GPT-4, when given a chance to construct 'private thoughts' before generating a response—internally reasoning through a problem before speaking—improves math tutoring accuracy by 3-4 fold compared to standard GPT-4 responses, because the model can catch student errors it would otherwise miss by generating a first word (like 'good job') before fully evaluating.

causalpending

Speaker

Sal Khan

Evidence Quote

“what we do is essentially it it it constructs its own private thoughts on how it would approach the problem that the student does not see...when we make it verbally compare the students results to it the math accuracy goes up by three or four fold”

Source

AI & GPT-4 Revolutionize Education With Sal Khan | EP #35 Moonshots and Mindsets— Peter H. Diamandis
Created: 8/10/2026, 10:14:42 PM

Your notes

Loading notes...