意力机制，只是如何让它更便宜。所以我认为人们目前专注于有效的方法，但我们在 2026 年可能会看到一个旗舰模型采用根本不同的方法。小公司——不是小公司，而是小改动——也在替代架构上有所作为。我们之前提到了 Qwen 3。Qwen 3 是旗舰模型之一；它可能不再是顶尖，因为它有点旧了，是夏天发布的。但 Qwen 模型发布时，通常都位居排行榜前列。他们有一个并行版本叫 Qwen 3 Next，尝试了不同的东西：一种带有门控 delta 网络的混合注意力机制，更像是一种基于状态的方法，更线性。但这不一定是他们的旗舰模型；他们并行尝试其他东西。这很有道理，因为你不想把所有鸡蛋放在一个篮子里。你想有一个好模型，然后在旁边尝试一些东西，如果效果好，以后再扩展。

So I think one thing we have to distinguish is between architectures that are more geared towards doing the same thing more efficiently and architecture changes that are geared towards getting more modeling performance accuracy out of the model. If we look at those models that push the state-of-the-art in modeling performance, there haven't been that many changes recently. In 2025, mixture of experts models have been making a comeback. There were other models like Mixtral and DeepSeek before, but they really became popular after DeepSeek V3 came out. DeepSeek V3 became popular because of DeepSeek R1, which is basically a fine-tuned or post-trained version of DeepSeek V3. A lot of companies adopted this architecture. I think Kimi straight up used that architecture and scaled it from 670 billion to 1 trillion parameters. Even the European company Mistral AI used the DeepSeek V3 architecture. So a lot of people are not gambling on trying something different; they take something that works and try to make progress in terms of changing the data and algorithms. But that doesn't mean there are no new ideas. DeepSeek V3, besides mixture of experts, had multi-head latent attention. It was also in one of the previous papers, but multi-head latent attention is essentially a tweak of the attention mechanism where you have an intermediate smaller compressed state of the keys and values. The keys and values are important to compress because then your KV cache becomes smaller. You don't store the full keys and values in the KV cache, but a compressed form, and then you reconstruct the keys and values from the compressed form during inference. So you are basically trading off compute with memory. To explain this better, you can think of it as a LoRA-like low-rank adaptation. Basically, you project it down into a compressed space and then project it up again. That's multi-head latent attention. That's an interesting tweak. In 2025-2026, people adopted it. Then DeepSeek V3.2 had another tweak: sparse attention. Sparse attention is also not new; there has always been research on how to make attention cheaper because it scales quadratically with sequence length. There have been hundreds if not thousands of papers. But with papers, I'm always a bit careful; I wait to see it in production, in a flagship model. The idea might work well for a small model, but things may fall apart once you scale to 500 billion, 600 billion, or 1 trillion parameters. DeepSeek is a nice case study because they have a flagship model, and if they use something in that flagship model, you know it works at scale. They have their own version of sparse attention, called Deep Sparse Attention. Instead of having one token pay attention to all previous tokens, it's more selective; it selects which tokens to pay attention to. It's kind of like a mask. You calculate a mask over all tokens to select a subset, making it cheaper and scale sub-quadratically. These are tweaks, but not fundamentally changing how attention works. It's still the same attention mechanism, but how to make it cheaper. So I think people are honing in on what works at the moment, but we might see in 2026 a flagship model with a fundamentally different approach. Small companies—not small, but small changes—have also been made in alternative architectures. We mentioned Qwen 3 earlier. Qwen 3 is one of the flagship models; it's maybe not at the top anymore because it's a bit older, released in summer. But when Qwen models come out, they are usually top of the leaderboards. They had a parallel version called Qwen 3 Next, which tried something different: a hybrid attention mechanism with a gated delta net, more like a state-based model approach that is more linear. But it's not necessarily their flagship model; they are trying other things in parallel. This makes sense because you don't want to put all your eggs in one basket. You want to have a good model and then try something on the side, scaling it up later if it works well.

