教育领域的Claude：Anthropic如何与Drew Bent共同塑造AI在学习中的角色

从学前到终身教育的创新动力源自杰出人才。十五年来，各种规模和阶段的EdTech公司都信赖高等教育领域为其输送推动变革的人才。当特定技能和经验对使命至关重要时，高等教育便是可靠的合作伙伴。提供全职、兼职和高管招聘服务，高等教育深谙您所需的市场化人才。了解更多请访问hireedu.com。

Innovation in pre k to gray learning is powered by exceptional people. For over fifteen years, EdTech companies of all sizes and stages have trusted higher education to find the talent that drives impact. When specific skills and experiences are mission critical, higher education is a partner that delivers. Offering permanent, fractional, and executive recruitment, higher education knows the go to market talent you need. Learn more at hireedu.com.

网址是hireedu.com。

That's hireedu.com.

人工智能能分析对话过程与辅导互动质量，随后为导师提供个性化反馈——通常会建议他们调整说话比例，花更多时间倾听并向学习者提问。

The AI can analyze how the conversation went, how the tutoring interaction went, and then get personalized feedback to the tutor, often encouraging them to have changed their talking ratio and spend more time listening and asking questions of their learners.

欢迎收听《EdTech内幕》，这是覆盖教育科技行业的顶级播客。从融资动态到影响力评估，从早教到K12、高等教育及职场领域的AI进展，您都能在EdTechInsiders找到相关内容。

Welcome to EdTech Insiders, the top podcast covering the education technology industry. From funding rounds to impact to AI developments across early childhood, k 12, higher ed, and work, you'll find it all here at EdTechInsiders.

记得订阅播客，查看我们的新闻通讯和活动日历。若想深度参与，请加入EdTech内幕Plus会员，获取优质内容、WhatsApp频道权限、活动优先参与权，以及来自Alex和Ben的幕后洞见。祝您享受本期节目。

Remember to subscribe to the pod, check out our newsletter, and also our event calendar. And to go deeper, check out EdTech Insiders Plus, where you can get premium content, access to our WhatsApp channel, early access to events, and back channel insights from Alex and Ben. Hope you enjoy today's pod.

今天我们访谈环节迎来一位特别嘉宾——Anthropic有益部署团队教育负责人Drew Bent。他同时联合创办了非营利辅导组织Schoolhouse.world（与Salcon合作）。此前他曾在可汗学院编写代码，教授高中数学，并拥有超过十年的学生辅导经验。

We have a very special guest today for our interview. We are here with Drew Bent, who leads education as part of Anthropics Beneficial Deployments team. He also cofounded the tutoring nonprofit schoolhouse. World with Salcon. And prior to that, he wrote code at Khan Academy, taught high school math, and has been tutoring students for over a decade.

Drew拥有MIT物理与计算机科学学位，以及斯坦福大学教育学硕士学位。相当厉害啊Drew。很高兴见到你。欢迎来到播客节目。

Drew has degrees in physics and computer science from MIT and an education master's from Stanford. Not bad, Drew. Nice to see you. Great to see you again. And to the podcast.

是的，非常高兴来到这里。

Yeah. I'm very excited to be here.

我特别期待与你的对话。首先让我们聊聊Claude教育版。Anthropic显然是前沿模型开发者，所做工作令人惊叹——我个人每天都把它当作全能伙伴来使用。但Anthropic在教育领域投入了大量精力。

I'm really excited to talk to you. So first, let's talk about Claude for Education. Anthropic is obviously frontier model, has been doing incredible work. I personally use it every day and talk into it just as a partner for everything I do. But Anthropic has put a lot of energy into education.

请谈谈为什么Anthropic选择将教育作为首批重要现实应用场景？这对你们的团队和正在构建的产品意味着什么？

Tell us about why Anthropic has chosen to focus on education as one of its first major real world applications, and what does that mean for your team and what you're building?

当然。我很乐意分享我们启动‘教育云’计划的初衷。退一步说，任何AI实验室为何首先关注教育领域确实是个合理问题。虽然无法代表其他实验室，但从Anthropic的角度来看，去年底我们分析了Claude在全球范围内的使用数据。

Of course. Yeah, I'm happy to share why we've started this Cloud for Education initiative. I think taking a step back, it's a fair question about why any of the AI labs are focused on education in the first place. Yeah. I can't speak to other labs, but what I can share from the anthropic perspective is it was late last year that we were looking at the data of how Claude was being used overall around the world.

编程是最主要的使用场景可能并不令人意外，但前四大用途中有两项——紧接其后的就是与教育相关的学习教学和学术研究。这对我们团队至关重要，因为我们不仅关注技术开发，更重视其负责任的应用方式。我们看到学生和教师正在全方位地使用这项技术，而教育科技领域对AI的潜力始终抱有诸多期待。

You may not be surprised to learn that coding was the top use case, but two out of the top four use cases. So right below there were education related learning and teaching, academic research. And so I think this was really important for our team to see because we built this general purpose technology, but we care not just about building the technology, but about the responsible deployment of it and how it gets used. And here it's getting used all throughout education by students and by teachers. And there's all sorts of hopes that we've all had in the ed tech field for what AI can do to education.

