Part VII: The Global Classroom
第七部分：全球课堂
Economics of AI in Education
AI 在教育中的经济学
A decent education is expensive anywhere in the world. In the US, Louisiana spends roughly $10,000 per student per year; New York spends $40,000. In India, government schools might spend anywhere between $500 and $1,200 per student per year. Despite the range in resources, the fundamental model is the same. Students are moved lockstep through curricula, oftentimes feeling lost or bored. If a student doesn’t keep pace in understanding a foundational concept, the class keeps moving. Limited support exists for personalization or for revisiting gaps, much less for one-on-one tutoring. This is despite the fact that many classrooms have students at a wide range of preparedness—some may be ahead of pace while others may be two or three grade levels behind.
世界上任何地方的优质教育都很昂贵。在美国，路易斯安那州每年每个学生的花费大约为 1 万美元；纽约州则为 4 万美元。在印度，公立学校每年每个学生的花费可能在 500 到 1200 美元之间。尽管资源差距很大，但基本模式是相同的。学生按照固定的课程进度推进，常常感到迷茫或无聊。如果学生在理解基础概念时跟不上，班级还是会继续前进。个性化支持或复习缺漏的支持有限，更不用说一对一辅导了。尽管许多课堂上的学生准备程度各不相同——有些可能进度超前，而有些可能落后两到三个年级。
The COVID-19 pandemic made things worse. During the 2020 school shutdowns, Black and Hispanic households with school-age children were 1.4 times as likely as white households to face limited access to computers and the internet, and more than two in five low-income households had only limited access. A bad prepandemic situation became downright dire. Consider that before 2020, 6 percent of Detroit eighth graders were performing at grade level; afterward, it dropped to 3 percent. The average American classroom in 2019 contained a spread of three grade levels of ability. After the pandemic, this spread expanded to six grade levels of ability. Put another way, in the same classroom of thirty students, teachers had to somehow support learners who were four-to-five grade levels behind while not boring the students who might have been ready to move ahead.
COVID-19 大流行使情况变得更糟。在 2020 年学校关闭期间，拥有学龄儿童的黑人和西班牙裔家庭面临有限的计算机和互联网访问的可能性是白人家庭的 1.4 倍，超过五分之二的低收入家庭只有有限的访问机会。疫情前的糟糕情况变得更加严峻。考虑到 2020 年之前，底特律只有 6% 的八年级学生达到年级水平；之后，这一比例下降到 3%。2019 年，普通美国课堂的能力范围跨越了三个年级。疫情之后，这一范围扩大到六个年级。换句话说，在一个有三十名学生的课堂中，教师必须以某种方式支持落后四到五个年级的学习者，同时不让那些可能已经准备好前进的学生感到无聊。
To address the situation, the US federal government funded $86 billion for elementary and secondary school emergency relief funds, amounting to $2,000 per American K–12 student. A lot of this money flowed into live tutoring programs, based on decades of research showing that tutoring can be an effective intervention for kids. Unfortunately, years later, most of the money was gone, with little to show for it. In hindsight, most experts believe this is because the tutoring was not connected to what was going on inside the classroom, and many students found it logistically hard to access. Students might have also run up against a stigma associated with going to tutoring in the first place.
为了应对这种情况，美国联邦政府提供了 860 亿美元用于小学和中学紧急救助基金，相当于每位美国 K-12 学生 2000 美元。大量资金流入了现场辅导计划，基于数十年的研究表明辅导对孩子来说是一种有效的干预措施。不幸的是，几年后，大部分资金已经用完，几乎没有成效。事后看来，大多数专家认为这是因为辅导没有与课堂内发生的事情相关联，而且许多学生发现它在后勤上难以获得。学生们也可能遇到了与接受辅导相关的耻辱感。
A platform like Khanmigo exists to bridge this gap—offering personalized, accessible, and high-quality education. Even before Khanmigo, efficacy studies had shown that classrooms using Khan Academy as little as thirty to sixty minutes a week during the pandemic not only avoided the COVID slide but outperformed pre-COVID standards by 20 to 40 percent. And this didn’t cost $2,000 per student. It was free.
像 Khanmigo 这样的平台就是为了弥合这一差距——提供个性化、可访问和高质量的教育。即使在 Khanmigo 之前，功效研究已经表明，在大流行期间每周使用可汗学院仅三十到六十分钟的课堂不仅避免了 COVID 下滑，而且比 COVID 之前的标准高出 20% 到 40%。而且这并不需要每个学生 2000 美元的花费。这是免费的。
Now, large language model platforms build off of those results to provide even richer support. An AI tutor is available whenever students need it, including in the classroom while they are doing their existing academic work. It can inform teachers and parents exactly what students are working on and where they need more help. Students who are further behind don’t need to feel shame or embarrassment in asking for assistance, since the AI isn’t a real person. Curious students can ask questions without feeling like they are wasting someone’s time.