嗯。那持续学习呢？这经常被提及为一个机会，尤其是在我们真正擅长整合工具和搜索能力之前，因为模型会很快过时。但人们仍然有兴趣让模型能够更新训练数据、删除内容、融入新知识。你预见到这个领域会有重大创新吗？

Yeah. What about continual learning? That comes up frequently as an opportunity, particularly before we got really good at incorporating tools and the ability to do searches, because models would get stale very quickly. But there's still this interest in having a model that we can keep its training data updated, delete things, incorporate new knowledge. Do you foresee significant innovation in that area?

是的，我认为这可能是最大的梦想，即如何让模型自我改进。现在能取得的最大成就，就是有人找到一种方法让它奏效。但我认为目前还没有通往这个目标的路径。没有什么东西能让你说，“哦，那就是能给我们带来可靠持续学习的东西。”话虽如此，我认为已经存在一些形式的持续学习。甚至像——嗯，它更受控制。不是模型自动更新自己，而是人们从最近的互联网或最近的任务中收集数据，然后仔细更新模型。我认为更像是这样。我们不是不更新模型，但也不是完全自动地更新。这是一种半自动的事情。我认为之所以如此，不仅是因为它更可靠。

Yeah, I think this is maybe the biggest dream in the sense of how can we make the model improve itself. The biggest achievement that could be made right now is if someone finds a way that this works. But I think right now there is no pathway to this. There is nothing where you would say, 'Oh, that's the thing that will give us reliable continual learning.' That being said, I think there are already forms of continual learning. Even something like—well, it's more controlled. Instead of the model automatically updating itself, people collect data from the recent internet or recent tasks and then carefully update the model. I think it's more like that. It's not that we don't update models, but we also don't do it fully automatically. It's a semi-automatic type of thing. I think it's like that not only because it's more reliable.

因为直接在新数据上更新模型是有风险的，但也因为资源限制——比如我不知道 OpenAI 有多少份模型副本，但肯定不能每个用户一个副本，那太贵了。每个人家里都得有一台小型超级计算机，或者一台价值 10 万美元的电脑才能运行一个大型旗舰模型。所以公司不能为每个用户实时更新所有内容，那根本不可行。基本上，除非模型只在个人设备上运行，否则我认为我们无法实现真正好的持续学习。另外，你更新时必须非常小心，不想让模型变差。因为这是一个非常重要且昂贵的产品，如果你只是考虑把数据反馈给 OpenAI，然后 OpenAI 自动更新模型，万一更新出错，就会扰乱所有人的使用。所以我认为这更多是一个基础设施安全方面的问题。但另一方面，如果你看推理训练，也就是基于可验证奖励的强化学习，如果基于正确答案运行并持续进行，这从某种意义上也是一种持续学习——你可以技术上一直运行它，只是需要更有选择性。

So because yeah it's risky to just update a model on new data but it's also because of resource constraints because for example I don't know how many copies of the model OpenAI has but I mean it's not you can't definitely not have a single copy per user that would be way too expensive. I mean everyone would have to have a little supercomputer at home or like a $100,000 computer to have like a big flagship model. And so companies can't just update everything on the fly for each user because that would be infeasible. Basically, and so unless we have ways that the models only run on the personal device, I don't think we have or can have really good continual learning essentially because yeah and then the other thing is yeah, you have to be really careful how you update it. You don't want to make the model worse. So because it's such an important expensive product, if you just even think about feeding the data back to OpenAI and then OpenAI automatically updates the model and maybe there's a bad update and it disrupts everything for everyone. And so I think it's more of an infrastructure security type of issue. But otherwise I mean if you look at the reasoning training, the reinforcement learning with verifiable rewards, if you run this based on correct answers and that type of setting and you just keep it running, it is kind of a form of continual learning in a sense where you can technically just keep running this and it's just like you want to be more selective basically.

