What if we replaced all teachers and professors with a virtual teacher? The challenges of education in the time of AI and what is happening in the real world today starting with Khan Academy.
A beautiful book about AI is AI 2041 written by Chen Qifan-a Chinese science fiction writer-and Kai-Fu Lee-a Taiwanese-American computer scientist, investor and writer.
The book is a kind of Black Mirror dedicated to artificial intelligence in which Chen Qifan tells, in 10 different episodes, stories of daily life in the world invaded by AIs and Kai-Fu Lee comments on them comparing them with the state of the art of AI evolution.
The third episode of the book is titled "Twin Sparrows" and tells the story of twins who have at their disposal a digital avatar, a Private AI, who follows them as early as elementary school throughout their school development.
A kind of private tutor, available 24/7, that the kids can customize in terms of appearance and manner of speech, and that they can modify as they grow up (to understand it, something like Peppa Pig when they are little and Leonardo da Vinci when they grow up), who learns everything they learn, understands how they are learning, senses their desires and inclinations, and helps them constantly.
The avatar is monitored by the teachers who supervise the children's activities and, as long as they are minors, by their parents, who meddle in too many things. But I reveal no more.
Can you imagine it? Science fiction?
These days I had the great opportunity to try out a preview of the beta version of Khanmigo, an AI developed by Khan Academy which, thanks to the power of Chat-GPT4 and the support of OPENAI, is about to release an AI that aims to be a Tutor for students but also an Assistant for teachers.
In other words, it is a still rudimentary version of the digital avatar concept from AI 2041, except that it's already happening in 2023!
What is Khan Academy?
AI avatar for students
In this post we will only discuss Khamigo for students, leaving further reasoning with teachers to later posts.
When I explained to my children during Christmas break 2022 that the Chat-GPT phenomenon was about to explode, I immediately told them that it was the perfect tool for copying, doing essays, summarizing texts, solving a few logic problems (very few at the time). They could have cheated in a way no one had done before. They could have remained ignorant and dishonest at very little cost.
Or they could have gone a different route, used him as a tutor to delve into subjects they liked, had explanations given on subjects hostile to them, learned any subject better. Surfing on the hallucinations that would come would have helped, moreover, their critical spirit, since the unreliability of the tool in some situations would have been sneaky.
I got down to it and, by dint of modifications, suggestions found on the web, personal thoughts, created a prompt that could simply help them to be questioned in any subject.
I realized early on, though, that:
the context needed to be explained in great, great detail;
the subject matter to be studied had to be framed in different ways;
my children have very different ways of learning;
my prompt did not know the state of the art of their school knowledge and in general, did not know how they learned most effectively, did not get much feedback from them;
already stressed by professors, having even an AI pressing you to check if you have studied generates emotions that AIs still don't understand.
An LLM-based generative AI helps a lot, but creating such an AI is much more complex than turning to the 'simple' Chat GPT, especially if we can't articulate our questions, if we can't connect it with vertical AIs, if we don't know anything about machine learning, programming, data analysis.
Also, to do these exercises, you need to have at least Chat GPT 4 to better articulate the conversations and reduce the hallucinations a bit.
Khanmigo promises to solve and improve all of these situations while being aware of the many limitations to date (including: hallucinations, use of incorrect training data, false quotes, and other inaccuracies that, however, help fact checking).
How does it work?
Access is very simple and looks like an evolved chat. For now the interface is only in English although, having Chat GPT 4 behind the scenes, we can speak to it in any language.
The system is very feedback-oriented, always proposing topics, asking for feedback in the form of OK/KO or more structured information, warning that the system may make mistakes and explaining why.
Then, as you see, the menu on the left allows for different types of interaction:
Brainstorming on entrance exams;
Tutoring on math, science, humanities areas;
Curiosity Exploration;
Practice exercises for elementary, middle, high school up to college tests;
Chats with famous people or historical figures;
Story creation;
A game on deconstructing words to better understand their meaning;
Up to suggest practical questions that a high school student might ask himself:
Right now, I believe these are areas in which Chat GPT has been trained by Khan Academy to operate with very structured instructions on behavior and goals for the student.
Khanmigo stores all the work that is done, all the conversations, and promises to use them on our side (even though they themselves ask that we not provide personal data to the system) to help us learn better and better with better orientation to our goals and abilities.
Case in point: Let's do our homework.
It would take a very long post to tell the whole story, but as you can see, the approach is very calm (I deliberately answered incorrectly to show you how Khanmigo reacts).
As you can see, the chat is structured to:
continually ask questions;
reassure when you are wrong without making judgments of merit;
explain in detail how the question is resolved;
make yourself available for further investigation.
Then, in response to a typically concise answer from a student, Khanmigo explains the steps to arrive at the solution.
Ignite your curiosity
(My favorite section).
This strikes me as a great companion for those dull moments when I prefer not to hang out on social media or read books. It reminds me a lot of the approach of the "Maybe Not Everyone Knows That..." column in Puzzles Week.
I'll stop, because teaching is not my topic and I don't want to teach anything, just show things that will help each person think.
So...
The topic of AI as a tutor for students is fascinating, profound, frightening.
However, we are in the presence of a technology that has together important advantages and challenges.
On the one hand, they could facilitate autonomous learning, keep students motivated and engaged, support teachers in planning lessons.
On the other, the risk of errors, cheating, reducing the role of teachers or, worse, atrophying students' critical thinking and their ability to solve problems.
Moreover, it is a completely digital tool: during the pandemic, we experienced how much damage kids suffered as a result of social withdrawal, lack of confrontation with their peers, and the mental confusion that resulted. Those of us who have children know what it means to leave them alone with digital tools.
Khanmigo at the moment is not the solution to all the problems; leaving our kids to deal with an AI requires supervision by people (teachers, tutors, parents) trained in AI issues and ready to change course constantly.
But we also experience (at least in Italy) a primitive school system that constantly lowers the level of interactions to facilitate kids, not understanding that Generation Y is not less intelligent than the previous ones but requires educational approaches in step with the inputs with which social, video games, culture media, constantly bombard them.
A few days ago, I was talking about this with a principal who felt strongly about the need for a change in the teaching paradigm even though he could not find an immediate solution.
Could this be the way ?
Addendum
We talk about the death of the metaverse, we struggle to make sense of living with virtual/augmented reality glasses on. Remember when 20 years ago, with the arrival of smartphones and Bluetooth headsets, it was said that they would not be successful and that they isolated too much from the world and ended up making people senile?
Well, now go look at Apple Vision Pro (yes I know it's super expensive today), then read Ready Player One (or watch the movie).
Done?
Imagine classrooms full of kids equipped with visors connected to multimodal generative AIs (i.e., processing text, images, sounds for both input and output), which are in turn connected to domain-specific AIs trained to teach our kids how to learn.
It will certainly be worth talking about in more depth.
Σχόλια