Harper Carroll came to AI education through a CS background at Stanford, machine learning engineering at Meta, and a brief stint at a small GPU compute startup in late 2023, where she noticed that almost no one understood how to fine-tune open source models. She started writing and teaching to help drive signups for the startup’s platform. Her first guide, posted right after Mistral 7B was released, when she had about 50 followers, got 50,000 views. In March 2024, a video explaining the difference between AI and machine learning got 5 million views, with 1 in 20 viewers following her afterward. She now has more than 500,000 followers across multiple platforms and is a full-time AI educator.Harper Carroll 进入 AI 教育领域有着深厚的背景：她在斯坦福大学学习计算机科学，曾在 Meta 从事机器学习工程，并于 2023 年底在一家小型 GPU 计算初创公司短暂工作过。在那里，她注意到几乎没有人了解如何微调开源模型。为了帮助初创公司推广其平台，她开始撰写文章并进行教学。她的第一篇指南发布于 Mistral 7B 发布后不久，当时她只有约 50 名粉丝，却获得了 5 万次浏览。2024 年 3 月，她发布了一段解释 AI 与机器学习区别的视频，获得了 500 万次观看，每 20 名观众中就有 1 人随后关注了她。目前，她在多个平台拥有超过 50 万名粉丝，是一名全职 AI 教育工作者。
We covered fine-tuning versus prompting, what it actually means to learn to code in 2025, and what the AI field gets wrong when it talks to the public.我们讨论了微调与提示词的区别、在 2025 年学习编程的真正含义，以及 AI 领域在与公众沟通时存在的误区。
Understanding the world with math用数学理解世界
We started with Harper’s own AI learning journey, and it contained a wonderful insight. She grew up loving math and came to computer science at Stanford because algorithms seemed like wonderful math puzzles. Eventually she realized that AI is “understand[ing] the world around us with math.” Text-based LLMs are only one branch. The field as a whole is “the math of the world.” That seems like a deep intuition that all of us need to internalize.我们从 Harper 自己的 AI 学习之旅开始，其中包含了一个精彩的见解。她从小就热爱数学，进入斯坦福大学学习计算机科学是因为算法看起来就像美妙的数学谜题。最终她意识到，AI 是“用数学理解我们周围的世界”。基于文本的 LLM 只是其中的一个分支。整个领域本质上是“世界的数学”。这似乎是一种我们所有人都需要内化的深刻直觉。
AI as a medium作为媒介的 AI
A study that circulated last year found that people who used AI to write essays showed reduced brain activity compared to people who write unaided. The reaction in many quarters was alarm. People said, “We’re outsourcing cognition and our brains will atrophy.” Harper’s smart response was that those users must have given the AI a one-sentence prompt and accepted whatever came back.去年流传的一项研究发现，与独立写作的人相比，使用 AI 撰写文章的人大脑活动有所减少。许多领域的反应是恐慌。人们说：“我们正在外包认知，我们的大脑将会萎缩。”Harper 的聪明回应是，那些用户肯定只是给 AI 发送了一个一句话的提示词，然后就接受了生成的任何结果。
As she put it, that’s the equivalent of just telling Alexa to order you the most popular book this week. Of course less brain activity is being measured! Contrast that with the difference between shopping for a book by browsing and searching at Amazon versus driving to a physical bookstore. There’s certainly a difference, but it isn’t outsourcing cognition. It’s saving time, and that time might well be spent on other demanding cognitive tasks.正如她所言，这相当于让 Alexa 帮你订购本周最畅销的书。当然，测量到的大脑活动会更少！对比一下在亚马逊上通过浏览和搜索买书，与开车去实体书店买书的区别。这当然有区别，但这并不是外包认知。这是在节省时间，而这些时间完全可以花在其他高要求的认知任务上。
