教育困境与AI的必然性
二十年前，我创立了一家社会企业，立志改变世界。我们为贫困人口提供了数百万份餐食，在非洲部分地区普及了数万剂疫苗，并在印度贫民窟资助了学校。我曾以为自己在这些领域做得相当不错，产生了巨大的影响力。然而，一次与英国政府官员的交流，让我意识到问题的严重性：20%的学生在完成中学教育后，仍无法达到足够的读写能力。我开始反思，如果连拥有实体学校和合格教师的英国都存在这样的问题，那么我在印度贫民窟学校的努力，其影响力又打了多少折扣？究竟是什么环节出了错？我们必须找到解决之道。
于是，我深入学校，与大量师生交流。我发现了教育一线存在的两个关键问题。第一，教育交付模式依然是“一刀切”，面对的是一个大约30到35人的班级。第二，一个应该每天都成为头条新闻的事实是：74%的教师计划在未来三年内辞职。原因何在？是工作量。他们花费大量时间进行微观标记、微观评估，试图弄清楚每个孩子的学习进度。他们白天是教师，晚上却成了数据分析师，而这绝非他们当初选择这份职业的初衷。
在走访学校的过程中，我手中拿着智能手机，而机器学习应用早已能指导我们如何购物、如何储蓄、如何睡眠。我由此想到：为什么教室里不能有这样的技术，来指导我们如何学习呢？我们必须构建这样的技术。但我们不能仅仅依赖任何现成的机器学习推荐引擎。我们需要将人工智能（Artificial Intelligence: 模拟人类智能的计算机系统）与神经科学理论（Neuroscientific Theory: 关于大脑结构、功能和学习机制的科学解释）以及学习科学（Learning Sciences: 研究人类如何学习的跨学科领域）相结合，去理解这个房间里每一个大脑的学习方式。因为如果我们能解决学习的根本问题，就能显著改善学习成果，为每个人实现个性化教育，并为教师提供智能化的洞察，从而减轻他们的工作负担。
十二年前，我组建了一个团队，他们开发出了这项技术。如今，这项技术已在140多个国家被学生使用，我们收集了超过400亿个关于儿童学习方式的数据点。接下来，我将分享一些我在学习过程中学到的关于学习的洞察。但在那之前，我认为分享一些学生在我们平台上的反馈至关重要，因为它能揭示孩子们在使用AI教育伙伴时的期望。
Original English
Twenty years ago, I founded a social enterprise. I wanted to change the world. And we were funding millions of meals to the underprivileged. We were providing tens of thousands of vaccines across parts of Africa. And we were funding schools in the slums of India. Now I thought that I was doing quite a good job and having a lot of impact in all of these areas, until one day I was working with ministers in the UK, and they said that 20 percent of students leave secondary schools in the UK and they're not able to read and write well enough. Now I thought with brick and mortar schools and qualified teachers, if they're not able to do that in the UK, then I'm not having the impact that I wanted to have in those schools, in the slums in India. So what's going on? What's the problem? We need to fix it. So I went to schools. I went to schools and I asked lots of questions. And I found two critical problems on the front line of education. The first is that they continue to have the one-size-fits-all delivery of education to a classroom of around 30 to 35 people. The second, I think you will agree with me, should be headline news every single day. 74 percent of teachers want to quit their jobs in the next three years. Why? It's because of workload. They spend so much time micro-marking, micro-assessing, trying to figure out where every child is at. They are teachers by day, and they are data analysts by night. And not one of them signed up to do that night job. So walking around schools, I had a smartphone in my hand, and we had machine-learning applications telling us how we should shop, how we should save and sleep. And I thought, why don't we have this technology in the classroom, telling us how we should learn? We’ve got to build that technology. But we can’t use any old machine learning recommendation engine. We need to combine artificial intelligence with neuroscientific theory and the learning sciences to learn how every single brain in this room learns. Because if we can fix learning, we can improve outcomes. We can personalize education for every single one of us and provide intelligent insights to teachers to reduce the workload. So 12 years ago, I built a team. They built the technology. It exists, students use it in over 140 countries. We've collected over 40 billion data points on how children learn. And I'm going to show you a couple of the things that I've learned about learning on the way. But before I do that, I thought it would be really important to share with you some student feedback that I have on our platform. It’s really important because it tells us what children’s expectations are when they use an AI education partner.AI捷径的陷阱与真实学习的本质
