A Microsoft scientist and researcher at the University of York, Adrian de Winter, recently decided to test GPT-4 by running Doom without prior training. The results of this experiment led to the conclusion that GPT-4 can play Doom with a certain degree of effectiveness without prior training. However, this requires a GPT-4V interface for artificial intelligence vision and the addition of Matplotlib and Python to enable the input of artificial intelligence.
The GPT-4 AI while playing Doom can open doors, fight enemies, and shoot weapons. Almost all the main functions of the gameplay, including world navigation, are used. However, the artificial intelligence does not have a sense of object permanence - once an enemy goes off-screen, the AI no longer recognizes it at all. Despite instructions being included on what to do if an off-screen enemy deals damage, the model cannot react to anything beyond its narrow visual range.
Speaking of the ethical implications of his research, De Winter notes:
"It's very alarming how easy it was for me to create code to make the model shoot at something, and for the model - to shoot at something with no real understanding of the instructions."
He believes that the model is actually not aware of what it is doing, so it is important to use such models with caution to avoid misuse.
The fact that GPT-4 did not hesitate to shoot at objects resembling humans is worrisome. It seems like it is time to implement Isaac Asimov's "Laws of Robotics" into all artificial intelligence models.
Source: tomshardware
Comments (0)
There are no comments for now