PhD holders waste their potential on teaching

[caspex]

2 hours ago, Carl-Richard said:

Please, no, just no. I would rather let a kid teach me than ShitGPT.

well chatgpt is a bad example because it's not trained to teach, nor is it trained in a narrow subject to have very specialized knowledge. If an AI was also trained, in addition to teaching, on the kinds of questions students ask for that particular subject, it'll be a godly teacher.

[Carl-Richard]

6 hours ago, Swarnim said:

well chatgpt is a bad example because it's not trained to teach, nor is it trained in a narrow subject to have very specialized knowledge. If an AI was also trained, in addition to teaching, on the kinds of questions students ask for that particular subject, it'll be a godly teacher.

What does the teaching involve which ChatGPT cannot already do and which doesn't require generalized AI?

[caspex]

Just now, Carl-Richard said:

What does the teaching involve which ChatGPT cannot already do and which doesn't require generalized AI?

It's like how ChatGPT is shit at math stuff but WolframAlpha, which is trained for that, is much better. I'd imagine AI trained for teaching particular subjects would be much better at answering questions and explaining nuances than ChatGPT, good enough that I think they can teach you the basic concepts very well. I'd still suggest studying intermediary and advance concepts from humans though, but I think AI should take off some load from the professors when it comes to the introductory stuff.

ChatGPT is taught basically everything, it can mix up views and easily contradict itself when it comes to technicalities. I'd assume an AI trained in a very narrow range of subjects and nuances (questions), will be able to teach much better. When the questions delve too deep into higher concepts, the AI can simply be trained to say that it's a higher concept that it's not equipped to explain. It could then give sources to learn about those higher concepts. It could also be trained to tell you what's relevant to the syllabus at hand and what's not. The possibilities are crazy.

ChatGPT was never made to teach you stuff, of course it's bad.

[Carl-Richard]

1 hour ago, Swarnim said:

It's like how ChatGPT is shit at math stuff but WolframAlpha, which is trained for that, is much better. I'd imagine AI trained for teaching particular subjects would be much better at answering questions and explaining nuances than ChatGPT, good enough that I think they can teach you the basic concepts very well. I'd still suggest studying intermediary and advance concepts from humans though, but I think AI should take off some load from the professors when it comes to the introductory stuff.

ChatGPT is taught basically everything, it can mix up views and easily contradict itself when it comes to technicalities. I'd assume an AI trained in a very narrow range of subjects and nuances (questions), will be able to teach much better. When the questions delve too deep into higher concepts, the AI can simply be trained to say that it's a higher concept that it's not equipped to explain. It could then give sources to learn about those higher concepts. It could also be trained to tell you what's relevant to the syllabus at hand and what's not. The possibilities are crazy.

ChatGPT was never made to teach you stuff, of course it's bad.

So it's still just a chatbot? I think a specialized bot actually needs general ChatGPT training as a support beam, or else it will be nothing like teacher. A teacher has general knowledge and special knowledge. The student also has some general knowledge, which determines which questions and concepts the student feeds the bot, and if the bot doesn't have any training in that general knowledge, it won't be able to unpack those questions (if it's even able to be coherent at all). So with a ChatGPT support beam, it will probably have just as many flaws as ChatGPT if not more.

Edited August 5, 2023 by Carl-Richard

[caspex]

3 minutes ago, Carl-Richard said:

So it's still just a chatbot? I think a specialized bot actually needs general ChatGPT training as a support beam, or else it will be nothing like teacher. A teacher has general knowledge and special knowledge. The student also has some general knowledge, which determines which questions and concepts the student feeds the bot, and if the bot doesn't have any training in that general knowledge, it won't be able to unpack those questions (let alone be coherent in any fashion). So it will probably have just as many flaws as ChatGPT if not more.

Of course, that's a given. I was stating that it could be fed more specialized knowledge regarding questions asked by students. It could be fed data that explains the same concept from many angles. Imagine if an AI was trained on all kinds of questions regarding a concept or topic or a subject, and tons of quality answers for those questions. ON TOP OF your usual specialized knowledge of that subject that it could also pull from.

I am sure there are problems with my approach, but regardless of that, my point is that while I am not quite sure how, I believe it's definitely possible to create an AI that is great at explaining stuff. Especially as AI advances. It's not wise to think that AI's teaching capability is maxed out in ChatGPT.

[Jodistrict]

ChatGPT is extremely useful for learning.  It’s like having an interactive textbook.  In the traditional method, you read a chapter and then answer the questions to test your knowledge.  It is static and inflexible.  With ChatGPT,  you can ask “But what about this?” questions to get a more complete understanding of the subject.  It is the analogy of having a list of questions to ask the professor after class.  As has already been mentioned, the quality improves dramatically by training the learning algorithm to a specific subject.   

[Carl-Richard]

3 hours ago, Swarnim said:

Of course, that's a given. I was stating that it could be fed more specialized knowledge regarding questions asked by students. It could be fed data that explains the same concept from many angles. Imagine if an AI was trained on all kinds of questions regarding a concept or topic or a subject, and tons of quality answers for those questions. ON TOP OF your usual specialized knowledge of that subject that it could also pull from.

I am sure there are problems with my approach, but regardless of that, my point is that while I am not quite sure how, I believe it's definitely possible to create an AI that is great at explaining stuff. Especially as AI advances. It's not wise to think that AI's teaching capability is maxed out in ChatGPT.

Sure. I was mostly just mocking the current state of ChatGPT.

 

8 minutes ago, Jodistrict said:

ChatGPT is extremely useful for learning.  It’s like having an interactive textbook.  In the traditional method, you read a chapter and then answer the questions to test your knowledge.  It is static and inflexible.  With ChatGPT,  you can ask “But what about this?” questions to get a more complete understanding of the subject.  It is the analogy of having a list of questions to ask the professor after class.  As has already been mentioned, the quality improves dramatically by training the learning algorithm to a specific subject.   

If you manage to bypass the fact that it spews complete bullshit in a non-zero percent of answers, sure, it's a great learning tool. But for me personally? No thanks.

[Jodistrict]

Coursera has a free course "Prompt Engineering for ChatGPT".   

https://www.coursera.org/

[Carl-Richard]

On 5.8.2023 at 8:09 PM, Jodistrict said:

Coursera has a free course "Prompt Engineering for ChatGPT".   

https://www.coursera.org/

One thing that ChatGPT is useful for is as an enhanced google directory in cases where you can easily control check if the information is true afterwards. For example, I asked it for a Bible verse that has a certain message, and I could easily control check if the verse actually exists, and I can also decide whether the interpretation of the message is OK.