我收到的反馈包括：“我试着说谢谢。”“这很棒。”“太精彩了。”“我认为Century将帮助我实现我曾认为不可能的事情。”——这听起来像是一个“金童”，我的毕生目标似乎已然实现。然而，这些天真可爱的孩子们也会发来这样的信息：“我不喜欢这个网站，它让我能够完成我的家庭作业。”（笑声）等等。
更甚者，我甚至收到了“贿赂”：“我给你10万英镑，我没开玩笑。你只需要给我布置很少的作业。给我一个按钮，帮我完成所有工作。”这些孩子们的想法，以及这种情绪，与一项近期调查的结果不谋而合。调查询问孩子们如何使用AI大语言模型（LLMs）和聊天机器人来完成家庭作业，结果令人震惊：五分之一的孩子承认，他们让AI代劳了所有功课。这意味着，他们并非利用AI来辅助学习，而是主动地规避学习。
我知道现在有些人可能在皱眉，心想：“他们怎么敢？”但我认为，他们的行为与我们自己并没有太大区别。回想一下我们第一次使用ChatGPT（一种大型语言模型聊天机器人）时的感受。这是我非常科学的图表（此处省略图表描述）。我想，你们当时都感到欣喜若狂，心想：“哇，我看起来会像个天才。我再也不用工作了。这太棒了。”是吧？
然后，它出现了幻觉和虚构（hallucinated and confabulated）。你们可能会说：“科技巨头，说真的，你们只有一件事要做，Sam Altman，拿着那么多钱，结果却在编造信息，对吧？”接着，那位在法庭上分享了AI生成内容而被罚款的律师，也经历了类似的羞辱和尴尬。我认为，我们最终都陷入了一种“沉没成本”式的接受状态，意识到捷径并不能真正取代努力。它们确实很有帮助，但我们仍然需要学习、需要产出、需要思考。
当我们阅读大语言模型聊天机器人给出的长篇答案时，读起来确实感觉非常流畅，不是吗？问题在于，我们常常将这种流畅性误认为是学习本身。这也是为什么我们身边的人（当然不是我们自己）会陷入一种“能力幻觉”（Illusion of Competence: 误以为自己掌握了某项知识或技能，但实际上只是表面熟悉），觉得自己无所不知。
然而，关于学习，我们真正知道的是：学习需要研究者称之为“有益的挣扎”（Productive Struggle: 指学习过程中所需的、能够促进理解和记忆的认知努力）。正是这种认知上的努力，构建了我们的理解。
我总结了四种最重要的学习技巧，它们都涉及“有益的挣扎”，并且能显著改善学习成果。我们已经验证了它们的效果。其中三项与记忆有关。这一点至关重要：记忆与理解是同一枚硬币的两面。仔细想想，我们依赖回忆起来的知识来塑造我们的思考。如果我们无法回忆，就无法使用。
因此，第一项重要技巧是检索（Retrieval: 从大脑中回忆信息）。这仅仅是从我们大脑中提取信息的过程。在一项研究中，学生们被给予一段文字。那些只阅读了一遍，然后尝试从记忆中回忆的学生，比那些反复阅读的学生，记忆效果要好得多。
第二项是间隔（Spacing: 将学习分散在一段时间内进行）。这意味着学生将学习分散在一段时间内进行。因此，不是一次性死记硬背，而是通过随时间推移进行的积极回忆过程，你实际上是在一次又一次地经历那种“有益的挣扎”。
第三项，我们可能不太喜欢它，但它是生成（Generation: 主动尝试产生答案或内容）。在一项研究中，学生们被给予了词对，如“快速-快”和“冷-热”。另一组学生只被给予第一个词，以及一个提示，比如“F”，他们必须想出“fast”（快）。那些必须自己生成答案的学生，即使一开始答错了，也会形成更强的记忆痕迹，最终记得更多。
第四项是反思（Reflection: 回顾学习过程、目标和差距）。当我们反思自己的工作，并以三种非常具体的方式获得结构化反馈时——“我现在是如何学习的？”“我的学习目标是什么？”“为了达到目标，还有哪些差距，我需要做什么？”——这些学生的学习成果会得到提升。
你会发现，这四种技巧有一个共同点：它们更难。它们都涉及“有益的挣扎”。我们知道，持续的脑力劳动能够强化大脑的相应区域，并且与大脑的生长呈正相关。
Original English
It’s really important because it tells us what children’s expectations are when they use an AI education partner. So I get feedback like this. "I'm trying to say thank you." "It's lovely." "It's brilliant." "I think Century will help me achieve things that I thought were impossible." It's a golden child, right? My life's purpose has been fulfilled. And then these sweet, lovely, innocent children send me messages like this. “I don’t like this website, it makes me able to do my homework.” (Laughter) Wait. And then I'm being bribed. "I will give you 100,000 pounds, I'm not joking. You just need to give me no work. Give me a button to do the work for me." Now these children and that sentiment very much ties in with a recent survey where children were asked, how do you use AI LLMs, chatbots, with your homework? A staggering fifth of children admitted they get AI to do all of their work for them. So they're not using AI to help them learn. They're using AI to actively avoid learning. Now I know some of you are frowning right now thinking, how dare they? I don't think they're that different from us. Think about how we felt when we first used ChatGPT. This is my very scientific chart. I think that you all felt euphoric. You thought, "Wow, I'm going to look like a genius. I never need to do any work ever again. This is amazing." Yeah? And then it hallucinated and confabulated. And you were, like, "Big tech, seriously, you had one job to do, Sam Altman, with all that money, and it's making stuff up," right? And then for the lawyer who shared it in a courtroom and got fined, sheer humiliation and embarrassment for those people. And I think we've ended up with this sort of sinking realization of acceptance, right, that the shortcuts don't really replace the work. They're very helpful. But we still need to learn, we need to produce, and we need to think. Now when