My framing is that AI is a medium, the way language is a medium, or photography. Anyone can take a photograph or write a book. The words available to every writer are the same; what differs is what they do with them, just as some photographers do something with it that others can’t. The same is true of software. There’s a line in Aaron Sorkin’s movie The Social Network where the Zuckerberg character says about the Winklevosses, “If you guys were the inventors of Facebook, you’d have invented Facebook.” An idea and its execution aren’t the same thing. One person gives AI a prompt and the output is bad. Another builds a process around AI and the output is great. What you bring to the medium is what determines the result. Harper agreed.我的观点是，AI 是一种媒介，就像语言或摄影是一种媒介一样。任何人都可以拍照或写书。每个作家可用的词汇都是一样的；不同的是他们如何运用这些词汇，就像有些摄影师能拍出别人拍不出的作品一样。软件也是如此。在 Aaron Sorkin 的电影《社交网络》中，扎克伯格这个角色在谈到温克莱沃斯兄弟时说：“如果你们是 Facebook 的发明者，你们早就发明出 Facebook 了。”想法和执行并不是一回事。一个人给 AI 一个提示词，输出结果很糟糕。另一个人围绕 AI 构建了一个流程，输出结果就很棒。你为媒介带来了什么，决定了最终的结果。Harper 对此表示赞同。
Fine-tuning is like psychedelics for AI微调就像是 AI 的迷幻药
I’ve been trying to figure out how we can use AI for writing and editing at O’Reilly. We want skills and workflows that accelerate our productivity but don’t produce copy that reads as whatever the base model sounds like when nobody’s putting in any effort.我一直在研究我们如何能在 O’Reilly 将 AI 用于写作和编辑。我们想要的是能够提高生产力，但又不会产生那种“没人用心时基础模型听起来的样子”的文案的技能和工作流程。
Takeaway posts like this one are a great use case for AI-assisted writing. As source material we have a transcript, with the actual conversation between the participants (or in the case of one of our online conferences, their presentations). We want a structured summary that captures the high points and suggests possible clips for social media. I (or whomever is using this AI-assisted workflow) can then rewrite, rearrange, elaborate, or delete from that first draft. It might not be as good as a draft written from scratch, but quite frankly, it’s far better than the alternative, which is no summary at all. I just don’t have time to write them all unaided.像这样的总结文章是 AI 辅助写作的一个绝佳用例。作为源材料，我们有参与者之间实际对话的转录稿（或者在我们的一次在线会议中，是他们的演讲稿）。我们需要一个结构化的摘要，捕捉重点并为社交媒体建议可能的剪辑片段。然后，我（或任何使用此 AI 辅助工作流程的人）可以对初稿进行重写、重新排列、润色或删除。它可能不如从零开始写的草稿好，但坦率地说，它远比另一种选择——完全没有摘要——要好得多。我实在没有时间在没有辅助的情况下写完所有这些内容。
When I’m writing an article, I generate a similar “transcript” by recording myself talking about the ideas I’m wrestling with and trying to put into the world. Then I ask Claude to put it together into something a bit more structured.当我写文章时，我会通过录下自己谈论正在思考并试图表达出来的想法，来生成类似的“转录稿”。然后我让 Claude 将其整理成结构更清晰的内容。
I’ve been improving Claude’s ability to produce prose that we can use by rewriting its output, showing it the differences, and then asking it to construct a skill that captures what it’s learned. Over time, it’s gotten closer and closer to something that I’m comfortable with, and I’m now generalizing that into a system that learns any author’s voice, respects the various conventions of the target content type (which can be very different across books, articles and blog posts, social media, and marketing materials like back cover copy and course descriptions), and applies editing suggestions from my favorite books on good writing, including Strunk and White and On Writing Well by William Zinsser.我一直在通过重写 Claude 的输出、向它展示差异，然后要求它构建一种捕捉所学内容的技能，来提高它生成可用散文的能力。随着时间的推移，它越来越接近我满意的程度。我现在正将其推广为一个系统，该系统可以学习任何作者的语调，尊重目标内容类型的各种惯例（书籍、文章、博客文章、社交媒体以及封底文案和课程描述等营销材料之间的差异可能非常大），并应用我最喜欢的写作书籍中的编辑建议，包括《Strunk and White》和 William Zinsser 的《On Writing Well》。
