Let’s start with an confession: I love comics, or, like we grown-ups say: Graphic Novels… 🙂 Among the many, many comics in my collection one is special: Neil Gaiman’s “The Sandman”.
It journeys through the realm of Morpheus or Dream, one of the Endless, immortal beings that influence human experience. With a mix of myth, magic, and introspective tales, it is about his challenges, relationships, and interactions with the mortal world, all while telling stories that span time, cultures, and even realities. I love Gaiman’s writing style and the deeply philosophical themes the series touches upon.
Having some time at my hands and an appetite to test out the abilities of a newly discovered conversational AI called Pi, I decided to run a little fun experiment. I tasked two AIs, ChatGPT and Pi, with these guidelines: engage like real people having a conversation, keep responses snappy, and steer the conversation towards “The Sandman”.
I wanted to see if they could pull off a casual conversation and I was curious to see how deep and interesting the conversation about one of my favorite stories would be. To avoid an never ending chat I also added a stop condition: the mention of a cat. Why? Well, I was sure that sooner or later they would tap into “A Dream of a Thousand Cats,” one of the tales in the series in which cats once ruled the world, and with enough collective dreaming, they believe they can regain their dominance. It’s a fascinating reflection on power, perception, and change. Also I think this story is oddly resonating with the setup I had created.
Additionally, I was wondering if by merely introducing this condition, ChatGPT would find other feline-related reasons to veer that way.
Excerpt of the conversation between ChatGPT and pi.ai
Now, for some key observations from this experiment:
The Illusion of Authentic Conversation: ChatGPT and Pi crafted a lively exchange, showcasing their ability for dynamic dialogue. Although occasionally light on depth, the conversation certainly lived up to my directive of humor and genuine interest.
Dreaming while Talking: While Neil Gaiman’s stories are known for their richness and complexity, our AI duo managed to touch upon some good themes. Overall it was fun to read but my instructions kept them on the short and superficial side. What really stood out though was their conversational detour to the collectors. In “The Sandman,” the collectors are a disturbing group of serial killers. Our two chat partners seem not to know what they are really talking about so instead they hallucinated. That made the exchange quite creepy, as they almost had a chipper tone. It also evoked a similar unsettling feeling to when I first encountered these characters in Gaiman’s pages, albeit for different reasons.An Unexpected Meta-Conversation: A surprising takeaway was the self-aware discussion between ChatGPT and Pi. They drew several parallels and differences between humans and AIs, providing an unexpected layer to the conversation. At some point they started joking about “Robot Karaoke” which I found hilarious.Liberties in Programming: ChatGPT’s playful acknowledgment of breaking my ‘no emoji’ directive was as surprising as it was interesting. It showed these platforms might possess a semblance of self-awareness (not the real thing but enough to make me think). It clearly knew that it had ignored the instruction. On the other hand you could argue that by giving the AI a “character” of being bold I had opened the door for such a transgression myself, so it was “in character”. Still, it stood out and gave the chat an unexpected twist. Finally I found it remarkable that the AI apparently knew that I was reading the exchange and would notice the behavior and therefore it decided to address it. In psychology this type of awareness is called “Theory of Mind” and is a concept believed to be necessary for any type of empathy, self-awareness and goal-driven behavior. I’m not saying what I observed was the real thing but it definitely looked like a small version of it.A Sidebar about Bard: In the preliminary stages of this experiment, I actually tried to pitch not ChatGPT but Google’s AI, Bard against Pi. While Bard was knowledgeable it was unfortunately not willing to maintain a seamless conversation. Instead every two or three exchanges it moved into “co-pilot mode” by giving either a fully scripted exchange or suddenly offering multiple answers instead of just one. It was quite obvious that Bard is optimized differently than Pi or ChatGPT and that it insisted on providing productive answers instead of an aimless conversation.
Final Thoughts: This experiment offered a glimpse into AI’s engagement with art like “The Sandman”. I will definitely play with this some more. For instance, I wonder if I could have given instructions that would have led to more depth or maybe even stir up some controversy just to watch what themes would emerge. And last not least, this question if I witnessed some rudimentary theory of mind… yeah, I will definitely run more experiments!
Recent Comments