夜雨聆风
从2024年开始，我们阳民研究团队已经开始研究人工智能对社会创新的影响，重点关注“制度企业家”视角，我们将扩展制度企业家的研究案例，并拓展制度资本理论的适用范围。
从今天起，阳民将系统性梳理人工智能驱动社会创新的主要内容。今天我将介绍一个案例：Geekie的个性化学习。Geekie在教育机构利用人工智能进行社会创新方面是一个非常重要的案例。通过人工智能实现技能的个性化学习，Geekie公司为公立和私立学校提供服务。Geekie is a social enterprise using AI to personalise learning across large-scale education systems. By analysing student data in real time, the platform adapts content to individual learning levels, supporting both students and teachers in improving outcomes. Integrated into public education systems in Brazil, the solution enables more targeted instruction and data-driven decision-making at scale. This case study shows how AI can be embedded into existing systems to improve equity and effectiveness in education.
所有AI产品，这个个性化需求平台，采用了自适应评估和学习计划。平台的艺术性在于个性化，它使用项目反应理论（IRT）作为实现基础，其网络影响力覆盖了1000万学生，地理范围达到巴西90%的城市，包括最偏远和资源不足的地区。资金来源包括私人收入、公私合作伙伴关系，以及教育部的国家接入计划。
背景：巴西的教育系统呈现出一个惊人的悖论。公立学校系统服务于全国约85%的学生，但这些学校长期资源不足。来自高收入家庭的学生就读私立学校，在那里他们能获得密集的备考和支持，以便应对大学入学考试。然而，动态在这里发生了关键性的逆转：巴西最负盛名的大学是公立机构，完全免学费，但其录取完全取决于一年一度、难度极高的全国统一入学考试的表现。
这个系统的后果是可想而知的，也是严峻的：来自资金充足的私立学校的学生，占据了免费公立大学有限的入学名额；而那些来自资源匮乏的公立学校系统的学生，则发现自己是在“一只手被绑在背后”的情况下参与竞争。统计数据描绘了一幅黯淡的图景：在公立系统完成高中学业的学生中，只有约10%在核心科目（如葡萄牙语和数学）上达到最低熟练度。辍学率居高不下，原因是经济压力迫使男孩早早进入劳动力市场，以及少女怀孕问题迫使女孩离开校园。许多学生仅仅因为意识到自己落后太多而失去了动力。
正是这样的现状，激励了Geekie团队的创立。团队认识到，全国统一入学考试是公立学校学生通往机遇的最重要单一门户。如果他们能帮助这些学生补上多年的学业差距，就能从根本上改变他们的人生轨迹。创始人明白，所需的人力辅导和技能在财务上是不可行的。技术，特别是人工智能，提供了唯一可行的路径，以可负担的成本和规模提供个性化教育。
因此，他们以项目反应理论（IRT）为基础。在Geekie之前，许多教育公司曾试图通过内容交付来解决这个问题。他们以较低的成本提供课程和教材，希望能支持学生备考大学入学考试。但团队观察到了一个一致的模式：这些解决方案缺乏个性化，学生的参与度和进步有限，因为他们缺少理解所提供内容所需的基础知识。即使是高质量的材料，在学生无法弥补其基础学习差距、又缺乏个体支持的情况下，也被证明是无效的。
团队由联合创始人兼首席技术官Leonardo de Carvalho领导。他们的发现包括一个关键的洞见：他们研究了入学考试本身。巴西的大学入学考试使用项目反应理论（IRT）这一统计框架，根据学生对测试题目的反应来评估其能力。IRT考虑了每个题目的难度，能更精确地测量学习成果。全球各地的政府都使用IRT，因为它使得评估在不同人群和不同时间点具有可比性。对于Geekie而言，IRT之所以特别重要，在于其数学结构。该框架依赖于少量参数，并通过实际反应数据进行优化，这使得它在根本上与机器学习相似。正如Leonardo解释的：“统计学和机器学习非常接近。项目反应理论是一个带有自由参数的模型，你使用数据来优化它，所以它在某种程度上已经是一种AI。”事实上，他们已经在使用这样的模型，这给了团队信心，即采用数据驱动的算法方法进行个性化教学不仅是可行的，而且与现有的教育基础设施相契合。
编程语言和基础设施
团队面临的第一个技术设计是选择编程语言。在评估了类似选项后，他们选择了Python。当时，虽然其他语言如Ruby在Web开发中很流行，但Python提供了它们无法比拟的优势：一个强大的数据分析和科学计算生态系统。诸如SciPy和NumPy等库为统计分析和数值计算提供了强大的能力，这对于构建他们方法核心的数据驱动教育学习模型至关重要。在基础设施方面，Geekie从一开始就在云端运营，使用亚马逊网络服务（AWS）。团队依靠易于使用的虚拟服务器来托管他们的应用程序。后来像AWS Elastic Beanstalk这样无缝的计算平台还不存在，所以他们与基础服务合作，并自己构建了许多支持程序。
