I’ve spent a lot of my life getting good at broadcast-style instruction, specifically lecturing and technical writing. These are both difficult skills because you have to learn to build DAGs. When teaching, you have to understand the delta between your understanding and your student’s understanding. When doing broadcast-style instruction, you do this for many students concurrently.

It’s not only that this is hard but also that it forces you to teach suboptimally to each student. When building a single piece of curriculum to teach everyone, you have to build up from your students’ lowest point of shared knowledge. This means you’ll go too slowly for the high-performers, and you’ll inevitably leave some students in the dust. In school we try to control for this with prerequisites and waitlist screening forms, but these tools are far too coarse-grained to be useful.

Written materials like textbooks solve some of this since students can move at their own pace, but they don’t solve the interactivity problem. The only way to learn at max difficulty & speed is to do so interactively. When you learn something, you must think hard about it and test your understanding by asking probing questions. Textbooks can teach you the information, but they can’t help you test the boundaries of your understanding.

We use broadcasting because it’s economically scalable. One teacher can broadcast to 30 students. One textbook can reach millions. In a perfect world, we’d be able to give each student their own PhD-level tutor who can tailor instruction to exactly their knowledge and skill level. Sadly, this is cost-prohibitive.

I have high hopes for AI tutors as a way to solve this problem. Seeding an LLM with high quality data will allow users to interactively learn about the topic of their choice in a high-bandwidth, personalized, and social way. Hopefully, my kids will be much smarter than I am.