但我们也清楚存在许多担忧。这本质上是我们对技术社会影响的思考实践——如何成为更负责任的技术管理者。与其简单指责学生用它作弊或教师自动化工作内容，我们AI实验室（包括Anthropic和其他机构）更有责任：当数百万人如此使用技术时，我们必须先研究现状，再优化其教育应用。

But we also know there's a lot of concerns. And so this is for us primarily an exercise in thinking about the societal impact of our technology and what can we do to be more responsible stewards of this technology. And I think it's very easy to just put the blame and the burden on, you know, why are students using it to cheat in these cases, or why are teachers automating parts of their jobs in these cases? But really we have a responsibility as the AI labs and Thropic and all the other ones that if we have a technology that's being used by millions of people in this way, that we have to first study it and then figure out what we can do to make it better for these educational purposes.

百分百同意。作为这项工作的一部分，你们四月份发布了Anthropic教育报告，通过独创的分类法研究了学生的实际使用情况——比如直接作业与协作作业、问题解决与答案生成等维度。能否谈谈报告中那些出乎意料的发现？特别是学生如何通过Claude进行教育相关的应用？

A 100%. And, you know, as part of that work, you've just put out a, in April, a Anthropics education report where you looked at how students are actually using it, and you have a really interesting taxonomy of sort of how you put it together. You have this direct work versus collaborative work, problem solving versus sort of asking for output. Tell us a little bit about the findings from that report, and what was surprising or counterintuitive for you about how students are using Claude and Tropic for education? This April education report that you're talking about here

这份四月报告是我们的首次尝试（后续会有更多研究），重点分析学生使用Claude的模式。我们认为公开这些实证数据至关重要——通过隐私保护工具，我们分析了50万条对话（不查看具体内容），先由Claude自行分析汇总，再由人工复核。

was our first attempt, and we will have more research on this, so stay tuned, to study how, in this case, students are using Claude. We thought it was really important to share some empirical data and to frankly make this open source and share it with, with the public, because there's been a lot of studies, of course, around how students and teachers are using AI. There's been sort of small, you know, maybe RCTs and sort of smaller case studies, but we of course have access to a lot of data. And because we have this privacy preserving tool where we're able to study, you know, 500,000 conversations without looking at any individual conversation. The interesting part of this is that we have Claude itself, go analyze all the conversations, summarize them, bring it up, and then humans review at the end.

数据显示了许多优秀用例：比如用Claude完成创作项目、练习口语考试等。但同样存在令人担忧的大规模使用模式。我们呼吁所有AI实验室共享类似数据，因为这些问题绝非我们独有。

We were able to share this data. And there were lots of great use cases, right? So people using it to build their own projects and artifacts with Claude, practicing for oral exams in school and all of these things. But importantly, there were also a lot of these worrisome use cases that people have been wondering about, but we could see it at scale on our platform. And we would love all the AI labs to share similar types of data because I don't think this is specific to us.

正如你所说，我们必须建立分类体系——单纯看对话数据很难判断学生是在寻求作业帮助、备考，还是自主学习。因此我们构建了全新的分类框架。

No. As you mentioned, we had to figure out a way to categorize this. It's hard if you just look at the AI conversation data to know exactly the context in which a student's using it. Are they asking help on a particular question or a homework problem or an exam they're preparing for, just some self learning they're doing? So that's why we built that taxonomy that you mentioned.

实际上你们和Ben在初期提供了很大帮助。由于现有分类法无法满足需求，我们自下而上创建了新体系。特别要提布鲁姆分类法——教育工作者都很熟悉这个从记忆事实到分析创造的认知技能金字塔。作为前高中数学教师，我经常运用这个框架。

And actually I will say you and Ben were very helpful as we reached out initially to you all to get a sense of what are the existing taxonomies out there to study what student usage looks like of these AI tools. We really couldn't find one that encompass what we were looking for. So that's why we built this new one in sort of a bottoms up way. But the part that I want to pull out is Bloom's taxonomy, which, you know, many of your educators, you know, listening to this will have thought about it in a classroom context. And, you know, I was a former teacher, high school math teacher, and I was often thinking about this taxonomy of these, you know, cognitive skills that I want my students to work their way up from just simply, you know, recalling a fact and remembering it all the way up to being able to analyze and create their own knowledge.

没错。

Right.

我们逆向分析了Claude在对话中展现布鲁姆分类法各层级的能力，结果令人惊讶：呈现倒金字塔结构——Claude更频繁地展现高阶认知技能（如创造分析）。从图灵测试角度看这很惊人，但从认知替代角度引发了担忧：如果学生依赖AI完成高阶思维，他们自身的思考能力会如何？好消息是当学生与Claude协同创作时，双方都能处于创造层级——但这不会自动发生。

And so what we did is we took all of these conversations and we actually sort of flipped it and said, can we analyze Claude's Claude's behavior in these conversations with students to see how well it's exhibiting different parts of the Bloom's taxonomy? What we found surprised us, which is that it was an inverted pyramid where Claude was often performing these top cognitive skills, creating and analyzing way more than it was at the bottom of the pyramid. And so from a turn test perspective, wow, this technology is incredible. But from a cognitive offloading perspective, started to raise concerns about, well, if the students are not, if they're able to work with an AI and have the AI exhibit these top cognitive domains, then what does that mean for their own thinking? The good news is that if a student and Claude are co creating, they can both be at that top level of creating together, but it doesn't necessarily happen out of the box.