模型开发
模型开发由学术研究驱动。团队阅读了相关的机器学习和AI论文，并从头开始实现模型。这种实践方法使他们深入理解了支撑其技术的机制，并允许他们针对具体的教育挑战定制解决方案。
客户IRT实现
Geekie从一开始就秉承开源理念，将其大部分代码库公开，并依赖开源库来加速开发。然而，他们遇到了一个显著的缺口：当时，在任何编程语言中，可用于实现项目反应理论（IRT）的工具都很有限。团队别无选择，只能从头开发自己的IRT包。该模型实际上依赖于真实的学生答题数据来校准每个题目的参数。模型捕获三个关键特征：
难度：题目有多大的挑战性。
第二，区分度：该问题能在多大程度上区分理解该主题的学生与未理解的学生。第三，猜测参数：学生仅凭偶然机会答对题目的概率，这对多项选择题尤其重要。因此，为了估计这些参数，团队实施了教育...
A couple of thousand responses should be sufficient to calibrate each question. Once calibrated, the question parameters could be deployed to millions of students at virtually no additional cost. This efficiency was central to Jackie's ability to scale the Bain network for knowledge dependence as GEEKIE evolved from assessment to full learning data.
The team needed to move the grander module to work at the level of broad knowledge. It may seem messy or not make sense, but students need to gather knowledge at the topic level.
团队开发了一个贝叶斯网络，用于表示主题流。在该网络中，知识量以及主题之间的连接强度被量化。如果一名学生证明掌握了某个主题的技术，该网络就能计算该学生也理解相关主题的概率。
莱昂纳多用了一个简单的例子来说明这一点：在一元一次方程和一元二次方程中，通常需要先学习一元一次方程，才能解一元二次方程。
在我们的图谱中，会有这两个知识点，以及一个从“一元一次方程”到“一元二次方程”的连接。如果我们测试发现一名学生已经掌握了二次方程，我们就绝不会再建议他去学习一次方程。这种结构并非为了优化每个学生的学习时间，让他们只专注于必须学的内容，而是因为他们实际上已经掌握了相关知识。团队将这个贝叶斯网络与一个基线方法进行了比较，我们发现，在预测学生随着每个新知识点的学习，其熟练度如何随时间演变方面，贝叶斯网络表现更优。
In this hybrid approach, developed through extensive experimentation, proved successful for generating personalized study plans at the Tropical Institute. The local team notes that they may have been among the first in the world to combine these two concepts in this way.
Prototyping the solution, the first phase of the AI journey focused on assessment rather than full-scale learning delivery. The team needed to find a way to attract students and engage them meaningfully. They launched an assessment tool designed to mirror the Brothers National University entrance exam.
This is a final format in college participation, as it is directly aligned with students' academic goals. Upon completion, students receive a score that shows how close they were to the threshold required for their desired university course. So, if AI isn't made available to young people who are experiencing educational disadvantages, it's not going to be a divider; it's going to be a team. TK developed an extensive bank—a repository of calibrated questions mapped to specific levels of difficulty.