Harper attacked the same problem from a different angle. She built a dataset of roughly 1,000 of her Instagram captions, video transcripts, and X posts, then fed them to Claude as context and asked it to write in her style. Unfortunately, the output tested 100% AI by a detection tool, even with 1,000 examples of her real voice in the prompt. She then fine-tuned an open source Llama model on the same data. The fine-tuned output tested 100% human. She gave a compelling demo at South by Southwest showing how easy this is to do. It took her about 20 minutes.Harper 从另一个角度解决了同样的问题。她建立了一个包含约 1000 条 Instagram 字幕、视频转录稿和 X 帖子的数据集，然后将它们作为上下文输入给 Claude，并要求它以她的风格写作。不幸的是，即使在提示词中提供了 1000 个她真实语调的示例，检测工具仍显示输出结果 100% 为 AI 生成。随后，她使用相同的数据对开源 Llama 模型进行了微调。微调后的输出结果显示 100% 为人类创作。她在西南偏南大会（South by Southwest）上做了一个引人注目的演示，展示了这有多容易实现。她只用了大约 20 分钟。
After Harper said that prompting doesn’t shift the output distribution the way fine-tuning does, I told her the story about the French writer Marcel Proust that I first used in my conversation with Steve Wilson, which I picked up from Alain de Botton’s How Proust Can Change Your Life. A friend comes to visit the bedridden Proust, and making polite conversation begins to tell him about the train trip to Paris. “More slowly,” Proust replies. This cycle repeats several times until the friend is telling him small details like the old man feeding pigeons on the steps of the station.在 Harper 说提示词无法像微调那样改变输出分布后，我给她讲了法国作家马塞尔·普鲁斯特的故事，我第一次在与 Steve Wilson 的谈话中使用了这个故事，它来自 Alain de Botton 的《普鲁斯特如何改变你的人生》。一位朋友去探望卧床不起的普鲁斯特，在礼貌交谈时开始向他讲述去巴黎的火车旅行。“慢一点，”普鲁斯特回答道。这个循环重复了几次，直到朋友开始向他讲述车站台阶上老人喂鸽子等微小的细节。
Harper got it, and broke it down more slowly in her inimitable way. Here’s why in-context prompting fails where fine-tuning succeeds:Harper 听懂了，并以她独特的方式更深入地剖析了这一点。以下是为什么上下文提示词在微调成功的地方会失败的原因：
Basically AI models are these massive mathematical equations, and the parameters are variables when you’re training, and then they become constants in those equations when you’re running inference . . .So what you’re doing when you’re training the model is you’re learning how to map, by adjusting those constants when they’re variables during training,. . .input to desired output.基本上，AI 模型是这些庞大的数学方程，参数在训练时是变量，而在运行推理时就变成了方程中的常量……所以当你训练模型时，你所做的就是通过在训练期间将常量作为变量进行调整，来学习如何将输入映射到期望的输出。
Once the model is deployed, the probability distribution over output tokens is fixed. You can put 1,000 examples in a prompt and ask the model to pattern-match, but you’re asking it to do that with frozen weights. The surface behavior bends a little, but the underlying distribution doesn’t shift. Fine-tuning lets you actually modify the weights and how the model wants to write.模型部署后，输出 token 的概率分布就固定了。你可以在提示词中放入 1000 个示例并要求模型进行模式匹配，但你是在要求它使用冻结的权重来完成这项工作。表面行为会发生一点弯曲，但底层分布并没有改变。微调让你能够真正修改权重以及模型想要如何写作的方式。
Her suggested approach for building the training dataset is to take your own writing, have AI rewrite it with its characteristic tics, then train with the AI version as input and your original as the target output. You’re teaching the model to undo the tells.她建议构建训练数据集的方法是：采用你自己的写作内容，让 AI 用其特有的习惯用语重写，然后以 AI 版本作为输入，以你的原始版本作为目标输出进行训练。你是在教模型如何撤销那些“AI 味”。
