Microsoft scientist and York University researcher Adrian de Winter recently decided to test GPT-4 by running Doom without prior training. The results of this experiment concluded that GPT-4 can play Doom with a certain degree of efficiency without prior training. Although this requires the GPT-4V interface to see the AI and the addition of Matplotlib and Python to enable AI input.
The GPT-4 AI can open doors, fight enemies, and fire weapons while playing Doom. The main gameplay functions, including world navigation, are used by almost everyone. However, the AI has no sense of object permanence — once the enemy moves out of the frame, the AI is no longer aware of it at all. Although instructions are included on what to do if an in-frame enemy causes damage, the model cannot react to anything outside of its narrow visual range.
Speaking about the ethical implications of his research, De Winter notes:
«It was very disturbing how easy it was for me to create code to make the model shoot at something, and for the model to shoot at something accurately without actually guessing the directions».
He believes that the model is not actually aware of what it is doing, so it is important to use such models with caution to avoid abuse.
The fact that the GPT-4 did not hesitate to shoot at human-shaped objects is troubling. It looks like it’s time to introduce «Robotics Laws» Isaac Asimov into all artificial intelligence models.
Source: tomshardware