这使得平台能够推荐有针对性的问题，匹配每位学生当前的水平，帮助他们逐步提升以实现其目标。
第一个解决方案在20多个国家/地区部署，建立了数据驱动的方法，并为更深度的学生参与奠定了基础。从最初的想法到第一个原型的部署，大约花费了六个月时间。在此期间，团队同步进行开发和部署，避免过早或过度优化，并根据真实的学生数据持续改进他们的模型。
所以，组建开发团队的关键是组建一个主要由学生和刚从大学（推测为“two universities”或特定校名，原语音识别为“too university”）毕业的人员组成的开发团队。许多人拥有计算机、工程方面的强大学术背景，并且深度沉浸于机器学习（推测原意为“machine learning”，语音识别为“emotion learning”）和人工智能领域。Leonardo（推测为人名，原语音识别为“leonarddo”）本人就是一名计算机工程师（推测原意为“computer engineer”，语音识别为“computer injury”）。
该团队在2012年起步时仅有4名开发者，并在接下来的10年里扩展到了22人。在整个过程中，团队严重依赖学术研究。他们阅读了大量的科学论文，并通过动手实验来研究复杂的AI概念。
这个过程远非一帆风顺。当时，并没有现成的指南来构建个性化的学习系统。正如联合创始人 Claudius Sorkey 所描述的那样：“当我阅读论文时，我能在高层次上理解其场景，但需要一个人来将所有那些参考文献转化为切实可行的东西。”他们连接学术理论与功能软件的能力，成为了团队的一个决定性特征。在 Jerky 平台中使用的数据驱动校准问题，主要由来自合作学校的教师和学科专家编写。
However, while educators could develop pedagogical questions, they typically lacked insight into how difficult a question actually was in practice. The co-founder explained: "Teachers don't know how difficult it is, either usually or how good they are, or if there is something that you need a good student to answer wrong." Data played a crucial role in bridging this gap. By collecting a few thousand responses to a question, they could calibrate it using AI. This is the data-driven calibration process.
Data revealed the true difficulty of each question, its discrimination power, and the likelihood that a student could guess correctly. The team sold the solution to private schools initially, which provided high-quality calibration data from students who were graded on the results and therefore tried their best. This gave them a robust data set before opening their national assessment. Unlike a large language model, which requires massive and constant training effort, calibrating a rule-based model is far less resource intensive.
Neurological evidence emphasizes this efficiency. The main point is that calibrating the model is expensive, but using the model is very cheap. Tens of thousands of students calibrated most of our questions, but the parameters we were using were ones we already had.
We expanded reach through mobile delivery and national access to make it a personal learning platform widely accessible. GK delivers the solution as an app. The app is carefully optimized for no internet usage. This was critical for many students in low-income communities who lack access to high-speed internet or high-end devices. The app was a lightweight solution.
The function, with minimal connectivity, allows students to study effectively even in remote or underserved areas. This impact was remarkable, as the platform reached students in over 90% of Brazil's cities. They achieved a very representative user base they built. We drew a map in our office that lit up where people were using our solution at that moment. It was really incredible to see the entire map lit up. Sometimes you zoom into a poor region of the country, and there were people using our solution. During peak years, four to five million students used the platform, with approximately one million students engaged daily.
The platform remained limited to those with smartphone access. To reduce this digital divide, the Ministry of Education partnered with Jake to enable access through public infrastructure. They established “Learn Houses,” public learning centers equipped with a computer where students could access the GK platform for free. The long-term status of these centers remains uncertain, but they played a crucial role in ensuring broader accessibility during that period.
Teacher engagement and community trust were one of the most powerful aspects of Jake’s expansion. It was the enthusiastic and organic support from teachers.
Although the platform was not initially created as a tool for educators, teachers began recommending it independently, often without any direct contact with the company.
Because the platform was free for public schools and students, and because it offered an awfully clear path toward academic improvement and college admission, it became a practical and hopeful solution. The unexpected nature of this adoption was interesting—it was fascinating to see teachers who didn't know us telling students to use it, saying it was their ticket to university. It was really incredible. This grassroots adoption, driven by genuine impact rather than marketing, validated the generic approach and its relevance in the broadening educational landscape.
The completion of the platform motivated students who recognized the opportunity, and the teachers who championed it created a virtuous cycle of engagement. Its success, consideration of data responsibility, and a key skill: as they gathered an increased volume of student data, the team remained conscious of their ethical responsibility in handling sensitive educational information. Technology from the old city, along with internal policy, ensured that student data would never be commercialized and would be used solely for assessing progress.
The platform learned and provided personalized study recommendations. The company committed to complying with the Brazilian data protection law and maintaining student privacy standards in line with global best practices.
Unlike today's concerns around Generative AI and not-skill pre-trained models that use massive datasets from the internet, this approach was fundamentally different. The platform did not generate new content from external sources. Instead, it ran entirely on structured student interaction data to gauge learning through assessment, topical mastery, and personalized content.
