Conversational AI is a great tool for education
twitter.com
twitter.com
I already used GPT to help me tutor my kids, and for my kids to use it when they get stuck. They get unstuck faster. They are critical but also more willing to accept response as fact, we discuss this regularly and they seem to be getting the point.
so many kids get left behind because a teacher is unable to spend time with them, how amazing will it not be for each student to have their own supporting teacher?
hopefully, we will be able to harness AI for the better and good.
In the future I expect AI tutors to ask more questions of students to quickly identify where their knowledge gaps are (so they won't even have to formulate the "bad" questions themselves).
I'd change that to "the future will be wonderful for whoever is able to afford this technology." AI has the power to give some people a massive competitive advantage over others, and some people will be willing to spend a lot to keep that to themselves. It'll be democratized to an extent but you can be absolutely certain "AI tutors" will be a market, and people at lower incomes won't be able to afford the best ones (which is fine, it's just like human tutors. That's capitalism.).
Also I personally think you are under a misconception. Commoditised high quality AI tutors would increase variance/inequality of outcomes, not decrease it. As it would increase the learning rate of gifted individuals even more drastically than the one of the general population.
The more edtech improves and becomes cheaper, the more the limits of individuals will be set by their biological constraints. Which are mostly genetics
--- Why is the sky red at dawn?
The angels are baking, honey! ---
You give a lot of credit to people
The op is saying “I learned <TECHNICAL_JARGON>” but did they? How are they quantifying learning? How do they know what they “learned” is even correct.
I agree with the headline but I think it needs a qualifier of “in the presence of an educator”. The educator, can be a technical text, is there to sanity check the conversational agent.
In my experience the best use of such agents at this time is in a domain where I’m already an expert and I want it to do some remedial work or lookups for me.
IMO, these Conv-AI tools should indicate to the user when they are hallucinating.
If they could do that, then it would be fairly easy to "not hallucinate".
To put it another way the "don't hallucinate" problem and the "warn me if you're hallucinating" problem are in the same difficulty class.
Some people are quite introspective and have pretty good meta-awareness, and hedge when they're not sure about something. While other people are known to just bullshit whenever they don't know something (or when they just feel like it), and don't change this behaviour even when lots of people have pointed it out.
How much more would you be worried if you were having a conversation with a known bullshitter?
If the LLM can give me a bird's eye view of the subject, then it enables me to go off and do my own research and come with my own conclusions, even if they don't align with what the LLM originally told me.
The fact is, there's a ton of misinformation on the Internet. Doesn't matter if you're getting your info from an LLM or not, you should almost always be trying to get your info from multiple sources if possible.
The friction involved is counterbalanced by the conversational aspect which makes it feel less tedious or even fun.
I often like to have the ai:
- provide exercises
- play the student asking me the e teacher to clarify and answer questions
- have it put on a little theater play with different students
I was talking about this earlier on mastodon:
- https://hachyderm.io/@mnl/111471750626975201
- https://hachyderm.io/@mnl/111467764783141450
- https://typeshare.co/go-go-golems/posts/ai-driven-self-educa...
I also find it one of the most effective ways of learning, and complement it with the usual book/video learning and (very importantly) writing down the important points learned during the conversation (potentially coming up with Flashcards which can then also be “cosplayed” by gpt), because gpt transcripts are just indigestible.
Generally - the areas where GPT starts to become unreliable are usually fairly "off-piste". My gut feeling is that in most standard educational contexts it would probably be fine.
Curious to hear specific examples where people have found this not to be the case.
- Using an LLM that hallucinates less. GPT-4 is more reliable than GPT-3.5, and the just-released Claude 2.1 is reportedly better than its predecessor. In my experience, Bard confabulates too much to be useful for many purposes.
- Using the AI to explore relatively general topics. In my tests, GPT-4 is excellent for getting an overview of, say, linguistic theories, the history of ethics, or the differences between quantum and classical physics. The more focused the topic is--how a particular verb conjugates in Romanian, what David Hume said about the death penalty, how gravity affects neutrinos--the more you need to double-check with other sources.
- Focusing not on learning facts but on the interactive exploration. One interesting exercise is to discuss counterfactuals: How might human civilization have developed if electricity had not been harnessed? What would have happened if a fifty-meter-diameter asteroid had struck the Rhine Valley in May 1944? There are no right or wrong answers to such questions, but exploring them with the AI can be very rewarding.
I agree with the OP: Interaction with LLMs can be a great way to learn, and it will get only better as their reliability improves further and they become more customizable for the individual learner. What I want most now is for them to have a persistent memory of our past conversations. Better multimodal capabilities would also be nice.
When combined with the ability to run code and read images I think it will really help with learning math. Show it your work, have it tell you why you got a wrong answer, and then it can tell you the concepts you need to review
The big disclaimer is that you can really only do that with subjects that you already know very well. Then again, precisely that can be part of the fun. Because it is fun to see where the model gets it right, vs where some details are in conflict with something you know from other sources.