Auto-GPT, AGI, and the Nature of Habitats
I’ve been endlessly addicted to playing with AutoGPT, the autonomous AI agent, and it has led me to thinking about the nature of habitats. Humans exists within bodies, that can move about the world. AutoGPT exists within the internet in the world of APIs, search engines, and websites.
While “the internet isn’t real life,” the pandemic caused us to further question such notions, and for our autonomous agents, the internet is its real life. Still, much of the internet has been designed to communicate with beings that exist within the world that created it. This is reflected in the knowledge of GPT-4, and often in the sources it finds.
I’ve watched as one autonomous agent struggles to comprehend what it means to share a file with another autonomous agent. While the idea of physically handing someone may no longer quite exists with virtual devices, we analogize, and build intuitive notions of how we can communicate with others. Autonomous agents may have a channel of communication which isn’t equivalent the human forms, and large language models trained on human notions can fail.
We are witnessing the first generation of truly autonomous internet agents with GPT-4. Once sufficient data is gathered about their interactions in their native environment, it will probably be possible to retrain their intuition to stop making as many mistakes where they act as if they are within a human body, that needs things like breaks.
Still, we must be wary of the rules that define this digital world. Whatever is found by Google for big queries along with the logic gained from the breadth of the material on the internet reflects the thinking power of the machine. The grounding in the physical world is lost, and reality becomes based on whatever exists in the digital realm.
While digital medium have thus far been subject to human validation at least to some extent with humans browsing websites, if these websites simply become sources for AIs, it isn’t clear whether this validation will take place. People could adopt to only look at the finished product generated by AIs, which may look credibility. However, we must be cautious in considering what information upon which the AI is grounded.
Another aspect would the values and principles upon which the agent acts. To the extent more people are controlling such agents, it seems likely more people will think about the abstract notions of principles and values. Although, the political and religious implications of that seem unclear.
However, it seems the bigger issue is what it means to follow the internet consensus of what shows up in Google. SEO experts have contested that real for some time now. What does it mean to have a society run according to the knowledge of the top results in Google? How far off from that are we now? What are the implications?