Should people still learn to code?人们还应该学习编程吗？
We also spent time on the inevitable question of whether people should still learn to code. We both agree they should, but not necessarily like they used to, by learning the detailed syntax of a programming language, then by trial and error as they painfully learn how hard it is to get the desired behavior.我们还花时间讨论了人们是否还应该学习编程这个不可避免的问题。我们都同意应该学习，但不一定像过去那样，通过学习编程语言的详细语法，然后通过反复试验，痛苦地了解获得预期行为有多么困难。
Harper’s take (which I also agree with) is that vibe coding has lowered the floor. People who could never afford to hire someone to build a product can now do so themselves. But it has also raised the ceiling, because people who actually understand systems can build vastly more sophisticated things with the same tools, which takes us back to the case for AI as a medium.Harper 的观点（我也同意）是，“氛围编程”（vibe coding）降低了门槛。那些以前雇不起人来开发产品的人现在可以自己动手了。但它也提高了上限，因为真正了解系统的人可以用同样的工具构建出复杂得多的东西，这又回到了 AI 作为媒介的论点。
Perhaps more importantly to the question of how much coding you should learn, experienced developers will also see failure modes that pure vibe coders miss. Harper gave an example that came from watching a friend using an agent tool that had, at some point, started storing its data in a Word document and using it as a makeshift database, probably because the session started with a Word doc. It was extremely slow and extremely inefficient. An engineer sees the problem immediately. A vibe coder might run that system for months before noticing something is wrong.对于你应该学习多少编程知识这个问题，也许更重要的是，经验丰富的开发人员会看到纯粹的“氛围程序员”所忽略的故障模式。Harper 举了一个例子，来自观察一位朋友使用代理工具，该工具在某个时候开始将其数据存储在 Word 文档中并将其用作临时数据库，这可能是因为会话是从 Word 文档开始的。它极其缓慢且极其低效。工程师一眼就能看出问题所在。而“氛围程序员”可能会运行该系统几个月才发现出了问题。
So yes, you should learn enough about coding to understand what’s happening. The art of teaching programming to the next generation will be developing useful projects that also highlight underlying concepts of software architecture and engineering.所以是的，你应该学习足够的编程知识来理解正在发生的事情。向下一代教授编程的艺术，将在于开发既实用又能突出软件架构和工程底层概念的项目。
Intuition as differentiator直觉作为差异化因素
Silicon Valley runs heavily on logic and on the idea that good decisions come from better data, more rigorous analysis, and sharper models. In this environment, intuition can get dismissed as something “soft and fuzzy,” Harper noted. And that’s the wrong mindset for AI.硅谷在很大程度上依赖逻辑，以及“好的决策来自更好的数据、更严谨的分析和更敏锐的模型”这一理念。Harper 指出，在这种环境下，直觉可能会被视为“软且模糊”的东西。而这对于 AI 来说是错误的思维方式。
AI is getting better and better at exactly the things the logical axis does well, but intuition remains a challenge because it often contradicts what the data says. Good intuition “goes against the input,” to use Harper’s phrase. A model that’s been trained to recognize patterns in data will, almost by definition, struggle with making decisions that run counter to those patterns. Just as skills-informed judgment supercharges AI-assisted engineers, intuition could be a uniquely human skill for a long time. Elevating it as a concern might bring the industry more of an attitude of humility towards ourselves and our place in the world.AI 在逻辑轴擅长的领域表现得越来越好，但直觉仍然是一个挑战，因为它往往与数据所说的相矛盾。用 Harper 的话来说，好的直觉是“违背输入”的。一个经过训练可以识别数据模式的模型，几乎在定义上就难以做出违背这些模式的决策。正如具备技能的判断力能增强 AI 辅助工程师的能力一样，直觉在很长一段时间内可能是一种独特的人类技能。将其提升为一个关注点，可能会给行业带来更多对我们自身以及我们在世界中地位的谦逊态度。