那你认为长上下文或更长的上下文是否能缓解持续学习的一些痛点或需求？比如在你提到的个性化模型案例中，一种方法是获取新信息并持续学习。另一种我见过有人尝试的是为模型创建个人 LoRA 适配器。但第三种方法就是把新信息直接放入上下文，在推理时使用。

And do you think that long context or longer contexts kind of alleviate some of the pain or need to do continual learning? Like you know in your case of the personal personalized models, one approach is to take new information and kind of continually learn against it. Another that I think folks have played around with is to create like personal LoRA adapters for a model. But then a third is to just put that new information into the context and use it at inference time.

我会说既肯定又否定。我确实认为长上下文语言模型最近带来了很多可能性——以前人们构建 RAG 系统（检索增强生成系统），现在它们几乎——我不会说过时了，如果你有一个固定的庞大数据库或文档集并反复使用，它们仍然非常有用。但如果你是普通用户，即使有一份一千页的 PDF，技术上大多数时候——一千页可能有点夸张，但比如两百页的 PDF，你可以把它放在上下文中。你不需要对 LLM 进行微调来训练它处理这些数据，也不需要 RAG，你可以在上下文中做很多事情。就像你说的，对于信息来说也是如此，你可以技术上把所有相关的新信息都放在上下文中。但我认为这只能做到一定程度，因为作为用户，你还得知道要提供什么信息。但如果你结合工具使用，比如你问一个历史事件，假设数据截止日期是 2025 年，而你问的是 2026 年的历史事件，语言模型仍然可以使用网络搜索。它可以使用工具在网上查找。所以你不需要为了那个特定历史事件更新 LLM。但如果这个历史事件有很多影响，波及周围很多其他事情，可能会被遗漏，那么你从工具调用中只能得到某些事实，而无法获得与其他数据点的完整交互。所以我认为它并没有完全取代更新，但可能降低了更新的必要性，或者不需要那么频繁地更新。

I would say yes and no. So I do think long context LMs have enabled so much recently, like where before people were building RAG systems, retrieval augmented generation systems, and now it's almost—I wouldn't say they are obsolete, they still are very useful if you have a fixed big database or document set and if you use it repeatedly. But if you're a regular user and you have even a thousand-page PDF, you can technically—most of the time, I mean thousand is maybe stretching it a bit much but like a 200-page PDF you can have it in context. You don't need to train the LLM in terms of fine-tune it on that data, you don't need to have a RAG, you can do a lot of stuff in context. And like you said, the same is maybe true for information where you could technically just provide all the relevant new information in context. But I think it only gets you so far because you also as a user have to know what to provide as information. But then if you couple that with tool use, for example if you ask about a historical event and let's say the data cutoff is 2025 and you ask about a 2026 historical event, the LM can still use a web search. It can still use a tool and look it up on the web. So you don't necessarily need to update the LLM for that particular historical event. But then if the historical event has a lot of ramifications and affects a lot of other things around it that might be missed, then you only get certain facts from a tool call but not that whole interaction with other data points. And I think so it's not fully replacing the updating but it is making it maybe less necessary to do it or it's not necessary to do it quite as often I think maybe.

那么你对未来一年领域重点的大致看法是——推理、推理时 Scaling、智能体。还有其他想法或预测吗？

So your kind of big picture thoughts on like where the field will be focused over the next year is you know again reasoning, inference time scaling, agents. Any other thoughts or predictions that come to mind for you?

