The Interplay Between Artificial Intelligence and Classic Video Games
Engineers, researchers, and hobbyists have delved into the realm of proving that the iconic first-person shooter game, Doom, can run on an array of devices ranging from lawnmowers to gut bacteria. However, on Wednesday, a principal applied scientist at Microsoft, Adrian de Wynter, unveiled a unique experiment that showcased the attempt to integrate Doom with an AI chatbot known as ChatGPT.
The Challenge of Integrating ChatGPT with Doom
De Wynter utilized OpenAI’s multimodal GPT-4V (Vision) to enable the chatbot to engage in gameplay. Despite the advancements in GPT-4 and its vision-enhanced variant, the AI model struggled to autonomously navigate Doom due to constraints in input processing and image rendering.
One significant issue that emerged during the experiment was the AI model’s tendency to hallucinate, fabricate explanations, or falsely claim completion of actions. Consequently, this led to unfavorable outcomes for Doom’s Space Marine as it battled against malicious monsters.
Insights and Reflections on the Experiment
Despite encountering challenges, GPT-4 managed to progress to the final room in the game on a single occasion. De Wynter acknowledged that Doom’s simplicity and open-source nature contributed to its suitability for testing intelligent agents due to the intricate reasoning demands posed by the game.
De Wynter noted that the project was a personal research endeavor unaffiliated with his role at Microsoft. He shared that debugging the experiment required meticulous attention to detail and occasional frame analysis to mitigate issues such as the AI model attempting to escape the game map through unconventional means.
Exploring New Horizons in AI Gaming Experiments
De Wynter’s project forms part of a series of innovative experiments aimed at pushing boundaries in unconventional gaming setups. Notably, the integration of Doom into the Bitcoin and Dogecoin blockchains showcased the versatile nature of the game.
Looking ahead, de Wynter expressed interest in leveraging large language models (LLMs) for future gaming experiments focused on enhancing reasoning and planning capabilities. While strategy games pose a challenge at present, de Wynter remains intrigued by the potential of simpler games or alternative models to yield promising outcomes.
Image/Photo credit: source url