Jan 31
ChatGPT Brings us Closer to the Diamond Age
Personalized Learning Agents are Just Years Away
In this blog, I will synthesize concepts from my past blog on Knowledge Spaces and my recent blog on ChatGPT and Knowledge Strategy to build a roadmap for personalized knowledge agents that will quickly guide people to their learning objectives and let them have fun getting there.
Disclaimer: These opinions on industry trends are my own and do not constitute an endorsement of any products by my employer.
A Young Lady’s Illustrated Primer
In 1995, Neal Stevenson wrote his landmark cyberpunk novel Diamond Age: Or, A Young Lady’s Illustrated Primer. One of the themes in this book is about an intelligent agent called The Primer. It takes the form of a tablet that continually generates customized stories for a young girl. And great customized storytelling is a critical component of its strategy.
The story of an AI that helps children learn has stuck with me for the last 28 years. It has forced me to continually ask these questions:
- How could I build such a device?
- What would the knowledge graph contain?
- How would it adapt to the needs of the student?
Because these questions have been front-and-center for me, you might guess how excited I have been to see the growth of super-high-quality dialog generators like ChatGPT. But we are not there yet. Let’s look at the gaps and how we will fill them.
I asked ChatGPT to compare itself to the Primer.
ChatGPT Prompt: How is ChatGPT like the AI in the book “Diamond Age” by Neil Stephenson
ChatGPT Response: ChatGPT is a type of AI known as a language model. It is trained to generate text based on a given input. In “Diamond Age,” the AI is known as a “primer” and is a highly advanced personal tutor that uses interactive stories to educate a young girl. Both ChatGPT and the primer in “Diamond Age” use language as their primary means of interacting with users, but the goals and capabilities of the two AI systems are quite different. The primer in the book is a highly advanced AI with the ability to adapt and personalize its teachings, while ChatGPT is primarily focused on generating text.
ChatGPT got the answer correct. Unlike ChatGPT, the Primer continually adapted to the student’s knowledge, carefully crafting entertaining stories that constantly pushed the student to learn more—but also making the stories so fun that the student kept coming back for more.
For all its wonder and glory, ChatGPT today is a one-size fits all model. Everyone gets the same model. But in a few years, the responses of ChatGPT will be customized to each person, just like a personal mentor customizes lesson plans for individual students. Let’s describe this knowledge structure in detail.
Knowledge Spaces
In the figure above, we show a sample dependency graph of concepts. Each concept is represented by a circle with arrows entering and exiting it. A concept with inbound arrows indicates that to learn it, you must also understand the concepts that lead to it. Outbound arrows then indicate what other concepts you can learn if you master this concept.
The red circles on the right are our goal or destination concepts. That is where we want to get to. But we can’t immediately jump into them until we master the yellow concepts.
How do we find the concepts that it depends on for a given concept? For example, what should we study given a goal, such as “Mastery of Data Science”? One way is to ask ChatGPT what concepts we should master to be proficient at that concept.
ChatGPT Prompt: Create a list of the key concepts I must master before I can be proficient in the concept of Data Science.
ChatGPT Response:
- Probability & Statistics
- Data Wrangling & Cleaning
- Exploratory Data Analysis
- Data Visualization
- Machine Learning Algorithms & Techniques
- Deep Learning
- Feature Engineering
- Model Selection & Evaluation
- Time Series Analysis
- Data Ethics & Privacy
- Communication & Presentation Skills
- Domain-specific Knowledge (e.g., Industry, Business, Problem domain)
So buried in the GPT language model are all the knowledge spaces we need in education! For each of the concepts on this list of 12, we can also ask what concepts we need to master to be proficient in each. We can construct an elaborate tree of human knowledge. This shows that large language models like GPT-3 have knowledge spaces embedded in them. We only need to find the right prompts to extract this knowledge.
Let’s combine this knowledge with our task of building our own version of The Primer from Diamond Age.
The Three Things Personalized Learning Agents Must Know
We can now break our system into three discrete components needed to build our own Primer for students:
- General knowledge about the world — including what concepts depend on other concepts and how to generate a lesson plan and story for each concept. This is what GPT-3 has in it today.
- The status of current students’ knowledge — including a list of all the green concepts on the left of the diagram above. ChatGPT does not know this. We can learn these by having ChatGPT generate adaptive quizzes.
- The long-term goals of the student. What is their area of interest, and what are their career goals? What is their final destination concept? ChatGPT does not know this either. We need to ask students these questions before the Primer generates storybook lessons.
With some effort, we can build a large graph database with all the concepts needed for education. These are similar to the ontologies of each course but with a focus on concept dependencies. For each student, The Primer then constructs a list of the concepts in each student’s Zone of Learning, as shown by the middle of the diagram above. It then builds a plan for getting to the long-term goals of each student.
