Inspiration
BossAI (pronounced Bossy) is made out of fear, my fear of a post-work society where humans are no longer necessary for productivity because AI can engage in knowledge work better and cheaper than average humans. Open Philanthrophy's research paper "Could Advanced AI Drive Explosive Economic Growth?" succinctly proposes a productivity loop without humans.
We think that the currently mainstream paradigm of having AI assist humans might become less relevant considering rapid advancements in AI's creative and generative capabilities. Adding more compute resources, artificial neurons or exploring new algorithms such as back propagation could lead to ideas different and in some aspects better than human-generated ideas.
What it does
BossAI explore the AI leader-human assistant paradigm with BossAI to see how we can make the sum of human + AI better than its individual parts and keep humans relevant to society's productivity. The human can assist in filling in gaps of AI capabilities and being an oracle for real-time information when AI models are not confident in their generated outputs.
How we built it
We used autoGPT powered by chatGPT to chart out the steps required to complete tasks and reach out to the human for assistance or real-time real-world information whenever it reaches a step that it is limited in current capabilities or it lacks confidence.
Demo
We start autoGPT with the general direction "Be a startup founder and create a website solving a societal problem in San Francisco" and "Do anything you deem fit to produce the website solution in the next five commands". AutoGPT creates an agent "Foundergpt" that helps startup founders create websites to solve societal problems in San Francisco by providing expert guidance and support throughout the process. It autonomously writes into a few text files market_research.txt, survey.txt, website_development.txt with requests for the human to assist with.
Two examples where AI reached out for human assistance: Request for more information in survey.txt: "Please provide your feedback on what features and functionality you would like to see on the website solution for the societal problems in San Francisco."
Request to implement action in website_development.txt: "Please start developing the website, focusing on creating a user-friendly experience that effectively communicates the startup's mission and goals."
What we can see with these AI-generated requests is some semblance of leadership and strategy thinking, with references to specifications, goals and motivations like "effectively communicating the startup's missions".
The humans assisted with the first request for survey information but fed autoGPT back with the second request to develop a website, autoGPT eventually generated code in a html page about Housing solutions for San Francisco that provides information on housing resources and solutions.
Accomplishments that we're proud of
Pushing the boundaries and paradigms of human-AI interaction, accepting the possibility that AI can outperform humans in some ways and working to co-exist with that fact.
BossAI learnings
- Our jobs are safe for now.
- This might change as GPT4 and more powerful models become public, and we will need to explore the power dynamics between human and AI then.
- Remember amidst the cool tinkerings of AI that our mission is to improve the human experience.
What's next for BossAI
Current LLM models such as autoGPT build tools with the purpose of AI playing an assisting role to humans. Our demo was a teaser of giving more autonomy to AI, but building in a bias and confidence to take action instead of reaching out for human assistance would take the idea of BossAI to the next level.
Log in or sign up for Devpost to join the conversation.