What the field gets wrong该领域的误区
I closed by asking Harper what the AI field most consistently gets wrong in how it talks to the public. She said that too much of the public-facing discourse leads with fear, of job displacement, of rapidly approaching AGI, and of a rocky transition that requires a universal basic income to cushion the blow. She’s not calling those impossible futures, but she thinks they’re the wrong introduction to the technology.最后，我问 Harper，AI 领域在与公众沟通时最常犯的错误是什么。她说，太多的公众话语以恐惧为导向，比如工作流失、即将到来的 AGI，以及需要全民基本收入来缓解冲击的艰难转型。她并不是说这些未来是不可能的，但她认为这是了解这项技术的错误切入点。
A lot of companies are using AI to ask how to do the same things at lower cost. The better question is how to raise ambitions. AI doesn’t just scale individual capabilities. It scales what organizations can attempt. But for it to work out that way, everybody has to actually learn AI. We can’t have AI haves and have-nots. That means lower-cost models, serious open source investment, and companies that don’t just become serfs to the major platforms.许多公司正在利用 AI 来询问如何以更低的成本做同样的事情。更好的问题是如何提高雄心。AI 不仅仅是扩展个人的能力。它扩展了组织可以尝试的范围。但要实现这一点，每个人都必须真正学习 AI。我们不能有 AI 的“拥有者”和“被剥夺者”。这意味着需要低成本的模型、严肃的开源投资，以及那些不仅仅沦为大型平台附庸的公司。
Harper has been making this point for a while, to audiences ranging from engineers to people who’ve never written a line of code. “There is not really much to fear right now,” she says. “AI is this incredible productivity tool.” The people who will struggle, in her view, are the ones who refuse to engage with it at all.Harper 提出这个观点已经有一段时间了，受众从工程师到从未写过一行代码的人都有。“现在真的没什么好怕的，”她说，“AI 是一种令人难以置信的生产力工具。”在她看来，那些会挣扎的人是拒绝接触它的人。
At O’Reilly, we’ve been working on a version of the same narrative at an organizational level. The fear-first narrative produces avoidance, and avoidance is the one thing that will actually leave someone behind. So we’re building a corporate AI transformation practice that starts with people’s existing jobs, and figures out how to “mix in” AI to make them more impactful. We’re learning how to teach both the humans and the agents at the same time to make them more productive together.在 O’Reilly，我们一直在组织层面致力于同样的叙事。恐惧优先的叙事会导致回避，而回避是唯一真正让人落后的因素。因此，我们正在建立一种企业 AI 转型实践，从人们现有的工作开始，找出如何“融入”AI 以使他们更具影响力。我们正在学习如何同时教导人类和代理，使他们共同提高生产力。
On July 9, I’ll be speaking with Trail of Bits cofounder and CEO Dan Guido about the playbook his company used to go AI native, which he first outlined at this year’s [un]prompted. He’ll give a version of the same talk, then take about 40 minutes of audience questions on what worked, what didn’t, and what is still unsolved. I hope you join us to find out what’s changed since [un]prompted and where the playbook is heading next. Register here; it’s free and open to all.7 月 9 日，我将与 Trail of Bits 的联合创始人兼首席执行官 Dan Guido 对话，探讨他公司实现 AI 原生化的手册，他曾在今年的 [un]prompted 大会上首次概述过。他将进行相同主题的演讲，然后花约 40 分钟回答观众关于哪些有效、哪些无效以及哪些尚未解决的问题。我希望你能加入我们，了解自 [un]prompted 以来发生了什么变化，以及这本手册的下一步方向。点击此处注册；活动免费并向所有人开放。
