Alexander Lewis 
Mar 13, 2024
A photo taken on October 4, 2023, in Manta, near Turin, shows a smartphone and a laptop displaying the logos of the artificial intelligence OpenAI research laboratory and ChatGPT robot. — AFP
Ever thought of playing a first-person shooting game with the help of OpenAI's GPT-45 model? Surely you did not, as the AI tool is not developed for this purpose.
But to anyone's surprise, Adrian de Wynter, a Microsoft scientist, has come up with a way for OpenAI’s GPT-4 model to play Doom. Accounting his efforts to assess if the AI tool possess the potential to play the classic PC game, he published a paper and a blog post. He did this by merging GPT-4 with Vision API, which lets the AI model receive visual images, prompting a corresponding action.
de Wynter then made a program to take screenshots from a computer running Doom, then added them on GPT-4, aiming to make it beat the game. The answers were then converted into outputs used to control the game.
Read more: Future of AI governance — Mapping regulatory landscape
GPT-4 can play Doom
In his experiment, de Wynter provided GPT-4 access to a set of prompt generators that were instructed to succeed in the game and could learn from past actions. The findings revealed that OpenAI's technology is capable of playing Doom to a “passable” extent, indicating that the AI can navigate the map and engage with enemies.
However, when compared to human performance, GPT-4 “kinda fails at the game and my grandma has played it wayyy better than this model,” de Wynter stated in his blog post. For instance, the AI model would sometimes exhibit foolish behavior such as getting stuck in corners and aimlessly punching the wall, or shooting explosive barrels at close range.
Furthermore, GPT-4 would frequently let go enemies and behave as though they did not exist when they were out of its field of view. “Most of the time the model died because it got stuck in a corner and ended up being shot in the back,” de Wynter added.
The results indicate that GPT-4 may encounter difficulties with long-term reasoning and "object permanence," which refers to the ability to recognise that an object still exists even when it is not visible. Another issue is that OpenAI's model sometimes fabricated reasons for needing to take certain actions while playing the game.
“So, while this is a very interesting exploration around planning and reasoning, and could have applications in automated videogame testing, it is quite obvious that this model is not aware of what it is doing,” he added. “I strongly urge everyone to think about what deployment of these models implies for society and their potential misuse.”