因此，我们在思考自身产品开发时，必须考虑如何促进这一点。关于这个问题最后要补充的是，克劳德在这个金字塔模型中存在一个明显短板——评估环节。克劳德本身并不擅长评估工作，而这正是人类不可替代的领域。当我们讨论AI素养，以及2025年如何成为AI的优秀使用者或协作者时，输出评估能力至关重要。所以在师生培训中，我们常常重点培养这项能力。

And so that's what we have to think about when we think about our own product development is how do we encourage that? And then one last thing I'll say on this is there was one gap in this pyramid where Claude was not doing a lot of it, which is evaluating. So Claude itself was not necessarily evaluating a lot of things, but that is where the humans, of course. And when we think about like AI fluency and what does it mean to be a good user of AI or collaborator of AI in 2025, being able to evaluate the output is so, so important. So when we work with students and teachers, it's often focusing on that competency.

这确实合理，因为包括克劳德在内的前沿通用模型，其内置行为机制本就不包含评估功能。比如当学生提出'我想写篇论文，这是我的论点'时，系统设计初衷就不是要评判'这个论点不够好，我认为你可以改进'这类反馈。所以评估环节成为分类体系中的薄弱链路，本质上是设计理念使然。

It also makes sense because the sort of built in behaviors of, especially in sort of general purpose mode of the frontier models, including Claude, are not to evaluate you know, if a student comes in and says, I'm thinking this and this. I want to write a paper, here's my thesis. It's not sort of designed to be like, I'm not sure that's a great thesis. If I'm evaluating it, and I think you could do better, that's sort of not you know? So it makes sense that the evaluation is sort of the weak the weak link in the taxonomy just because that's not how it's designed.

关键点在于，大语言模型最初就是作为应答机器被构建的。

The key points as well is that, like, LLMs were originally built to be answer machines.

确实如此。

Exactly.

它们被设计来回答问题。但优秀的教育者从不直接给答案，教育场景中的大语言模型也不应该如此。这正是我们去年四月发布的学习模式重点突破的方向——相信后续我们会深入探讨——现在行业已涌现多种学习模式，这些尝试都在表明：这项技术本非为教育而生，却被广泛应用。

Answer questions. But great tutors and teachers don't just give answers, and neither should LLMs in the educational context. And so that is the key thing that we have been, and I'm sure we'll talk more about it, but that we've been working on in our learning mode that we released last April, and there's a lot of these learning modes that we're starting to see. All of this is trying to like, say this technology wasn't built for the education purpose. It's being used so much.

它蕴含巨大潜力，但我们需要引导其展现苏格拉底式的教学行为，让更多学生受益。

It has so much potential, but we need to help pull out these strong Socratic pedagogical behaviors, so that more students can access that.

完全正确。在Anthropic内部专注教育领域必定充满挑战，作为能与任何对手抗衡的顶尖前沿模型创造者，其工具和模型体系堪称卓越。但这些技术本质上并非为教育定制。

Exactly. So, I mean, it must be such an interesting role being inside Anthropic and really focusing on education because Anthropic is truly cutting edge, world class, one of the very, very few frontier LLM creators that can compete with absolutely anything. It's an incredible tool and set of models and all of these things. But it isn't necessarily custom built for education. It, in fact, is not built for education by nature.

你们如何平衡大语言模型通用应答器的强大功能与课堂安全有效的学习工具需求？这种既要避免成为答题机器，又要符合教学原理、促进学生思考而非外包其批判性思维的权衡，想必需要极其精妙的拿捏。

How do you sort of balance the general purpose answering machine power of LLMs with with at Anthropic with safe, useful, you know, tools for learning in the classroom? I think this is like that it must be such a tricky push and pull trying to think of how to put the pieces together in a way that that isn't an answering machine that is actually, you know, pedagogically sound and can actually help students think and not offload as they like to say or outsource their critical thinking. How do you balance it?

广告之后马上回来。

We'll be right back.

本季《教育科技内参》由Starbridge赞助播出。每年K12学区与高校支出超5000亿美元，但多数销售团队错失商机信号。Starbridge追踪董事会纪要、预算草案等早期迹象，助您快速转化为精准营销。在招标阶段前锁定胜局——这正是顶尖教育科技团队保持领先的秘诀。

This season of EdTech insiders is brought to you by Starbridge. Every year k 12 districts and higher ed institutions spend over half a trillion dollars, but most sales teams miss the signals. Starbridge tracks early signs like board minutes, budget drafts, and strategic plans, and then helps you turn them into personalized outreach fast. Win the deal before it hits the RFP stage. That's how top ed tech teams stay ahead.

这是个很好的问题。我要说的是，我们目前正为教育领域定制开发，因为我们意识到这是必须优先考虑的领域。但正如你所言，我们首先是一个AI实验室，团队里大多是技术专家。我们也明白教师参与的重要性。你知道，我提到过，我曾站在讲台上。

It's a great question. And I will say we are custom building it now for education because we realize that this is an area we have to prioritize. But to your point, we are an AI lab first and foremost, and we are many technologists here. We also realize it's important to have teachers. You know, I mentioned, I was in the classroom.

