Curation: The Last Bastion of Teaching

22 Dec, 2024

I've been experimenting with teaching throughout my career. And while teaching has many components, I think only one of them will survive the rise of AI tools over the next ~2 years: curation.

AI can already do the following:

• Explain concepts in exquisite detail. Answer questions with infinite patience
• Integrate feedback and tailor explanations to each student
• Analyze text, images, audio and video and turn them into explanations
• Create and grade exercises. Provide feedback to the student
• Create slides, visualizations, simulations and interactive apps
• Inspire curiosity and provide emotional support (Claude is especially good at this)

What is left for humans? Curation. Which means:

• Choosing what to teach and what not to teach. Identifying the essential load-bearing concepts and discarding the rest.
• Guessing what the students should do with guidance vs. what they should attempt on their own.
• Identifying the best order for teaching something. Textbooks tend to present knowledge in logical sequence, but often when teaching you move in a roundabout way, based on what is needed to support intution and understanding.
• Knowing what's useful in the real world. If you ask AI about my job (data engineering) it can explain any individual concept, but it doesn't necessarily know what we're doing in the industry today, which technologies and practices are actually useful to learn. Much of that knowledge is never written down on the public web [1].
• Identifying high-quality sources of information and filtering out bad resources.

Today, AI is extremely advanced in terms of raw intelligence, but two limitations persist:

1. It doesn't know things that are not written on the web. This includes a lot of real world knowledge, which is either oral or locked in private files. I call this 'dark knowledge'.
2. It is heavily bottlenecked on context. You can (laboriously) try to insert the missing information in its limited context window, but it will take you time and you'll soon run out of tokens.

The teaching functions I listed above are curently resistant to AI because they need a large context and rely on dark knowledge [2].

So what will teaching at the frontier look like in the next 2 years? The most successful teachers will be curators. They will heavily leverage AI to automate and scale their work while they focus on curation. They will spend most of their energy to:

• Create tailored syllabi and learning paths
• Turn dark knowledge into visible knowledge
• Provide AI with the rich context that it needs to excel
• Review AI outputs for correctness and safety

This new form of human-AI collaboration will greatly increase the quantity and quality of learning experiences. It could also turn more experts into teachers, who previously didn't have the time to do the grunt work that comes with the more boring/tiresome tasks (like writing tutorials by hand or grading papers). But it might negatively affect professional teachers, who are largely paid to do said grunt work, and are not necessarily great curators [3].

[1] To be fair, professional teachers face the same limitation. You can see this in computer science: those who are teaching (professors etc.) don't know what the industry is doing, while those who know what the industry is doing are too busy to teach.

[2] Additionally, we need experts to review the AI's output for correctness and reliability. But this is a more tractable problem.

[3] What about human connection? Wouldn't people prefer to be taught by other people rather than machines? Maybe; but I expect that AI tutors will be so much cheaper and yet better than humans, that the economics of hiring human teachers will not make sense for the vast majority of the population. Hence I don't see this as a determining factor.