现在，大型语言模型平台在这些结果的基础上提供更丰富的支持。AI 导师在学生需要时随时可用，包括在学生进行现有学业工作的课堂上。它可以准确地告知教师和家长学生在做什么以及他们需要更多帮助的地方。落后的学生不需要感到羞耻或尴尬，因为 AI 不是一个真人。好奇的学生可以问问题而不觉得他们在浪费别人的时间。
Providing scaled support like this is incredibly cost-effective and accessible, but it isn’t free. Even before considering generative AI, our annual budget as a nonprofit is more than $70 million. That’s a significant number, but it is also equivalent to the budget of a large high school in many parts of the United States—and Khan Academy reaches more than a hundred million learners a year. We need to raise a large chunk of this money every year from philanthropists to keep the content and software free to users. These resources are necessary for content development, product development, and server costs, among other things.
提供这样的规模化支持是极具成本效益和可访问的，但它并不是免费的。即使在考虑生成性 AI 之前，我们作为非营利组织的年度预算超过 7000 万美元。这是一个很大的数字，但它也相当于美国许多地区的一所大型高中的预算——而可汗学院每年覆盖超过一亿名学习者。我们需要每年从慈善家那里筹集大量资金，以保持内容和软件对用户免费。这些资源对于内容开发、产品开发和服务器成本等是必要的。
Generative AI adds a new layer of expenses beyond the cost of paying the salaries of engineers, designers, product managers, and content developers to iteratively improve a platform like Khanmigo. This is because the computation costs of a large language model like GPT-4 are significant. Right now, our best estimate of the computation costs of average usage of Khanmigo is between five and fifteen dollars a month per user. Assuming that we will have millions of users—which would cost tens of millions of dollars in computation costs—it is unlikely that we can raise enough money from philanthropy alone to offer the service for free. While dramatically cheaper than live tutoring, which can easily cost thirty dollars an hour, the platform does become less accessible than our free resources since we will need to charge school districts for access.
生成性 AI 增加了额外的支出，除了支付工程师、设计师、产品经理和内容开发人员的工资以迭代改进像 Khanmigo 这样的平台的成本之外。这是因为像 GPT-4 这样的大型语言模型的计算成本是巨大的。目前，我们对 Khanmigo 平均使用的计算成本的最佳估计是每用户每月五到十五美元。假设我们将拥有数百万用户——这将花费数千万美元的计算成本——仅靠慈善筹款是不太可能筹集到足够的资金以免费提供服务的。虽然比每小时轻松花费三十美元的现场辅导便宜得多，但该平台确实不如我们的免费资源容易获得，因为我们需要向学区收取访问费用。
That said, between philanthropy and funding by local school districts, the cost to the students in those districts is, and will remain, free. However, this still doesn’t address accessibility for poorer countries where thirty dollars a year could make up a significant portion of total education costs. The good news is that the computation will become cheaper and we will get better at using it more efficiently. These two trends should help bring the cost down by a factor of ten in the next few years. If we can reduce the costs by a factor of one hundred, which should happen in the next five to ten years, it will become comparable to the cost of using nongenerative web-based applications today
也就是说，介于慈善和地方学区的资助之间，这些学区的学生成本是且将继续是免费的。然而，这仍然无法解决在贫穷国家的可访问性问题，在这些国家中每年三十美元可能占总教育成本的很大一部分。好消息是计算成本将变得更便宜，我们将更好地高效利用它。这两个趋势应该有助于在未来几年内将成本降低十倍。如果我们能将成本降低一百倍，这应该会在未来五到十年内发生，它将与今天使用非生成性网络应用程序的成本相当。
At that point, the only real limits to access are the same ones we face with traditional Khan Academy: students would need access to the internet and devices, which aren’t reliably available to everyone. Nevertheless, I am hopeful that between devices getting cheaper and providers like SpaceX’s Starlink using swarms of satellites to provide low-cost broadband, nearly universal access will become a reality.
到那时，访问的唯一真正限制与我们在传统可汗学院中面临的相同：学生需要访问互联网和设备，这些对每个人来说都不是可靠的。然而，我希望随着设备变得更便宜以及像 SpaceX 的 Starlink 这样的提供商使用一群卫星提供低成本宽带，几乎普遍的访问将成为现实。
One major barrier to access in the early days of online learning was language. Now, large language models like GPT-4 can operate in every major language. Here, an English-language learner working on word problems in English could get support in his or her native language, or even in a mix of languages like Spanglish. The large language model’s conversational abilities make it feel like a real-time interaction, fostering a sense of connection. Even more, it can be used to do much of the translation work of the core platform.
在线学习早期的一个主要障碍是语言。现在，像 GPT-4 这样的大型语言模型可以用每种主要语言操作。在这里，正在用英语解决文字问题的英语学习者可以用他的母语甚至是像“Spanglish”这样的混合语言获得支持。大型语言模型的对话能力使其感觉像实时互动，培养了一种联系感。更重要的是，它可以用于进行核心平台的许多翻译工作。
It is exactly that sense of connection and support that made finding a low-cost, multilingual, and scalable solution to bringing high-quality access to education across the globe so very vital, says Stanford’s Susanna Loeb. “I’m optimistic and excited by what we can do now. In places where access to resources and pedagogies has been a real constraint, this technology can be transformational.”
斯坦福大学的 Susanna Loeb 表示，正是这种联系和支持感使得找到一种低成本、多语言和可扩展的解决方案，以将高质量的教育带到全球各地显得如此重要。“我对我们现在能做的事情感到乐观和兴奋。在资源和教学方法获取受限的地方，这项技术可以是变革性的。”
“点赞有美意，赞赏是鼓励”