Anthropic有很多教师参与这些项目工作。我认为其他AI实验室也开始出现这种趋势，这很棒。当所有AI实验室都开始招聘更多前教师、教育工作者和教育科技从业者时，这非常重要。当然，应用层正在发生许多变革，我们有许多优秀的教育科技机构基于我们的API开发，深入解决课堂中的具体用例。

We have a lot of teachers here at Anthropic, particularly working on these initiatives. And so I think you're starting to see this at other AI labs as well. And I think it's great. I think when all the AI labs start to hire more former teachers, educators, people who've worked in ed tech, that's really important. Of course, there's a lot of stuff happening at the application layer and we have a lot of amazing ed tech organizations that build on top of our API that go much deeper into particular use cases in the classroom.

全球的魔法学校都在使用Claude，他们喜爱它正是因为我们对隐私的严格保护以及API自带的安全护栏。这些都很重要，但我认为让教育工作者直接加入AI实验室同样关键——因为即使有专门的教育工具，人们仍会使用Claude这类通用聊天机器人咨询教育问题。这不是关于如何解决这种矛盾的直接答案，我们始终面临这种张力：有人（包括教师）会从效率角度使用工具，比如快速生成教学大纲初稿再修改。对正在编写大纲的教师来说，苏格拉底式对话未必是最佳形式。

Magic schools of the world use Claude's, and a lot of them love it because of the, you know, our strong focus on privacy and all of the guardrails that come with API. So all of that's really important, but I think it's also important to have some of these educators in the AI labs themselves, because even when you do have these specific tools for education, you will still find that people come on to Claude or these other chatbots, the general purpose tool and ask education questions of it. So this isn't direct answer question about how to deal with that tension. Other than we deal with the tension all the time of, you know, some people are trying to use things in a productivity perspective, even teachers, sometimes they just, you know, they want to create a syllabus for their class as a first pass that they can then revise. And so having a Socratic dialogue for a teacher, who's trying to create a syllabus.

他们可能只需要生成对课堂有帮助的内容。核心问题始终是：如何在满足用户现有需求的同时，鼓励更多主动学习？

May not be the best format. They may just want to create some output that will be helpful for their class. There's always this question around how do you encourage people to have more of this like active learning while also meeting them where they're at?

这种张力很有趣。你认为学习层应该发生在哪里？是在核心模型中，还是在魔法学校这类应用层？不同层级的平衡与评估持续进行着，教育科技在其中扮演重要角色。你曾担任教师、导师、非营利组织创始人，现在又任职可汗学院...

It's a fascinating tension. Think your point about sort of where does the learning layer happen? Is it happening within the core model or is it happening within the application layer, like a magic school or what is happening at what level is, is something continuously being balanced and measured, and ed tech is is a huge part of that. And, you know, you you've been a teacher, a tutor, a nonprofit founder. You're working in Khan Academy.

你深耕教育科技领域，提到Anthropic有许多教育工作者和行业老兵，这令人振奋。我们也采访过可汗学院联合创始人、现就职Google AI的Shantanu Sinha。你们是好友对吧？请谈谈你十余年的导师经历...

You've been in the ed tech world, and you mentioned that a lot of educators and ed tech veterans work at Anthropic, which is amazing to hear. We've also talked to, you know, Shantanu Sinha, one of the cofounders of of Khan Academy who now is at doing Google AI. He's good friends. Right. Tell us how your personal experiences you said a tutor for over a decade.

我也有教学经验。这些经历如何影响你在Claude设计测试学习功能和能力时的视角？你认为哪些经验对当前讨论最具贡献？

I I have that experience too, a teacher. How do they influence the way that you personally look at designing and testing the learning capabilities and features at Claude? What do you bring from those experiences that I think that you feel like add to the conversation?

是的，如果你有辅导经验应该能理解。教师、导师、科技非营利工作者这些身份塑造了我的世界观。其中一对一辅导经历尤为珍贵——那种全身心支持学生的理想化互动，在课堂环境中往往难以完全复现。

Yeah, I mean, I think you can relate to this if it's through all your tutoring and my tutoring as well. All of those identities of being a teacher, a tutor, working on a tech nonprofit, all of that are key to how I see the world. But the tutoring one in particular, think is, especially in this idealistic one on one tutoring interaction, which I love, if you can get time to just work one on one with the student, really support them is like, it's just so precious. And so often in the classroom, you're trying to approximate some of that, right. As a teacher.

辅导经历让我意识到：当要求中学生互相辅导时，他们本能会花90%时间讲解概念，只留10%时间提问。但研究表明，优秀导师的时间分配恰恰相反。虽然实践起来很困难，但当导师只讲20%，学生表达占80%，并逐渐形成苏格拉底式对话时，这种比例调整至关重要。

But one thing that you realize from tutoring is that if you ask a student, and I did this a lot in my classroom where I would ask my middle and high schoolers to peer tutor each other, and often their initial instinct was to go explain the concept in-depth to their fellow student. And they would spend 90% of the time talking and then they would ask some questions. And then the person they were tutoring, the tutee, would spend 10% of their time talking. And then when you look at the research and you look at just some amazing tutors, it's exactly flipped. And it is pretty painful and hard to do well.

我们在World Schoolhouse项目中也遵循类似理念：如何将这种技巧传授给全球15,000名高中生导师？这不仅是培养对话能力，更是要把握这种黄金比例。

But if you have someone where the tutor is talking 20% of the time and the students that they're tutoring is talking 80% of the time, and then you start to see maybe some of the Socratic dialogue, but not just that, but really trying to get that ratio right, at least as a proxy. That's really important. And so when we're working on Schoolhouse of World, it's a similar type of thing. It's like, how do we teach that skill to the tutors? We have 15,000 high school tutors around the world tutoring.

