Generative AI works by selecting the most likely next words.

But students misunderstand in many different ways. How would it be tailored instead of another way of delivering lesson plans?

Even assuming you already had the topology of the problem space, you'd have to interrogate the student with all possible ways for each question to determine what's missing for them.

But part of the problem is impedance matching: failing students get little positive feedback because they have to correct many mistakes before they start getting things right. Interrogation makes the impedance-matching problem much, much worse (no matter how sweet and enticing the voice).

Sure, I believe in success patterns and strong leaders. But I'd need to see a sketch of how Khan understands and addresses the problems before I'd invest or rely on this as anything other than a stop-gap that continues the under-funding of mass education.