> AI tools and tutors never can and never should replace teachers.
This part confused me on several levels.
Let's break it down.
---
1. Tutors (human) can not replace teachers on a mass scale
True, and the article touches on this: human tutors simply don't scale!
2. Tutors (human) should not replace teachers
False. Bloom's 2-sigma problem [0] states that on average, 1-1 tutored students outperform 98% of classroom-educated students. (!!!)
Therefore, if we could replace 1:30 education with 1:1 education, the evidence very strongly says that we should.
In other words, the only reason we have teachers, instead of tutors, is because we don't have enough tutors. If we had enough tutors, we would simply 1-1 tutor everyone -- assuming our goal is educational achievement.
3. AI tutors can never replace human tutors
This is a strange claim. It seems to imply that all technological progress will suddenly cease after today -- not just in AI, but human-computer interfaces, robotics, etc.? (Note: the article doesn't make this specific claim, so perhaps I'm straw-manning here.)
I don't expect convincingly human androids for a while. But are androids really necessary for teaching math? GPT-4o seems to do fine! [1]
---
As far as the tech goes, I think the only thing missing is memory and personalization. Currently GPT doesn't know what I know, and what I don't know. What is the dependency graph of knowledge? (KhanAcademy had one, but they've taken it down...) Where in that graph are there gaps in my understanding? What are my specific goals and interests?
Can we present the most relevant learning materials to me in alignment with all these parameters? Or else, can we generate them on the fly? It's all a solvable engineering problem.
This part confused me on several levels.
Let's break it down.
---
1. Tutors (human) can not replace teachers on a mass scale
True, and the article touches on this: human tutors simply don't scale!
2. Tutors (human) should not replace teachers
False. Bloom's 2-sigma problem [0] states that on average, 1-1 tutored students outperform 98% of classroom-educated students. (!!!)
Therefore, if we could replace 1:30 education with 1:1 education, the evidence very strongly says that we should.
In other words, the only reason we have teachers, instead of tutors, is because we don't have enough tutors. If we had enough tutors, we would simply 1-1 tutor everyone -- assuming our goal is educational achievement.
3. AI tutors can never replace human tutors
This is a strange claim. It seems to imply that all technological progress will suddenly cease after today -- not just in AI, but human-computer interfaces, robotics, etc.? (Note: the article doesn't make this specific claim, so perhaps I'm straw-manning here.)
I don't expect convincingly human androids for a while. But are androids really necessary for teaching math? GPT-4o seems to do fine! [1]
---
As far as the tech goes, I think the only thing missing is memory and personalization. Currently GPT doesn't know what I know, and what I don't know. What is the dependency graph of knowledge? (KhanAcademy had one, but they've taken it down...) Where in that graph are there gaps in my understanding? What are my specific goals and interests?
Can we present the most relevant learning materials to me in alignment with all these parameters? Or else, can we generate them on the fly? It's all a solvable engineering problem.
---
[0] https://nintil.com/bloom-sigma/
[1] Math Problems with GPT-4o [video]: https://news.ycombinator.com/item?id=40347320
[1] Math Problems with GPT-4o: https://www.youtube.com/watch?v=_nSmkyDNulk
See also: One study predicted human-level AI tutors by 2025
[2] Achieving Bloom's Two-Sigma Goal Using Intelligent Tutoring Systems (2020) - https://www.researchgate.net/publication/339246677_Achieving...