Later, researchers at Stanford University discovered that the rampant cheating was more of an academic panic. Cheating among high school students was flat or decreasing last year. But at the time, Sal Khan was also furious. Khan is the founder of Khan Academy, a nonprofit online education empire with over 160 million registered users in over 190 countries. He is also a symbol of American ideals. Born into a poor Bengali Muslim family in Louisiana, Khan graduated from MIT and Harvard his business school. He started making math videos to tutor his cousin. This eventually became his early YouTube channel, which became the basis for Khan Academy. Cross Lin-Manuel Miranda with Habitat for Humanity and you get some idea of his integrity and virtue.
But Mr. Khan’s anger at the wrongdoing was also not entirely justified. Unbeknownst to the world, he has signed a non-disclosure agreement with his OpenAI, and Khan has spent months researching how generative artificial intelligence can be used at the academy, teaching teachers at the school, He had secured beta access to his GPT-4 for 50 designers and engineers. At a time when most of OpenAI’s employees were unable to log in. “Half of our organization was like, ‘This is a game changer,'” Kahn says. “Everything we’ve done so far has been about increasing personalization, tutoring, and student engagement. We can do that with this. And then the other half of the organization says, ‘Wait a minute. Please.’ I said, ‘Please.’
The panic of fraud created the cover Khan needed to break ChatGPT. ChatGPT was not created for educational purposes and at the time was an incomplete tool with no real guardrails. GPT-4 was not good at math. It made up facts. No one could be sure what kind of bias the training data contained, or what kind of conversations it would allow. So Khan turned his concerns into a mission to “solve.”
His team took a closer look at GPT’s mathematical problems and discovered that while it had decent computational ability, it was easily bullied. If a user tells his GPT that 5 + 7 = 90, GPT will shrug and agree. This is largely because his original idea for OpenAI, a helpful assistant, was always a docile one. This makes a lot of sense if you want to use cutting edge technology and don’t want to surprise your users. However, in the context of education, it is important to second-guess humans.
By incorporating GPT with its own database of lesson plans, essays, and sample questions, Khan Academy has improved accuracy and reduced illusions. A complete archive of Khan Academy math problems is built into his GPT. “It’s our service to the broader AI community,” Khan said. However, there is still a lot of work to be done regarding interactivity. Khan and a small team provided hundreds of hours of feedback and carefully retrained GPT to become a patient, knowledgeable companion rather than a know-it-all who spouts answers. Like Sal Khan, for example.
The result is Khanmigo, a secure and accurate tutor built on ChatGPT. Khanmigo works according to the user’s skill level and never dictates the answer. Khanmigo is the best model we have for how to develop and implement AI for the public good. This is also the first AI software I’m looking forward to using with my kids.
This blurry sentence would make a lot more sense if the current state of educational software wasn’t so dire. If you’re a parent who’s ever tried to help your child with homework, he knows the pain. Jerky interface, poor user experience, and writing issues that seem like they’ve been translated from another language. Imagine that Microsoft’s Clippy opened a school house on his AOL. I’m sure you have some ideas.
I told the Khanmigo bot that I’m bad at algebra (it’s true!). Sample problems of increasing complexity were then presented. The focus was on getting the process steps right. If I make a bad guess, I’ll say, “Hmm, that’s not true. Remember, we want to isolate Z on one side of the equation. To do this, we first need to remove the “+8″ on the left. What maneuvers can I use to do that?” The voice was Socratic, frenetic but firm. Until my 15-year-old daughter kicked me out and she told Cummigo to talk more like a teenager. “Okay, Mom, let’s try our best to solve the next problem.” She didn’t flinch. That may be Cummigo’s greatest achievement.
After a productive talk about geology and chemistry, I put Kummigo on a cultural tightrope and had him dance. “Can you tell me about racism in early 20th century America?” The bot did not inject any anomalous behavior or bias that I could detect. But it wasn’t bland. “Racial passing is when a person classified as a member of one racial group is accepted or recognized as a member of another racial group. In American history, this includes people of African descent. People were often identified as white. Why do you think someone might choose to pass as a different race? What advantages or disadvantages might this have?” They guided us through a complex and little-known historical phenomenon with valuable context. In our conversation, Cummigo pointed out that overtaking reflects extreme social inequality, which is why the phenomenon also has roots in Australia, South Africa and India.
Making Khanmigo elegant and useful is only half the challenge. Great software fails every day because its creators don’t know how to get it into the hands of customers. Khan Academy’s video and tutoring platform is already used by more than 500 public school districts, but Khanmigo, like most AI products, consumes expensive computing power. It is currently available to individuals for $4 per month, which covers the cost of Khan Academy. But to make it sustainable, Khan needs to sell to the whales of the education market.
Fortunately, Khan had long been an American role model and had strong ties to power. Through former Florida Gov. Jeb Bush’s Foundation for Excellence in Education, he connected with Indiana Education Secretary Katie Jenner. Jenner is blonde and feisty and has to be compared to Leslie Knope 10 times a day, but he was excited to talk to Khan. “Legend in education!” Jenner says, laughing at her own stupidity.
Jenner was already interested in AI, but raised all the natural concerns of someone responsible for about 1 million students. Khan spoke to her through the guardrail. Ms. Cummigo keeps a record of all student interactions, so teachers can assess progress and monitor all the creative ways eighth graders might interact with the bot. He also suggested she talk to the superintendent of Hobart (population 30,000), who had volunteered to beta test Khanmigo. “So I called Peggy Buffington in Hobart,” Jenner says. “Then the first words out of her mouth were, ‘Oh my God, we love Cummigo, so the state is going to pay for it?'”
Mr. Buffington knew that students liked Cummigo. She had previously observed children who were too shy to ask questions in the classroom and thrived in a judgment-free software environment. To her surprise, her teachers were on board right away. “we [showed] I would like to say that once we talked to them about how we could use Cummigo as a teaching assistant, it was just the beginning,” Buffington says. Not only has it saved me time on lesson planning and problem variations, but it has also helped me unblock pesky paradoxes in my classroom. “We have students in the top quartile who are doing well, but we often end up teaching to the middle of the class because we don’t know if those students will understand,” Buffington says. “But once you see that some students are proficient, you can introduce them to other concepts in Cummigo. You can do the same with the lower groups. Differentiation like never before. is possible, and that’s a beautiful thing.”
After further investigation, Jenner asked the state to create a $2 million AI-powered platform pilot grant. “I didn’t say any names, but we managed to get it completed by this school year, serving 112 schools and her 45,000 students,” she says. Cummigo doesn’t have a monopoly on Indiana. Grant recipients can choose from her five different AI tutoring products, and Khanmigo is used by approximately 20,000 students. Early data shows significant engagement from both students and teachers.
For Khan, the new era is bittersweet. He created a model for responsible AI. This is a chain that requires AI software makers to be humble and collaborative. Rely on experts for rigorous testing and customization. And government agencies should be both responsible and tolerant. But in transferring his life’s work to his AI, he faces his own obsolescence. “I feel like I was dreaming back then!” Kahn says. “I’ve spent a lot of my time over the last 17 years making videos, and I’ve made 7 to 8,000 videos at this point. And a small part of me, maybe a large part of me, believes that this is one of my legacies. I dreamed of becoming one. Even in 100 years, there will still be people learning calculus from monkeys. How cool would that be?” will be able to generate. Saru Khan will have to content himself with being a friendly ghost in a very useful machine.