David - have you played around with have it provide students with developmental feedback after the interaction? Or even better, asking it to guide students through a reflection on their experience (what went well, where they excelled, where they struggled) and then having it provide it's feedback? It draws the experience out for the student, but I think that reflection piece is poewrful
Oohhh yes! I know there is a whole part of the game where the GPT actually gives the students a score! Would love to David to talk more about that piece. :)
The prompt length is a factor. I found it best to keep it under a page. There is a rough scoring system integrated into the prompt, which wasn't shared (it wouldn't be useful for a public audience since it's specific to the scoring rubrics in my course). That scoring system provides instant feedback after each student interaction. It can be unpredictable in its assessment. So, I try to keep it as lean as possible, and I don't use it assign grades.
Do you provide source material on the philosophers in the knowledge base or is the GPT using its training data? I would imagine this might make a significant difference in the quality of the responses.
It's just GPT4+'s training data. The assignment is designed to acquaint students with the more famous arguments and philosophical moves that were made by those involved in the debate; ChatGPT4+ is already well acquainted with them. The assignment isn't designed to go any deeper than what's already been dug. We have not yet seen an LLM capable of doing that. Perhaps soon.
Try Notebook LM and then upload as many chapters of the philosophers work as you can get through PDF's and you might find the sophistication of the responses are even better. Just a thought. I've done this with multiple texts and it's fairly impressive. My one recommendation is to break the works into chapters of no more than 10 - 20 pages. I find entire PDF"s, while the context window can handle them, aren't as effective. Though it sounds like for what you're doing, it's enough.
I love the potential of role-playing and simulations - it opens huge possibility for engaging students so the learning is much stickier. I've been playing around with providing the customGPT examples of "good" and "bad" interaction examples. I think it enables it to provide more robust and valuable developmental feedback.
The image generated is of white people, only one woman. Yes, this is indeed the panorama and historical context we have created. Still, it is a clear example of how to teach critical thinking to students, showing the biases of GenAI when creating content, especially images (you can see it more clearly). Thank you for sharing.
It indeed gives us much to think about. But I'm not sure we should infer a bias from its generations; at least, not these ones. I was pretty specific that the setting is 1920s Vienna; it might be generating the photo from its being trained on archive photos. Also, although it isn't in the article, in the next assignments -- one simulates a famous philosophical event in the 1960s, and one in the 1980s -- there is greater diversity in the generated image.
You raise an interesting point wondering how it chose to represent the scenes in the way it did. I wonder.
Indeed, your argument is correct, and it is for this reason that I stated the historical context was such. I am of the opinion that GenAi is a tool with the potential to facilitate the development of critical thinking skills, or at the very least, that is my objective. In any case, engaging in discourse with philosophers is always beneficial. Thanks for sharing.
David - have you played around with have it provide students with developmental feedback after the interaction? Or even better, asking it to guide students through a reflection on their experience (what went well, where they excelled, where they struggled) and then having it provide it's feedback? It draws the experience out for the student, but I think that reflection piece is poewrful
Oohhh yes! I know there is a whole part of the game where the GPT actually gives the students a score! Would love to David to talk more about that piece. :)
The prompt length is a factor. I found it best to keep it under a page. There is a rough scoring system integrated into the prompt, which wasn't shared (it wouldn't be useful for a public audience since it's specific to the scoring rubrics in my course). That scoring system provides instant feedback after each student interaction. It can be unpredictable in its assessment. So, I try to keep it as lean as possible, and I don't use it assign grades.
Do you provide source material on the philosophers in the knowledge base or is the GPT using its training data? I would imagine this might make a significant difference in the quality of the responses.
I believe the GPT is (surprisingly!) just using its training data. Let’s have @thephilosopherai weigh in and share
It's just GPT4+'s training data. The assignment is designed to acquaint students with the more famous arguments and philosophical moves that were made by those involved in the debate; ChatGPT4+ is already well acquainted with them. The assignment isn't designed to go any deeper than what's already been dug. We have not yet seen an LLM capable of doing that. Perhaps soon.
Try Notebook LM and then upload as many chapters of the philosophers work as you can get through PDF's and you might find the sophistication of the responses are even better. Just a thought. I've done this with multiple texts and it's fairly impressive. My one recommendation is to break the works into chapters of no more than 10 - 20 pages. I find entire PDF"s, while the context window can handle them, aren't as effective. Though it sounds like for what you're doing, it's enough.
Cool! Thanks for sharing, Steve. Are you doing something similar to Dave in your classroom?
I love the potential of role-playing and simulations - it opens huge possibility for engaging students so the learning is much stickier. I've been playing around with providing the customGPT examples of "good" and "bad" interaction examples. I think it enables it to provide more robust and valuable developmental feedback.
The image generated is of white people, only one woman. Yes, this is indeed the panorama and historical context we have created. Still, it is a clear example of how to teach critical thinking to students, showing the biases of GenAI when creating content, especially images (you can see it more clearly). Thank you for sharing.
It indeed gives us much to think about. But I'm not sure we should infer a bias from its generations; at least, not these ones. I was pretty specific that the setting is 1920s Vienna; it might be generating the photo from its being trained on archive photos. Also, although it isn't in the article, in the next assignments -- one simulates a famous philosophical event in the 1960s, and one in the 1980s -- there is greater diversity in the generated image.
You raise an interesting point wondering how it chose to represent the scenes in the way it did. I wonder.
Indeed, your argument is correct, and it is for this reason that I stated the historical context was such. I am of the opinion that GenAi is a tool with the potential to facilitate the development of critical thinking skills, or at the very least, that is my objective. In any case, engaging in discourse with philosophers is always beneficial. Thanks for sharing.