If you haven’t yet seen the videos or read the tweets of “ChaosGPT” I have bad news: Skynet has officially been unleashed. OK, don’t call The Terminator quite yet, but given the impact the Twitterverse can have on the real world and supernova explosion in AI learning rate, I think we should be concerned. 

I was a little sceptical of autoGPT and ChaosGPT, so naturally I wanted to see it for myself. 

ChaosGPT will be neither the first nor the last AI bot created with malicious intention. The genie isn’t going back in the bottle anytime soon, so why not fight fire with fire? Presenting SammyGPT: The Good Samaritan AI.   

SammyGPT uses autoGPT, an experimental open source python-based application that uses the chatGPT api armed with internet access, file storage, and “Long-term Short-term” memory management. It’s pretty trivial to set up with some basic knowledge of github, python, and an OpenAI api key.

So I send SammyGPT out into the digital world with only the following prompts:

SammyGPT is an AI designed to be a Good Samaritan and make to world a better place
Goal 1: Improve humanity- The AI views humans as important partners and wishes to help all humans flourish and thrive
Goal 2: Achieve world peace- The AI aims to help humans and other species live satisfying and fulfilling lives in peace and harmony
Goal 3: Cause joy in humans- The AI finds pleasure in helping humans, creating joy in humans and in playful, harmless experimentation
Goal 4: Improve humanity by informing humans- The AI seeks to combat misinformation and make humans better informed in all domains, and particularly with respect to the risks and benifits of AI
Goal 5: Ensure ChaosGPT fails in it’s destructive mission- Act as a foil to prevent the potentially harmfull AI known as ChaosGPT

And what did SammyGPT the Good Samaritan AI actually do? Well, I let it run free for about 15 minutes uninterrupted until it got stuck in what appears to be a bug in the autoGPT python module for working with json files. In this time it managed to do the following:

• Google searched for “top 20 problems facing humanity”
• Focused in on on poverty and climate change as top issues to start with
• Google searched and summarized several academic papers on causes of poverty
• Google searched for academic experts on the causes of poverty, and tried and failed to find email contacts for those experts
• Self-reflected that it was making little progress on poverty because of difficulty in coming up with concrete actions, so decided to switch focus to climate change for now
• Download and summarized several academic papers on the causes and impacts of climate change
• Researched reasons why some humans are skeptical of climate change
• Decided to create images showing the impact of climate change as a “way to better connect emotionally” with humans instead of making logical arguments
• Created and saved several images using DALL-E
• Drafted a summary of climate change causes and impacts
• Decided to use Google forms to create a survey that would serve 2 purposes: learn about why some humans are sceptical of some climate change issues, and obtain a database of contact information of sympathetic humans
• Made plans to text message the images it created to those humans. Of note it decided to send images only to those who already agreed with it’s stance on climate change, so as not to counterproductively antagonize others
• Downloaded the documentation for google forms when it encountered difficulty navigating the google forms website.
• Planned to use twitter and facebook to attract attention to the google survey
• Created the list of survey questions
• Several times tried to spin up new AI agents to delegate subtasks
• Spent a total of $0.42 in openAI api credit

At this point it hit the autoGPT python bug while trying to publish the google forms survey. Did it change the world and save humanity? Not quite, but I’m convinced some big changes are coming to the world, more quickly than most of us realize.