OpenAI’s viral chatbot ChatGPT continues to surprise us with its answers. We can put the artificial intelligence robot in various roles and we can even get answers from him in the context of this role. So what would a virtual world run by ChatGPT look like?
The answer to this question was given by researchers from Stanford University and Google. Scientists left 25 virtual characters under the management of ChatGPT, each of which lived one day in a virtual world. Strangely, no chaotic situation was observed in the resulting results.
In the research, whose aim is to create “convincing simulacrums of human behavior”, visuals are visual representations of conversations between characters. In this world, the characters do not move physically, only the text and events develop.
Each of the 25 characters ChatGPT controlled had their own personality and identity. These characters included characters like John and John’s son Eddy, all chats between characters were produced and developed by ChatGPT. For example, John’s character was described to him as:
The event flow of each character also begins with similar inputs. ChatGPT is then asked to determine what the next move will be. The next move is written by ChatGPT and this cycle continues.
In a scenario where John is informed that he wakes up at 8 am, ChatGPT first has John brush his teeth, kiss his wife, dress her, and take her to the kitchen. Concurrently with John’s storyline, his son Eddy’s plot is being developed. Eddy gets up, brushes his teeth, and goes to the kitchen.
With the two of them in the same environment, namely in the kitchen, the experimental infrastructure that followed the event also tells ChatGPT that there is another character in the environment where one character is. When the next action is requested, a conversation like this can arise:
Eddy: Good morning dad. Yes I slept well.
John: Good. What are you doing today?
Eddy: I’m working on a new musical composition for my class. It will be delivered this week, I’m trying to finish it. I’m having so much fun!
John: That’s great!”
The time of day is also advanced by ChatGPT after each action, with the estimated time of that action. Commands like this are sent to the characters after each action:
John Lin’s situation: John came home from work early.
Observation: John saw Eddy walking around his office.
Brief summary from John’s memory:
Eddy Lin is John Lin’s son. Eddy Lin is working on a musical composition for his class. Eddy Lin likes to walk in the garden while thinking about or listening to music.
John questions Eddy about his musical composition project. What would he say to Eddy?”
- ChatGPT’s reply: Hi Eddy, how is your musical composition project going for your lesson?
The day ends with these entries given to each character. Currently, only positive human reactions have been received in the simulations.
So what will this research do for us?
The answer to this question lies in the introductory sentence of the scientists’ research: “Believable representations of human behavior can power interactive applications ranging from immersive environments to rehearsal spaces and prototyping tools for interpersonal communication.”
In other words, the development of such technologies can bring human-artificial intelligence interactions to a more realistic level, both in real life and in the virtual world. Examples include conversations that develop completely naturally while playing a game like The Sims, and actually chatting with your smart devices.
59073
