A clever person has already found a unique application for ChatGPT-4, the newest artificial intelligence language model from the creators of ChatGPT. As it turns out, GPT-4 can read the code for the Game Boy Advance game Pokémon Emerald and turn it into a fully interactive text adventure, with the mechanics of a Pokémon adventure.
Twitter user Dan Dangond discovered that the AI model could run Pokémon Emerald, a popular Game Boy Advance game released stateside in 2005. Dangond shared a screenshot-by-screenshot thread that contains their conversation with the chatbot. The game starts more or less as it would if you were playing it on the handheld — save for some some features, like setting up the clock at the beginning — but works as a fully interactive text adventure. Dangond “plays” the game by responding to the AI in chat.
As always with mainline Pokémon games, in Pokémon Emerald you explore a pixelated world and battle other Pokémon in turn-based combat, so the text largely describes battles and calculates them. As Dangond begins, the AI offers the choice between the three starters of the region: Treecko, Torchic, and Mudkip. From there, the Mudkip is sent out to battle a Poochyena, which is an actual Pokémon you can battle on the first route.
Starts up the game without a problem, pretty accurate so far, skipping the moving van and clock setting scene which honestly isn’t the most riveting thing in this game anyway pic.twitter.com/o4nqtMdCHH
— Dan Dangond (@DanDangond) March 15, 2023
The rest of the game runs like a rough text simulation that incorporates all the key mechanics of Pokémon, like type strengths and weaknesses. However, there are several missteps, for example, it fails to calculate that the bug/ground type Nincada is weak to water. It also randomly decided to start including accuracy calculations after playing through a portion of the game, and makes a Rock Tomb attack miss. It also skips several battles and misses certain beats.
This text-based game comes at a time when many people are still feeling out what the actual capabilities of generative AI are. According to OpenAI, GPT-4 is a multimodal model, that is “ less capable than humans in many real-world scenarios,” but can exhibit “human-level performance on various professional and academic benchmarks.” According to other tech journalists who have covered the technology, it’s supposedly more capable and sophisticated that its predecessor, ChatGPT, and it is acing standardized tests left and right. Now we can add running a Pokémon game to this list of skills.
ChatGPT-4 wasn’t created with running video games in mind, and it shows. Any person who would want to play the game like this would greatly benefit from having played the original game. Without visuals and a general knowledge of the world, it would be nearly impossible to play. Still, it’s a fun discovery that GPT-4 can run this game in any capacity, and that’s where the potential for fun and more tinkering lies.