所以我们正在尝试教授这一点。现在有趣的是在AI实验室里，我们也在思考如何让AI也具备这种行为。没错，因为AI尤其是大型语言模型会讲很多话，但通过恰当的提示和产品功能设计，可以引导它们。

So we're trying to teach that. What's interesting now is in an AI lab. It's also like, how do we get that behavior in the AI as well? Exactly. Because the AI also will talk all the LLMs will talk a lot, but again, with the right prompting and with the right product affordances.

如果不仅仅是聊天机器人，而是有各种交互元素让学生共同创造，我们看到学生甚至使用Claude代码。在那里更像是双向对话，但我觉得这就像我亲身一对一辅导经历中的要素，既适用于教学场景，也适用于学校教育，现在对Anthropic同样重要。

And if it's not just solely a chat bot, but if you have artifacts and all of these different interactive elements as well, where students can be co creating, we see students even using Claude code. And so there it's much more of a conversation back and forth, but I think that's like an element that I've just had with my own one on one tutoring experiences that then was relevant in teaching, relevant in a school style world, and now relevant at Anthropic.

你说得太对了，有意思的是我一直在试用各种前沿模型的不同学习模式。确实如你所说，要不是你点出来我都没意识到——即使在学习模式下它们也特别话多。

You make such a good point, and it's funny. I've been playing with the different study modes and and learning modes in all of the frontier models. And you're right. Something that I don't think I would have noticed it until you said it, but it's really true. They even in the learning modes, they're very garrulous.

对吧？你说一句，它就会抛出一大堆：'这里有几种思考方式，你对哪种感兴趣？你可以这样做或那样做，我来解释...'

Right? You say one thing, and it sort of gives you a whole slew. It's like, well, here here's a whole bunch of ways to think about that. Which one are you interested in? And you could do this or this or this, or let me explain it.

我知道这比通用模式要好，学习模式确实更专业。

And I know it's it's better than you know, the learning modes are better than the off the shelf general purpose mode.

但还是太啰嗦了。

But it's still too much.

但还是太啰嗦了。

But it's still too much.

没错，确实。

Right. Right.

真有意思，我们总能拿到这些播客的文字稿。我个人的目标是尽量少说话，但从没联想到辅导场景。你说得对，最好的辅导者应该多倾听，尽可能少说。

And it's so funny. We always get the transcripts of all of these podcasts, a personal goal for myself is to sort of talk as little as possible in these. And I never connected it to tutoring, but I think you're right. The best tutors do a lot of listening and that much talking as much as possible.

我喜欢听你说话，这是对话嘛。如果这次对话比例不同也别太苛责自己，毕竟是交流。不过我明白你的意思。

I like when you talk, it's a conversation. So don't be too hard on yourself if if the ratio in this conversation looks different because it is a conversation. But I I hear your I hear your point.

我对此表示赞同。确实如此。虽然这是一场对话，但我认为工作面试是另一个例子，你知道，当面试官说得越多，对候选人来说并不总是好事。通常面试官说得更多反而显得古怪。

It it I appreciate that. True. It is a conversation, but I think job interviews are another example of, you know, when when the the more that the interviewer speaks, they sort of it's not always that's good usually for the candidate of the interviewer speaking more. It's a wacky situation.

哦，

Oh,

有意思。你提到了Schoolhouse World。它完全建立在同伴辅导的基础上，围绕人际连接构建。

interesting. So you mentioned Schoolhouse. World. It's built all around peer tutoring. It's built around human connection.

我很好奇，当你思考Anthropic和Claude的未来时，你们这周刚宣布了一个新模型。另外我想稍微提一下Andrew Ng的新课程，关于深度学习和Cloud Code的，据说是目前最尖端的编码工具。我自己不是程序员，无法评价，但听很多人这么说。你们Cloud领域有很多很棒的发展。

I'm curious, as you think about sort of the future of Anthropic and Claude, you just announced a new model this week. There's also a I I wanna give a little pitch for there's a new Andrew Ng class, you know, for deep learning about Cloud Code, which is, by all accounts, the absolute cutting edge best coding tool. I'm not a coder myself, so I can't speak to it, but I've heard that from many, many people. Lots of great stuff happening in your Cloud

Code让每个人都成为程序员。

Code makes everyone a coder.

所以他们

So They they

之后，你和我可以再一起捣鼓点什么。

after this, you and I again hack on something together.

就这么办。我们来做点

Let's do it. Let's do some

但说到你的观点，我会

But to your point, I'll

继续。但当你展望AI辅导或AI对话的未来时，你认为AI是否会有机会支持更多社交和多人互动？没有人希望看到未来每个学生只与电脑直接互动。我们看到一些教育科技初创公司开始在应用层引入促进、小组支持或同时为多个孩子提供辅导。我好奇这是否是你在基础层会考虑的事情。

go ahead. But when you look to the sort of future of AI tutoring or AI conversations, do you think there's going to be an opportunity for AI to support more social and multi human interactions? None of us are want a future where every student is only interacting directly with a computer. We've seen some ed tech startups that start to bring facilitation or small group support or tutoring for multiple kids at the same time, they've done it at the application layer. I'm curious if it's something you ever think about at the foundational layer.