是的，我很想看看——虽然是小事情，但你知道我们讨论过，Transformer 架构没有大的替代品，但有一些东西比如文本扩散模型。谷歌有一个等待页面，他们计划推出一个文本扩散模型，一个小的——不是小，而是一个替代方案，我很好奇。也许它会取代免费层的语言模型。这很有趣。我主要感兴趣是因为文本扩散模型有很多研究。这是一种不同的思路：不是顺序生成文本，而是更像 BERT 模型，有掩码，然后逐渐去噪或用文本替换掩码。我想看看它在规模上的表现，因为目前大多数模型都是研究模型之类的。这不是什么前沿性能，但可能更便宜、更快，甚至可能改善日常体验，比如谷歌搜索中的摘要，那也是基于语言模型的，但并非最佳。这些小的生活质量改进。嗯，我们录制的时候是春节前，历史上春节期间总会有很多模型发布，开源权重模型发布。所以可能有个小惊喜。也许我们会看到 DeepSeek 版本 4，也许会有更大的变化。所以我挺感兴趣跟进看看。但嗯，目前我想到的，我们基本都覆盖了。

Yeah, I will be curious to see—I mean it's a little thing but you know like we talked, there is no big alternative to the Transformer architecture but there is for example things like text diffusion models. And Google for example they had like a waiting page, they are planning to launch a text diffusion model like a small—I mean not small but like an alternative where I'm really curious. Maybe that's going to be replacing like the free tier of LMs. That's a really interesting thing. I'm mainly interested in that because there's been a lot of research on text diffusion models. So it's like a different take: instead of generating the text sequentially, it's more like a BERT model where you have masks and then you gradually denoise or replace the masks by text. I just want to see how it performs at scale because right now most of the models are research models and just things like that. It's nothing you know I think people should get excited about in terms of cutting edge performance but it will be maybe cheaper and faster and maybe that will be making the like it's like an everyday maybe improvement for even like the Google summary in Google search which is also LM based but it's not the best. And these little quality of life improvements I think. Well, why we're recording this is it's before the Chinese New Year, and historically around the Chinese New Year, there have been always a lot of model releases, open weight model releases. So maybe there's like a little surprise in there. Maybe we'll see DeepSeek version 4 and maybe there is a bigger change. So I'm kind of like interested in following that and seeing that. But yeah, right now top of my head I think we covered pretty much everything.

我们换个话题，聊聊你个人最近在忙什么。你提到了书的章节。说说你现在的书，以及人们在哪里可以了解更多。

Let's maybe switch gears a little bit and update us on what you've been working on personally. You've kind of referenced chapters of the book. Talk a little bit about your current book and where folks can learn more about it.

是的，上次我上你的播客时，我们聊了我的《从零开始构建大语言模型》这本书。它涵盖了从构建架构到预训练模型，再到指令微调的整个过程。目标不是构建一个能帮你做所有家务的个人助手，因为那需要花费 5 万到 10 万美元，而且工作量巨大。即使现在训练自己的 LLM 比以前简单了，也不是周末就能常规完成的事情。

Yeah, so I think last time I was on your podcast, we talked about my Build a Large Language Model from Scratch book. So it's basically the whole journey from building the architecture to pre-training a model and then also doing instruction fine-tuning. And the goal of that was not to build your personal assistant that does all the things at home for you. Because that would cost like $50,000 to $100,000 and be a lot of work. And it's—I mean even though it's simpler nowadays to train your own LLM, it's not something you can do routinely on a weekend.

但那本书的目标仍然是教人们理解 LLM 的工作流程，因为这能帮助你更好地使用 LLM，理解什么是上下文、上下文的限制是什么，或者注意力机制是如何工作的，以及为什么输入变长时成本更高。就像你自己构建 LLM 一样，相比于更自由形式的讲解，你能获得非常清晰的理解。所以很多人喜欢那本书，它现在也是一本很受欢迎的教科书。我当时非常兴奋能写续作。所以现在我正在从零构建一个推理模型，这有点像续集。两本书没有重叠，它可以作为独立书籍阅读，但主要聚焦于我们讨论过的推理技术：可验证奖励的强化学习、GRPO 算法、推理时 Scaling（规模扩张），所有这些技术都是在你有预训练 LLM 之后使用的。所以这本书从给定一个预训练 LLM 开始，我们使用 Qwen 3，最小的 Qwen 3 模型，然后加入推理时 Scaling（规模扩张）和强化学习。前 360 页已经进入早期访问阶段，我希望在四月前完成，只剩最后一章了。不过那一章工作量很大，因为要运行所有实验。所以我一直在运行大量实验，尤其是 GRPO 算法，因为有很多不同的论文和改进，需要在实践中尝试。这很有趣，但也很辛苦。所以过去几周和几个月我主要在做实验，实际上相当令人兴奋。

