Hi John, thank you for sharing this fascinating approach to leveraging AI for low-stakes, high-availability practice in education. Your example of using virtual patients for medical training is inspiring, and the detailed steps you provided to build the interview practice tool are incredibly helpful. I’m especially drawn to how this can increase student confidence while reducing costs and logistical challenges.
I’ll definitely explore creating something similar! For now, I’ve been using ChatGPT’s customization options to create chatbots, but your approach with AI MicroApps has me thinking of the possibilities. I wanted to ask:
- Is there a way to collect user input (e.g., for analysis or to improve the tool) using the chatbot framework you described?
- If I have about 500 users for my app, do you have an estimate of what the usage cost might be, considering the approximate cost of $0.03 per run using GPT-4?
With that volume of users, do you recommend building and deploying the app from a local machine, or would hosting it via GitHub suffice?
I know I could do more research on these questions, but I thought checking with you might provide the most accurate and practical insights.
Thank you so much for sharing your expertise! Looking forward to your thoughts.
- My approach to pricing right now is to simply get the average price I expect for an interaction (In this case, maybe you try to have some longer conversations to get the upper bound) and then multiply it by the number of expected uses. For 500 users, maybe it's an average of 1.5 uses per user, and an upper bound of 5 cents per use. Budget of about $40. I'd advise someone building something like this is set up a key with a $40 max budget and observe how quickly that depletes.
- Hosting via Streamlit would be fine for 500 users. I've got these deployed in some MOOCs with about those numbers and no issues.
Hi John, thank you for sharing this fascinating approach to leveraging AI for low-stakes, high-availability practice in education. Your example of using virtual patients for medical training is inspiring, and the detailed steps you provided to build the interview practice tool are incredibly helpful. I’m especially drawn to how this can increase student confidence while reducing costs and logistical challenges.
I’ll definitely explore creating something similar! For now, I’ve been using ChatGPT’s customization options to create chatbots, but your approach with AI MicroApps has me thinking of the possibilities. I wanted to ask:
- Is there a way to collect user input (e.g., for analysis or to improve the tool) using the chatbot framework you described?
- If I have about 500 users for my app, do you have an estimate of what the usage cost might be, considering the approximate cost of $0.03 per run using GPT-4?
With that volume of users, do you recommend building and deploying the app from a local machine, or would hosting it via GitHub suffice?
I know I could do more research on these questions, but I thought checking with you might provide the most accurate and practical insights.
Thank you so much for sharing your expertise! Looking forward to your thoughts.
Best regards,
Minh
Hi Minh! Glad to hear you are still working on your own apps.
- I'm building support for SQL or Google Sheets storage for each run. It won't be turned on by default but it'll be available for exactly these kinds of cases. Probably ready in early January, but this is the WIP: https://github.com/jswope00/AI-MicroApps/commit/465aec15899bea2e79988c48faf74c42ad91c35b
- My approach to pricing right now is to simply get the average price I expect for an interaction (In this case, maybe you try to have some longer conversations to get the upper bound) and then multiply it by the number of expected uses. For 500 users, maybe it's an average of 1.5 uses per user, and an upper bound of 5 cents per use. Budget of about $40. I'd advise someone building something like this is set up a key with a $40 max budget and observe how quickly that depletes.
- Hosting via Streamlit would be fine for 500 users. I've got these deployed in some MOOCs with about those numbers and no issues.