我百分百同意。这正是目标所在——促进更多且更高质量的人际互动。以我联合创办Skull Star World的经历为例，与Sal Khan、Shashira、Moriah等人共事时，我们核心的洞见就是：教育本质在于人际互动。许多可汗学院用户都渴望这种互动，比如‘看完Sal的视频后如何与他交流？’

I agree 100%. That's the goal. The goal is to have more human to human interaction and more quality human to human interaction. And I can speak to when I co founded Skull Star World with Sal Khan, Shashira, Moriah, like all of us, our key insight here was that human to human interaction is at the core of education. And a lot of people were craving it on Khan Academy of like, how do I chat with Sal after he does one of the videos?

这种对话需求难以规模化。因此在学校之家项目中，我们致力于构建完全免费的志愿者同伴辅导系统，连接成千上万学生实现人际互动。后来AI技术兴起，我们开始将其用于后台支持——比如至今仍在实施的：记录辅导对话并保存30天。

You know, I want to have this conversation and that's hard to scale up. And so with schoolhouse, our world, it was about how can we create this entirely free peer to peer volunteer tutoring system where we could connect many thousands of students to have these human to human interactions. And then of course, AI came out and we started to use AI, but it was often behind the scenes. It was using the AI to, for example, and we still do this now at school house. We record the conversations, save them for thirty days.

AI会分析这些辅导对话的质量，为导师提供个性化反馈，通常建议他们调整说话比例，多倾听和提问。这是AI促进人际互动的典范。如今在Anthropic也是如此，这也是我加入的重要原因——价值观高度契合，尤其体现在公司名‘Anthropic’（人类中心）中。

And then from these tutoring interactions, and then the AI can analyze how the conversation went, how the tutoring interaction went, and then give personalized feedback to the tutor, often encouraging them to have changed their talking ratio and spend more time listening and asking questions of their learners. So that's a great example of an AI model encouraging humans to spend more time with each other. And now it's the same thing with anthropic. And this is a big part of my reason to also come to anthropic. Lot of alignment with the values of anthropic, but one of them being it's in our name, anthropic.

人的因素至关重要。比如教师使用Claude的主要场景是处理行政事务，节省数小时工时后，他们就能在办公时间多与学生互动。学生也是如此——他们用Claude学习概念后，理想状态是回到课堂进行实践项目合作。当然，这说起来容易...

The human element matters a lot here. And so when we see, for example, teachers using Claude, Some of the top use cases are helping them do sort of administrative tasks and things that would have taken them hours before so that they can then spend more time like in office hours with students, you know, same with the students. Like we see them learning about a concept on Claude. Of course you are interacting with an AI there, but the ideal thing is then you can come into the classroom and spend more time on hands on projects, working with other students. So this is easier said than done.

我并非轻视此事。强大技术必然会被高频使用，但目标不是让人沉迷AI，而是解放时间用于人际互动。

I'm not trying to like trivialize this because if you do build a technology and it is very powerful, like people will use it a lot, but the goal is not to have everyone spend all their time in the AI. It's really to open up and free their time to have more of these human to human interactions.

完全同意。我曾与Claude头脑风暴过AI辅助 facilitation（引导）的场景。比如小组讨论时，AI可以负责破冰、确保任务进度、明确角色分工——这种‘地下城主’式的引导模式，能让学习体验既丰富又充满互动性。

A 100%. I mean, one use case, I was brainstorming with Claude about this for a while at one point, we were thinking about different ways that AI could support facilitation. And, you know, something as simple as when you have a small group breakout, AI could be the one do all the icebreakers and keep people on task and make sure that everybody is aware of the instructions and what their roles are and sort of keeping things moving. Like, there's a lot of really interesting opportunities for AI to sort of serve as a facilitator for multiple people and make pure learning into something that's just really rich and exciting. Almost like they can almost be like a dungeon master.

没错。

Yeah.

这种轻度引导下人们仍保持协作交流的模式非常激动人心。关于Claude的学习模式——它源自真实学生行为数据，你们是如何将学生直接输出、协作等使用模式转化为这个产品功能的？

In a, learning experience where it sort of is slightly it's just guiding everybody, but they're still working together, conversing, talking. It's a really exciting model. So I wanna ask more about the Claude's learning mode. It's been really interesting, and it was inspired by real student behavior. You obviously, as you said, you did aggregated data analysis to see how students were using Claude.

能否详细说明这个从50万次对话数据分析到学习项目功能落地的过程？

Can you walk us through that sort of process of taking all of that data that you you were finding from how students were using Claude for direct output, for collaboration, and sort of baked it into this product feature about these learning projects and this learning mode?

今年春季我们发布学习模式前，定量分析了50万次Claude对话数据，发现需要超越问答的苏格拉底式对话功能。虽然教师群体确实强烈建议——他们关心学生的长期学习效果——但焦点小组显示，这需求其实植根于更广泛的使用行为。

So, yeah, we released our learning mode this past spring after we saw a lot of the data from our research at the quantitative level in terms of studying these half a million conversations of Claude. And we saw the need for the ability for Claude to not just answer questions, but also to ask you questions and to have the Socratic dialogue and all the best practices that teachers and tutors know, but to bring that into cloth. The thing is we also did a lot of focus groups and people would often think that it was the teachers that were telling us, oh, you need the learning mode. And we did, of course, hear this from teachers. Like, you know, they want what's best for their students' long term learning.

