We explain things by trying to reach a shared state of ‘imagination’, where the other person is also ‘seeing’ what I am. This can be achieved by conversation where students try to visualize on their own, or by diagramming, where a visual is forced and both teacher and student work on \same visual.
differential : https://www.youtube.com/watch?v=F40ZBDAG8-o
photon emission from atom : https://www.youtube.com/watch?v=sj52It93sWY
- Whiteboard with continuous editing marker
- Pointer to point at things on board
- Speaking to indicate information
- Knowledge of students state of mind
Diagramming
Real world Interactive diagramming
HCVerma explaining : https://youtu.be/clgV9QtKtPE?t=660
Attempts to diagram with current tools
- 24 kg nails to 9kg with a balancing pan
https://claude.ai/chat/f91b3b4c-d0f8-4606-923a-7adad82fc72e
-
Chess board with 62 pieces
https://claude.ai/chat/ebd63536-953e-473c-bcb4-a334fa0273da
-
Chatgpt teaching (non-interactive diagrams)
https://youtu.be/_nSmkyDNulk?t=58
Problems
- Diagrams can be drawn with code, but are not very fine grained.
- people draw while speaking, and use pointers to point to the diagram
- Most explanations are visual, even in text heavy subjects such as mathematics, computer science, finance etc.
Students level of understanding
Expressing state can be hard with LLMs : https://claude.ai/chat/b38140ac-841f-4b00-aa62-41fa6f2429bc
- If LLMs can eventually do everything a person does on a computer, can they become a good tutor ? : many proof points, physics wallah, coursera, mit ocw etc.
- Need to validate if online learning sitting on a computer is a valid alternative to
Path to AI tutor
- Need data of tutors teaching students to make a good tutor
- Students in coaching classes prepare for exams, because there are very less good schools and colleges in india. Exam matters a lot and so does exam prep.
- Exam prep requires a lot of question solving and students solve many mock papers
- Although coaching centres help a lot with tutoring, they don’t help a lot with personal question solving
POC
- Solve a set of math, physics and chemistry questions on the app
- Add few concepts with details of their content
- Meter bridge question
- Api wrapper for gpt
- Content for meter bridge, fixed diagrams etc.
- Audio recording and transcription
- Hint after transcription + check api
- Populate results in bottom drawer
- get question api, question sent as json, rendered on page as html
- display area for llm output
- send question + current answer and get hint and check api working
- write a doc
- question bank and book scans + index
- audio of user alongwith screen
- simplest way for user to indicate what they need [highlight information on page directly]
- audio + gestures
- store audio and send audio to get transcript. integrate one of live tts apis