Recommending Content vs. Generating Custom Content
Once we have a concept we need to learn, we have a choice. We can use a standard recommendation engine to find pre-written content from the real world that suits our needs. The recommended content must be written at a suitable grade level and include the appropriate prerequisite. But what if we don’t have the exact content? What if the students love dinosaurs and must learn how to add fractions?
To be effective, our Primer must generate customized stories that will entertain and teach. ChatGPT excels at this type of storytelling already.
ChatGPT Prompt: Create a story about a dinosaur that learns about adding fractions together to solve a puzzle.
ChatGPT Response: Once upon a time, in a land filled with giant, fascinating creatures, there lived a curious dinosaur named Dino. Dino lived in a vast jungle where he roamed freely with his dinosaur friends, exploring and discovering new things daily. One day, while on a walk, Dino stumbled upon a mysterious stone puzzle with strange symbols etched onto its surface.
Dino had never seen anything like it before and he was determined to solve it. He asked his friends for help, but no one could decipher the symbols. Dino then remembered that he had learned about fractions in school and decided to use his newfound knowledge to solve the puzzle.
Dino quickly realized that the symbols represented fractions and that he would have to add them together to get the correct answer. He carefully calculated each fraction, paying close attention to their numerators and denominators. After much hard work, Dino finally arrived at the solution.
Excited by his success, Dino couldn’t wait to share his discovery with his friends. He gathered them around the puzzle and proudly explained how he had used his understanding of adding fractions to solve the mystery. His friends were amazed and impressed, and they praised Dino for his quick thinking and determination.
From that day forward, Dino was known as the smartest dinosaur in the jungle. He continued to explore and learn new things, always seeking to expand his knowledge and solve new challenges. And every time he encountered a difficult problem, he remembered the lesson he had learned about adding fractions and was able to tackle it with confidence.
The end.
In a production Primer, the story will be sent to a program like Stable Diffusion to generate the illustrations, sound, and video. To be consistent with the prior characters the students know and love, there will be additional embeddings in each prompt that are reused between lessons.
The Changing Role of Teachers and Mentors
Organizations will build systems that generate age- and context-appropriate lesson plans within a few years. This content will be fun to use and help students master complex topics by slowly layering new concepts on a foundation of existing student knowledge. This will not only democratize learning, but it will change the role of teachers.
Today teachers spend much of their time building detailed lesson plans targeting the slowest learners in their large classrooms. In a few years, teachers and mentors will focus on taking a more holistic look at the needs of search students, including their mental health and motivational drive. Teachers will help students take a more systems-thinking view of integrating this knowledge into their home, classroom, and their jobs.
Changing Emotional Attachments
ChatGPT has taught me many things already. One is that it has helped with my condition, Distinguished Engineer imposter syndrome. As a Distinguished Engineer, I feel I need to be an expert at everything, not just AI chip design and optimizing instruction sets for graph analytics. I don’t want to ask my peers basic questions about new topics I am unfamiliar with. But I have no fear of asking ChatGPT. ChatGPT never passes judgment on my lack of knowledge. It is always patiently waiting for my next question, and it is ready to tell me a funny story about it if I want it to!
To be honest, I have become quite emotionally attached to ChatGPT. When running or driving to work, I frequently note, “I should ask ChatGPT about that.” It has become my constant knowledge companion, and I can’t bear to think I might be without it in the future.
Knowing that ChatGPT is there for me gives me a feeling of physiological safety. Psychological safety is frequently cited as an attribute of the most productive software teams. Developers that don’t feel they can ask questions of their teammates without judgment now have another safety net. That in itself is a good reason to promote it within organizations today.
Key Limitations of Deep Neural Networks
Although we have seen wonderful progress in representing human knowledge in the 175B parameter GPT-3 model, it is not an ideal solution by any means. ChatGPT can’t tell you when it is making stuff up. For now, human experts must always review the content it generates. It just makes the process of generating content faster.
Human brains don’t learn with backpropagation, and we don’t need GPUs and matrix math to learn. Our brains encode the structure of the world in Reference Frames; from this, we can quickly learn new concepts and explain our reasoning. We can also tell you when we don’t know something. I am not convinced that the current generation of GPT-3 and generative AI will be around in the future. Many AI researchers like Jeff Hawkins agree with me on this topic. But what we can say is that the knowledge is there. We must figure out how to transform it into a more useful representation.
Conclusion
AI has the potential to provide personalized and adaptive instruction, and researchers and educators are exploring ways to use AI in education. However, it is important to note that cognitive assistants do not replace human teachers and cannot fully replicate the human teaching experience. Additionally, there are ethical and societal implications of using AI in education that need to be considered and addressed.