But the goal of that book was to teach people still how that workflow works to understand how LLMs work because that helps you to use LLM better, to understand what is the context, what's the limitation of the context, or how does attention work and why is it more expensive if my input gets longer. And it's just like if you build the LLM yourself, you kind of get a real clear understanding compared to just explaining it in a more free-form based approach. So yeah, a lot of people like that and it's like a very popular textbook also for teaching now. And I was then really excited to work on the sequel. So right now I'm working on building a reasoning model from scratch, which is kind of like the sequel. There's no overlap between the books. It can be read as a standalone book, but it's mainly focused on the reasoning techniques we talked about: the reinforcement learning with verifiable rewards, the GRPO algorithm, inference scaling, all these techniques that once you have a pre-trained LLM. So the book starts by given a pre-trained LLM, we use Qwen 3, the smallest Qwen 3 model, and then adding inference scaling and the reinforcement learning. So the first 360 pages are already in the early access and I'm hoping to finish, I mean there's only one more chapter left, by April. Well, I mean the chapter is a lot of work because you have to run all the experiments. So I've been running a lot of experiments especially for the GRPO algorithms because there have been so many different papers and improvements and trying them out in practice. It's been a lot of fun but it's also a lot of work. So I've been mostly running experiments in the last couple of weeks and months and yeah, it's quite exciting actually.

那么读者可以直接拿起第二本书开始读吗，还是你希望他们在读第二本之前先读完第一本？

And so can folks pick up the second book and run with that or do you expect folks to have read the entire first book before they start with the second?

我觉得两种方式都可以。你不必先读第一本书。第二本书使用预训练 LLM，所以你不需要自己预训练 LLM，也不需要第一本书来为第二本书训练 LLM。所以它是相对独立的。但第二本书没有详细解释预训练或架构。我有一个附录解释架构，但不如第一本书详细。所以如果人们想了解 LLM 从预训练到后训练的整个生命周期，我认为按顺序阅读是合理的。但你也可以从第二本书开始，学习推理时 Scaling（规模扩张）和推理，然后如果你对预训练感兴趣，可以之后再补上。基本上两种方式都行。

I would say either way works. You don't have to read the first book. The second book uses a pre-trained LLM, so you don't have to pre-train your own LLM or you don't need the first book to train the LLM for the second book. So it's kind of independent like that. But the second book doesn't explain in detail the pre-training or the architecture. I have an appendix on explaining the architecture but it's not quite as detailed as the first book. So I think if people want to understand the whole life cycle of an LLM from pre-training to post-training, I think it would make sense to read them sequentially. But you could also start with the second book, learn about inference scaling and reasoning, and then if you're interested in the pre-training you can fill in the gaps later on. I think either way works basically.

太棒了，Sebastian。很高兴和你叙旧，我们得比每三年一次更频繁地聊聊。非常感谢你参与节目，分享你对现状和未来走向的看法。

Well, very cool, Sebastian. It's been great catching up with you and we need to do it more often than every three years. But thanks so much for jumping on and sharing a bit of your perspective on kind of where things are and where things are going.

是的，非常感谢你的邀请，Sam。我过得很愉快。我喜欢谈论 LLM 和 AI。所以，这真是一种享受，谢谢你的邀请。

Yeah, thank you so much for the invitation, Sam. I had a great time. I love talking about LLMs and AI. So, well, that was a treat and thanks for having me on.

谢谢。

Thank you.