Talking to APF Compass: Making a Foresight ChatGPT Bot
NOTE: While I usually use Plural Futures to focus and explore alternative futures, especially non-technology focused or led futures, I felt like I should to explore and share ways tools like ChatGPT could help support foresight work.
As part of the recent 20th Anniversary of the Association of Professional Futurists at the Museum of the Future in Dubai, Wendy Shultz provided a delightful gift of an APF-branded USB drive with the 11 years worth of APF’s magazine Compass. That’s over a decade of history of the APF and the foresight field.
I thought it would be interesting if I uploaded all that wonderful writing and wisdom into ChatGPT and see how well LLM-based AI could search, summerize and cite a library of information provided by its user.
Oh Compass Bot, Tell me a Foresight Tale
So I signed-up for a ChatGPT bot (using ChatGPT 3.5) and uploaded the Compass PDF files. After a few minutes, I had CompassBot up-and-running trained on 33 issues of the APF Compass magazine. I started asking CompassBot a few basic questions:
The results were mixed. In almost all cases, CompassBot could give me a response that did address the particular question, but the quality of the responses varied. The majority of the time it was able to select an edition of the Compass with a relevant article, but sometimes it felt random. It would be interesting if it could list different relevant Compass articles and either create a grand summary or compare and contrasts between them.
So far this looks like a good alternative to searching a library of PDFs if you only want a surface understanding of a topic without reading an entire article.
There is likely a lot more ChatGPT could do. I used a simple software to set-up this chat bot in a couple of minutes. Perhaps I could do more with my own version? Or perhaps it needs more data? Indeed, in the realm of LLM, 33 PDFs are a measly sum. Perhaps, my Bot yearns for more data. So I quickly asked myself what other data sources could I use?
What’s next? A DatorBot? Journal of Futures Studies Bot?
I was awfully tempted to make a DatorBot using his publicly available published papers as sourced by Scholars.Google.com but that felt inappropriate, like building a shrine that was also blasphemous (as AI could ‘hallucinate’ things he never said). As a guest editor of a special issue of Journal of Futures Studies on exploring the potential impact of AI on Futures Studies, I knew the next path I needed to take: Train ChatGPT on all the publicly available Journal of Futures Studies published papers.
Asking ChatGPT how to create code to download all PDFs from JFSDigital.org
I realised that the Journal of Futures Studies has easily over 500 published articles available digitally, but no easy way to download them all. So I asked ChatGPT for some code in creating a program to download all PDF documents on the JFS website. And in less than a minute, ChatGPT gave me the code and in a few minutes I had up and running on my computer.
Using the PDF download code that ChatGPT wrote for me, I was able to download around 1000 PDFs of publicly available published articles from the JFS.
It’s taking a while to upload all these PDFs. So I will have to share my findings in another blog post.
But what could be useful in having an AI chatbot respond to your Foresight questions using 1000 published papers? I assume this experiment will yield limited use case at best, and more than anything reveal LLM’s limitations. It might show a good cautionary tale of realism amid this wave of Generative AI hype.
No, I’m not making this bot public
There are, of course, lots of questions about using ChatGPT or any LLM / generative AI for this. I’m still wrangling them myself. There are questions around data privacy (see ChatGPT’s policy here) or what happens if ChatGPT summarises information incorrectly or with false attribution. This is why I’m, for now, not making either chatbots public but I am looking for ways to inspire thinking around this as thoughtfully as possible.
It’s not difficult to set-up your own ChatGPT or other AI-bot that uses data you give it. There are free guides of how to do this online. Fiverr, a freelancing website, has people offering to set-up custom ChatGPT bots for less than 100 USD. I made mine for less than 20 EUR. It took me longer to write this blog post than to set-up either APF Compass or JFS chatbots.
So you may want to give this a spin on your own time. For now, I’ll be playing around with the chatbots and sharing interesting insights as I get them.
Talking to APF Compass: Making a Foresight ChatGPT Bot
NOTE: While I usually use Plural Futures to focus and explore alternative futures, especially non-technology focused or led futures, I felt like I should to explore and share ways tools like ChatGPT could help support foresight work.
As part of the recent 20th Anniversary of the Association of Professional Futurists at the Museum of the Future in Dubai, Wendy Shultz provided a delightful gift of an APF-branded USB drive with the past 11 years worth of APF’s magazine Compass. That’s over a decade of history of the APF and the foresight field.
I thought it would be interesting if I uploaded all that wonderful writing and wisdom into ChatGPT and see how well LLM-based AI could search, summerize and cite a library of information provided by its user.
Oh Compass Bot, Tell me a Foresight Tale
So I signed-up for a ChatGPT bot (using ChatGPT 3.5) and uploaded the Compass PDF files. After a few minutes, I had CompassBot up-and-running trained on 33 issues of the APF Compass magazine. I started asking CompassBot a few basic questions:
The results were mixed. In almost all cases, CompassBot could give me a response that did address the particular question, but the quality of the responses varied. The majority of the time it was able to select an edition of the Compass with a relevant article, but sometimes it felt random. It would be interesting if it could list different relevant Compass articles and either create a grand summary or compare and contrasts between them.
So far this looks like a good alternative to searching a library of PDFs if you only want a surface understanding of a topic without reading an entire article.
There is likely a lot more ChatGPT could do. I used a simple software to set-up this chat bot in a couple of minutes. Perhaps I could my own version? Or perhaps it needs more data? Indeed, in the realm of LLM, 33 PDFs are a measly sum. Perhaps, my Bot yearns for more data. So I quickly asked myself what other data sources could I use?
What’s next? A DatorBot? Journal of Futures Studies Bot?
I was awfully tempted to make a DatorBot using his publicly available published papers as sourced by Scholars.Google.com but that felt inappropriate, like building a shrine that was also blasphemous (as AI could ‘hallucinate’ things he never said). As a guest editor of a special issue of Journal of Futures Studies on the impact of AI on Futures Studies, I knew the next path I needed to take: Train ChatGPT on all the publicly available Journal of Futures Studies published papers.
Asking ChatGPT how to create code to download all PDFs from JFSDigital.org
I realised that the Journal of Futures Studies has easily over 500 published articles available digitally, but no easy way to download to download it. So I asked ChatGPT for some code in creating a program to download all PDF documents on the JFS website. And in less than a minute, ChatGPT gave me the code and in a few minutes I had up and running.
Using the PDF download code that ChatGPT wrote for me, I was able to download around 1000 PDFs of publicly available published articles from the JFS.
What could be useful in having an AI chatbot respond to your Foresight questions using 1000 published papers?
No, I’m not making this bot public
There are, of course, lots of questions about using ChatGPT or any LLM / generative AI for this. I’m still wrangling them myself. There are questions around data privacy(see here) or what happens if ChatGPT summarises information incorrectly or with false attribution. This is why I’m, for now, not making either chatbots public but I am looking for ways to inspire thinking around this thoughtfully as possible.
It’s not difficult to set-up your own ChatGPT or other AI-bot that uses data you give it. There are free guides of how to do this online. Fiverr, a freelancing website, has people offering to set-up custom ChatGPT bots using the data you share with it for less than 100 USD. I made mine for less than 20 EUR. It took me longer to write this blog post than to set-up either chatbots.
So you may want to give this a spin on your own time. For now, I’ll be playing around with the chatbots and sharing interesting insights as I get them.