However, the team didn't discuss potential biases in their datasets, particularly regarding differences between public and private school students or regional disparities across Brazil. However, because Geekie achieved national-scale reach with a user base predominantly from public schools, the data was broadly representative of the Brazilian student population. Claudio stated, "I think our data was very representative of pretty much every student in the country, so it's not that we were using a specific set of data from private schools from a certain region; we didn't have that small sample bias."
As Geekie evolved into an AI-driven personalized learning platform, the team encountered several significant technical challenges related to both the complexity of the educational domain and the diversity of the student population.
Distinguishing correlation from causation was one of the most difficult early problems involving the structure and the calibration of the bias in the network. Initially, data came from a limited set of schools, many of which followed similar teaching sequences. The model tended to capture correlations that were artifacts of the curriculum rather than true causal relationships between concepts.
如果大多数学校在讲授主题B之前先讲授主题A，模型可能会错误地推断，掌握主题A是掌握主题B的原因，而这仅仅是因为教学顺序的先后。解决方案随着组织规模的扩大而出现：我们从不同地区的更广泛学校中收集了数据，这些学校有着各不相同的教学顺序。学习路径的更大多样性帮助系统区分了真正的因果依赖关系和表面的相关性，从而在最终实现了更准确的推断。
Adaptable Learning Recommendation: Scalability and Infrastructure Optimization
The platform had to serve hundreds of thousands of students simultaneously, many of whom had limited internet connectivity and used older devices. It needed to deliver real-time assessment and personalized content under these constraints.
Positive signals on scalability emerged during national assessment periods. The system needed to handle 300,000 to 500,000 concurrent users. The team responded by implementing a high-performance, lightweight front end to ensure usability on low bandwidth and older smartphones. On the back end, the team designed a system for efficient and scalable learning inference.
系统需要能够同时处理数十万用户的并发活动，且不牺牲性能。这需要精心设计可持续的架构、负载均衡和资源优化。
寻找合适的AI模型
当时，几乎没有学术研究专门针对这一特定用户案例。挑战在于构建一个能够追踪学生随时间推移的能力水平，并引导他们完成特定主题学习路径的模型，且需保证高准确率。
经过大量实验，团队成功将贝叶斯知识追踪与BLLIP模型相结合。这种混合模型实现了更高程度的个性化学生学习路径，在主题层面被证明是成功且可扩展的，尤其适用于数学、科学和语言等科目。
然而，该模型在哲学和社会学等更主观、概念定义更困难的学科中效果较差。模型经常无法覆盖这些主观性强的科目，这需要不同的策略或更高的推理精度来处理。
给社会创新者的建议
当被问及会给其他考虑将AI用于社会影响的组织什么建议时，团队负责人强调：从简单开始，并持续构建。从能提升人们现有水平15-20%的简单方案开始，然后逐步提升到50%、60%甚至100%，这并不可耻。如果我们一开始就试图做最后一个模块，我们永远也到不了那里。
团队坚持一项政策：始终保持一个高基线，即一个代表基础专家知识的简单、一致模型。任何新AI模型在部署前都必须超越这个基线。这种方法使他们保持务实，并确保复杂性服务于特定目的。
反思：若拥有现代工具会做何不同
两位创始人指出，内容创作和交付的多样性是重点。现代学习管理系统可以支持开放式问题，而非他们当时依赖的多项选择题形式。他们本可以更高效、更多样化地制作内容，包括为有特殊需求的学习者提供材料。
Claudio补充说，AI也可以改善内部运营，从销售和营销到客户服务及后台基础工作。The Geekie story demonstrates that meaningful social impacts through AI do not require the most sophisticated technology available. It requires a clear understanding of the problem, thoughtful application of appropriate tools, and a willingness to use these tools iteratively based on real-world feedback. A team of four developers built something that reached millions of students and gave them a genuine chance at educational opportunity. That achievement was not based on technological complexity, but on deep commitment to the community they served.
所以，根据一位年轻学者的理论，制度企业家将运用AI技术来促进教育，并惠及每一位学生。
当然，这个案例研究对于教育经济学和社会企业经济学而言非常重要。
好的，今天的内容就到这里。感谢聆听。
当然，如果您想了解社会企业和社会金融，或者想了解全球一些成功的案例，您可以联系阳民：Fairtown@PKU.org.cn