因此他们担心作弊问题。学习模式在那时确实有帮助，但关键是在那个特定时刻。我记得我们与一群大学生焦点小组的对话，他们可能是最早提出‘学习模式’这个概念的人。他们担心——现在依然担心——‘脑力退化’，这是他们的表述方式。

And so they were worried about cheating. And so learning mode was helpful there, but it was the specific moment. I can remember the conversation we had was with a focus group of college students. They were the ones who I think first maybe even coined the term learning mode And they were worried and they are worried about brain rot. And that's the way they frame it.

有意思。

Interesting.

这种担忧当然合理。学生最清楚自己需要什么，但长期学习效果与短期需求总有矛盾。比如深夜赶作业时，如果工具让他们轻易依赖AI作为拐杖，所有人都会这么做——不仅是学生，成年人也一样。但令人鼓舞的是，学生们也渴望产品能提供促进真实学习的功能设计。

And of course it makes sense. I think students know what's best for themselves, but there's always this challenge of what's best for their long term learning. And then in the short term, you know, it's late at night and they need to get ready for finish some assignment. And if the tools make it so easy for them to just steer in the direction of just using the AI as a crutch, then we will all do that, not just students, adults as well. And so, but it was very promising that it was, you know, students also want to have these product affordances and product features that encourage more of learning.

这正是我们努力的方向。值得高兴的是，所有AI实验室——我们既非首创也非唯一——最近几个月都在开发各类学习模式，OpenAI和Gemini都在做。我们称之为‘向上竞争’，尤其在教育这类有益领域，各实验室竞相投入更多教育资源、开展相关研究、开发学习功能，并直接与师生合作开发——这种良性竞争非常可贵。

And so that's what we've been working on. And I'll just say, like, I've been very delighted to see that all of the AI labs and we weren't the first and we are not the last either. There's been a lot over the last few months that have been working on different types of learning modes and OpenAI and Gemini. And for us, we call that race to the top where we think, especially in an area like education or beneficial deployments, having a competitive pressure for all of the AI labs to be doing more in education, more research around it, more product features to support learning and really working directly with students and teachers to build these features. Like that's really good.

确实。

Yeah.

这与其他行业可能截然不同，但这是Anthropic的核心原则：向上竞争，而非向下沉沦。

So it's very different maybe from other areas of industry. We but but this is a general principle of anthropic. Race to the top, not race to the bottom.

说得好。这也源于教育政策，但确实如此。从我观察者的视角看，这种关于‘谁能打造最佳学习模式、最佳AI导师、最佳对话建构工具’的竞争——减少脑力退化，融入学习科学——简直是场精彩的竞赛。

I love it. Yeah. And and it's also something comes out of the school policy, but it's it's true. I felt that as well from my perch, you know, looking at these tools from the outside is it feels like the idea of there being a competitive who can have the best learning mode, the best AI tutor, the best conversational constructive, you know, reducing brain rot and incorporating learning science. Like, what an amazing competition to have.

我们正处在一个美好的时代。

Like This is a great universe that we're in right now.

没错。在这个时间线上，这些本可专注任何领域的大型科技公司都聚焦教育。快速聊聊学习科学研究——我知道时间有限，但我非常好奇。比如谷歌专门开发了LearnLM教育模型并整合进Gemini核心，成为其学习产品的导向。

Exactly. It's a good timeline where where all of these amazing big tech companies who frankly can focus on anything are focusing on education. Let's talk quickly about the the research and the learning science. I know I only have you for a few more minutes, but this is something I'm so curious about. You know, we saw in Google's case, they did this very specific thing where they made this LearnLM, this sort of model for education, and then folded it back into core Gemini, and then it became sort of the orientation that fuels their learning products.

我猜想Anthropic可能——当然不涉及商业机密——你们如何将苏格拉底对话法、布鲁姆分类法等已知的学习科学原理，融入Claude学习模式的人格设定中？

And I would imagine that Anthropic may you know, without giving any industry secrets away, I'm not asking for them, but, you know, I'm curious how you think about getting the research. You've mentioned Socratic dialogue or the Bloom's taxonomy. You know, some of the things that we we know about learning science from the research and tutors know. How are you thinking about getting that into the mind frame of Anthropix, into Claude's sort of personality when it comes to learning mode?

是的。我认为答案是以上所有因素的综合。在模型层面需要做些工作，确保这些模型不仅训练过问答、生产力等用例，还要考虑教育场景的需求。然后在提示词层面也有很多工作，你提到的所有这些学习模式，我认为提示词设计是关键部分——我们发现大约5%（随便说个数字）的学生能以非常复杂的方式使用Claude，将其转化为苏格拉底式的导师互动。但并非所有人都会花时间写长篇系统提示。

Yeah. I think it's the answer is all of the above. There's some stuff at the model level, making sure that these models have been trained on not just, you know, question answer, productivity use cases, but thinking about things that also make sense for the education context. Then there's stuff at the prompting level, which you see a lot of, all of these learn modes that you've mentioned, I think have some pretty key part at the prompting level, which is, again, we found that 5%, know, just throw out a number there, maybe 5% or so of our students were using Claude in these very sophisticated ways and turning it into Socratic tutor type of interactions. But not everyone is taking the time to write a long lengthy system prompt.