we read those long answers that an LLM chatbot gives us, it feels very fluent when you read it, doesn't it? The problem is, is that fluency we often mistake for learning, and that is why people we know, not us, of course, but they end up with this sort of illusion of competence, like they know everything, right? What we actually know about learning is that learning requires what researchers call a “productive struggle.” It's this sort of mental effort, right, that builds understanding. Now my top learning techniques, I’ve got four of them that all involve a productive struggle, and they improve outcomes. We've seen them work. Three of them are about memory. This is really important. Memory and understanding are two sides of the same coin. If you think about it, we draw on what we remember in order to shape what we think. If we can't recall it, we can't use it. So the first important one is retrieval. This is simply the act of recalling from our brains. The students in a study were given a passage like this one. And it's the students who only read it once, but then tried to recall it from their memory, who could remember it far better than students who just read it over and over and over again. The second is spacing, and this is essentially students who then space their learning over time. So rather than cramming things all in one go, students that can do that active process of retrieval over time, because then you're essentially going through that productive struggle over and over again. The third, we don't like this one, but it's just generation, right? So students in a study were given word pairs like rapid-fast and cold and hot. But then another set of students were just given the first word. And then a cue like the F, they had to come up with "fast." Students who have to generate the answers themselves, even if they get them wrong initially, create a stronger memory trace. They remember more in the end. And then the fourth is reflection. When we reflect on our work and we are given structured feedback in three very specific ways: How am I learning right now? What is my learning goal, and then what are the gaps to get to that goal, what do I need to do? Those students improve their outcomes. Now you'll find that these four techniques have something in common. They are harder. They all involve a productive struggle. We know sustained mental effort strengthens the parts of the brain, and it's positively correlated with growth in the brain.AI作为人类智慧的催化剂，而非替代品
在我家乡伦敦，有一项关于黑色出租车司机的惊人研究。在伦敦当出租车司机，必须通过一项名为“知识”（The Knowledge）的考试，你需要记住伦敦市26,000条街道。而且，你不能使用导航应用程序。这简直太疯狂了，对吧？他们没有Uber司机，对吧？神经科学家扫描了他们的大脑，发现经验丰富的出租车司机海马体（大脑中负责空间记忆和导航的部分）的某些区域更大。这是因为他们必须构建所有这些心智模型，每次载新乘客时都要生成新的路线。因此，这种增长，由于与他们必须做的事情呈正相关，是非常有意义且发人深省的，而这对于学习来说也同样适用。
持久的学习并非来自捷径，它源于特定类型的努力。而这正是AI在教育领域大放异彩的原因。因为AI能够识别我们所有人学习的模式，它能识别跨学科概念之间的联系模式。它能预测你何时会忘记某事，并在恰当的时机为你提供相关材料；它能提供及时、有针对性的干预措施，并赋予教师这些洞察。它能迫使你生成答案，而不是仅仅揭示答案。并且，它能根据教师精心设计的评分标准，提供惊人的结构化反馈。因此，设计精良的AI在教育领域可以发挥巨大的作用，我们已经看到了它的成效。
现在，很多人，包括学生和成年人，会来问我：“但是，为什么还要费力去学呢？我们有GPS，有AI，可以谷歌搜索任何问题的答案，所以我们不再需要这样做。”这种想法是错误的。
思考一下AI，AI是我们历史预测我们未来的能力。它在识别数据模式方面表现出色，在药物发现、蛋白质折叠、新材料和晶体等领域取得了卓越的突破，并且AI一直是这些突破的优秀伙伴。但关键在于，这些成就都不是AI孤立完成的。是我们人类，定义了问题，设定了目标，选择了数据集，决定了哪些发现是重要的。我们的知识不仅仅是琐碎的信息，它是思考和发现的原材料。AI并非要取代我们的专业知识，而是要让我们的专业知识得以扩展。
回想一下动力飞行、青霉素、电力，乃至AI本身。人类在学习过程中经历了“有益的挣扎”，建立了领域内的专业知识，并在此基础上，凭借想象力的飞跃创造了创新。因此，对于那些想通过作弊、想利用AI完成家庭作业的学生，以及对于我们这些终身学习者，那些不断阅读、不断巩固“能力幻觉”的人们，请记住：除非经历挣扎，否则你无法获得成长。
所以，AI对教育是好是坏，完全取决于你。是我们设计得好，还是你利用它来补充或取代人类认知？下一次当你学习、想要投资自己、自我教育、渴望成长，并可能实现那次想象力的飞跃时，请记住：脑力劳动并非过程中的缺陷，它是一个关键的特征，它让学习得以巩固，让我们能够建立专业知识，并为人类的创造力注入动力。
非常感谢您的聆听，祝您在AI之旅中一切顺利。 （掌声和欢呼声）