Hacker News Reader

Comment by 11thEarlOfMar | Hacker News Reader

IMHO, we should be able to train an AI to be a constant in-the-student's-ear teacher. The AI should be able to learn about the student's specific learning style and hence, the student and AI teach each other.

If every student is taught in the style that they learn the most efficiently with, could we expect performance to increase dramatically overall? Would it work to close the socio-economic performance gap among schools? How would the role of human teacher change?

I am sure there are people who've been focusing on this, it can't be novel, would appreciate to know what's being done in this realm.

View on HN

No. LLMs aren't capable of learning how a student best learns. Nor are they capable of really "teaching".

At best, using an LLM this way would get it to parrot out existing, common explanations for some math topics.

If you asked it something outside its existing knowledge base (or asked it a question from a different perspective), it would make up fake or logically inconsistent, but very reasonable-sounding explanations which would be horrific for a student trying to learn.

It cannot actually learn mathematics or logic itself, so it has no way to ensure that its explanations are logically or mathmatically sound.

idk. I think if I were a kid provided with a curriculum of math, I could have grinded out many years of progress in a fraction of the time using an LLM. I would probably not have practiced the execution of curriculum, but I would have been hitting a lot of different reasoning models. The ability to ask followup questions in LLMs is great. And its proficiency in grade school math is very good.

I'd rather be able to think through vector spaces than be able to solve algebraic equations effectively and efficiently.

Sounds a lot like the "Mrs Davis" TV Show.

Anyhow, this is not a good idea. We are having issues with their scores and performance declining. We have to find the root cause, not try to patch it with another band-aid by saying "Here, lets try this OTHER NEW THING", on top of the hundreds of other tools students already have.

How many years should we spend on this 'find the root cause' exercise without acting to intervene?

But instead we will train AIs to be constant in-the-person's-ear salesmen and propagandists.

Yes, I’m adamant. But schools are very slow to react and commercial initiatives are hardly a substitute for school curriculum at least where I’m from. School teach and then teach to test. So even a little bit different way of phrasing could throw a students ability to score the local test. In the long run you’d be learning but the signaling function of education is then missing. Plus what I notice in one of my kids: tests below his level get abysmal performance due to lack of interest. Again teaching to test in his class with him lacking the grit to just perform regardless of his feelings regarding the difficulty of the test. (He is nine so nothing lost yet.)

This is the goal of Khanmigo.

https://www.khanacademy.org/khan-labs

Is there some principled reason to believe that an AI trained on 8th grades will emergently perform better than 8th graders?

Hmm... that's a clue to how I may have poorly constructed my post.

The point is that the AI is already fully trained on teaching. But it uses the student's learning progress to determine how that specific student learns the most efficiently.

Yes, possible dystopian outcomes abound, maybe they can be mitigated.

CGP Grey talked about it. "Digital Aristotle"

If you were extremely rich and lucky in the time of Aristotle, you could have Aristotle as a tutor. Why can't everyone have their own personal digital 24/7 tutor?

https://www.youtube.com/watch?v=7vsCAM17O-M

Reminds me of the "Young Lady's Illustrated Primer" from Neal Stephenson's "Diamond Age", too:

https://en.wikipedia.org/wiki/The_Diamond_Age

Reply on news.ycombinator.com