所以这正是AI实验室可以大展身手的地方。此外产品层面也需要考虑如何呈现这些功能。这是全方位的工程。我认为谷歌的LearnLM项目就是个很好的范例，他们在这方面已经探索很久了。

So we, that's where the, I think the AI labs can help a lot. And then there's stuff at the product level as well to like how to surface that. So it's all of the above. And I think, you know, Google with LearnLM, they've been doing this for a while, think is is a great example for for the broader field.

这个区分非常精准。这种变革发生在体验的多个维度——提示词、产品、模型，最终形成完整拼图。看到这个领域持续演进真是令人振奋。

Yeah. That's a really great distinction. It happens with is within multiple aspects of the experience, the prompting, with the product, with the model, and you can sort of put all the pieces together. And I'm just excited to see it continue to evolve. It's really exciting.

Anthropic的学习模式非常出色，更激动人心的是最近几周，我们看到多家主流模型供应商和实验室都推出了自己的学习模式版本，现在完全是场良性竞赛。

The learning mode in Enthropic is really great, and it's so exciting that, you know, in the last few weeks, we've seen several of the major model providers and labs present their version of learning mode, and now it's just a race to the top.

我想提前透露个消息——不知道这个功能会在视频发布前还是后上线——很多用户提到用Claude Code编程。我们收到大量CS专业学生、初级程序员和从业者的反馈：'我不想让编程能力退化，不想失去这项技能'。

I'll I'll I'll preview something that's actually coming I don't know if it'll come before or after this gets released, but we've you know, a lot of people, you mentioned, use Claude code and Claude for coding. And so we hear this a lot on the programming, the CS student side, and really just everyone, junior programmers, people in the industry saying, I don't want my programming skills to atrophy. I don't want to lose, lose that.

或者当我正在学习...

Or if I'm learning for the

编程时，面对能完成20分钟任务的编码助手，我该如何学习？所以先剧透下：我们即将为Claude Code推出专属学习模式。

first time, like how do I learn when that coding agent can do twenty minutes of a task? And so the little teaser here is that we will shortly be launching a learning mode for Claude code as well.

而且

And

我认为就像其他学习模式一样，这只是起点。我们需要持续迭代，目前收到很多反馈，还需要更多。但这个问题至关重要——当人们问'学编程的人该怎么办？该不该学编程？'

so I think this is like all these other learning modes. Just the start. Like we need to iterate on this and we've been getting a lot of feedback. We need more feedback, but I think this is going to be really, really important when everyone is asking, well, what about someone who's learning to program? Should they learn to program?

他们该如何学习编程？我们正在所有产品界面层面思考这个问题。

How can they learn to program? And so we're thinking about this at the level of all of our different product surfaces.

这太不可思议了，而且本该如此。你可以直接开始创建自己的应用程序，还不需要懂编程，但过程中应该边做边学。因为如果只是让工具代劳，你自己就学不到真本事。所以这真是个非常非常振奋人心的消息。

That's incredible. And that's exactly how it should be. You can jump right into creating your own applications without having to know how to code yet, but then you should be learning along the way. Because if it's if you're just letting letting the the tools do it for you, you're not learning how to do it yourself. So that's very, very exciting news.

我们会看看时间是否合适，希望这期播客发布时它也能上线，我们会在节目备注里附上链接。如果还没发布，那也是近期值得期待的事。真的很令人兴奋。Drew Bent，这次对话非常愉快，随时欢迎你再来。

And we'll we'll see if the dates work out, but, hopefully, that will be out by the time this podcast is out, and we will link to it in the show notes. If not, that's something to look forward to very soon. That's really exciting. Drew Bent, this has been a a pleasure. You're welcome back anytime.

我真的是Claude的日常用户，几乎什么事都用它。Drew Bend负责Anthropic教育领域的良性部署，还是与Sal Khan共同创立辅导非营利组织Schoolhouse的联合创始人，工作成果非常令人振奋。

I I'm literally a a daily user of Claude. I use it for almost everything. And Drew Bend leads education as part of Anthropic's beneficial deployments. He was also the cofounder of the tutoring nonprofit schoolhouse with Sal Khan. Really exciting work.

非常感谢你来到《教育科技内幕》节目。

And thank you so much for being here with us on EdTech Insiders.

太感谢了Alex，我是这档播客的忠实粉丝，能来做客真是荣幸。

Thanks so much, Alex. Big fan of the podcast. It's an honor to be here. Aw.

感谢收听本期《教育科技内幕》。如果喜欢我们的节目，请记得评分并分享给教育科技界的同仁。想获取更多内容，可订阅Substaff上的免费《教育科技内幕》电子报。再次感谢本季赞助商Tuck Advisors，他们很高兴推出首款并购AI工具——并购矩阵。

Thanks for listening to this episode of EdTech insiders. If you like the podcast, remember to rate it and share it with others in the EdTech community. For those who want even more EdTech insider, subscribe to the free EdTech insiders newsletter on Substaff. Thanks again to our season sponsor, Tuck Advisors. Tuck Advisors is pleased to introduce the first ever m and a AI tool, the m and a matrix.

这个定制版Chat GBT能帮你探究任何过往或未来并购交易背后的原因。欢迎试用体验并反馈意见，联系邮箱confidential@TuckAdvisors.com交流见解。

This custom chat GBT helps you explore the reasons behind any past or future m and a transaction. Play, test, and enjoy it, and they welcome your feedback. Contact them at confidential@TuckAdvisors.com to discuss your insights.